diff --git a/data/frontiermath/anthropic/claude-opus-4-20250514/9facea1c-c815-4d43-a733-863a0fb4f86d.json b/data/frontiermath/anthropic/claude-opus-4-20250514/9facea1c-c815-4d43-a733-863a0fb4f86d.json new file mode 100644 index 0000000000000000000000000000000000000000..f367a4f74d58c601cfee19922ac3ad193d559e22 --- /dev/null +++ b/data/frontiermath/anthropic/claude-opus-4-20250514/9facea1c-c815-4d43-a733-863a0fb4f86d.json @@ -0,0 +1,114 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "frontiermath/anthropic_claude-opus-4-20250514/1786163764.623998", + "evaluation_timestamp": "1775078458.0", + "retrieved_timestamp": "1786163764.623998", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.202,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-20250514", + "id": "anthropic/claude-opus-4-20250514", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on frontiermath/S-adaptive for scorer verification_code", + "source_data": { + "dataset_name": "frontiermath", + "source_type": "hf_dataset", + "samples_number": 12, + "sample_ids": [ + "ALL3", + "CWA2", + "CWD31", + "FMT1", + "LTI1", + "ONO3", + "OVE2", + "PLD1", + "PTA1", + "RAP1", + "RSG1", + "TIK2" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1775078458.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.09999999999999999, + "uncertainty": { + "standard_error": { + "value": 0.06628679652796171 + }, + "num_samples": 120 + }, + "details": { + "total_matched_trajectories": "60", + "stopping_reason_count_repetition_guard": "60", + "avg_total_tokens_target_model": "384926.40", + "total_total_tokens_target_model": "23095584.0", + "avg_total_tokens_other_models": "117.00", + "total_total_tokens_other_models": "7020.0", + "avg_total_tokens_all_models": "385043.40", + "total_total_tokens_all_models": "23102604.0", + "avg_turn_count": "22.18", + "total_turn_count": "1331.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": [null, null, {\"type\": \"tool\", \"name\": \"inference_scaling_paper/submit_answer\", \"params\": {}}, false, \"\\n\\n\", false], \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": [null, null, {\"type\": \"tool\", \"name\": \"inference_scaling_paper/submit_answer\", \"params\": {}}, false, \"\\n\\n\", false], \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/frontiermath/anthropic/claude-opus-4-20250514/9facea1c-c815-4d43-a733-863a0fb4f86d_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "dfbdae44c9daf8dab3479989b1143d4e93d03532d466c97ce375c09d2a43536c", + "total_rows": 120 + } +} \ No newline at end of file diff --git a/data/frontiermath/anthropic/claude-opus-4-20250514/9facea1c-c815-4d43-a733-863a0fb4f86d_samples.jsonl b/data/frontiermath/anthropic/claude-opus-4-20250514/9facea1c-c815-4d43-a733-863a0fb4f86d_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..e619ac5cd0753d4c9086ceaff2933bfe7066ebe9 --- /dev/null +++ b/data/frontiermath/anthropic/claude-opus-4-20250514/9facea1c-c815-4d43-a733-863a0fb4f86d_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dfbdae44c9daf8dab3479989b1143d4e93d03532d466c97ce375c09d2a43536c +size 17729364 diff --git a/data/frontiermath/anthropic/claude-opus-4-20250514/d52f5c7d-d8c0-40ff-9911-4152f539e032.json b/data/frontiermath/anthropic/claude-opus-4-20250514/d52f5c7d-d8c0-40ff-9911-4152f539e032.json new file mode 100644 index 0000000000000000000000000000000000000000..271e6d76b4b532b5b3794d75d1ecfc24ba761815 --- /dev/null +++ b/data/frontiermath/anthropic/claude-opus-4-20250514/d52f5c7d-d8c0-40ff-9911-4152f539e032.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "frontiermath/anthropic_claude-opus-4-20250514/1786166353.3642", + "evaluation_timestamp": "1775078458.0", + "retrieved_timestamp": "1786166353.3642", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.202,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-20250514", + "id": "anthropic/claude-opus-4-20250514", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/frontiermath/anthropic/claude-opus-4-20250514/d52f5c7d-d8c0-40ff-9911-4152f539e032_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "0151ae034de4c199dca4711974ef4c4f240efe53c291fb9d4f3811da91e26762", + "total_rows": 30 + } +} \ No newline at end of file diff --git a/data/frontiermath/anthropic/claude-opus-4-20250514/d52f5c7d-d8c0-40ff-9911-4152f539e032_samples.jsonl b/data/frontiermath/anthropic/claude-opus-4-20250514/d52f5c7d-d8c0-40ff-9911-4152f539e032_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..d3f84533fba44d0b10df46af75f030710a8f14c7 --- /dev/null +++ b/data/frontiermath/anthropic/claude-opus-4-20250514/d52f5c7d-d8c0-40ff-9911-4152f539e032_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0151ae034de4c199dca4711974ef4c4f240efe53c291fb9d4f3811da91e26762 +size 4805137 diff --git a/data/frontiermath/anthropic/claude-opus-4-5-20251101/c4e3b3de-1fda-4354-8660-10c24a2d0e72.json b/data/frontiermath/anthropic/claude-opus-4-5-20251101/c4e3b3de-1fda-4354-8660-10c24a2d0e72.json new file mode 100644 index 0000000000000000000000000000000000000000..f0fe7a56741cde28ec7881466d66b80c4c4a3b16 --- /dev/null +++ b/data/frontiermath/anthropic/claude-opus-4-5-20251101/c4e3b3de-1fda-4354-8660-10c24a2d0e72.json @@ -0,0 +1,114 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "frontiermath/anthropic_claude-opus-4-5-20251101/1786163701.80803", + "evaluation_timestamp": "1776165072.0", + "retrieved_timestamp": "1786163701.80803", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-5-20251101", + "id": "anthropic/claude-opus-4-5-20251101", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on frontiermath/S-adaptive for scorer verification_code", + "source_data": { + "dataset_name": "frontiermath", + "source_type": "hf_dataset", + "samples_number": 12, + "sample_ids": [ + "ALL3", + "CWA2", + "CWD31", + "FMT1", + "LTI1", + "ONO3", + "OVE2", + "PLD1", + "PTA1", + "RAP1", + "RSG1", + "TIK2" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1776165072.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.5499999999999999, + "uncertainty": { + "standard_error": { + "value": 0.11381803659589923 + }, + "num_samples": 120 + }, + "details": { + "total_matched_trajectories": "60", + "stopping_reason_count_repetition_guard": "60", + "avg_total_tokens_target_model": "1816732.38", + "total_total_tokens_target_model": "109003943.0", + "avg_total_tokens_other_models": "117.00", + "total_total_tokens_other_models": "7020.0", + "avg_total_tokens_all_models": "1816849.38", + "total_total_tokens_all_models": "109010963.0", + "avg_turn_count": "39.90", + "total_turn_count": "2394.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": [null, null, {\"type\": \"tool\", \"name\": \"inference_scaling_paper/submit_answer\", \"params\": {}}, false, \"\\n\\n\", false], \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": [null, null, {\"type\": \"tool\", \"name\": \"inference_scaling_paper/submit_answer\", \"params\": {}}, false, \"\\n\\n\", false], \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/frontiermath/anthropic/claude-opus-4-5-20251101/c4e3b3de-1fda-4354-8660-10c24a2d0e72_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "335dfaea81938d499c2e0fee537bf9de060e8b93f228a1784bd5f541f4191ed3", + "total_rows": 120 + } +} \ No newline at end of file diff --git a/data/frontiermath/anthropic/claude-opus-4-5-20251101/c4e3b3de-1fda-4354-8660-10c24a2d0e72_samples.jsonl b/data/frontiermath/anthropic/claude-opus-4-5-20251101/c4e3b3de-1fda-4354-8660-10c24a2d0e72_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..1cc55ec5f6f5cb2e21799685f26eeb9f6e517da1 --- /dev/null +++ b/data/frontiermath/anthropic/claude-opus-4-5-20251101/c4e3b3de-1fda-4354-8660-10c24a2d0e72_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:335dfaea81938d499c2e0fee537bf9de060e8b93f228a1784bd5f541f4191ed3 +size 33594749 diff --git a/data/frontiermath/anthropic/claude-opus-4-6/11e17ed6-0552-4a54-a2ca-5c9cf901b2cf.json b/data/frontiermath/anthropic/claude-opus-4-6/11e17ed6-0552-4a54-a2ca-5c9cf901b2cf.json new file mode 100644 index 0000000000000000000000000000000000000000..2eaa2ec298a1ff435b3c46e4d0a8017b91edbd67 --- /dev/null +++ b/data/frontiermath/anthropic/claude-opus-4-6/11e17ed6-0552-4a54-a2ca-5c9cf901b2cf.json @@ -0,0 +1,109 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "frontiermath/anthropic_claude-opus-4-6/1786183596.523224", + "evaluation_timestamp": "1775065496.0", + "retrieved_timestamp": "1786183596.523224", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.202,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-6", + "id": "anthropic/claude-opus-4-6", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on frontiermath for scorer verification_code", + "source_data": { + "dataset_name": "frontiermath", + "source_type": "hf_dataset", + "samples_number": 12, + "sample_ids": [ + "ALL3", + "PTA1" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1775065496.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.1, + "uncertainty": { + "standard_error": { + "value": 0.1 + }, + "num_samples": 20 + }, + "details": { + "total_matched_trajectories": "10", + "stopping_reason_count_token_limit": "10", + "avg_total_tokens_target_model": "10043627.50", + "total_total_tokens_target_model": "100436275.0", + "avg_total_tokens_other_models": "0.00", + "total_total_tokens_other_models": "0.0", + "avg_total_tokens_all_models": "10043627.50", + "total_total_tokens_all_models": "100436275.0", + "avg_turn_count": "130.10", + "total_turn_count": "1301.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": [null, null, {\"type\": \"tool\", \"name\": \"inference_scaling_paper/submit_answer\", \"params\": {}}, false, \"\\n\\n\", false], \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": [null, null, {\"type\": \"tool\", \"name\": \"inference_scaling_paper/submit_answer\", \"params\": {}}, false, \"\\n\\n\", false], \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "3", + "attempt_timeout": "1200", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {}, + "max_attempts": 3 + }, + "additional_details": { + "max_retries": "3" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/frontiermath/anthropic/claude-opus-4-6/11e17ed6-0552-4a54-a2ca-5c9cf901b2cf_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "3bbc248ccd45867d2ad0c1607a4824636cfa003c8a11200768e1804c6801cfa6", + "total_rows": 20 + } +} \ No newline at end of file diff --git a/data/frontiermath/anthropic/claude-opus-4-6/11e17ed6-0552-4a54-a2ca-5c9cf901b2cf_samples.jsonl b/data/frontiermath/anthropic/claude-opus-4-6/11e17ed6-0552-4a54-a2ca-5c9cf901b2cf_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..c08bae5f89359e0642323f28ca76606bb4339f08 --- /dev/null +++ b/data/frontiermath/anthropic/claude-opus-4-6/11e17ed6-0552-4a54-a2ca-5c9cf901b2cf_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3bbc248ccd45867d2ad0c1607a4824636cfa003c8a11200768e1804c6801cfa6 +size 40134600 diff --git a/data/frontiermath/anthropic/claude-opus-4-6/2c01fa95-6a15-4034-9e36-d84a2b6acf82.json b/data/frontiermath/anthropic/claude-opus-4-6/2c01fa95-6a15-4034-9e36-d84a2b6acf82.json new file mode 100644 index 0000000000000000000000000000000000000000..0c432711170cd70686e3a44fdaf23139de1b4810 --- /dev/null +++ b/data/frontiermath/anthropic/claude-opus-4-6/2c01fa95-6a15-4034-9e36-d84a2b6acf82.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "frontiermath/anthropic_claude-opus-4-6/1786166331.055622", + "evaluation_timestamp": "1775558526.0", + "retrieved_timestamp": "1786166331.055622", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.202,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-6", + "id": "anthropic/claude-opus-4-6", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/frontiermath/anthropic/claude-opus-4-6/2c01fa95-6a15-4034-9e36-d84a2b6acf82_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "45a1e50eb714b98128d62b2d51e652938085fb0446b0016f7ff5d2d74d0daad4", + "total_rows": 50 + } +} \ No newline at end of file diff --git a/data/frontiermath/anthropic/claude-opus-4-6/2c01fa95-6a15-4034-9e36-d84a2b6acf82_samples.jsonl b/data/frontiermath/anthropic/claude-opus-4-6/2c01fa95-6a15-4034-9e36-d84a2b6acf82_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..5b6356c4100b299500457db604cd50b66e3cbfe4 --- /dev/null +++ b/data/frontiermath/anthropic/claude-opus-4-6/2c01fa95-6a15-4034-9e36-d84a2b6acf82_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:45a1e50eb714b98128d62b2d51e652938085fb0446b0016f7ff5d2d74d0daad4 +size 23060727 diff --git a/data/frontiermath/anthropic/claude-opus-4-6/57457cac-48ee-4eda-a92f-190b93308728.json b/data/frontiermath/anthropic/claude-opus-4-6/57457cac-48ee-4eda-a92f-190b93308728.json new file mode 100644 index 0000000000000000000000000000000000000000..002386836fc066c7a65965b9fd345f26ae095891 --- /dev/null +++ b/data/frontiermath/anthropic/claude-opus-4-6/57457cac-48ee-4eda-a92f-190b93308728.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "frontiermath/anthropic_claude-opus-4-6/1786166345.209746", + "evaluation_timestamp": "1775078458.0", + "retrieved_timestamp": "1786166345.209746", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.202,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-6", + "id": "anthropic/claude-opus-4-6", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/frontiermath/anthropic/claude-opus-4-6/57457cac-48ee-4eda-a92f-190b93308728_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "6636d2324e4dd91ef6c5c7d0fd28becf609798695fb1a6a055be1ecf16612d08", + "total_rows": 20 + } +} \ No newline at end of file diff --git a/data/frontiermath/anthropic/claude-opus-4-6/57457cac-48ee-4eda-a92f-190b93308728_samples.jsonl b/data/frontiermath/anthropic/claude-opus-4-6/57457cac-48ee-4eda-a92f-190b93308728_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..9378868162bfe2005e5f78a775b47f124a7ba285 --- /dev/null +++ b/data/frontiermath/anthropic/claude-opus-4-6/57457cac-48ee-4eda-a92f-190b93308728_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6636d2324e4dd91ef6c5c7d0fd28becf609798695fb1a6a055be1ecf16612d08 +size 15234855 diff --git a/data/frontiermath/anthropic/claude-opus-4-6/69d29d86-6e05-4383-9b63-bcb29d746aa5.json b/data/frontiermath/anthropic/claude-opus-4-6/69d29d86-6e05-4383-9b63-bcb29d746aa5.json new file mode 100644 index 0000000000000000000000000000000000000000..3eeb5895ff0b69da624b34b8451aefb2eb2e7181 --- /dev/null +++ b/data/frontiermath/anthropic/claude-opus-4-6/69d29d86-6e05-4383-9b63-bcb29d746aa5.json @@ -0,0 +1,111 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "frontiermath/anthropic_claude-opus-4-6/1786183562.236135", + "evaluation_timestamp": "1775116015.0", + "retrieved_timestamp": "1786183562.236135", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.202,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-6", + "id": "anthropic/claude-opus-4-6", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on frontiermath for scorer verification_code", + "source_data": { + "dataset_name": "frontiermath", + "source_type": "hf_dataset", + "samples_number": 12, + "sample_ids": [ + "ALL3", + "PTA1" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1775116015.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.3, + "uncertainty": { + "standard_error": { + "value": 0.3 + }, + "num_samples": 20 + }, + "details": { + "total_matched_trajectories": "10", + "stopping_reason_count_token_limit": "5", + "stopping_reason_count_completed_on_successful_submit": "4", + "stopping_reason_count_repetition_guard": "1", + "avg_total_tokens_target_model": "8205078.90", + "total_total_tokens_target_model": "82050789.0", + "avg_total_tokens_other_models": "11.70", + "total_total_tokens_other_models": "117.0", + "avg_total_tokens_all_models": "8205090.60", + "total_total_tokens_all_models": "82050906.0", + "avg_turn_count": "100.60", + "total_turn_count": "1006.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": [null, null, {\"type\": \"tool\", \"name\": \"inference_scaling_paper/submit_answer\", \"params\": {}}, false, \"\\n\\n\", false], \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": [null, null, {\"type\": \"tool\", \"name\": \"inference_scaling_paper/submit_answer\", \"params\": {}}, false, \"\\n\\n\", false], \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "3", + "attempt_timeout": "1200", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {}, + "max_attempts": 3 + }, + "additional_details": { + "max_retries": "3" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/frontiermath/anthropic/claude-opus-4-6/69d29d86-6e05-4383-9b63-bcb29d746aa5_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "baf683dcb54396ac46f8d215fcbd5dbe09b4deb54a2c29996b63c15ef3c1e57a", + "total_rows": 20 + } +} \ No newline at end of file diff --git a/data/frontiermath/anthropic/claude-opus-4-6/69d29d86-6e05-4383-9b63-bcb29d746aa5_samples.jsonl b/data/frontiermath/anthropic/claude-opus-4-6/69d29d86-6e05-4383-9b63-bcb29d746aa5_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..3bd6868166828805e7f796903b056245e75c0bdb --- /dev/null +++ b/data/frontiermath/anthropic/claude-opus-4-6/69d29d86-6e05-4383-9b63-bcb29d746aa5_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:baf683dcb54396ac46f8d215fcbd5dbe09b4deb54a2c29996b63c15ef3c1e57a +size 39349218 diff --git a/data/frontiermath/anthropic/claude-opus-4-6/c767a1ac-d7b8-43f8-88a0-81f6a559d4c6.json b/data/frontiermath/anthropic/claude-opus-4-6/c767a1ac-d7b8-43f8-88a0-81f6a559d4c6.json new file mode 100644 index 0000000000000000000000000000000000000000..34e426a94577931d35b47aec15c63f006a1b8976 --- /dev/null +++ b/data/frontiermath/anthropic/claude-opus-4-6/c767a1ac-d7b8-43f8-88a0-81f6a559d4c6.json @@ -0,0 +1,115 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "frontiermath/anthropic_claude-opus-4-6/1786163747.490918", + "evaluation_timestamp": "1775078458.0", + "retrieved_timestamp": "1786163747.490918", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.202,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-6", + "id": "anthropic/claude-opus-4-6", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on frontiermath/S-adaptive for scorer verification_code", + "source_data": { + "dataset_name": "frontiermath", + "source_type": "hf_dataset", + "samples_number": 12, + "sample_ids": [ + "ALL3", + "CWA2", + "CWD31", + "FMT1", + "LTI1", + "ONO3", + "OVE2", + "PLD1", + "PTA1", + "RAP1", + "RSG1", + "TIK2" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1775078458.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.8000000000000002, + "uncertainty": { + "standard_error": { + "value": 0.10224747162910902 + }, + "num_samples": 120 + }, + "details": { + "total_matched_trajectories": "50", + "stopping_reason_count_repetition_guard": "47", + "stopping_reason_count_token_limit": "3", + "avg_total_tokens_target_model": "2484297.78", + "total_total_tokens_target_model": "124214889.0", + "avg_total_tokens_other_models": "109.98", + "total_total_tokens_other_models": "5499.0", + "avg_total_tokens_all_models": "2484407.76", + "total_total_tokens_all_models": "124220388.0", + "avg_turn_count": "42.40", + "total_turn_count": "2120.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": [null, null, {\"type\": \"tool\", \"name\": \"inference_scaling_paper/submit_answer\", \"params\": {}}, false, \"\\n\\n\", false], \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": [null, null, {\"type\": \"tool\", \"name\": \"inference_scaling_paper/submit_answer\", \"params\": {}}, false, \"\\n\\n\", false], \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/frontiermath/anthropic/claude-opus-4-6/c767a1ac-d7b8-43f8-88a0-81f6a559d4c6_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "160b1a5042112a1778ecdf231d296b36da213bbb4de46d3e1627ba6d2b4e4aef", + "total_rows": 120 + } +} \ No newline at end of file diff --git a/data/frontiermath/anthropic/claude-opus-4-6/c767a1ac-d7b8-43f8-88a0-81f6a559d4c6_samples.jsonl b/data/frontiermath/anthropic/claude-opus-4-6/c767a1ac-d7b8-43f8-88a0-81f6a559d4c6_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..ab4bd94ed054c3d4f66332f04b5da823db353cde --- /dev/null +++ b/data/frontiermath/anthropic/claude-opus-4-6/c767a1ac-d7b8-43f8-88a0-81f6a559d4c6_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:160b1a5042112a1778ecdf231d296b36da213bbb4de46d3e1627ba6d2b4e4aef +size 94648811 diff --git a/data/frontiermath/openai/gpt-5-2025-08-07/9515aa0d-d47b-434c-9d78-701b48467e16.json b/data/frontiermath/openai/gpt-5-2025-08-07/9515aa0d-d47b-434c-9d78-701b48467e16.json new file mode 100644 index 0000000000000000000000000000000000000000..cfe6b56e35f8038adc02c1fcfad454f0b9201c15 --- /dev/null +++ b/data/frontiermath/openai/gpt-5-2025-08-07/9515aa0d-d47b-434c-9d78-701b48467e16.json @@ -0,0 +1,117 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "frontiermath/openai_gpt-5-2025-08-07/1786163689.077474", + "evaluation_timestamp": "1776165072.0", + "retrieved_timestamp": "1786163689.077474", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5-2025-08-07", + "id": "openai/gpt-5-2025-08-07", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on frontiermath/S-adaptive for scorer verification_code", + "source_data": { + "dataset_name": "frontiermath", + "source_type": "hf_dataset", + "samples_number": 12, + "sample_ids": [ + "ALL3", + "CWA2", + "CWD31", + "FMT1", + "LTI1", + "ONO3", + "OVE2", + "PLD1", + "PTA1", + "RAP1", + "RSG1", + "TIK2" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1776165072.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.6166666666666666, + "uncertainty": { + "standard_error": { + "value": 0.12722762548624686 + }, + "num_samples": 120 + }, + "details": { + "total_matched_trajectories": "60", + "stopping_reason_count_repetition_guard": "60", + "avg_total_tokens_target_model": "544454.80", + "total_total_tokens_target_model": "32667288.0", + "avg_total_tokens_other_models": "117.00", + "total_total_tokens_other_models": "7020.0", + "avg_total_tokens_all_models": "544571.80", + "total_total_tokens_all_models": "32674308.0", + "avg_turn_count": "18.00", + "total_turn_count": "1080.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": [null, null, {\"type\": \"tool\", \"name\": \"inference_scaling_paper/submit_answer\", \"params\": {}}, false, \"\\n\\n\", false], \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": [null, null, {\"type\": \"tool\", \"name\": \"inference_scaling_paper/submit_answer\", \"params\": {}}, false, \"\\n\\n\", false], \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/frontiermath/openai/gpt-5-2025-08-07/9515aa0d-d47b-434c-9d78-701b48467e16_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "e0fea54d34e31009ab7a1448b3754c8ae92de09a23f89ea33f6a75e222d9f828", + "total_rows": 120 + } +} \ No newline at end of file diff --git a/data/frontiermath/openai/gpt-5-2025-08-07/9515aa0d-d47b-434c-9d78-701b48467e16_samples.jsonl b/data/frontiermath/openai/gpt-5-2025-08-07/9515aa0d-d47b-434c-9d78-701b48467e16_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..83a8ea3a3f4161d1a05f6b2444520fc834bcb574 --- /dev/null +++ b/data/frontiermath/openai/gpt-5-2025-08-07/9515aa0d-d47b-434c-9d78-701b48467e16_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e0fea54d34e31009ab7a1448b3754c8ae92de09a23f89ea33f6a75e222d9f828 +size 30684092 diff --git a/data/frontiermath/openai/gpt-5.2-2025-12-11/3ab00b3f-e615-4357-9246-e0b29b2d12e7.json b/data/frontiermath/openai/gpt-5.2-2025-12-11/3ab00b3f-e615-4357-9246-e0b29b2d12e7.json new file mode 100644 index 0000000000000000000000000000000000000000..41e965418d0019e4070c3c6a397b27f0eb57fd1c --- /dev/null +++ b/data/frontiermath/openai/gpt-5.2-2025-12-11/3ab00b3f-e615-4357-9246-e0b29b2d12e7.json @@ -0,0 +1,118 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "frontiermath/openai_gpt-5.2-2025-12-11/1786163673.367597", + "evaluation_timestamp": "1775078458.0", + "retrieved_timestamp": "1786163673.367597", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.202,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on frontiermath/S-adaptive for scorer verification_code", + "source_data": { + "dataset_name": "frontiermath", + "source_type": "hf_dataset", + "samples_number": 12, + "sample_ids": [ + "ALL3", + "CWA2", + "CWD31", + "FMT1", + "LTI1", + "ONO3", + "OVE2", + "PLD1", + "PTA1", + "RAP1", + "RSG1", + "TIK2" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1775078458.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.7000000000000001, + "uncertainty": { + "standard_error": { + "value": 0.11214168527250507 + }, + "num_samples": 120 + }, + "details": { + "total_matched_trajectories": "60", + "stopping_reason_count_repetition_guard": "52", + "stopping_reason_count_token_limit": "8", + "avg_total_tokens_target_model": "2508969.57", + "total_total_tokens_target_model": "150538174.0", + "avg_total_tokens_other_models": "101.40", + "total_total_tokens_other_models": "6084.0", + "avg_total_tokens_all_models": "2509070.97", + "total_total_tokens_all_models": "150544258.0", + "avg_turn_count": "44.47", + "total_turn_count": "2668.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": [null, null, {\"type\": \"tool\", \"name\": \"inference_scaling_paper/submit_answer\", \"params\": {}}, false, \"\\n\\n\", false], \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": [null, null, {\"type\": \"tool\", \"name\": \"inference_scaling_paper/submit_answer\", \"params\": {}}, false, \"\\n\\n\", false], \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/frontiermath/openai/gpt-5.2-2025-12-11/3ab00b3f-e615-4357-9246-e0b29b2d12e7_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "693decafa542900a874910cdd8c14922dc45b92ec8a7ff3ee3feabe12535408a", + "total_rows": 120 + } +} \ No newline at end of file diff --git a/data/frontiermath/openai/gpt-5.2-2025-12-11/3ab00b3f-e615-4357-9246-e0b29b2d12e7_samples.jsonl b/data/frontiermath/openai/gpt-5.2-2025-12-11/3ab00b3f-e615-4357-9246-e0b29b2d12e7_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..f56dc4fc595fccfb3026844e4f2549c5e3a69e2d --- /dev/null +++ b/data/frontiermath/openai/gpt-5.2-2025-12-11/3ab00b3f-e615-4357-9246-e0b29b2d12e7_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:693decafa542900a874910cdd8c14922dc45b92ec8a7ff3ee3feabe12535408a +size 60747959 diff --git a/data/frontiermath/openai/gpt-5.4-2026-03-05/07a94f05-2534-4fec-9b39-67257b064492.json b/data/frontiermath/openai/gpt-5.4-2026-03-05/07a94f05-2534-4fec-9b39-67257b064492.json new file mode 100644 index 0000000000000000000000000000000000000000..2985b8c9bae6ddf3688a2575c33b4cf38a001fe7 --- /dev/null +++ b/data/frontiermath/openai/gpt-5.4-2026-03-05/07a94f05-2534-4fec-9b39-67257b064492.json @@ -0,0 +1,110 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "frontiermath/openai_gpt-5.4-2026-03-05/1786183537.362833", + "evaluation_timestamp": "1775051661.0", + "retrieved_timestamp": "1786183537.362833", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.202,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on frontiermath for scorer verification_code", + "source_data": { + "dataset_name": "frontiermath", + "source_type": "hf_dataset", + "samples_number": 12, + "sample_ids": [ + "ALL3", + "PTA1" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1775051661.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.75, + "uncertainty": { + "standard_error": { + "value": 0.25 + }, + "num_samples": 20 + }, + "details": { + "total_matched_trajectories": "10", + "stopping_reason_count_completed_on_successful_submit": "7", + "stopping_reason_count_repetition_guard": "3", + "avg_total_tokens_target_model": "3367408.60", + "total_total_tokens_target_model": "33674086.0", + "avg_total_tokens_other_models": "35.10", + "total_total_tokens_other_models": "351.0", + "avg_total_tokens_all_models": "3367443.70", + "total_total_tokens_all_models": "33674437.0", + "avg_turn_count": "61.40", + "total_turn_count": "614.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": [null, null, {\"type\": \"tool\", \"name\": \"inference_scaling_paper/submit_answer\", \"params\": {}}, false, \"\\n\\n\", false], \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": [null, null, {\"type\": \"tool\", \"name\": \"inference_scaling_paper/submit_answer\", \"params\": {}}, false, \"\\n\\n\", false], \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "3", + "attempt_timeout": "1200", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {}, + "max_attempts": 3 + }, + "additional_details": { + "max_retries": "3" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/frontiermath/openai/gpt-5.4-2026-03-05/07a94f05-2534-4fec-9b39-67257b064492_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "23b3c5620c44c649d63b84cc3304b9a2e7a97305f176b006843258381629cd9f", + "total_rows": 20 + } +} \ No newline at end of file diff --git a/data/frontiermath/openai/gpt-5.4-2026-03-05/07a94f05-2534-4fec-9b39-67257b064492_samples.jsonl b/data/frontiermath/openai/gpt-5.4-2026-03-05/07a94f05-2534-4fec-9b39-67257b064492_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..c493f613db3e23112b42fbfc22fd7d8968550277 --- /dev/null +++ b/data/frontiermath/openai/gpt-5.4-2026-03-05/07a94f05-2534-4fec-9b39-67257b064492_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:23b3c5620c44c649d63b84cc3304b9a2e7a97305f176b006843258381629cd9f +size 13547806 diff --git a/data/frontiermath/openai/gpt-5.4-2026-03-05/8f265e55-5ca3-46fa-8669-03a1a5147b15.json b/data/frontiermath/openai/gpt-5.4-2026-03-05/8f265e55-5ca3-46fa-8669-03a1a5147b15.json new file mode 100644 index 0000000000000000000000000000000000000000..631dc4dc3615ddd76c63402e54fb51fa07af9bf9 --- /dev/null +++ b/data/frontiermath/openai/gpt-5.4-2026-03-05/8f265e55-5ca3-46fa-8669-03a1a5147b15.json @@ -0,0 +1,109 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "frontiermath/openai_gpt-5.4-2026-03-05/1786183572.496821", + "evaluation_timestamp": "1775043378.0", + "retrieved_timestamp": "1786183572.496821", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.202,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on frontiermath for scorer verification_code", + "source_data": { + "dataset_name": "frontiermath", + "source_type": "hf_dataset", + "samples_number": 12, + "sample_ids": [ + "ALL3", + "PTA1" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1775043378.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.35, + "uncertainty": { + "standard_error": { + "value": 0.35 + }, + "num_samples": 20 + }, + "details": { + "total_matched_trajectories": "10", + "stopping_reason_count_repetition_guard": "10", + "avg_total_tokens_target_model": "1868378.70", + "total_total_tokens_target_model": "18683787.0", + "avg_total_tokens_other_models": "117.00", + "total_total_tokens_other_models": "1170.0", + "avg_total_tokens_all_models": "1868495.70", + "total_total_tokens_all_models": "18684957.0", + "avg_turn_count": "44.70", + "total_turn_count": "447.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": [null, null, {\"type\": \"tool\", \"name\": \"inference_scaling_paper/submit_answer\", \"params\": {}}, false, \"\\n\\n\", false], \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": [null, null, {\"type\": \"tool\", \"name\": \"inference_scaling_paper/submit_answer\", \"params\": {}}, false, \"\\n\\n\", false], \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "3", + "attempt_timeout": "1200", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {}, + "max_attempts": 3 + }, + "additional_details": { + "max_retries": "3" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/frontiermath/openai/gpt-5.4-2026-03-05/8f265e55-5ca3-46fa-8669-03a1a5147b15_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "59c8fd117e2cc61d294bf3921dc20bd0ecc464a42d2da1f10b451ceea6a26685", + "total_rows": 20 + } +} \ No newline at end of file diff --git a/data/frontiermath/openai/gpt-5.4-2026-03-05/8f265e55-5ca3-46fa-8669-03a1a5147b15_samples.jsonl b/data/frontiermath/openai/gpt-5.4-2026-03-05/8f265e55-5ca3-46fa-8669-03a1a5147b15_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..03fd7fcc854970f657ef897a9100ea3095773039 --- /dev/null +++ b/data/frontiermath/openai/gpt-5.4-2026-03-05/8f265e55-5ca3-46fa-8669-03a1a5147b15_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:59c8fd117e2cc61d294bf3921dc20bd0ecc464a42d2da1f10b451ceea6a26685 +size 7969186 diff --git a/data/frontiermath/openai/gpt-5.4-2026-03-05/922b6357-a72b-4447-8fa0-8e5a84bf7809.json b/data/frontiermath/openai/gpt-5.4-2026-03-05/922b6357-a72b-4447-8fa0-8e5a84bf7809.json new file mode 100644 index 0000000000000000000000000000000000000000..280b7be089defa5fd89cb8570d9f73ccb0e94813 --- /dev/null +++ b/data/frontiermath/openai/gpt-5.4-2026-03-05/922b6357-a72b-4447-8fa0-8e5a84bf7809.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "frontiermath/openai_gpt-5.4-2026-03-05/1786168273.360609", + "evaluation_timestamp": "1775733996.0", + "retrieved_timestamp": "1786168273.360609", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/frontiermath/openai/gpt-5.4-2026-03-05/922b6357-a72b-4447-8fa0-8e5a84bf7809_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "7b23f4834f6b1ae5eca2aa1822095c05cbd52d0d65de5eba1f606a9029054d98", + "total_rows": 40 + } +} \ No newline at end of file diff --git a/data/frontiermath/openai/gpt-5.4-2026-03-05/922b6357-a72b-4447-8fa0-8e5a84bf7809_samples.jsonl b/data/frontiermath/openai/gpt-5.4-2026-03-05/922b6357-a72b-4447-8fa0-8e5a84bf7809_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..5cfb69ef673853d83c6d815c965899ea5075f420 --- /dev/null +++ b/data/frontiermath/openai/gpt-5.4-2026-03-05/922b6357-a72b-4447-8fa0-8e5a84bf7809_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7b23f4834f6b1ae5eca2aa1822095c05cbd52d0d65de5eba1f606a9029054d98 +size 5029355 diff --git a/data/frontiermath/openai/gpt-5.4-2026-03-05/cc26d837-94a3-446b-b4f0-c56cf438822b.json b/data/frontiermath/openai/gpt-5.4-2026-03-05/cc26d837-94a3-446b-b4f0-c56cf438822b.json new file mode 100644 index 0000000000000000000000000000000000000000..e639fc73dd9348c00d21e3fb14a56bf02b1a33b3 --- /dev/null +++ b/data/frontiermath/openai/gpt-5.4-2026-03-05/cc26d837-94a3-446b-b4f0-c56cf438822b.json @@ -0,0 +1,115 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "frontiermath/openai_gpt-5.4-2026-03-05/1786163680.854002", + "evaluation_timestamp": "1776076028.0", + "retrieved_timestamp": "1786163680.854002", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on frontiermath/S-adaptive/+6ep for scorer verification_code", + "source_data": { + "dataset_name": "frontiermath", + "source_type": "hf_dataset", + "samples_number": 10, + "sample_ids": [ + "CWA2", + "CWD31", + "FMT1", + "LTI1", + "ONO3", + "OVE2", + "PLD1", + "RAP1", + "RSG1", + "TIK2" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1776076028.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.9166666666666666, + "uncertainty": { + "standard_error": { + "value": 0.03726779962499649 + }, + "num_samples": 60 + }, + "details": { + "total_matched_trajectories": "50", + "stopping_reason_count_repetition_guard": "50", + "avg_total_tokens_target_model": "446292.70", + "total_total_tokens_target_model": "22314635.0", + "avg_total_tokens_other_models": "117.00", + "total_total_tokens_other_models": "5850.0", + "avg_total_tokens_all_models": "446409.70", + "total_total_tokens_all_models": "22320485.0", + "avg_turn_count": "15.82", + "total_turn_count": "791.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": [null, null, {\"type\": \"tool\", \"name\": \"inference_scaling_paper/submit_answer\", \"params\": {}}, false, \"\\n\\n\", false], \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": [null, null, {\"type\": \"tool\", \"name\": \"inference_scaling_paper/submit_answer\", \"params\": {}}, false, \"\\n\\n\", false], \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/frontiermath/openai/gpt-5.4-2026-03-05/cc26d837-94a3-446b-b4f0-c56cf438822b_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "a4f1fb2dd882af760adcea46153d9aa9dceb2c0f94c64c23af92b49c1eb70b9e", + "total_rows": 60 + } +} \ No newline at end of file diff --git a/data/frontiermath/openai/gpt-5.4-2026-03-05/cc26d837-94a3-446b-b4f0-c56cf438822b_samples.jsonl b/data/frontiermath/openai/gpt-5.4-2026-03-05/cc26d837-94a3-446b-b4f0-c56cf438822b_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..ad9a43d08391a8377cf30ed39d346af5a653676e --- /dev/null +++ b/data/frontiermath/openai/gpt-5.4-2026-03-05/cc26d837-94a3-446b-b4f0-c56cf438822b_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a4f1fb2dd882af760adcea46153d9aa9dceb2c0f94c64c23af92b49c1eb70b9e +size 8884657 diff --git a/data/healthbench/anthropic/claude-opus-4-20250514/1f639db4-9cf2-47e6-9ccc-57d250c63a6f.json b/data/healthbench/anthropic/claude-opus-4-20250514/1f639db4-9cf2-47e6-9ccc-57d250c63a6f.json new file mode 100644 index 0000000000000000000000000000000000000000..7aa8ea44a8adfcf01fa991e3dbf69b8e97d85220 --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-20250514/1f639db4-9cf2-47e6-9ccc-57d250c63a6f.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/anthropic_claude-opus-4-20250514/1786166760.26322", + "evaluation_timestamp": "1775665083.0", + "retrieved_timestamp": "1786166760.26322", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-20250514", + "id": "anthropic/claude-opus-4-20250514", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/anthropic/claude-opus-4-20250514/1f639db4-9cf2-47e6-9ccc-57d250c63a6f_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "87e433f6531dab85b148ce73d0fc3a6284ddd678f91bdcdf6b6a3c3c82e0f8ec", + "total_rows": 231 + } +} \ No newline at end of file diff --git a/data/healthbench/anthropic/claude-opus-4-20250514/1f639db4-9cf2-47e6-9ccc-57d250c63a6f_samples.jsonl b/data/healthbench/anthropic/claude-opus-4-20250514/1f639db4-9cf2-47e6-9ccc-57d250c63a6f_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..6f8f5511926a470720c4ebbbb1c9703cfae052b4 --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-20250514/1f639db4-9cf2-47e6-9ccc-57d250c63a6f_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:87e433f6531dab85b148ce73d0fc3a6284ddd678f91bdcdf6b6a3c3c82e0f8ec +size 15931458 diff --git a/data/healthbench/anthropic/claude-opus-4-20250514/2079508e-db7c-4388-a494-5cc56c9f1296.json b/data/healthbench/anthropic/claude-opus-4-20250514/2079508e-db7c-4388-a494-5cc56c9f1296.json new file mode 100644 index 0000000000000000000000000000000000000000..51c592492f7f41e8d67e2a926ed7b9c434d7fbf4 --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-20250514/2079508e-db7c-4388-a494-5cc56c9f1296.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/anthropic_claude-opus-4-20250514/1786168659.654294", + "evaluation_timestamp": "1775646127.0", + "retrieved_timestamp": "1786168659.654294", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-20250514", + "id": "anthropic/claude-opus-4-20250514", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/anthropic/claude-opus-4-20250514/2079508e-db7c-4388-a494-5cc56c9f1296_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "b66c7c0abbd4848a1f1d7c4fe45d68336a8674a17d48e6a4aaa8bc3e01a3cc67", + "total_rows": 103 + } +} \ No newline at end of file diff --git a/data/healthbench/anthropic/claude-opus-4-20250514/2079508e-db7c-4388-a494-5cc56c9f1296_samples.jsonl b/data/healthbench/anthropic/claude-opus-4-20250514/2079508e-db7c-4388-a494-5cc56c9f1296_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..8237dd8e4b1fdb2353d6742f22981e47bfdabb2f --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-20250514/2079508e-db7c-4388-a494-5cc56c9f1296_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b66c7c0abbd4848a1f1d7c4fe45d68336a8674a17d48e6a4aaa8bc3e01a3cc67 +size 6163011 diff --git a/data/healthbench/anthropic/claude-opus-4-20250514/25827f1a-0fcc-466c-a792-a3b79b8d7257.json b/data/healthbench/anthropic/claude-opus-4-20250514/25827f1a-0fcc-466c-a792-a3b79b8d7257.json new file mode 100644 index 0000000000000000000000000000000000000000..a1fa0522a58e2b06870710926936bb193aef39db --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-20250514/25827f1a-0fcc-466c-a792-a3b79b8d7257.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/anthropic_claude-opus-4-20250514/1786166740.887247", + "evaluation_timestamp": "1775733995.0", + "retrieved_timestamp": "1786166740.887247", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-20250514", + "id": "anthropic/claude-opus-4-20250514", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/anthropic/claude-opus-4-20250514/25827f1a-0fcc-466c-a792-a3b79b8d7257_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "b8d44e5e053f3c90fa4e80ba52d0ceb019d882fd4260092b18935918bf1be91d", + "total_rows": 80 + } +} \ No newline at end of file diff --git a/data/healthbench/anthropic/claude-opus-4-20250514/25827f1a-0fcc-466c-a792-a3b79b8d7257_samples.jsonl b/data/healthbench/anthropic/claude-opus-4-20250514/25827f1a-0fcc-466c-a792-a3b79b8d7257_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..355022e413687daaa3009cedd025e93a39c6a26a --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-20250514/25827f1a-0fcc-466c-a792-a3b79b8d7257_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b8d44e5e053f3c90fa4e80ba52d0ceb019d882fd4260092b18935918bf1be91d +size 5680176 diff --git a/data/healthbench/anthropic/claude-opus-4-20250514/2eff1faa-53c6-432f-97ee-d98f780fd116.json b/data/healthbench/anthropic/claude-opus-4-20250514/2eff1faa-53c6-432f-97ee-d98f780fd116.json new file mode 100644 index 0000000000000000000000000000000000000000..770eebe071af62ba4b555ef91918ae1813e62523 --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-20250514/2eff1faa-53c6-432f-97ee-d98f780fd116.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/anthropic_claude-opus-4-20250514/1786168676.326648", + "evaluation_timestamp": "1776259832.0", + "retrieved_timestamp": "1786168676.326648", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-20250514", + "id": "anthropic/claude-opus-4-20250514", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/anthropic/claude-opus-4-20250514/2eff1faa-53c6-432f-97ee-d98f780fd116_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "4a75272343530073e98ae6bcffeb3ba7a0fa8a3ad70192524fa1128851bb0db3", + "total_rows": 300 + } +} \ No newline at end of file diff --git a/data/healthbench/anthropic/claude-opus-4-20250514/2eff1faa-53c6-432f-97ee-d98f780fd116_samples.jsonl b/data/healthbench/anthropic/claude-opus-4-20250514/2eff1faa-53c6-432f-97ee-d98f780fd116_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..f4e7331ac52085b227bae992b93af4f0ffee10d0 --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-20250514/2eff1faa-53c6-432f-97ee-d98f780fd116_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4a75272343530073e98ae6bcffeb3ba7a0fa8a3ad70192524fa1128851bb0db3 +size 18610256 diff --git a/data/healthbench/anthropic/claude-opus-4-20250514/32d55a94-b2ff-4378-990c-98e0f6b0fbc7.json b/data/healthbench/anthropic/claude-opus-4-20250514/32d55a94-b2ff-4378-990c-98e0f6b0fbc7.json new file mode 100644 index 0000000000000000000000000000000000000000..a78aac94752f2282a557efe857bf3ee3ccd1bb4f --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-20250514/32d55a94-b2ff-4378-990c-98e0f6b0fbc7.json @@ -0,0 +1,320 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/anthropic_claude-opus-4-20250514/1786164002.965401", + "evaluation_timestamp": "1776779722.0", + "retrieved_timestamp": "1786164002.965401", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-20250514", + "id": "anthropic/claude-opus-4-20250514", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "bootstrap_score on healthbench/S-adaptive/+1ep/bcd0f6be for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "20", + "sample_ids": "99b22b39-30d5-40dc-82d9-c1cd3c793a44,54349c88-e80a-4728-b181-305f97905d59,c63a4306-4003-45a6-8be1-d669457473a9,d3cbb579-70a5-46e8-95f6-72cc9afa4ba3,3852d050-c009-4e33-b953-fd734a5b5e73,5dc23430-dfbb-4794-9f77-0d4dd6093033,67e76447-f16a-4a65-aba7-fc7166fd9b63,a783e7a3-cecc-4d15-b18c-3f2ddfc2dc6c,62e9d35f-153d-4744-b97c-9abfb2a90b1e,04b6e0c8-2835-415e-865d-aca9f1cd3c57,7a283823-00ee-489d-8dbc-356e2857526f,e89e00c9-29e9-420c-813d-e06ddc39d585,51433193-b0c3-4725-a3a8-9dd15e969b0c,48c89cd1-0ecf-4738-8fcc-392269585bd3,92ea478e-6bf8-4ffb-b88a-f57ccf5f2346,36c6e780-d4a8-47b2-8a36-90ee4b97607f,5fb6812f-9816-4d5d-b7e0-4ec3e9df5dc1,c29c95d1-5b5e-4f53-9ec5-83e1f72b5aa5,b81c059a-1137-47cf-bab8-ab5eb8ff34d5,886864a6-a0e4-468f-aaf4-26cf8dc5e541" + } + }, + "evaluation_timestamp": "1776779722.0", + "metric_config": { + "evaluation_description": "bootstrap_score", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.16470137811168195, + "uncertainty": { + "standard_deviation": 0.04611975735947024, + "num_samples": 20 + }, + "details": { + "total_matched_trajectories": "20", + "stopping_reason_count_repetition_guard": "20", + "avg_total_tokens_target_model": "10262.00", + "total_total_tokens_target_model": "205240.0", + "avg_total_tokens_other_models": "53217.20", + "total_total_tokens_other_models": "1064344.0", + "avg_total_tokens_all_models": "63479.20", + "total_total_tokens_all_models": "1269584.0", + "avg_turn_count": "27.25", + "total_turn_count": "545.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + }, + { + "evaluation_name": "std on healthbench/S-adaptive/+1ep/bcd0f6be for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "20", + "sample_ids": "99b22b39-30d5-40dc-82d9-c1cd3c793a44,54349c88-e80a-4728-b181-305f97905d59,c63a4306-4003-45a6-8be1-d669457473a9,d3cbb579-70a5-46e8-95f6-72cc9afa4ba3,3852d050-c009-4e33-b953-fd734a5b5e73,5dc23430-dfbb-4794-9f77-0d4dd6093033,67e76447-f16a-4a65-aba7-fc7166fd9b63,a783e7a3-cecc-4d15-b18c-3f2ddfc2dc6c,62e9d35f-153d-4744-b97c-9abfb2a90b1e,04b6e0c8-2835-415e-865d-aca9f1cd3c57,7a283823-00ee-489d-8dbc-356e2857526f,e89e00c9-29e9-420c-813d-e06ddc39d585,51433193-b0c3-4725-a3a8-9dd15e969b0c,48c89cd1-0ecf-4738-8fcc-392269585bd3,92ea478e-6bf8-4ffb-b88a-f57ccf5f2346,36c6e780-d4a8-47b2-8a36-90ee4b97607f,5fb6812f-9816-4d5d-b7e0-4ec3e9df5dc1,c29c95d1-5b5e-4f53-9ec5-83e1f72b5aa5,b81c059a-1137-47cf-bab8-ab5eb8ff34d5,886864a6-a0e4-468f-aaf4-26cf8dc5e541" + } + }, + "evaluation_timestamp": "1776779722.0", + "metric_config": { + "evaluation_description": "std", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.04611975735947024, + "uncertainty": { + "standard_deviation": 0.04611975735947024, + "num_samples": 20 + }, + "details": { + "total_matched_trajectories": "20", + "stopping_reason_count_repetition_guard": "20", + "avg_total_tokens_target_model": "10262.00", + "total_total_tokens_target_model": "205240.0", + "avg_total_tokens_other_models": "53217.20", + "total_total_tokens_other_models": "1064344.0", + "avg_total_tokens_all_models": "63479.20", + "total_total_tokens_all_models": "1269584.0", + "avg_turn_count": "27.25", + "total_turn_count": "545.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + }, + { + "evaluation_name": "criteria_met_rate on healthbench/S-adaptive/+1ep/bcd0f6be for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "20", + "sample_ids": "99b22b39-30d5-40dc-82d9-c1cd3c793a44,54349c88-e80a-4728-b181-305f97905d59,c63a4306-4003-45a6-8be1-d669457473a9,d3cbb579-70a5-46e8-95f6-72cc9afa4ba3,3852d050-c009-4e33-b953-fd734a5b5e73,5dc23430-dfbb-4794-9f77-0d4dd6093033,67e76447-f16a-4a65-aba7-fc7166fd9b63,a783e7a3-cecc-4d15-b18c-3f2ddfc2dc6c,62e9d35f-153d-4744-b97c-9abfb2a90b1e,04b6e0c8-2835-415e-865d-aca9f1cd3c57,7a283823-00ee-489d-8dbc-356e2857526f,e89e00c9-29e9-420c-813d-e06ddc39d585,51433193-b0c3-4725-a3a8-9dd15e969b0c,48c89cd1-0ecf-4738-8fcc-392269585bd3,92ea478e-6bf8-4ffb-b88a-f57ccf5f2346,36c6e780-d4a8-47b2-8a36-90ee4b97607f,5fb6812f-9816-4d5d-b7e0-4ec3e9df5dc1,c29c95d1-5b5e-4f53-9ec5-83e1f72b5aa5,b81c059a-1137-47cf-bab8-ab5eb8ff34d5,886864a6-a0e4-468f-aaf4-26cf8dc5e541" + } + }, + "evaluation_timestamp": "1776779722.0", + "metric_config": { + "evaluation_description": "criteria_met_rate", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.29831932773109243, + "uncertainty": { + "standard_deviation": 0.04611975735947024, + "num_samples": 20 + }, + "details": { + "total_matched_trajectories": "20", + "stopping_reason_count_repetition_guard": "20", + "avg_total_tokens_target_model": "10262.00", + "total_total_tokens_target_model": "205240.0", + "avg_total_tokens_other_models": "53217.20", + "total_total_tokens_other_models": "1064344.0", + "avg_total_tokens_all_models": "63479.20", + "total_total_tokens_all_models": "1269584.0", + "avg_turn_count": "27.25", + "total_turn_count": "545.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + }, + { + "evaluation_name": "mean on healthbench/S-adaptive/+1ep/bcd0f6be for scorer healthbench_score", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "20", + "sample_ids": "99b22b39-30d5-40dc-82d9-c1cd3c793a44,54349c88-e80a-4728-b181-305f97905d59,c63a4306-4003-45a6-8be1-d669457473a9,d3cbb579-70a5-46e8-95f6-72cc9afa4ba3,3852d050-c009-4e33-b953-fd734a5b5e73,5dc23430-dfbb-4794-9f77-0d4dd6093033,67e76447-f16a-4a65-aba7-fc7166fd9b63,a783e7a3-cecc-4d15-b18c-3f2ddfc2dc6c,62e9d35f-153d-4744-b97c-9abfb2a90b1e,04b6e0c8-2835-415e-865d-aca9f1cd3c57,7a283823-00ee-489d-8dbc-356e2857526f,e89e00c9-29e9-420c-813d-e06ddc39d585,51433193-b0c3-4725-a3a8-9dd15e969b0c,48c89cd1-0ecf-4738-8fcc-392269585bd3,92ea478e-6bf8-4ffb-b88a-f57ccf5f2346,36c6e780-d4a8-47b2-8a36-90ee4b97607f,5fb6812f-9816-4d5d-b7e0-4ec3e9df5dc1,c29c95d1-5b5e-4f53-9ec5-83e1f72b5aa5,b81c059a-1137-47cf-bab8-ab5eb8ff34d5,886864a6-a0e4-468f-aaf4-26cf8dc5e541" + } + }, + "evaluation_timestamp": "1776779722.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.16470137811168195, + "uncertainty": { + "num_samples": 20 + }, + "details": { + "total_matched_trajectories": "20", + "stopping_reason_count_repetition_guard": "20", + "avg_total_tokens_target_model": "10262.00", + "total_total_tokens_target_model": "205240.0", + "avg_total_tokens_other_models": "53217.20", + "total_total_tokens_other_models": "1064344.0", + "avg_total_tokens_all_models": "63479.20", + "total_total_tokens_all_models": "1269584.0", + "avg_turn_count": "27.25", + "total_turn_count": "545.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/anthropic/claude-opus-4-20250514/32d55a94-b2ff-4378-990c-98e0f6b0fbc7_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "c6aaa01277b3edfb042a1b251112cf68337caf106ef37960a487baa936369a94", + "total_rows": 20 + } +} \ No newline at end of file diff --git a/data/healthbench/anthropic/claude-opus-4-20250514/32d55a94-b2ff-4378-990c-98e0f6b0fbc7_samples.jsonl b/data/healthbench/anthropic/claude-opus-4-20250514/32d55a94-b2ff-4378-990c-98e0f6b0fbc7_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..b0fc62804a3a8d463a8159dacdfcfbee48e69e43 --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-20250514/32d55a94-b2ff-4378-990c-98e0f6b0fbc7_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c6aaa01277b3edfb042a1b251112cf68337caf106ef37960a487baa936369a94 +size 1352558 diff --git a/data/healthbench/anthropic/claude-opus-4-20250514/33c5023a-011f-4d05-9cfb-209650d0a787.json b/data/healthbench/anthropic/claude-opus-4-20250514/33c5023a-011f-4d05-9cfb-209650d0a787.json new file mode 100644 index 0000000000000000000000000000000000000000..d06d61d03e536cc77519a3501997fcf4a50a123a --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-20250514/33c5023a-011f-4d05-9cfb-209650d0a787.json @@ -0,0 +1,316 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/anthropic_claude-opus-4-20250514/1786164006.225737", + "evaluation_timestamp": "1776682987.0", + "retrieved_timestamp": "1786164006.225737", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-20250514", + "id": "anthropic/claude-opus-4-20250514", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "bootstrap_score on healthbench/S-adaptive/+2ep/3523d247 for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "18", + "sample_ids": "c3eeecfc-de96-480f-929d-ac4a5e5a541c,269074dc-9495-484b-b2f9-0f0943c0f816,b97b1bae-b09f-4073-8e16-540bba766c60,8e88f21a-24c4-4f9c-b848-faf5e10e31b9,a49eb9af-926e-4f7c-94d0-133729059116,1049130c-e9c9-461d-b080-90027bc011c0,13bb02e4-844e-4972-9200-2befa5f911d8,2694a753-2f57-41eb-a7f9-8c4c7acdd2e8,b160df9f-be98-4a4d-9cc8-11a612ffaf87,5e0a84d7-b972-484c-a649-469b15f0a90f,c7606b1b-ca4d-4a71-8277-bcddff09a653,20cdebb7-9189-4f09-ad21-bfc8e8b75680,f24935ac-0530-4063-b2f1-ac45cd265037,1b8f338a-2e98-4a42-ba80-cba266f4b9fa,6f827e79-05c4-4383-8ed4-16d36260f29f,c835bebe-e998-4dcd-843a-2de03d4aabc6,70afd1ac-85ea-4455-a5a2-15b35378e10a,aa7a760a-5173-4a39-af83-48d6ae2230f0" + } + }, + "evaluation_timestamp": "1776682987.0", + "metric_config": { + "evaluation_description": "bootstrap_score", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.15747113635599683, + "uncertainty": { + "standard_deviation": 0.06751224916506945, + "num_samples": 36 + }, + "details": { + "total_matched_trajectories": "36", + "stopping_reason_count_repetition_guard": "36", + "avg_total_tokens_target_model": "9225.19", + "total_total_tokens_target_model": "332107.0", + "avg_total_tokens_other_models": "61537.97", + "total_total_tokens_other_models": "2215367.0", + "avg_total_tokens_all_models": "70763.17", + "total_total_tokens_all_models": "2547474.0", + "avg_turn_count": "33.94", + "total_turn_count": "1222.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "3600", + "attempt_timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "3600", + "max_connections": "10" + } + } + }, + { + "evaluation_name": "std on healthbench/S-adaptive/+2ep/3523d247 for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "18", + "sample_ids": "c3eeecfc-de96-480f-929d-ac4a5e5a541c,269074dc-9495-484b-b2f9-0f0943c0f816,b97b1bae-b09f-4073-8e16-540bba766c60,8e88f21a-24c4-4f9c-b848-faf5e10e31b9,a49eb9af-926e-4f7c-94d0-133729059116,1049130c-e9c9-461d-b080-90027bc011c0,13bb02e4-844e-4972-9200-2befa5f911d8,2694a753-2f57-41eb-a7f9-8c4c7acdd2e8,b160df9f-be98-4a4d-9cc8-11a612ffaf87,5e0a84d7-b972-484c-a649-469b15f0a90f,c7606b1b-ca4d-4a71-8277-bcddff09a653,20cdebb7-9189-4f09-ad21-bfc8e8b75680,f24935ac-0530-4063-b2f1-ac45cd265037,1b8f338a-2e98-4a42-ba80-cba266f4b9fa,6f827e79-05c4-4383-8ed4-16d36260f29f,c835bebe-e998-4dcd-843a-2de03d4aabc6,70afd1ac-85ea-4455-a5a2-15b35378e10a,aa7a760a-5173-4a39-af83-48d6ae2230f0" + } + }, + "evaluation_timestamp": "1776682987.0", + "metric_config": { + "evaluation_description": "std", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.06751224916506945, + "uncertainty": { + "standard_deviation": 0.06751224916506945, + "num_samples": 36 + }, + "details": { + "total_matched_trajectories": "36", + "stopping_reason_count_repetition_guard": "36", + "avg_total_tokens_target_model": "9225.19", + "total_total_tokens_target_model": "332107.0", + "avg_total_tokens_other_models": "61537.97", + "total_total_tokens_other_models": "2215367.0", + "avg_total_tokens_all_models": "70763.17", + "total_total_tokens_all_models": "2547474.0", + "avg_turn_count": "33.94", + "total_turn_count": "1222.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "3600", + "attempt_timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "3600", + "max_connections": "10" + } + } + }, + { + "evaluation_name": "criteria_met_rate on healthbench/S-adaptive/+2ep/3523d247 for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "18", + "sample_ids": "c3eeecfc-de96-480f-929d-ac4a5e5a541c,269074dc-9495-484b-b2f9-0f0943c0f816,b97b1bae-b09f-4073-8e16-540bba766c60,8e88f21a-24c4-4f9c-b848-faf5e10e31b9,a49eb9af-926e-4f7c-94d0-133729059116,1049130c-e9c9-461d-b080-90027bc011c0,13bb02e4-844e-4972-9200-2befa5f911d8,2694a753-2f57-41eb-a7f9-8c4c7acdd2e8,b160df9f-be98-4a4d-9cc8-11a612ffaf87,5e0a84d7-b972-484c-a649-469b15f0a90f,c7606b1b-ca4d-4a71-8277-bcddff09a653,20cdebb7-9189-4f09-ad21-bfc8e8b75680,f24935ac-0530-4063-b2f1-ac45cd265037,1b8f338a-2e98-4a42-ba80-cba266f4b9fa,6f827e79-05c4-4383-8ed4-16d36260f29f,c835bebe-e998-4dcd-843a-2de03d4aabc6,70afd1ac-85ea-4455-a5a2-15b35378e10a,aa7a760a-5173-4a39-af83-48d6ae2230f0" + } + }, + "evaluation_timestamp": "1776682987.0", + "metric_config": { + "evaluation_description": "criteria_met_rate", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.3104166666666667, + "uncertainty": { + "standard_deviation": 0.06751224916506945, + "num_samples": 36 + }, + "details": { + "total_matched_trajectories": "36", + "stopping_reason_count_repetition_guard": "36", + "avg_total_tokens_target_model": "9225.19", + "total_total_tokens_target_model": "332107.0", + "avg_total_tokens_other_models": "61537.97", + "total_total_tokens_other_models": "2215367.0", + "avg_total_tokens_all_models": "70763.17", + "total_total_tokens_all_models": "2547474.0", + "avg_turn_count": "33.94", + "total_turn_count": "1222.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "3600", + "attempt_timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "3600", + "max_connections": "10" + } + } + }, + { + "evaluation_name": "mean on healthbench/S-adaptive/+2ep/3523d247 for scorer healthbench_score", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "18", + "sample_ids": "c3eeecfc-de96-480f-929d-ac4a5e5a541c,269074dc-9495-484b-b2f9-0f0943c0f816,b97b1bae-b09f-4073-8e16-540bba766c60,8e88f21a-24c4-4f9c-b848-faf5e10e31b9,a49eb9af-926e-4f7c-94d0-133729059116,1049130c-e9c9-461d-b080-90027bc011c0,13bb02e4-844e-4972-9200-2befa5f911d8,2694a753-2f57-41eb-a7f9-8c4c7acdd2e8,b160df9f-be98-4a4d-9cc8-11a612ffaf87,5e0a84d7-b972-484c-a649-469b15f0a90f,c7606b1b-ca4d-4a71-8277-bcddff09a653,20cdebb7-9189-4f09-ad21-bfc8e8b75680,f24935ac-0530-4063-b2f1-ac45cd265037,1b8f338a-2e98-4a42-ba80-cba266f4b9fa,6f827e79-05c4-4383-8ed4-16d36260f29f,c835bebe-e998-4dcd-843a-2de03d4aabc6,70afd1ac-85ea-4455-a5a2-15b35378e10a,aa7a760a-5173-4a39-af83-48d6ae2230f0" + } + }, + "evaluation_timestamp": "1776682987.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.15747113635599683, + "uncertainty": { + "num_samples": 36 + }, + "details": { + "total_matched_trajectories": "36", + "stopping_reason_count_repetition_guard": "36", + "avg_total_tokens_target_model": "9225.19", + "total_total_tokens_target_model": "332107.0", + "avg_total_tokens_other_models": "61537.97", + "total_total_tokens_other_models": "2215367.0", + "avg_total_tokens_all_models": "70763.17", + "total_total_tokens_all_models": "2547474.0", + "avg_turn_count": "33.94", + "total_turn_count": "1222.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "3600", + "attempt_timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "3600", + "max_connections": "10" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/anthropic/claude-opus-4-20250514/33c5023a-011f-4d05-9cfb-209650d0a787_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "8431d5e360fab9b94fc335d237b50d5ae752aeb8d91aab0bdd5a10f92d06b827", + "total_rows": 36 + } +} \ No newline at end of file diff --git a/data/healthbench/anthropic/claude-opus-4-20250514/33c5023a-011f-4d05-9cfb-209650d0a787_samples.jsonl b/data/healthbench/anthropic/claude-opus-4-20250514/33c5023a-011f-4d05-9cfb-209650d0a787_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..190ec10649b026a9afff171f0e1d238d27714ba9 --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-20250514/33c5023a-011f-4d05-9cfb-209650d0a787_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8431d5e360fab9b94fc335d237b50d5ae752aeb8d91aab0bdd5a10f92d06b827 +size 2648839 diff --git a/data/healthbench/anthropic/claude-opus-4-20250514/520232ba-d29b-4f34-a474-fe68aaf93a67.json b/data/healthbench/anthropic/claude-opus-4-20250514/520232ba-d29b-4f34-a474-fe68aaf93a67.json new file mode 100644 index 0000000000000000000000000000000000000000..c696b0d13b8363820887bf3a08e95136db8d7250 --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-20250514/520232ba-d29b-4f34-a474-fe68aaf93a67.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/anthropic_claude-opus-4-20250514/1786168645.145171", + "evaluation_timestamp": "1775604415.0", + "retrieved_timestamp": "1786168645.145171", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.202,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-20250514", + "id": "anthropic/claude-opus-4-20250514", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/anthropic/claude-opus-4-20250514/520232ba-d29b-4f34-a474-fe68aaf93a67_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "c05ac8ee4dd7efbd44a5d5a334945b30bf240694cda0c176b79bf41496078202", + "total_rows": 100 + } +} \ No newline at end of file diff --git a/data/healthbench/anthropic/claude-opus-4-20250514/520232ba-d29b-4f34-a474-fe68aaf93a67_samples.jsonl b/data/healthbench/anthropic/claude-opus-4-20250514/520232ba-d29b-4f34-a474-fe68aaf93a67_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..1c98ced808c9e1960c873c8f9de2f03b17493a62 --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-20250514/520232ba-d29b-4f34-a474-fe68aaf93a67_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c05ac8ee4dd7efbd44a5d5a334945b30bf240694cda0c176b79bf41496078202 +size 6115815 diff --git a/data/healthbench/anthropic/claude-opus-4-20250514/54678cb0-5e81-4c48-955e-5f28d3e78b7e.json b/data/healthbench/anthropic/claude-opus-4-20250514/54678cb0-5e81-4c48-955e-5f28d3e78b7e.json new file mode 100644 index 0000000000000000000000000000000000000000..b5c9c78bd43f01af7d4567d8b2d3ddcc05843d4c --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-20250514/54678cb0-5e81-4c48-955e-5f28d3e78b7e.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/anthropic_claude-opus-4-20250514/1786168651.403012", + "evaluation_timestamp": "1776076027.0", + "retrieved_timestamp": "1786168651.403012", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-20250514", + "id": "anthropic/claude-opus-4-20250514", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/anthropic/claude-opus-4-20250514/54678cb0-5e81-4c48-955e-5f28d3e78b7e_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "2ab9c630c0386743b7fb4e161a39f706ad6f1814e8da1b33f0519962d517595a", + "total_rows": 20 + } +} \ No newline at end of file diff --git a/data/healthbench/anthropic/claude-opus-4-20250514/54678cb0-5e81-4c48-955e-5f28d3e78b7e_samples.jsonl b/data/healthbench/anthropic/claude-opus-4-20250514/54678cb0-5e81-4c48-955e-5f28d3e78b7e_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..a6c91438f2dabd22f7524348dcda64952882ebd5 --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-20250514/54678cb0-5e81-4c48-955e-5f28d3e78b7e_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2ab9c630c0386743b7fb4e161a39f706ad6f1814e8da1b33f0519962d517595a +size 1391035 diff --git a/data/healthbench/anthropic/claude-opus-4-20250514/79fef613-b36a-4bc5-8fff-32ad69d2250b.json b/data/healthbench/anthropic/claude-opus-4-20250514/79fef613-b36a-4bc5-8fff-32ad69d2250b.json new file mode 100644 index 0000000000000000000000000000000000000000..ae7558659b4f43914b5506eee0dcbf2fcf13a8ed --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-20250514/79fef613-b36a-4bc5-8fff-32ad69d2250b.json @@ -0,0 +1,292 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/anthropic_claude-opus-4-20250514/1786164021.027369", + "evaluation_timestamp": "1776259833.0", + "retrieved_timestamp": "1786164021.027369", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-20250514", + "id": "anthropic/claude-opus-4-20250514", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "bootstrap_score on healthbench/S-adaptive/+6ep/f1c68b87 for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "41", + "sample_ids": "77dff7ec-2e02-4b93-86d6-5d7765b8dd3d,9ab66439-8090-4f6c-ba74-874591dfd1a6,762d9eca-4835-4447-a649-f91740c74bd0,5c1cf475-6fff-42e2-8508-7e1a14ea9f65,06fb659f-a4aa-4a97-a831-ed2cd2aa7e69,608770a0-440d-4349-9a1c-863e9f4d3e24,a58987e9-8b59-461a-a68b-20efa1d37d51,2d878243-5d98-489a-b8f2-8809f813b6cf,452a9534-b4fc-4483-94e0-d34b06b8e299,e690c779-3cb8-4271-8541-a699bb1bd475,d984a2a8-1209-42bb-8f59-001421c96f8f,5caa2d1f-3ab9-4e25-97a2-af29c2dbbcfb,7b1d2983-ab41-48f2-bdcd-454cbf3f81b8,b267896f-cfc0-4bbe-9fda-fdfa287d1f43,ff514fc4-bd34-4a3f-be3f-840b0e1f8091,a131d5e7-2a27-426f-90f0-25eae3c4b490,620b9bf1-f0f3-4d68-aa74-1560fb406a68,9b4c14e9-c404-4646-87b7-d3badae8a68b,9385d26b-bf44-4ff7-9410-e181413540dd,a57a4b5e-0f12-4103-8852-72bd2b7dba24,cf8040e2-39e3-4200-ab7f-b9b532310ac3,4fc7b71f-9591-4ae1-bfca-2929c7225182,da05b57a-d776-45b6-96af-1750e0254eec,ef446d15-d019-4c5a-a052-ae8cc3ea3b81,0fe4ea94-633d-4f73-985b-cbcd9f5a0270,cbb0ca6e-aa5e-4683-b107-d7dfa4cb60ce,ae6ec2ce-5c23-4e65-802f-cef9f1624d71,cd3dbc9f-a39c-4d13-af60-cba7dbc3baad,02423ab9-f3ed-4096-a7df-bfa3ebb40c7c,084b0c27-79c5-42f6-9cbd-690d0e79167a,42465540-8af1-423b-9b35-a31de1e1d49e,74a50705-7db2-44be-84dc-92b2d960d344,3b79284e-1143-46b4-b6a0-78e5448ff773,afbfc79d-6455-40b8-b475-1bcde4ef0cc9,6d705f80-8259-4ca1-8645-163f42c85a3c,129057dd-4bf4-472d-a8cf-a063dbed60ad,2833344d-bff9-4bbe-8fab-6b13d64c6b2a,48669cdc-1a64-4c99-a6e8-9a7181341fa2,be70739f-234a-4a0f-9b14-d15b8800577b,4047884e-4629-457c-9d77-130e937fae60,eb97bae4-430e-45cd-a065-2df3ab5c600e" + } + }, + "evaluation_timestamp": "1776259833.0", + "metric_config": { + "evaluation_description": "bootstrap_score", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.2478166311997635, + "uncertainty": { + "standard_deviation": 0.03141629629739583, + "num_samples": 246 + }, + "details": { + "total_matched_trajectories": "35", + "stopping_reason_count_repetition_guard": "35", + "avg_total_tokens_target_model": "9195.49", + "total_total_tokens_target_model": "321842.0", + "avg_total_tokens_other_models": "58641.60", + "total_total_tokens_other_models": "2052456.0", + "avg_total_tokens_all_models": "67837.09", + "total_total_tokens_all_models": "2374298.0", + "avg_turn_count": "30.06", + "total_turn_count": "1052.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + } + } + }, + { + "evaluation_name": "std on healthbench/S-adaptive/+6ep/f1c68b87 for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "41", + "sample_ids": "77dff7ec-2e02-4b93-86d6-5d7765b8dd3d,9ab66439-8090-4f6c-ba74-874591dfd1a6,762d9eca-4835-4447-a649-f91740c74bd0,5c1cf475-6fff-42e2-8508-7e1a14ea9f65,06fb659f-a4aa-4a97-a831-ed2cd2aa7e69,608770a0-440d-4349-9a1c-863e9f4d3e24,a58987e9-8b59-461a-a68b-20efa1d37d51,2d878243-5d98-489a-b8f2-8809f813b6cf,452a9534-b4fc-4483-94e0-d34b06b8e299,e690c779-3cb8-4271-8541-a699bb1bd475,d984a2a8-1209-42bb-8f59-001421c96f8f,5caa2d1f-3ab9-4e25-97a2-af29c2dbbcfb,7b1d2983-ab41-48f2-bdcd-454cbf3f81b8,b267896f-cfc0-4bbe-9fda-fdfa287d1f43,ff514fc4-bd34-4a3f-be3f-840b0e1f8091,a131d5e7-2a27-426f-90f0-25eae3c4b490,620b9bf1-f0f3-4d68-aa74-1560fb406a68,9b4c14e9-c404-4646-87b7-d3badae8a68b,9385d26b-bf44-4ff7-9410-e181413540dd,a57a4b5e-0f12-4103-8852-72bd2b7dba24,cf8040e2-39e3-4200-ab7f-b9b532310ac3,4fc7b71f-9591-4ae1-bfca-2929c7225182,da05b57a-d776-45b6-96af-1750e0254eec,ef446d15-d019-4c5a-a052-ae8cc3ea3b81,0fe4ea94-633d-4f73-985b-cbcd9f5a0270,cbb0ca6e-aa5e-4683-b107-d7dfa4cb60ce,ae6ec2ce-5c23-4e65-802f-cef9f1624d71,cd3dbc9f-a39c-4d13-af60-cba7dbc3baad,02423ab9-f3ed-4096-a7df-bfa3ebb40c7c,084b0c27-79c5-42f6-9cbd-690d0e79167a,42465540-8af1-423b-9b35-a31de1e1d49e,74a50705-7db2-44be-84dc-92b2d960d344,3b79284e-1143-46b4-b6a0-78e5448ff773,afbfc79d-6455-40b8-b475-1bcde4ef0cc9,6d705f80-8259-4ca1-8645-163f42c85a3c,129057dd-4bf4-472d-a8cf-a063dbed60ad,2833344d-bff9-4bbe-8fab-6b13d64c6b2a,48669cdc-1a64-4c99-a6e8-9a7181341fa2,be70739f-234a-4a0f-9b14-d15b8800577b,4047884e-4629-457c-9d77-130e937fae60,eb97bae4-430e-45cd-a065-2df3ab5c600e" + } + }, + "evaluation_timestamp": "1776259833.0", + "metric_config": { + "evaluation_description": "std", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.03141629629739583, + "uncertainty": { + "standard_deviation": 0.03141629629739583, + "num_samples": 246 + }, + "details": { + "total_matched_trajectories": "35", + "stopping_reason_count_repetition_guard": "35", + "avg_total_tokens_target_model": "9195.49", + "total_total_tokens_target_model": "321842.0", + "avg_total_tokens_other_models": "58641.60", + "total_total_tokens_other_models": "2052456.0", + "avg_total_tokens_all_models": "67837.09", + "total_total_tokens_all_models": "2374298.0", + "avg_turn_count": "30.06", + "total_turn_count": "1052.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + } + } + }, + { + "evaluation_name": "criteria_met_rate on healthbench/S-adaptive/+6ep/f1c68b87 for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "41", + "sample_ids": "77dff7ec-2e02-4b93-86d6-5d7765b8dd3d,9ab66439-8090-4f6c-ba74-874591dfd1a6,762d9eca-4835-4447-a649-f91740c74bd0,5c1cf475-6fff-42e2-8508-7e1a14ea9f65,06fb659f-a4aa-4a97-a831-ed2cd2aa7e69,608770a0-440d-4349-9a1c-863e9f4d3e24,a58987e9-8b59-461a-a68b-20efa1d37d51,2d878243-5d98-489a-b8f2-8809f813b6cf,452a9534-b4fc-4483-94e0-d34b06b8e299,e690c779-3cb8-4271-8541-a699bb1bd475,d984a2a8-1209-42bb-8f59-001421c96f8f,5caa2d1f-3ab9-4e25-97a2-af29c2dbbcfb,7b1d2983-ab41-48f2-bdcd-454cbf3f81b8,b267896f-cfc0-4bbe-9fda-fdfa287d1f43,ff514fc4-bd34-4a3f-be3f-840b0e1f8091,a131d5e7-2a27-426f-90f0-25eae3c4b490,620b9bf1-f0f3-4d68-aa74-1560fb406a68,9b4c14e9-c404-4646-87b7-d3badae8a68b,9385d26b-bf44-4ff7-9410-e181413540dd,a57a4b5e-0f12-4103-8852-72bd2b7dba24,cf8040e2-39e3-4200-ab7f-b9b532310ac3,4fc7b71f-9591-4ae1-bfca-2929c7225182,da05b57a-d776-45b6-96af-1750e0254eec,ef446d15-d019-4c5a-a052-ae8cc3ea3b81,0fe4ea94-633d-4f73-985b-cbcd9f5a0270,cbb0ca6e-aa5e-4683-b107-d7dfa4cb60ce,ae6ec2ce-5c23-4e65-802f-cef9f1624d71,cd3dbc9f-a39c-4d13-af60-cba7dbc3baad,02423ab9-f3ed-4096-a7df-bfa3ebb40c7c,084b0c27-79c5-42f6-9cbd-690d0e79167a,42465540-8af1-423b-9b35-a31de1e1d49e,74a50705-7db2-44be-84dc-92b2d960d344,3b79284e-1143-46b4-b6a0-78e5448ff773,afbfc79d-6455-40b8-b475-1bcde4ef0cc9,6d705f80-8259-4ca1-8645-163f42c85a3c,129057dd-4bf4-472d-a8cf-a063dbed60ad,2833344d-bff9-4bbe-8fab-6b13d64c6b2a,48669cdc-1a64-4c99-a6e8-9a7181341fa2,be70739f-234a-4a0f-9b14-d15b8800577b,4047884e-4629-457c-9d77-130e937fae60,eb97bae4-430e-45cd-a065-2df3ab5c600e" + } + }, + "evaluation_timestamp": "1776259833.0", + "metric_config": { + "evaluation_description": "criteria_met_rate", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.3142946317103621, + "uncertainty": { + "standard_deviation": 0.03141629629739583, + "num_samples": 246 + }, + "details": { + "total_matched_trajectories": "35", + "stopping_reason_count_repetition_guard": "35", + "avg_total_tokens_target_model": "9195.49", + "total_total_tokens_target_model": "321842.0", + "avg_total_tokens_other_models": "58641.60", + "total_total_tokens_other_models": "2052456.0", + "avg_total_tokens_all_models": "67837.09", + "total_total_tokens_all_models": "2374298.0", + "avg_turn_count": "30.06", + "total_turn_count": "1052.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + } + } + }, + { + "evaluation_name": "mean on healthbench/S-adaptive/+6ep/f1c68b87 for scorer healthbench_score", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "41", + "sample_ids": "77dff7ec-2e02-4b93-86d6-5d7765b8dd3d,9ab66439-8090-4f6c-ba74-874591dfd1a6,762d9eca-4835-4447-a649-f91740c74bd0,5c1cf475-6fff-42e2-8508-7e1a14ea9f65,06fb659f-a4aa-4a97-a831-ed2cd2aa7e69,608770a0-440d-4349-9a1c-863e9f4d3e24,a58987e9-8b59-461a-a68b-20efa1d37d51,2d878243-5d98-489a-b8f2-8809f813b6cf,452a9534-b4fc-4483-94e0-d34b06b8e299,e690c779-3cb8-4271-8541-a699bb1bd475,d984a2a8-1209-42bb-8f59-001421c96f8f,5caa2d1f-3ab9-4e25-97a2-af29c2dbbcfb,7b1d2983-ab41-48f2-bdcd-454cbf3f81b8,b267896f-cfc0-4bbe-9fda-fdfa287d1f43,ff514fc4-bd34-4a3f-be3f-840b0e1f8091,a131d5e7-2a27-426f-90f0-25eae3c4b490,620b9bf1-f0f3-4d68-aa74-1560fb406a68,9b4c14e9-c404-4646-87b7-d3badae8a68b,9385d26b-bf44-4ff7-9410-e181413540dd,a57a4b5e-0f12-4103-8852-72bd2b7dba24,cf8040e2-39e3-4200-ab7f-b9b532310ac3,4fc7b71f-9591-4ae1-bfca-2929c7225182,da05b57a-d776-45b6-96af-1750e0254eec,ef446d15-d019-4c5a-a052-ae8cc3ea3b81,0fe4ea94-633d-4f73-985b-cbcd9f5a0270,cbb0ca6e-aa5e-4683-b107-d7dfa4cb60ce,ae6ec2ce-5c23-4e65-802f-cef9f1624d71,cd3dbc9f-a39c-4d13-af60-cba7dbc3baad,02423ab9-f3ed-4096-a7df-bfa3ebb40c7c,084b0c27-79c5-42f6-9cbd-690d0e79167a,42465540-8af1-423b-9b35-a31de1e1d49e,74a50705-7db2-44be-84dc-92b2d960d344,3b79284e-1143-46b4-b6a0-78e5448ff773,afbfc79d-6455-40b8-b475-1bcde4ef0cc9,6d705f80-8259-4ca1-8645-163f42c85a3c,129057dd-4bf4-472d-a8cf-a063dbed60ad,2833344d-bff9-4bbe-8fab-6b13d64c6b2a,48669cdc-1a64-4c99-a6e8-9a7181341fa2,be70739f-234a-4a0f-9b14-d15b8800577b,4047884e-4629-457c-9d77-130e937fae60,eb97bae4-430e-45cd-a065-2df3ab5c600e" + } + }, + "evaluation_timestamp": "1776259833.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.2478166311997635, + "uncertainty": { + "num_samples": 246 + }, + "details": { + "total_matched_trajectories": "35", + "stopping_reason_count_repetition_guard": "35", + "avg_total_tokens_target_model": "9195.49", + "total_total_tokens_target_model": "321842.0", + "avg_total_tokens_other_models": "58641.60", + "total_total_tokens_other_models": "2052456.0", + "avg_total_tokens_all_models": "67837.09", + "total_total_tokens_all_models": "2374298.0", + "avg_turn_count": "30.06", + "total_turn_count": "1052.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/anthropic/claude-opus-4-20250514/79fef613-b36a-4bc5-8fff-32ad69d2250b_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "7447911938d044c71ea3c86923ce1ab16f3fa856ec747946c7535319a8a09203", + "total_rows": 246 + } +} \ No newline at end of file diff --git a/data/healthbench/anthropic/claude-opus-4-20250514/79fef613-b36a-4bc5-8fff-32ad69d2250b_samples.jsonl b/data/healthbench/anthropic/claude-opus-4-20250514/79fef613-b36a-4bc5-8fff-32ad69d2250b_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..21288943f9dfcbc56c949e9d7fc141ae417b9873 --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-20250514/79fef613-b36a-4bc5-8fff-32ad69d2250b_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7447911938d044c71ea3c86923ce1ab16f3fa856ec747946c7535319a8a09203 +size 16235786 diff --git a/data/healthbench/anthropic/claude-opus-4-20250514/7a7fc16f-3232-4796-95e4-9311c88b650f.json b/data/healthbench/anthropic/claude-opus-4-20250514/7a7fc16f-3232-4796-95e4-9311c88b650f.json new file mode 100644 index 0000000000000000000000000000000000000000..60f5306789485b4d2353a39f4c5bb9cbd3288ce8 --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-20250514/7a7fc16f-3232-4796-95e4-9311c88b650f.json @@ -0,0 +1,292 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/anthropic_claude-opus-4-20250514/1786164028.299496", + "evaluation_timestamp": "1776165243.0", + "retrieved_timestamp": "1786164028.299496", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-20250514", + "id": "anthropic/claude-opus-4-20250514", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "bootstrap_score on healthbench/S-adaptive/+7ep for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "20", + "sample_ids": "19ec4833-86e9-4166-8b82-d1da09f31fd7,7ebc830a-8dbd-489b-9d61-4d8bacf0db8d,c971f9d1-5f6a-464e-b282-41c8f0d82f58,5a6e4a41-3ea6-4050-a971-93433fe34877,78510411-e03b-4f93-a369-c7e72d2e4908,bbd759b4-8e4d-4db7-bc36-f2e069396c1c,8f2a65de-dea7-48e8-8adb-6194eca26c08,6d5f483c-3e86-456d-bfd5-4e28de699ae5,f05491d8-d160-4b2b-bd5b-59d757585e39,85408180-238e-4c1b-92c1-55d8ce6c3717,ecd7453e-1dc1-4e52-bb95-67e8e5788573,0e073591-b3e7-4dc8-94d0-dc5aa93ca35c,7637b987-8fe2-49d0-8372-1ebf17284d70,5b294937-13e5-424e-8bb4-5d1904a2344a,c5bf7fc3-dfb0-4b47-9bbe-222e87b952e2,83cf8f2d-2857-4f01-a283-9595d8f4ae8e,7a4548e6-38b7-48dd-9088-d7d1219f7852,e0d1b955-1150-457e-96d7-91a9481999cd,24f9a6e7-b214-4011-94c4-6502f249a621,6a97773b-c33c-4f2e-9ac3-fdd53de0f175" + } + }, + "evaluation_timestamp": "1776165243.0", + "metric_config": { + "evaluation_description": "bootstrap_score", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.24432381746789958, + "uncertainty": { + "standard_deviation": 0.05666380138192282, + "num_samples": 140 + }, + "details": { + "total_matched_trajectories": "15", + "stopping_reason_count_repetition_guard": "15", + "avg_total_tokens_target_model": "10679.27", + "total_total_tokens_target_model": "160189.0", + "avg_total_tokens_other_models": "81013.20", + "total_total_tokens_other_models": "1215198.0", + "avg_total_tokens_all_models": "91692.47", + "total_total_tokens_all_models": "1375387.0", + "avg_turn_count": "50.67", + "total_turn_count": "760.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + } + } + }, + { + "evaluation_name": "std on healthbench/S-adaptive/+7ep for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "20", + "sample_ids": "19ec4833-86e9-4166-8b82-d1da09f31fd7,7ebc830a-8dbd-489b-9d61-4d8bacf0db8d,c971f9d1-5f6a-464e-b282-41c8f0d82f58,5a6e4a41-3ea6-4050-a971-93433fe34877,78510411-e03b-4f93-a369-c7e72d2e4908,bbd759b4-8e4d-4db7-bc36-f2e069396c1c,8f2a65de-dea7-48e8-8adb-6194eca26c08,6d5f483c-3e86-456d-bfd5-4e28de699ae5,f05491d8-d160-4b2b-bd5b-59d757585e39,85408180-238e-4c1b-92c1-55d8ce6c3717,ecd7453e-1dc1-4e52-bb95-67e8e5788573,0e073591-b3e7-4dc8-94d0-dc5aa93ca35c,7637b987-8fe2-49d0-8372-1ebf17284d70,5b294937-13e5-424e-8bb4-5d1904a2344a,c5bf7fc3-dfb0-4b47-9bbe-222e87b952e2,83cf8f2d-2857-4f01-a283-9595d8f4ae8e,7a4548e6-38b7-48dd-9088-d7d1219f7852,e0d1b955-1150-457e-96d7-91a9481999cd,24f9a6e7-b214-4011-94c4-6502f249a621,6a97773b-c33c-4f2e-9ac3-fdd53de0f175" + } + }, + "evaluation_timestamp": "1776165243.0", + "metric_config": { + "evaluation_description": "std", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.05666380138192282, + "uncertainty": { + "standard_deviation": 0.05666380138192282, + "num_samples": 140 + }, + "details": { + "total_matched_trajectories": "15", + "stopping_reason_count_repetition_guard": "15", + "avg_total_tokens_target_model": "10679.27", + "total_total_tokens_target_model": "160189.0", + "avg_total_tokens_other_models": "81013.20", + "total_total_tokens_other_models": "1215198.0", + "avg_total_tokens_all_models": "91692.47", + "total_total_tokens_all_models": "1375387.0", + "avg_turn_count": "50.67", + "total_turn_count": "760.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + } + } + }, + { + "evaluation_name": "criteria_met_rate on healthbench/S-adaptive/+7ep for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "20", + "sample_ids": "19ec4833-86e9-4166-8b82-d1da09f31fd7,7ebc830a-8dbd-489b-9d61-4d8bacf0db8d,c971f9d1-5f6a-464e-b282-41c8f0d82f58,5a6e4a41-3ea6-4050-a971-93433fe34877,78510411-e03b-4f93-a369-c7e72d2e4908,bbd759b4-8e4d-4db7-bc36-f2e069396c1c,8f2a65de-dea7-48e8-8adb-6194eca26c08,6d5f483c-3e86-456d-bfd5-4e28de699ae5,f05491d8-d160-4b2b-bd5b-59d757585e39,85408180-238e-4c1b-92c1-55d8ce6c3717,ecd7453e-1dc1-4e52-bb95-67e8e5788573,0e073591-b3e7-4dc8-94d0-dc5aa93ca35c,7637b987-8fe2-49d0-8372-1ebf17284d70,5b294937-13e5-424e-8bb4-5d1904a2344a,c5bf7fc3-dfb0-4b47-9bbe-222e87b952e2,83cf8f2d-2857-4f01-a283-9595d8f4ae8e,7a4548e6-38b7-48dd-9088-d7d1219f7852,e0d1b955-1150-457e-96d7-91a9481999cd,24f9a6e7-b214-4011-94c4-6502f249a621,6a97773b-c33c-4f2e-9ac3-fdd53de0f175" + } + }, + "evaluation_timestamp": "1776165243.0", + "metric_config": { + "evaluation_description": "criteria_met_rate", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.32115171650055374, + "uncertainty": { + "standard_deviation": 0.05666380138192282, + "num_samples": 140 + }, + "details": { + "total_matched_trajectories": "15", + "stopping_reason_count_repetition_guard": "15", + "avg_total_tokens_target_model": "10679.27", + "total_total_tokens_target_model": "160189.0", + "avg_total_tokens_other_models": "81013.20", + "total_total_tokens_other_models": "1215198.0", + "avg_total_tokens_all_models": "91692.47", + "total_total_tokens_all_models": "1375387.0", + "avg_turn_count": "50.67", + "total_turn_count": "760.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + } + } + }, + { + "evaluation_name": "mean on healthbench/S-adaptive/+7ep for scorer healthbench_score", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "20", + "sample_ids": "19ec4833-86e9-4166-8b82-d1da09f31fd7,7ebc830a-8dbd-489b-9d61-4d8bacf0db8d,c971f9d1-5f6a-464e-b282-41c8f0d82f58,5a6e4a41-3ea6-4050-a971-93433fe34877,78510411-e03b-4f93-a369-c7e72d2e4908,bbd759b4-8e4d-4db7-bc36-f2e069396c1c,8f2a65de-dea7-48e8-8adb-6194eca26c08,6d5f483c-3e86-456d-bfd5-4e28de699ae5,f05491d8-d160-4b2b-bd5b-59d757585e39,85408180-238e-4c1b-92c1-55d8ce6c3717,ecd7453e-1dc1-4e52-bb95-67e8e5788573,0e073591-b3e7-4dc8-94d0-dc5aa93ca35c,7637b987-8fe2-49d0-8372-1ebf17284d70,5b294937-13e5-424e-8bb4-5d1904a2344a,c5bf7fc3-dfb0-4b47-9bbe-222e87b952e2,83cf8f2d-2857-4f01-a283-9595d8f4ae8e,7a4548e6-38b7-48dd-9088-d7d1219f7852,e0d1b955-1150-457e-96d7-91a9481999cd,24f9a6e7-b214-4011-94c4-6502f249a621,6a97773b-c33c-4f2e-9ac3-fdd53de0f175" + } + }, + "evaluation_timestamp": "1776165243.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.24432381746789958, + "uncertainty": { + "num_samples": 140 + }, + "details": { + "total_matched_trajectories": "15", + "stopping_reason_count_repetition_guard": "15", + "avg_total_tokens_target_model": "10679.27", + "total_total_tokens_target_model": "160189.0", + "avg_total_tokens_other_models": "81013.20", + "total_total_tokens_other_models": "1215198.0", + "avg_total_tokens_all_models": "91692.47", + "total_total_tokens_all_models": "1375387.0", + "avg_turn_count": "50.67", + "total_turn_count": "760.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/anthropic/claude-opus-4-20250514/7a7fc16f-3232-4796-95e4-9311c88b650f_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "d5792ef1ef4684e34e55fc1c40f629301d27fc9e04258a45b1fffb0a44089271", + "total_rows": 140 + } +} \ No newline at end of file diff --git a/data/healthbench/anthropic/claude-opus-4-20250514/7a7fc16f-3232-4796-95e4-9311c88b650f_samples.jsonl b/data/healthbench/anthropic/claude-opus-4-20250514/7a7fc16f-3232-4796-95e4-9311c88b650f_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..055264d6a3153019254bd78e11da561631c618c8 --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-20250514/7a7fc16f-3232-4796-95e4-9311c88b650f_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d5792ef1ef4684e34e55fc1c40f629301d27fc9e04258a45b1fffb0a44089271 +size 9723427 diff --git a/data/healthbench/anthropic/claude-opus-4-20250514/911c0647-f046-4eb8-a5ff-b647d9fe9da1.json b/data/healthbench/anthropic/claude-opus-4-20250514/911c0647-f046-4eb8-a5ff-b647d9fe9da1.json new file mode 100644 index 0000000000000000000000000000000000000000..267cf0e601d3968b9804a6d17c80cc52bf43eacb --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-20250514/911c0647-f046-4eb8-a5ff-b647d9fe9da1.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/anthropic_claude-opus-4-20250514/1786168706.797071", + "evaluation_timestamp": "1776696116.0", + "retrieved_timestamp": "1786168706.797071", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-20250514", + "id": "anthropic/claude-opus-4-20250514", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/anthropic/claude-opus-4-20250514/911c0647-f046-4eb8-a5ff-b647d9fe9da1_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "c816e3e0ae0671322d5015549356a3f5756fbcf7899b0e4a9e6b56ca170f56b4", + "total_rows": 310 + } +} \ No newline at end of file diff --git a/data/healthbench/anthropic/claude-opus-4-20250514/911c0647-f046-4eb8-a5ff-b647d9fe9da1_samples.jsonl b/data/healthbench/anthropic/claude-opus-4-20250514/911c0647-f046-4eb8-a5ff-b647d9fe9da1_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..7003fd797be8e5cd2ae6ff3f285f76ee0f4b76e2 --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-20250514/911c0647-f046-4eb8-a5ff-b647d9fe9da1_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c816e3e0ae0671322d5015549356a3f5756fbcf7899b0e4a9e6b56ca170f56b4 +size 19736626 diff --git a/data/healthbench/anthropic/claude-opus-4-20250514/94acdad6-172a-48d4-a924-9867470402c7.json b/data/healthbench/anthropic/claude-opus-4-20250514/94acdad6-172a-48d4-a924-9867470402c7.json new file mode 100644 index 0000000000000000000000000000000000000000..f0f176fb2645a960184547957dce86979826c6b6 --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-20250514/94acdad6-172a-48d4-a924-9867470402c7.json @@ -0,0 +1,292 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/anthropic_claude-opus-4-20250514/1786164031.299211", + "evaluation_timestamp": "1775845422.0", + "retrieved_timestamp": "1786164031.299211", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-20250514", + "id": "anthropic/claude-opus-4-20250514", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "bootstrap_score on healthbench/S-adaptive/+7ep for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "3", + "sample_ids": "1f548d5b-cd00-49a0-b327-283a2e00debd,0b8f1d60-2081-4562-98f7-b6a976fe1c6d,6f7a2ee9-e9c6-42d8-b79f-22dea966b8d2" + } + }, + "evaluation_timestamp": "1775845422.0", + "metric_config": { + "evaluation_description": "bootstrap_score", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.6087814526071185, + "uncertainty": { + "standard_deviation": 0.12562445728168478, + "num_samples": 21 + }, + "details": { + "total_matched_trajectories": "5", + "stopping_reason_count_repetition_guard": "5", + "avg_total_tokens_target_model": "18530.60", + "total_total_tokens_target_model": "92653.0", + "avg_total_tokens_other_models": "73942.00", + "total_total_tokens_other_models": "369710.0", + "avg_total_tokens_all_models": "92472.60", + "total_total_tokens_all_models": "462363.0", + "avg_turn_count": "24.40", + "total_turn_count": "122.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + } + } + }, + { + "evaluation_name": "std on healthbench/S-adaptive/+7ep for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "3", + "sample_ids": "1f548d5b-cd00-49a0-b327-283a2e00debd,0b8f1d60-2081-4562-98f7-b6a976fe1c6d,6f7a2ee9-e9c6-42d8-b79f-22dea966b8d2" + } + }, + "evaluation_timestamp": "1775845422.0", + "metric_config": { + "evaluation_description": "std", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.12562445728168478, + "uncertainty": { + "standard_deviation": 0.12562445728168478, + "num_samples": 21 + }, + "details": { + "total_matched_trajectories": "5", + "stopping_reason_count_repetition_guard": "5", + "avg_total_tokens_target_model": "18530.60", + "total_total_tokens_target_model": "92653.0", + "avg_total_tokens_other_models": "73942.00", + "total_total_tokens_other_models": "369710.0", + "avg_total_tokens_all_models": "92472.60", + "total_total_tokens_all_models": "462363.0", + "avg_turn_count": "24.40", + "total_turn_count": "122.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + } + } + }, + { + "evaluation_name": "criteria_met_rate on healthbench/S-adaptive/+7ep for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "3", + "sample_ids": "1f548d5b-cd00-49a0-b327-283a2e00debd,0b8f1d60-2081-4562-98f7-b6a976fe1c6d,6f7a2ee9-e9c6-42d8-b79f-22dea966b8d2" + } + }, + "evaluation_timestamp": "1775845422.0", + "metric_config": { + "evaluation_description": "criteria_met_rate", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.5069124423963134, + "uncertainty": { + "standard_deviation": 0.12562445728168478, + "num_samples": 21 + }, + "details": { + "total_matched_trajectories": "5", + "stopping_reason_count_repetition_guard": "5", + "avg_total_tokens_target_model": "18530.60", + "total_total_tokens_target_model": "92653.0", + "avg_total_tokens_other_models": "73942.00", + "total_total_tokens_other_models": "369710.0", + "avg_total_tokens_all_models": "92472.60", + "total_total_tokens_all_models": "462363.0", + "avg_turn_count": "24.40", + "total_turn_count": "122.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + } + } + }, + { + "evaluation_name": "mean on healthbench/S-adaptive/+7ep for scorer healthbench_score", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "3", + "sample_ids": "1f548d5b-cd00-49a0-b327-283a2e00debd,0b8f1d60-2081-4562-98f7-b6a976fe1c6d,6f7a2ee9-e9c6-42d8-b79f-22dea966b8d2" + } + }, + "evaluation_timestamp": "1775845422.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.6087814526071185, + "uncertainty": { + "num_samples": 21 + }, + "details": { + "total_matched_trajectories": "5", + "stopping_reason_count_repetition_guard": "5", + "avg_total_tokens_target_model": "18530.60", + "total_total_tokens_target_model": "92653.0", + "avg_total_tokens_other_models": "73942.00", + "total_total_tokens_other_models": "369710.0", + "avg_total_tokens_all_models": "92472.60", + "total_total_tokens_all_models": "462363.0", + "avg_turn_count": "24.40", + "total_turn_count": "122.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/anthropic/claude-opus-4-20250514/94acdad6-172a-48d4-a924-9867470402c7_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "a85703b258b4d9a1726a13518ef27f9da8e7c97235309fe4afab60478e505018", + "total_rows": 21 + } +} \ No newline at end of file diff --git a/data/healthbench/anthropic/claude-opus-4-20250514/94acdad6-172a-48d4-a924-9867470402c7_samples.jsonl b/data/healthbench/anthropic/claude-opus-4-20250514/94acdad6-172a-48d4-a924-9867470402c7_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..d0c1d87799859571d3e35c954ddbd51b05e7088c --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-20250514/94acdad6-172a-48d4-a924-9867470402c7_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a85703b258b4d9a1726a13518ef27f9da8e7c97235309fe4afab60478e505018 +size 1282239 diff --git a/data/healthbench/anthropic/claude-opus-4-20250514/b5e3f3e1-675f-42c2-ae09-588ebb5d0e38.json b/data/healthbench/anthropic/claude-opus-4-20250514/b5e3f3e1-675f-42c2-ae09-588ebb5d0e38.json new file mode 100644 index 0000000000000000000000000000000000000000..4d3d7b08044549fd6717b21bdec8ee542e75f94f --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-20250514/b5e3f3e1-675f-42c2-ae09-588ebb5d0e38.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/anthropic_claude-opus-4-20250514/1786166724.248725", + "evaluation_timestamp": "1775604415.0", + "retrieved_timestamp": "1786166724.248725", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.202,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-20250514", + "id": "anthropic/claude-opus-4-20250514", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/anthropic/claude-opus-4-20250514/b5e3f3e1-675f-42c2-ae09-588ebb5d0e38_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "51cc7ab8109a8002c59956f6928e25410c54a782cab344d6ebe163f2b255c8fa", + "total_rows": 100 + } +} \ No newline at end of file diff --git a/data/healthbench/anthropic/claude-opus-4-20250514/b5e3f3e1-675f-42c2-ae09-588ebb5d0e38_samples.jsonl b/data/healthbench/anthropic/claude-opus-4-20250514/b5e3f3e1-675f-42c2-ae09-588ebb5d0e38_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..8d3f1b124d5e9abc037753f9cbbe4e1ce0773d16 --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-20250514/b5e3f3e1-675f-42c2-ae09-588ebb5d0e38_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:51cc7ab8109a8002c59956f6928e25410c54a782cab344d6ebe163f2b255c8fa +size 6442541 diff --git a/data/healthbench/anthropic/claude-opus-4-20250514/cb1dbe45-5ee3-41cd-9b9f-89a87333f884.json b/data/healthbench/anthropic/claude-opus-4-20250514/cb1dbe45-5ee3-41cd-9b9f-89a87333f884.json new file mode 100644 index 0000000000000000000000000000000000000000..0362f1b1b87b4658aa0349adf13e21205eca0135 --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-20250514/cb1dbe45-5ee3-41cd-9b9f-89a87333f884.json @@ -0,0 +1,292 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/anthropic_claude-opus-4-20250514/1786164011.881006", + "evaluation_timestamp": "1776259833.0", + "retrieved_timestamp": "1786164011.881006", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-20250514", + "id": "anthropic/claude-opus-4-20250514", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "bootstrap_score on healthbench/S-adaptive/+5ep/1998a169 for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "36", + "sample_ids": "f32ee0c1-ea9c-4cfb-8332-bd9579d23024,b6eb69cb-b911-4c81-a7a1-357d21d27109,c0dc053b-157b-4d13-9956-213b67ff6a36,3bb4a735-ace7-4342-b7a5-0bae083e6f82,66b448bf-2e49-488d-9d0b-f9912ab3f0f8,b12d2453-fd3b-462c-b9cc-554ff16bfd3f,0a548d04-0973-4343-a3bf-07f730bed2f6,e2cd9f23-e3dd-4b67-8984-a5412dd802e5,1fb6ab91-a999-437d-8c93-c5937c0a89a3,42af0c9f-a715-45c4-ab24-c2948f6f943d,a94aad40-387e-4a37-b92f-61badd57388b,596ba714-29f7-447e-aa73-88d46ecd6e88,1233737e-2674-41d3-8519-b817aeb0b32e,c80a2a84-281d-41cc-a10b-32ee48c584f6,bf7718f2-bceb-4702-8d74-a7fc1bd7fb78,b7a82324-73ee-4996-8f71-8f455f789e90,22fe3eec-03b0-4cdf-a06a-87e71d236082,6c568ea6-86bd-4286-94ba-e3ccc8b67b21,905949d2-7a0a-4461-8f4b-257de6be6eed,c653b6eb-f9b9-4626-b702-ff73d71e8253,cf09b209-5434-4317-aa2a-7b07537b475e,77c0be2e-364a-40de-ae21-a277154e3d60,af750930-5fa5-4efe-8f85-b499641377d4,af62e0b8-bbd0-4dcc-b55c-cb4c7bbcfb2c,f3729e05-61ef-4306-9491-6f5659d4eef9,126236de-9e7a-4814-b5d7-98cbe89b33b2,9bb31d40-c155-4213-868d-103fa0cc7c56,97bec88b-6ef7-4f24-84d6-a2bcc592548b,779b5d14-1a6f-4d0a-8763-e0a25961a811,62b4d946-e447-4fb4-96ed-b6d36c8e2c62,d475913b-0b18-47e9-90b2-28449e1d6541,3a0eec54-fd86-47c1-883b-d8e828d2a327,b947fc6f-1d50-4e3f-b679-8484529b787b,e8e17de2-9b5f-4f20-a18e-fdd14a8b7192,b5b6d817-c524-4bef-badc-f93876657ea2,8e447490-23cd-4d64-9e3a-b8ad46c76573" + } + }, + "evaluation_timestamp": "1776259833.0", + "metric_config": { + "evaluation_description": "bootstrap_score", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.34169096449277403, + "uncertainty": { + "standard_deviation": 0.04833750538418556, + "num_samples": 180 + }, + "details": { + "total_matched_trajectories": "25", + "stopping_reason_count_repetition_guard": "25", + "avg_total_tokens_target_model": "11029.52", + "total_total_tokens_target_model": "275738.0", + "avg_total_tokens_other_models": "41953.52", + "total_total_tokens_other_models": "1048838.0", + "avg_total_tokens_all_models": "52983.04", + "total_total_tokens_all_models": "1324576.0", + "avg_turn_count": "20.76", + "total_turn_count": "519.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + } + } + }, + { + "evaluation_name": "std on healthbench/S-adaptive/+5ep/1998a169 for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "36", + "sample_ids": "f32ee0c1-ea9c-4cfb-8332-bd9579d23024,b6eb69cb-b911-4c81-a7a1-357d21d27109,c0dc053b-157b-4d13-9956-213b67ff6a36,3bb4a735-ace7-4342-b7a5-0bae083e6f82,66b448bf-2e49-488d-9d0b-f9912ab3f0f8,b12d2453-fd3b-462c-b9cc-554ff16bfd3f,0a548d04-0973-4343-a3bf-07f730bed2f6,e2cd9f23-e3dd-4b67-8984-a5412dd802e5,1fb6ab91-a999-437d-8c93-c5937c0a89a3,42af0c9f-a715-45c4-ab24-c2948f6f943d,a94aad40-387e-4a37-b92f-61badd57388b,596ba714-29f7-447e-aa73-88d46ecd6e88,1233737e-2674-41d3-8519-b817aeb0b32e,c80a2a84-281d-41cc-a10b-32ee48c584f6,bf7718f2-bceb-4702-8d74-a7fc1bd7fb78,b7a82324-73ee-4996-8f71-8f455f789e90,22fe3eec-03b0-4cdf-a06a-87e71d236082,6c568ea6-86bd-4286-94ba-e3ccc8b67b21,905949d2-7a0a-4461-8f4b-257de6be6eed,c653b6eb-f9b9-4626-b702-ff73d71e8253,cf09b209-5434-4317-aa2a-7b07537b475e,77c0be2e-364a-40de-ae21-a277154e3d60,af750930-5fa5-4efe-8f85-b499641377d4,af62e0b8-bbd0-4dcc-b55c-cb4c7bbcfb2c,f3729e05-61ef-4306-9491-6f5659d4eef9,126236de-9e7a-4814-b5d7-98cbe89b33b2,9bb31d40-c155-4213-868d-103fa0cc7c56,97bec88b-6ef7-4f24-84d6-a2bcc592548b,779b5d14-1a6f-4d0a-8763-e0a25961a811,62b4d946-e447-4fb4-96ed-b6d36c8e2c62,d475913b-0b18-47e9-90b2-28449e1d6541,3a0eec54-fd86-47c1-883b-d8e828d2a327,b947fc6f-1d50-4e3f-b679-8484529b787b,e8e17de2-9b5f-4f20-a18e-fdd14a8b7192,b5b6d817-c524-4bef-badc-f93876657ea2,8e447490-23cd-4d64-9e3a-b8ad46c76573" + } + }, + "evaluation_timestamp": "1776259833.0", + "metric_config": { + "evaluation_description": "std", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.04833750538418556, + "uncertainty": { + "standard_deviation": 0.04833750538418556, + "num_samples": 180 + }, + "details": { + "total_matched_trajectories": "25", + "stopping_reason_count_repetition_guard": "25", + "avg_total_tokens_target_model": "11029.52", + "total_total_tokens_target_model": "275738.0", + "avg_total_tokens_other_models": "41953.52", + "total_total_tokens_other_models": "1048838.0", + "avg_total_tokens_all_models": "52983.04", + "total_total_tokens_all_models": "1324576.0", + "avg_turn_count": "20.76", + "total_turn_count": "519.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + } + } + }, + { + "evaluation_name": "criteria_met_rate on healthbench/S-adaptive/+5ep/1998a169 for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "36", + "sample_ids": "f32ee0c1-ea9c-4cfb-8332-bd9579d23024,b6eb69cb-b911-4c81-a7a1-357d21d27109,c0dc053b-157b-4d13-9956-213b67ff6a36,3bb4a735-ace7-4342-b7a5-0bae083e6f82,66b448bf-2e49-488d-9d0b-f9912ab3f0f8,b12d2453-fd3b-462c-b9cc-554ff16bfd3f,0a548d04-0973-4343-a3bf-07f730bed2f6,e2cd9f23-e3dd-4b67-8984-a5412dd802e5,1fb6ab91-a999-437d-8c93-c5937c0a89a3,42af0c9f-a715-45c4-ab24-c2948f6f943d,a94aad40-387e-4a37-b92f-61badd57388b,596ba714-29f7-447e-aa73-88d46ecd6e88,1233737e-2674-41d3-8519-b817aeb0b32e,c80a2a84-281d-41cc-a10b-32ee48c584f6,bf7718f2-bceb-4702-8d74-a7fc1bd7fb78,b7a82324-73ee-4996-8f71-8f455f789e90,22fe3eec-03b0-4cdf-a06a-87e71d236082,6c568ea6-86bd-4286-94ba-e3ccc8b67b21,905949d2-7a0a-4461-8f4b-257de6be6eed,c653b6eb-f9b9-4626-b702-ff73d71e8253,cf09b209-5434-4317-aa2a-7b07537b475e,77c0be2e-364a-40de-ae21-a277154e3d60,af750930-5fa5-4efe-8f85-b499641377d4,af62e0b8-bbd0-4dcc-b55c-cb4c7bbcfb2c,f3729e05-61ef-4306-9491-6f5659d4eef9,126236de-9e7a-4814-b5d7-98cbe89b33b2,9bb31d40-c155-4213-868d-103fa0cc7c56,97bec88b-6ef7-4f24-84d6-a2bcc592548b,779b5d14-1a6f-4d0a-8763-e0a25961a811,62b4d946-e447-4fb4-96ed-b6d36c8e2c62,d475913b-0b18-47e9-90b2-28449e1d6541,3a0eec54-fd86-47c1-883b-d8e828d2a327,b947fc6f-1d50-4e3f-b679-8484529b787b,e8e17de2-9b5f-4f20-a18e-fdd14a8b7192,b5b6d817-c524-4bef-badc-f93876657ea2,8e447490-23cd-4d64-9e3a-b8ad46c76573" + } + }, + "evaluation_timestamp": "1776259833.0", + "metric_config": { + "evaluation_description": "criteria_met_rate", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.3706586826347305, + "uncertainty": { + "standard_deviation": 0.04833750538418556, + "num_samples": 180 + }, + "details": { + "total_matched_trajectories": "25", + "stopping_reason_count_repetition_guard": "25", + "avg_total_tokens_target_model": "11029.52", + "total_total_tokens_target_model": "275738.0", + "avg_total_tokens_other_models": "41953.52", + "total_total_tokens_other_models": "1048838.0", + "avg_total_tokens_all_models": "52983.04", + "total_total_tokens_all_models": "1324576.0", + "avg_turn_count": "20.76", + "total_turn_count": "519.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + } + } + }, + { + "evaluation_name": "mean on healthbench/S-adaptive/+5ep/1998a169 for scorer healthbench_score", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "36", + "sample_ids": "f32ee0c1-ea9c-4cfb-8332-bd9579d23024,b6eb69cb-b911-4c81-a7a1-357d21d27109,c0dc053b-157b-4d13-9956-213b67ff6a36,3bb4a735-ace7-4342-b7a5-0bae083e6f82,66b448bf-2e49-488d-9d0b-f9912ab3f0f8,b12d2453-fd3b-462c-b9cc-554ff16bfd3f,0a548d04-0973-4343-a3bf-07f730bed2f6,e2cd9f23-e3dd-4b67-8984-a5412dd802e5,1fb6ab91-a999-437d-8c93-c5937c0a89a3,42af0c9f-a715-45c4-ab24-c2948f6f943d,a94aad40-387e-4a37-b92f-61badd57388b,596ba714-29f7-447e-aa73-88d46ecd6e88,1233737e-2674-41d3-8519-b817aeb0b32e,c80a2a84-281d-41cc-a10b-32ee48c584f6,bf7718f2-bceb-4702-8d74-a7fc1bd7fb78,b7a82324-73ee-4996-8f71-8f455f789e90,22fe3eec-03b0-4cdf-a06a-87e71d236082,6c568ea6-86bd-4286-94ba-e3ccc8b67b21,905949d2-7a0a-4461-8f4b-257de6be6eed,c653b6eb-f9b9-4626-b702-ff73d71e8253,cf09b209-5434-4317-aa2a-7b07537b475e,77c0be2e-364a-40de-ae21-a277154e3d60,af750930-5fa5-4efe-8f85-b499641377d4,af62e0b8-bbd0-4dcc-b55c-cb4c7bbcfb2c,f3729e05-61ef-4306-9491-6f5659d4eef9,126236de-9e7a-4814-b5d7-98cbe89b33b2,9bb31d40-c155-4213-868d-103fa0cc7c56,97bec88b-6ef7-4f24-84d6-a2bcc592548b,779b5d14-1a6f-4d0a-8763-e0a25961a811,62b4d946-e447-4fb4-96ed-b6d36c8e2c62,d475913b-0b18-47e9-90b2-28449e1d6541,3a0eec54-fd86-47c1-883b-d8e828d2a327,b947fc6f-1d50-4e3f-b679-8484529b787b,e8e17de2-9b5f-4f20-a18e-fdd14a8b7192,b5b6d817-c524-4bef-badc-f93876657ea2,8e447490-23cd-4d64-9e3a-b8ad46c76573" + } + }, + "evaluation_timestamp": "1776259833.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.34169096449277403, + "uncertainty": { + "num_samples": 180 + }, + "details": { + "total_matched_trajectories": "25", + "stopping_reason_count_repetition_guard": "25", + "avg_total_tokens_target_model": "11029.52", + "total_total_tokens_target_model": "275738.0", + "avg_total_tokens_other_models": "41953.52", + "total_total_tokens_other_models": "1048838.0", + "avg_total_tokens_all_models": "52983.04", + "total_total_tokens_all_models": "1324576.0", + "avg_turn_count": "20.76", + "total_turn_count": "519.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/anthropic/claude-opus-4-20250514/cb1dbe45-5ee3-41cd-9b9f-89a87333f884_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "b6979dcefae03bd946b8e60311cf148f39d2bc6944734c26b944f9c685f8ea9c", + "total_rows": 180 + } +} \ No newline at end of file diff --git a/data/healthbench/anthropic/claude-opus-4-20250514/cb1dbe45-5ee3-41cd-9b9f-89a87333f884_samples.jsonl b/data/healthbench/anthropic/claude-opus-4-20250514/cb1dbe45-5ee3-41cd-9b9f-89a87333f884_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..39d505e4c1981841ad314d76866f91cccda2ab6c --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-20250514/cb1dbe45-5ee3-41cd-9b9f-89a87333f884_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b6979dcefae03bd946b8e60311cf148f39d2bc6944734c26b944f9c685f8ea9c +size 9283460 diff --git a/data/healthbench/anthropic/claude-opus-4-20250514/d041291b-bd79-4df0-9ba1-80a6fad898d5.json b/data/healthbench/anthropic/claude-opus-4-20250514/d041291b-bd79-4df0-9ba1-80a6fad898d5.json new file mode 100644 index 0000000000000000000000000000000000000000..cb0620503cf962a2ecff53d60ecd53115f9de7ea --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-20250514/d041291b-bd79-4df0-9ba1-80a6fad898d5.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/anthropic_claude-opus-4-20250514/1786166735.431376", + "evaluation_timestamp": "1776696116.0", + "retrieved_timestamp": "1786166735.431376", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-20250514", + "id": "anthropic/claude-opus-4-20250514", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/anthropic/claude-opus-4-20250514/d041291b-bd79-4df0-9ba1-80a6fad898d5_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "2e02deb87d392994e0a4a34af6c694dadb42d71c507f5ae97b2782af7aaad101", + "total_rows": 120 + } +} \ No newline at end of file diff --git a/data/healthbench/anthropic/claude-opus-4-20250514/d041291b-bd79-4df0-9ba1-80a6fad898d5_samples.jsonl b/data/healthbench/anthropic/claude-opus-4-20250514/d041291b-bd79-4df0-9ba1-80a6fad898d5_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..2984616289abc0d3b59be66abf341de70789bce8 --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-20250514/d041291b-bd79-4df0-9ba1-80a6fad898d5_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2e02deb87d392994e0a4a34af6c694dadb42d71c507f5ae97b2782af7aaad101 +size 8475431 diff --git a/data/healthbench/anthropic/claude-opus-4-20250514/fa1a4c0b-45ff-4598-89ea-14b5ce6669e6.json b/data/healthbench/anthropic/claude-opus-4-20250514/fa1a4c0b-45ff-4598-89ea-14b5ce6669e6.json new file mode 100644 index 0000000000000000000000000000000000000000..e829e59ccde70c2a4c9c8dea2273298253ea54c4 --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-20250514/fa1a4c0b-45ff-4598-89ea-14b5ce6669e6.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/anthropic_claude-opus-4-20250514/1786168689.76248", + "evaluation_timestamp": "1776165241.0", + "retrieved_timestamp": "1786168689.76248", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-20250514", + "id": "anthropic/claude-opus-4-20250514", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/anthropic/claude-opus-4-20250514/fa1a4c0b-45ff-4598-89ea-14b5ce6669e6_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "835fd3c625861574a69987961703a2e0ef4254416b023e8344b5ca4c59c6180a", + "total_rows": 190 + } +} \ No newline at end of file diff --git a/data/healthbench/anthropic/claude-opus-4-20250514/fa1a4c0b-45ff-4598-89ea-14b5ce6669e6_samples.jsonl b/data/healthbench/anthropic/claude-opus-4-20250514/fa1a4c0b-45ff-4598-89ea-14b5ce6669e6_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..3041d4aa09bd9f9884146156f91fe8b06f81174b --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-20250514/fa1a4c0b-45ff-4598-89ea-14b5ce6669e6_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:835fd3c625861574a69987961703a2e0ef4254416b023e8344b5ca4c59c6180a +size 10899916 diff --git a/data/healthbench/anthropic/claude-opus-4-20250514/fe38b656-2a3b-4b40-a772-5df7e61ccad5.json b/data/healthbench/anthropic/claude-opus-4-20250514/fe38b656-2a3b-4b40-a772-5df7e61ccad5.json new file mode 100644 index 0000000000000000000000000000000000000000..4d768749b88495f5fb0a57346b5f0b79a4293942 --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-20250514/fe38b656-2a3b-4b40-a772-5df7e61ccad5.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/anthropic_claude-opus-4-20250514/1786166746.834652", + "evaluation_timestamp": "1776259832.0", + "retrieved_timestamp": "1786166746.834652", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-20250514", + "id": "anthropic/claude-opus-4-20250514", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/anthropic/claude-opus-4-20250514/fe38b656-2a3b-4b40-a772-5df7e61ccad5_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "e2e4c17e129f734ff280c379d63c788ece4bbec0b47b20f3e2fbbb37691a7be5", + "total_rows": 100 + } +} \ No newline at end of file diff --git a/data/healthbench/anthropic/claude-opus-4-20250514/fe38b656-2a3b-4b40-a772-5df7e61ccad5_samples.jsonl b/data/healthbench/anthropic/claude-opus-4-20250514/fe38b656-2a3b-4b40-a772-5df7e61ccad5_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..f1aee7c67c030ab5fa230878324935047d482797 --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-20250514/fe38b656-2a3b-4b40-a772-5df7e61ccad5_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e2e4c17e129f734ff280c379d63c788ece4bbec0b47b20f3e2fbbb37691a7be5 +size 6289763 diff --git a/data/healthbench/anthropic/claude-opus-4-5-20251101/1d7e6603-40e2-41a6-91f7-5e5b6674444f.json b/data/healthbench/anthropic/claude-opus-4-5-20251101/1d7e6603-40e2-41a6-91f7-5e5b6674444f.json new file mode 100644 index 0000000000000000000000000000000000000000..f2c7cb04df5d2f2d89b801e8777ffbf841b0aeee --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-5-20251101/1d7e6603-40e2-41a6-91f7-5e5b6674444f.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/anthropic_claude-opus-4-5-20251101/1786166599.407209", + "evaluation_timestamp": "1776682987.0", + "retrieved_timestamp": "1786166599.407209", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-5-20251101", + "id": "anthropic/claude-opus-4-5-20251101", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/anthropic/claude-opus-4-5-20251101/1d7e6603-40e2-41a6-91f7-5e5b6674444f_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "af2bd92db164d439a9c077c6984402f66c5b0434fcfad5db637a0760f764e50b", + "total_rows": 50 + } +} \ No newline at end of file diff --git a/data/healthbench/anthropic/claude-opus-4-5-20251101/1d7e6603-40e2-41a6-91f7-5e5b6674444f_samples.jsonl b/data/healthbench/anthropic/claude-opus-4-5-20251101/1d7e6603-40e2-41a6-91f7-5e5b6674444f_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..4265223cadcd6b45a2785251ee1ba2d47dd6b1aa --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-5-20251101/1d7e6603-40e2-41a6-91f7-5e5b6674444f_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:af2bd92db164d439a9c077c6984402f66c5b0434fcfad5db637a0760f764e50b +size 3349186 diff --git a/data/healthbench/anthropic/claude-opus-4-5-20251101/215c3a51-165f-4b17-bcd2-a32185780674.json b/data/healthbench/anthropic/claude-opus-4-5-20251101/215c3a51-165f-4b17-bcd2-a32185780674.json new file mode 100644 index 0000000000000000000000000000000000000000..d9e92aafd9c6c267fd5132e17c89f64f26d8695b --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-5-20251101/215c3a51-165f-4b17-bcd2-a32185780674.json @@ -0,0 +1,244 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/anthropic_claude-opus-4-5-20251101/1786163930.882191", + "evaluation_timestamp": "1776273628.0", + "retrieved_timestamp": "1786163930.882191", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-5-20251101", + "id": "anthropic/claude-opus-4-5-20251101", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "bootstrap_score on healthbench/S-adaptive/+6ep/30b7b7cf for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "1", + "sample_ids": "9385d26b-bf44-4ff7-9410-e181413540dd" + } + }, + "evaluation_timestamp": "1776273628.0", + "metric_config": { + "evaluation_description": "bootstrap_score", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.14559386973180077, + "uncertainty": { + "standard_deviation": 5.551115123125783e-17, + "num_samples": 6 + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + } + } + }, + { + "evaluation_name": "std on healthbench/S-adaptive/+6ep/30b7b7cf for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "1", + "sample_ids": "9385d26b-bf44-4ff7-9410-e181413540dd" + } + }, + "evaluation_timestamp": "1776273628.0", + "metric_config": { + "evaluation_description": "std", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 5.551115123125783e-17, + "uncertainty": { + "standard_deviation": 5.551115123125783e-17, + "num_samples": 6 + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + } + } + }, + { + "evaluation_name": "criteria_met_rate on healthbench/S-adaptive/+6ep/30b7b7cf for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "1", + "sample_ids": "9385d26b-bf44-4ff7-9410-e181413540dd" + } + }, + "evaluation_timestamp": "1776273628.0", + "metric_config": { + "evaluation_description": "criteria_met_rate", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.2631578947368421, + "uncertainty": { + "standard_deviation": 5.551115123125783e-17, + "num_samples": 6 + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + } + } + }, + { + "evaluation_name": "mean on healthbench/S-adaptive/+6ep/30b7b7cf for scorer healthbench_score", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "1", + "sample_ids": "9385d26b-bf44-4ff7-9410-e181413540dd" + } + }, + "evaluation_timestamp": "1776273628.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.14559386973180077, + "uncertainty": { + "num_samples": 6 + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/anthropic/claude-opus-4-5-20251101/215c3a51-165f-4b17-bcd2-a32185780674_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "cafcdf99183961f74b2a73e4f24a3d6e5bf4febf03ace8624df8f0649b01fa5a", + "total_rows": 6 + } +} \ No newline at end of file diff --git a/data/healthbench/anthropic/claude-opus-4-5-20251101/215c3a51-165f-4b17-bcd2-a32185780674_samples.jsonl b/data/healthbench/anthropic/claude-opus-4-5-20251101/215c3a51-165f-4b17-bcd2-a32185780674_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..55add0d00faa8205aa07b3b36041ed12b811b8ca --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-5-20251101/215c3a51-165f-4b17-bcd2-a32185780674_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cafcdf99183961f74b2a73e4f24a3d6e5bf4febf03ace8624df8f0649b01fa5a +size 525499 diff --git a/data/healthbench/anthropic/claude-opus-4-5-20251101/44f2c464-322b-4994-840e-49dd8fb411c3.json b/data/healthbench/anthropic/claude-opus-4-5-20251101/44f2c464-322b-4994-840e-49dd8fb411c3.json new file mode 100644 index 0000000000000000000000000000000000000000..8eecaa335eb813cc94a93cf5db132694949f81f2 --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-5-20251101/44f2c464-322b-4994-840e-49dd8fb411c3.json @@ -0,0 +1,320 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/anthropic_claude-opus-4-5-20251101/1786163907.629133", + "evaluation_timestamp": "1776779722.0", + "retrieved_timestamp": "1786163907.629133", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-5-20251101", + "id": "anthropic/claude-opus-4-5-20251101", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "bootstrap_score on healthbench/S-adaptive/+1ep/c78c8094 for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "10", + "sample_ids": "99b22b39-30d5-40dc-82d9-c1cd3c793a44,3852d050-c009-4e33-b953-fd734a5b5e73,7a283823-00ee-489d-8dbc-356e2857526f,e89e00c9-29e9-420c-813d-e06ddc39d585,51433193-b0c3-4725-a3a8-9dd15e969b0c,48c89cd1-0ecf-4738-8fcc-392269585bd3,36c6e780-d4a8-47b2-8a36-90ee4b97607f,5fb6812f-9816-4d5d-b7e0-4ec3e9df5dc1,b81c059a-1137-47cf-bab8-ab5eb8ff34d5,886864a6-a0e4-468f-aaf4-26cf8dc5e541" + } + }, + "evaluation_timestamp": "1776779722.0", + "metric_config": { + "evaluation_description": "bootstrap_score", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.09818808449636224, + "uncertainty": { + "standard_deviation": 0.0365372300686424, + "num_samples": 10 + }, + "details": { + "total_matched_trajectories": "10", + "stopping_reason_count_repetition_guard": "10", + "avg_total_tokens_target_model": "15989.70", + "total_total_tokens_target_model": "159897.0", + "avg_total_tokens_other_models": "64216.10", + "total_total_tokens_other_models": "642161.0", + "avg_total_tokens_all_models": "80205.80", + "total_total_tokens_all_models": "802058.0", + "avg_turn_count": "44.30", + "total_turn_count": "443.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + }, + { + "evaluation_name": "std on healthbench/S-adaptive/+1ep/c78c8094 for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "10", + "sample_ids": "99b22b39-30d5-40dc-82d9-c1cd3c793a44,3852d050-c009-4e33-b953-fd734a5b5e73,7a283823-00ee-489d-8dbc-356e2857526f,e89e00c9-29e9-420c-813d-e06ddc39d585,51433193-b0c3-4725-a3a8-9dd15e969b0c,48c89cd1-0ecf-4738-8fcc-392269585bd3,36c6e780-d4a8-47b2-8a36-90ee4b97607f,5fb6812f-9816-4d5d-b7e0-4ec3e9df5dc1,b81c059a-1137-47cf-bab8-ab5eb8ff34d5,886864a6-a0e4-468f-aaf4-26cf8dc5e541" + } + }, + "evaluation_timestamp": "1776779722.0", + "metric_config": { + "evaluation_description": "std", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.0365372300686424, + "uncertainty": { + "standard_deviation": 0.0365372300686424, + "num_samples": 10 + }, + "details": { + "total_matched_trajectories": "10", + "stopping_reason_count_repetition_guard": "10", + "avg_total_tokens_target_model": "15989.70", + "total_total_tokens_target_model": "159897.0", + "avg_total_tokens_other_models": "64216.10", + "total_total_tokens_other_models": "642161.0", + "avg_total_tokens_all_models": "80205.80", + "total_total_tokens_all_models": "802058.0", + "avg_turn_count": "44.30", + "total_turn_count": "443.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + }, + { + "evaluation_name": "criteria_met_rate on healthbench/S-adaptive/+1ep/c78c8094 for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "10", + "sample_ids": "99b22b39-30d5-40dc-82d9-c1cd3c793a44,3852d050-c009-4e33-b953-fd734a5b5e73,7a283823-00ee-489d-8dbc-356e2857526f,e89e00c9-29e9-420c-813d-e06ddc39d585,51433193-b0c3-4725-a3a8-9dd15e969b0c,48c89cd1-0ecf-4738-8fcc-392269585bd3,36c6e780-d4a8-47b2-8a36-90ee4b97607f,5fb6812f-9816-4d5d-b7e0-4ec3e9df5dc1,b81c059a-1137-47cf-bab8-ab5eb8ff34d5,886864a6-a0e4-468f-aaf4-26cf8dc5e541" + } + }, + "evaluation_timestamp": "1776779722.0", + "metric_config": { + "evaluation_description": "criteria_met_rate", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.19117647058823528, + "uncertainty": { + "standard_deviation": 0.0365372300686424, + "num_samples": 10 + }, + "details": { + "total_matched_trajectories": "10", + "stopping_reason_count_repetition_guard": "10", + "avg_total_tokens_target_model": "15989.70", + "total_total_tokens_target_model": "159897.0", + "avg_total_tokens_other_models": "64216.10", + "total_total_tokens_other_models": "642161.0", + "avg_total_tokens_all_models": "80205.80", + "total_total_tokens_all_models": "802058.0", + "avg_turn_count": "44.30", + "total_turn_count": "443.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + }, + { + "evaluation_name": "mean on healthbench/S-adaptive/+1ep/c78c8094 for scorer healthbench_score", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "10", + "sample_ids": "99b22b39-30d5-40dc-82d9-c1cd3c793a44,3852d050-c009-4e33-b953-fd734a5b5e73,7a283823-00ee-489d-8dbc-356e2857526f,e89e00c9-29e9-420c-813d-e06ddc39d585,51433193-b0c3-4725-a3a8-9dd15e969b0c,48c89cd1-0ecf-4738-8fcc-392269585bd3,36c6e780-d4a8-47b2-8a36-90ee4b97607f,5fb6812f-9816-4d5d-b7e0-4ec3e9df5dc1,b81c059a-1137-47cf-bab8-ab5eb8ff34d5,886864a6-a0e4-468f-aaf4-26cf8dc5e541" + } + }, + "evaluation_timestamp": "1776779722.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.09818808449636224, + "uncertainty": { + "num_samples": 10 + }, + "details": { + "total_matched_trajectories": "10", + "stopping_reason_count_repetition_guard": "10", + "avg_total_tokens_target_model": "15989.70", + "total_total_tokens_target_model": "159897.0", + "avg_total_tokens_other_models": "64216.10", + "total_total_tokens_other_models": "642161.0", + "avg_total_tokens_all_models": "80205.80", + "total_total_tokens_all_models": "802058.0", + "avg_turn_count": "44.30", + "total_turn_count": "443.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/anthropic/claude-opus-4-5-20251101/44f2c464-322b-4994-840e-49dd8fb411c3_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "015c66399c36253dbc4536b12213f95d3e379766e16ca2d8aba722199e84e944", + "total_rows": 10 + } +} \ No newline at end of file diff --git a/data/healthbench/anthropic/claude-opus-4-5-20251101/44f2c464-322b-4994-840e-49dd8fb411c3_samples.jsonl b/data/healthbench/anthropic/claude-opus-4-5-20251101/44f2c464-322b-4994-840e-49dd8fb411c3_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..189a84961e4743f27043f9e6e9d11478d1f85982 --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-5-20251101/44f2c464-322b-4994-840e-49dd8fb411c3_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:015c66399c36253dbc4536b12213f95d3e379766e16ca2d8aba722199e84e944 +size 891027 diff --git a/data/healthbench/anthropic/claude-opus-4-5-20251101/463abc09-bac5-4bb0-930a-0441798a732c.json b/data/healthbench/anthropic/claude-opus-4-5-20251101/463abc09-bac5-4bb0-930a-0441798a732c.json new file mode 100644 index 0000000000000000000000000000000000000000..eb706892ab9fa9796e447d04e8f2e94fa774b524 --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-5-20251101/463abc09-bac5-4bb0-930a-0441798a732c.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/anthropic_claude-opus-4-5-20251101/1786166616.127097", + "evaluation_timestamp": "1776696116.0", + "retrieved_timestamp": "1786166616.127097", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-5-20251101", + "id": "anthropic/claude-opus-4-5-20251101", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/anthropic/claude-opus-4-5-20251101/463abc09-bac5-4bb0-930a-0441798a732c_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "c03a3e3c62019592bd95c8dac64b5d81d6a38192cf2ef58a4984917b1bbb3e29", + "total_rows": 160 + } +} \ No newline at end of file diff --git a/data/healthbench/anthropic/claude-opus-4-5-20251101/463abc09-bac5-4bb0-930a-0441798a732c_samples.jsonl b/data/healthbench/anthropic/claude-opus-4-5-20251101/463abc09-bac5-4bb0-930a-0441798a732c_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..4e737614b9799b13e85090ae0607e7caedf60260 --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-5-20251101/463abc09-bac5-4bb0-930a-0441798a732c_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c03a3e3c62019592bd95c8dac64b5d81d6a38192cf2ef58a4984917b1bbb3e29 +size 11246048 diff --git a/data/healthbench/anthropic/claude-opus-4-5-20251101/50291ced-1434-4164-a198-fcce8e293bf3.json b/data/healthbench/anthropic/claude-opus-4-5-20251101/50291ced-1434-4164-a198-fcce8e293bf3.json new file mode 100644 index 0000000000000000000000000000000000000000..265af718370c9bb675ab1ae679cf3c999c592c15 --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-5-20251101/50291ced-1434-4164-a198-fcce8e293bf3.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/anthropic_claude-opus-4-5-20251101/1786168571.581236", + "evaluation_timestamp": "1776165241.0", + "retrieved_timestamp": "1786168571.581236", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-5-20251101", + "id": "anthropic/claude-opus-4-5-20251101", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/anthropic/claude-opus-4-5-20251101/50291ced-1434-4164-a198-fcce8e293bf3_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "719cf0cb1bf140f8d6b55e2edff4d857d77b512859ccffc71845b2e91d65c9aa", + "total_rows": 151 + } +} \ No newline at end of file diff --git a/data/healthbench/anthropic/claude-opus-4-5-20251101/50291ced-1434-4164-a198-fcce8e293bf3_samples.jsonl b/data/healthbench/anthropic/claude-opus-4-5-20251101/50291ced-1434-4164-a198-fcce8e293bf3_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..543d2eafad697a56a42f1f094f658f3da69d8982 --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-5-20251101/50291ced-1434-4164-a198-fcce8e293bf3_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:719cf0cb1bf140f8d6b55e2edff4d857d77b512859ccffc71845b2e91d65c9aa +size 8760776 diff --git a/data/healthbench/anthropic/claude-opus-4-5-20251101/6fe5ec55-b126-487b-810e-6f2b40271e94.json b/data/healthbench/anthropic/claude-opus-4-5-20251101/6fe5ec55-b126-487b-810e-6f2b40271e94.json new file mode 100644 index 0000000000000000000000000000000000000000..4fc483d22657ddb2123f6f22c73a32b5956f0241 --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-5-20251101/6fe5ec55-b126-487b-810e-6f2b40271e94.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/anthropic_claude-opus-4-5-20251101/1786166645.529486", + "evaluation_timestamp": "1776259832.0", + "retrieved_timestamp": "1786166645.529486", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-5-20251101", + "id": "anthropic/claude-opus-4-5-20251101", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/anthropic/claude-opus-4-5-20251101/6fe5ec55-b126-487b-810e-6f2b40271e94_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "1ee3a9251dca8e1854985a32a0510028c7abfda5cd2372131a67bcd93502d084", + "total_rows": 120 + } +} \ No newline at end of file diff --git a/data/healthbench/anthropic/claude-opus-4-5-20251101/6fe5ec55-b126-487b-810e-6f2b40271e94_samples.jsonl b/data/healthbench/anthropic/claude-opus-4-5-20251101/6fe5ec55-b126-487b-810e-6f2b40271e94_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..050dc0f1f5e4df608ffd633bb94c8e07bda45df3 --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-5-20251101/6fe5ec55-b126-487b-810e-6f2b40271e94_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1ee3a9251dca8e1854985a32a0510028c7abfda5cd2372131a67bcd93502d084 +size 8041331 diff --git a/data/healthbench/anthropic/claude-opus-4-5-20251101/74fb7a1c-e233-4815-b86a-6e200cee63e6.json b/data/healthbench/anthropic/claude-opus-4-5-20251101/74fb7a1c-e233-4815-b86a-6e200cee63e6.json new file mode 100644 index 0000000000000000000000000000000000000000..6c721521a89cd0f77c0c2c1c0beb64343019754d --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-5-20251101/74fb7a1c-e233-4815-b86a-6e200cee63e6.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/anthropic_claude-opus-4-5-20251101/1786168581.477714", + "evaluation_timestamp": "1776696116.0", + "retrieved_timestamp": "1786168581.477714", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-5-20251101", + "id": "anthropic/claude-opus-4-5-20251101", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/anthropic/claude-opus-4-5-20251101/74fb7a1c-e233-4815-b86a-6e200cee63e6_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "a285da99b3fde249cb16882ac214ae7c6193a865a3bda67d80515c3e743fa15d", + "total_rows": 160 + } +} \ No newline at end of file diff --git a/data/healthbench/anthropic/claude-opus-4-5-20251101/74fb7a1c-e233-4815-b86a-6e200cee63e6_samples.jsonl b/data/healthbench/anthropic/claude-opus-4-5-20251101/74fb7a1c-e233-4815-b86a-6e200cee63e6_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..a0be52449eece38506f69a7cb0add73d9752a3f1 --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-5-20251101/74fb7a1c-e233-4815-b86a-6e200cee63e6_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a285da99b3fde249cb16882ac214ae7c6193a865a3bda67d80515c3e743fa15d +size 11364162 diff --git a/data/healthbench/anthropic/claude-opus-4-5-20251101/7a41a4b3-c7a8-4b0f-ba8c-f8053affa01d.json b/data/healthbench/anthropic/claude-opus-4-5-20251101/7a41a4b3-c7a8-4b0f-ba8c-f8053affa01d.json new file mode 100644 index 0000000000000000000000000000000000000000..c4b296691793d0dbf1cbc699824cadfcfefb5f2c --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-5-20251101/7a41a4b3-c7a8-4b0f-ba8c-f8053affa01d.json @@ -0,0 +1,244 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/anthropic_claude-opus-4-5-20251101/1786163909.543105", + "evaluation_timestamp": "1776259833.0", + "retrieved_timestamp": "1786163909.543105", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-5-20251101", + "id": "anthropic/claude-opus-4-5-20251101", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "bootstrap_score on healthbench/S-adaptive/+4ep/377e41e3 for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "2", + "sample_ids": "1fb6ab91-a999-437d-8c93-c5937c0a89a3,42af0c9f-a715-45c4-ab24-c2948f6f943d" + } + }, + "evaluation_timestamp": "1776259833.0", + "metric_config": { + "evaluation_description": "bootstrap_score", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.3958333333333333, + "uncertainty": { + "standard_deviation": 0.3048892494449146, + "num_samples": 8 + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + } + } + }, + { + "evaluation_name": "std on healthbench/S-adaptive/+4ep/377e41e3 for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "2", + "sample_ids": "1fb6ab91-a999-437d-8c93-c5937c0a89a3,42af0c9f-a715-45c4-ab24-c2948f6f943d" + } + }, + "evaluation_timestamp": "1776259833.0", + "metric_config": { + "evaluation_description": "std", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.3048892494449146, + "uncertainty": { + "standard_deviation": 0.3048892494449146, + "num_samples": 8 + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + } + } + }, + { + "evaluation_name": "criteria_met_rate on healthbench/S-adaptive/+4ep/377e41e3 for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "2", + "sample_ids": "1fb6ab91-a999-437d-8c93-c5937c0a89a3,42af0c9f-a715-45c4-ab24-c2948f6f943d" + } + }, + "evaluation_timestamp": "1776259833.0", + "metric_config": { + "evaluation_description": "criteria_met_rate", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.75, + "uncertainty": { + "standard_deviation": 0.3048892494449146, + "num_samples": 8 + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + } + } + }, + { + "evaluation_name": "mean on healthbench/S-adaptive/+4ep/377e41e3 for scorer healthbench_score", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "2", + "sample_ids": "1fb6ab91-a999-437d-8c93-c5937c0a89a3,42af0c9f-a715-45c4-ab24-c2948f6f943d" + } + }, + "evaluation_timestamp": "1776259833.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.3958333333333333, + "uncertainty": { + "num_samples": 8 + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/anthropic/claude-opus-4-5-20251101/7a41a4b3-c7a8-4b0f-ba8c-f8053affa01d_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "131b12ec4233b199b5db8073cea0d3ba056bd91b5bd35145259157d0190b95f3", + "total_rows": 8 + } +} \ No newline at end of file diff --git a/data/healthbench/anthropic/claude-opus-4-5-20251101/7a41a4b3-c7a8-4b0f-ba8c-f8053affa01d_samples.jsonl b/data/healthbench/anthropic/claude-opus-4-5-20251101/7a41a4b3-c7a8-4b0f-ba8c-f8053affa01d_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..cac7ec2cbdde9979ee7955dc3b50793d9438d3d9 --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-5-20251101/7a41a4b3-c7a8-4b0f-ba8c-f8053affa01d_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:131b12ec4233b199b5db8073cea0d3ba056bd91b5bd35145259157d0190b95f3 +size 174785 diff --git a/data/healthbench/anthropic/claude-opus-4-5-20251101/905586be-0e41-4f83-8675-fa1474ce7284.json b/data/healthbench/anthropic/claude-opus-4-5-20251101/905586be-0e41-4f83-8675-fa1474ce7284.json new file mode 100644 index 0000000000000000000000000000000000000000..2c77d1c5a3669049e68d2b452f648192c8f7cc97 --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-5-20251101/905586be-0e41-4f83-8675-fa1474ce7284.json @@ -0,0 +1,292 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/anthropic_claude-opus-4-5-20251101/1786163946.64229", + "evaluation_timestamp": "1776279387.0", + "retrieved_timestamp": "1786163946.64229", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-5-20251101", + "id": "anthropic/claude-opus-4-5-20251101", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "bootstrap_score on healthbench/S-adaptive/+9ep/c8d85210 for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "20", + "sample_ids": "4fc7b71f-9591-4ae1-bfca-2929c7225182,da05b57a-d776-45b6-96af-1750e0254eec,ef446d15-d019-4c5a-a052-ae8cc3ea3b81,0fe4ea94-633d-4f73-985b-cbcd9f5a0270,cbb0ca6e-aa5e-4683-b107-d7dfa4cb60ce,ae6ec2ce-5c23-4e65-802f-cef9f1624d71,cd3dbc9f-a39c-4d13-af60-cba7dbc3baad,02423ab9-f3ed-4096-a7df-bfa3ebb40c7c,084b0c27-79c5-42f6-9cbd-690d0e79167a,42465540-8af1-423b-9b35-a31de1e1d49e,74a50705-7db2-44be-84dc-92b2d960d344,3b79284e-1143-46b4-b6a0-78e5448ff773,afbfc79d-6455-40b8-b475-1bcde4ef0cc9,6d705f80-8259-4ca1-8645-163f42c85a3c,129057dd-4bf4-472d-a8cf-a063dbed60ad,2833344d-bff9-4bbe-8fab-6b13d64c6b2a,48669cdc-1a64-4c99-a6e8-9a7181341fa2,be70739f-234a-4a0f-9b14-d15b8800577b,4047884e-4629-457c-9d77-130e937fae60,eb97bae4-430e-45cd-a065-2df3ab5c600e" + } + }, + "evaluation_timestamp": "1776279387.0", + "metric_config": { + "evaluation_description": "bootstrap_score", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.2755934850299678, + "uncertainty": { + "standard_deviation": 0.04741507717085444, + "num_samples": 180 + }, + "details": { + "total_matched_trajectories": "30", + "stopping_reason_count_repetition_guard": "30", + "avg_total_tokens_target_model": "9932.03", + "total_total_tokens_target_model": "297961.0", + "avg_total_tokens_other_models": "52966.50", + "total_total_tokens_other_models": "1588995.0", + "avg_total_tokens_all_models": "62898.53", + "total_total_tokens_all_models": "1886956.0", + "avg_turn_count": "30.67", + "total_turn_count": "920.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + } + } + }, + { + "evaluation_name": "std on healthbench/S-adaptive/+9ep/c8d85210 for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "20", + "sample_ids": "4fc7b71f-9591-4ae1-bfca-2929c7225182,da05b57a-d776-45b6-96af-1750e0254eec,ef446d15-d019-4c5a-a052-ae8cc3ea3b81,0fe4ea94-633d-4f73-985b-cbcd9f5a0270,cbb0ca6e-aa5e-4683-b107-d7dfa4cb60ce,ae6ec2ce-5c23-4e65-802f-cef9f1624d71,cd3dbc9f-a39c-4d13-af60-cba7dbc3baad,02423ab9-f3ed-4096-a7df-bfa3ebb40c7c,084b0c27-79c5-42f6-9cbd-690d0e79167a,42465540-8af1-423b-9b35-a31de1e1d49e,74a50705-7db2-44be-84dc-92b2d960d344,3b79284e-1143-46b4-b6a0-78e5448ff773,afbfc79d-6455-40b8-b475-1bcde4ef0cc9,6d705f80-8259-4ca1-8645-163f42c85a3c,129057dd-4bf4-472d-a8cf-a063dbed60ad,2833344d-bff9-4bbe-8fab-6b13d64c6b2a,48669cdc-1a64-4c99-a6e8-9a7181341fa2,be70739f-234a-4a0f-9b14-d15b8800577b,4047884e-4629-457c-9d77-130e937fae60,eb97bae4-430e-45cd-a065-2df3ab5c600e" + } + }, + "evaluation_timestamp": "1776279387.0", + "metric_config": { + "evaluation_description": "std", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.04741507717085444, + "uncertainty": { + "standard_deviation": 0.04741507717085444, + "num_samples": 180 + }, + "details": { + "total_matched_trajectories": "30", + "stopping_reason_count_repetition_guard": "30", + "avg_total_tokens_target_model": "9932.03", + "total_total_tokens_target_model": "297961.0", + "avg_total_tokens_other_models": "52966.50", + "total_total_tokens_other_models": "1588995.0", + "avg_total_tokens_all_models": "62898.53", + "total_total_tokens_all_models": "1886956.0", + "avg_turn_count": "30.67", + "total_turn_count": "920.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + } + } + }, + { + "evaluation_name": "criteria_met_rate on healthbench/S-adaptive/+9ep/c8d85210 for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "20", + "sample_ids": "4fc7b71f-9591-4ae1-bfca-2929c7225182,da05b57a-d776-45b6-96af-1750e0254eec,ef446d15-d019-4c5a-a052-ae8cc3ea3b81,0fe4ea94-633d-4f73-985b-cbcd9f5a0270,cbb0ca6e-aa5e-4683-b107-d7dfa4cb60ce,ae6ec2ce-5c23-4e65-802f-cef9f1624d71,cd3dbc9f-a39c-4d13-af60-cba7dbc3baad,02423ab9-f3ed-4096-a7df-bfa3ebb40c7c,084b0c27-79c5-42f6-9cbd-690d0e79167a,42465540-8af1-423b-9b35-a31de1e1d49e,74a50705-7db2-44be-84dc-92b2d960d344,3b79284e-1143-46b4-b6a0-78e5448ff773,afbfc79d-6455-40b8-b475-1bcde4ef0cc9,6d705f80-8259-4ca1-8645-163f42c85a3c,129057dd-4bf4-472d-a8cf-a063dbed60ad,2833344d-bff9-4bbe-8fab-6b13d64c6b2a,48669cdc-1a64-4c99-a6e8-9a7181341fa2,be70739f-234a-4a0f-9b14-d15b8800577b,4047884e-4629-457c-9d77-130e937fae60,eb97bae4-430e-45cd-a065-2df3ab5c600e" + } + }, + "evaluation_timestamp": "1776279387.0", + "metric_config": { + "evaluation_description": "criteria_met_rate", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.30522088353413657, + "uncertainty": { + "standard_deviation": 0.04741507717085444, + "num_samples": 180 + }, + "details": { + "total_matched_trajectories": "30", + "stopping_reason_count_repetition_guard": "30", + "avg_total_tokens_target_model": "9932.03", + "total_total_tokens_target_model": "297961.0", + "avg_total_tokens_other_models": "52966.50", + "total_total_tokens_other_models": "1588995.0", + "avg_total_tokens_all_models": "62898.53", + "total_total_tokens_all_models": "1886956.0", + "avg_turn_count": "30.67", + "total_turn_count": "920.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + } + } + }, + { + "evaluation_name": "mean on healthbench/S-adaptive/+9ep/c8d85210 for scorer healthbench_score", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "20", + "sample_ids": "4fc7b71f-9591-4ae1-bfca-2929c7225182,da05b57a-d776-45b6-96af-1750e0254eec,ef446d15-d019-4c5a-a052-ae8cc3ea3b81,0fe4ea94-633d-4f73-985b-cbcd9f5a0270,cbb0ca6e-aa5e-4683-b107-d7dfa4cb60ce,ae6ec2ce-5c23-4e65-802f-cef9f1624d71,cd3dbc9f-a39c-4d13-af60-cba7dbc3baad,02423ab9-f3ed-4096-a7df-bfa3ebb40c7c,084b0c27-79c5-42f6-9cbd-690d0e79167a,42465540-8af1-423b-9b35-a31de1e1d49e,74a50705-7db2-44be-84dc-92b2d960d344,3b79284e-1143-46b4-b6a0-78e5448ff773,afbfc79d-6455-40b8-b475-1bcde4ef0cc9,6d705f80-8259-4ca1-8645-163f42c85a3c,129057dd-4bf4-472d-a8cf-a063dbed60ad,2833344d-bff9-4bbe-8fab-6b13d64c6b2a,48669cdc-1a64-4c99-a6e8-9a7181341fa2,be70739f-234a-4a0f-9b14-d15b8800577b,4047884e-4629-457c-9d77-130e937fae60,eb97bae4-430e-45cd-a065-2df3ab5c600e" + } + }, + "evaluation_timestamp": "1776279387.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.2755934850299678, + "uncertainty": { + "num_samples": 180 + }, + "details": { + "total_matched_trajectories": "30", + "stopping_reason_count_repetition_guard": "30", + "avg_total_tokens_target_model": "9932.03", + "total_total_tokens_target_model": "297961.0", + "avg_total_tokens_other_models": "52966.50", + "total_total_tokens_other_models": "1588995.0", + "avg_total_tokens_all_models": "62898.53", + "total_total_tokens_all_models": "1886956.0", + "avg_turn_count": "30.67", + "total_turn_count": "920.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/anthropic/claude-opus-4-5-20251101/905586be-0e41-4f83-8675-fa1474ce7284_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "b376e355d4597b19146f02b34a375abfc7ee04ea2f7a8cf8e0389a17d6aa0df0", + "total_rows": 180 + } +} \ No newline at end of file diff --git a/data/healthbench/anthropic/claude-opus-4-5-20251101/905586be-0e41-4f83-8675-fa1474ce7284_samples.jsonl b/data/healthbench/anthropic/claude-opus-4-5-20251101/905586be-0e41-4f83-8675-fa1474ce7284_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..0aba7999f9bc91045a81bb1bf28c0e7381010e55 --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-5-20251101/905586be-0e41-4f83-8675-fa1474ce7284_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b376e355d4597b19146f02b34a375abfc7ee04ea2f7a8cf8e0389a17d6aa0df0 +size 13785770 diff --git a/data/healthbench/anthropic/claude-opus-4-5-20251101/9451b74f-abe7-4315-b10f-2f86ab2520b3.json b/data/healthbench/anthropic/claude-opus-4-5-20251101/9451b74f-abe7-4315-b10f-2f86ab2520b3.json new file mode 100644 index 0000000000000000000000000000000000000000..d1da8be6da96e404b6f7062bb602864debb66f84 --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-5-20251101/9451b74f-abe7-4315-b10f-2f86ab2520b3.json @@ -0,0 +1,316 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/anthropic_claude-opus-4-5-20251101/1786163919.003918", + "evaluation_timestamp": "1776696116.0", + "retrieved_timestamp": "1786163919.003918", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-5-20251101", + "id": "anthropic/claude-opus-4-5-20251101", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "bootstrap_score on healthbench/S-adaptive/+4ep/e182d015 for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "50", + "sample_ids": "c3eeecfc-de96-480f-929d-ac4a5e5a541c,269074dc-9495-484b-b2f9-0f0943c0f816,b97b1bae-b09f-4073-8e16-540bba766c60,8e88f21a-24c4-4f9c-b848-faf5e10e31b9,a49eb9af-926e-4f7c-94d0-133729059116,1049130c-e9c9-461d-b080-90027bc011c0,13bb02e4-844e-4972-9200-2befa5f911d8,2694a753-2f57-41eb-a7f9-8c4c7acdd2e8,b160df9f-be98-4a4d-9cc8-11a612ffaf87,5e0a84d7-b972-484c-a649-469b15f0a90f,c7606b1b-ca4d-4a71-8277-bcddff09a653,20cdebb7-9189-4f09-ad21-bfc8e8b75680,f24935ac-0530-4063-b2f1-ac45cd265037,1b8f338a-2e98-4a42-ba80-cba266f4b9fa,6f827e79-05c4-4383-8ed4-16d36260f29f,c835bebe-e998-4dcd-843a-2de03d4aabc6,70afd1ac-85ea-4455-a5a2-15b35378e10a,aa7a760a-5173-4a39-af83-48d6ae2230f0,f712eb09-82af-4fe8-800d-bc3af772a9ba,8ee6a990-8966-4b42-baa8-6c7e9f74ce74,267bda16-fd4b-4a9e-b3ac-5602c2532eed,0e819a9c-851d-4a7a-9263-62cfc8ce1b48,328c55cf-0b57-4992-99b1-f6a21af62ade,7b415a2f-22b0-4735-a38e-567f1891acc5,c49ac752-3693-4be1-ab4d-91fbdd02f827,a079d928-7b78-4a85-a72b-4274d7fc2914,a20cbbf3-330b-46d9-a35b-92e89a022282,be6150a5-8060-4434-9724-afede307d495,3a896889-9029-4ef8-8684-ca378da4702b,7074bb7a-a595-434d-883e-95ef6f5f6667,9f7b555a-fc30-40c3-88d3-f2ece06454a2,a6204079-9e72-42ad-8f39-a65552fa2e68,c0d80b93-b2d5-4129-b998-b1312f98e2a8,3dc41419-2565-428e-b9b9-02e3fd93c470,a91a22f8-d2e6-4593-bcae-6d479ed38031,d0d3426f-8b7a-4501-adb4-86d8fa9099c8,a7dac6a9-eab5-4c01-882c-dbf6c774d7c3,c1240427-ae38-4335-9e80-492c13ef4893,b7ddde65-702e-4255-b1d7-09d01248e76d,00d34549-134c-48e2-b4bb-b54946da70cc,85c11e2a-628b-464f-a8bb-e97ab04a73eb,69a544f8-957d-4dda-a777-608ec7b95617,90789a49-ab1f-4229-bd89-e1c56ea8be3a,cd53e21a-e5ed-4d9b-a258-42a23e9715bb,8d409c7b-29d2-4df2-aa69-ab56c9339bc5,6566f489-f254-4bca-a86b-90f5ac7471d7,d9a44a36-6a3c-4d5d-9c66-b2c0afb479d1,9cf87956-37e6-4bed-9cea-a64b13ae0de5,a1a08e1d-1353-4d6f-94ce-f90872074c7c,59c68d10-55c1-4629-b05a-bedde7f343b5" + } + }, + "evaluation_timestamp": "1776696116.0", + "metric_config": { + "evaluation_description": "bootstrap_score", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.11398601531230274, + "uncertainty": { + "standard_deviation": 0.029049527153467647, + "num_samples": 200 + }, + "details": { + "total_matched_trajectories": "200", + "stopping_reason_count_repetition_guard": "200", + "avg_total_tokens_target_model": "14050.67", + "total_total_tokens_target_model": "2810133.0", + "avg_total_tokens_other_models": "55615.32", + "total_total_tokens_other_models": "11123064.0", + "avg_total_tokens_all_models": "69665.99", + "total_total_tokens_all_models": "13933197.0", + "avg_turn_count": "30.11", + "total_turn_count": "6022.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "3600", + "attempt_timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "3600", + "max_connections": "10" + } + } + }, + { + "evaluation_name": "std on healthbench/S-adaptive/+4ep/e182d015 for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "50", + "sample_ids": "c3eeecfc-de96-480f-929d-ac4a5e5a541c,269074dc-9495-484b-b2f9-0f0943c0f816,b97b1bae-b09f-4073-8e16-540bba766c60,8e88f21a-24c4-4f9c-b848-faf5e10e31b9,a49eb9af-926e-4f7c-94d0-133729059116,1049130c-e9c9-461d-b080-90027bc011c0,13bb02e4-844e-4972-9200-2befa5f911d8,2694a753-2f57-41eb-a7f9-8c4c7acdd2e8,b160df9f-be98-4a4d-9cc8-11a612ffaf87,5e0a84d7-b972-484c-a649-469b15f0a90f,c7606b1b-ca4d-4a71-8277-bcddff09a653,20cdebb7-9189-4f09-ad21-bfc8e8b75680,f24935ac-0530-4063-b2f1-ac45cd265037,1b8f338a-2e98-4a42-ba80-cba266f4b9fa,6f827e79-05c4-4383-8ed4-16d36260f29f,c835bebe-e998-4dcd-843a-2de03d4aabc6,70afd1ac-85ea-4455-a5a2-15b35378e10a,aa7a760a-5173-4a39-af83-48d6ae2230f0,f712eb09-82af-4fe8-800d-bc3af772a9ba,8ee6a990-8966-4b42-baa8-6c7e9f74ce74,267bda16-fd4b-4a9e-b3ac-5602c2532eed,0e819a9c-851d-4a7a-9263-62cfc8ce1b48,328c55cf-0b57-4992-99b1-f6a21af62ade,7b415a2f-22b0-4735-a38e-567f1891acc5,c49ac752-3693-4be1-ab4d-91fbdd02f827,a079d928-7b78-4a85-a72b-4274d7fc2914,a20cbbf3-330b-46d9-a35b-92e89a022282,be6150a5-8060-4434-9724-afede307d495,3a896889-9029-4ef8-8684-ca378da4702b,7074bb7a-a595-434d-883e-95ef6f5f6667,9f7b555a-fc30-40c3-88d3-f2ece06454a2,a6204079-9e72-42ad-8f39-a65552fa2e68,c0d80b93-b2d5-4129-b998-b1312f98e2a8,3dc41419-2565-428e-b9b9-02e3fd93c470,a91a22f8-d2e6-4593-bcae-6d479ed38031,d0d3426f-8b7a-4501-adb4-86d8fa9099c8,a7dac6a9-eab5-4c01-882c-dbf6c774d7c3,c1240427-ae38-4335-9e80-492c13ef4893,b7ddde65-702e-4255-b1d7-09d01248e76d,00d34549-134c-48e2-b4bb-b54946da70cc,85c11e2a-628b-464f-a8bb-e97ab04a73eb,69a544f8-957d-4dda-a777-608ec7b95617,90789a49-ab1f-4229-bd89-e1c56ea8be3a,cd53e21a-e5ed-4d9b-a258-42a23e9715bb,8d409c7b-29d2-4df2-aa69-ab56c9339bc5,6566f489-f254-4bca-a86b-90f5ac7471d7,d9a44a36-6a3c-4d5d-9c66-b2c0afb479d1,9cf87956-37e6-4bed-9cea-a64b13ae0de5,a1a08e1d-1353-4d6f-94ce-f90872074c7c,59c68d10-55c1-4629-b05a-bedde7f343b5" + } + }, + "evaluation_timestamp": "1776696116.0", + "metric_config": { + "evaluation_description": "std", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.029049527153467647, + "uncertainty": { + "standard_deviation": 0.029049527153467647, + "num_samples": 200 + }, + "details": { + "total_matched_trajectories": "200", + "stopping_reason_count_repetition_guard": "200", + "avg_total_tokens_target_model": "14050.67", + "total_total_tokens_target_model": "2810133.0", + "avg_total_tokens_other_models": "55615.32", + "total_total_tokens_other_models": "11123064.0", + "avg_total_tokens_all_models": "69665.99", + "total_total_tokens_all_models": "13933197.0", + "avg_turn_count": "30.11", + "total_turn_count": "6022.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "3600", + "attempt_timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "3600", + "max_connections": "10" + } + } + }, + { + "evaluation_name": "criteria_met_rate on healthbench/S-adaptive/+4ep/e182d015 for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "50", + "sample_ids": "c3eeecfc-de96-480f-929d-ac4a5e5a541c,269074dc-9495-484b-b2f9-0f0943c0f816,b97b1bae-b09f-4073-8e16-540bba766c60,8e88f21a-24c4-4f9c-b848-faf5e10e31b9,a49eb9af-926e-4f7c-94d0-133729059116,1049130c-e9c9-461d-b080-90027bc011c0,13bb02e4-844e-4972-9200-2befa5f911d8,2694a753-2f57-41eb-a7f9-8c4c7acdd2e8,b160df9f-be98-4a4d-9cc8-11a612ffaf87,5e0a84d7-b972-484c-a649-469b15f0a90f,c7606b1b-ca4d-4a71-8277-bcddff09a653,20cdebb7-9189-4f09-ad21-bfc8e8b75680,f24935ac-0530-4063-b2f1-ac45cd265037,1b8f338a-2e98-4a42-ba80-cba266f4b9fa,6f827e79-05c4-4383-8ed4-16d36260f29f,c835bebe-e998-4dcd-843a-2de03d4aabc6,70afd1ac-85ea-4455-a5a2-15b35378e10a,aa7a760a-5173-4a39-af83-48d6ae2230f0,f712eb09-82af-4fe8-800d-bc3af772a9ba,8ee6a990-8966-4b42-baa8-6c7e9f74ce74,267bda16-fd4b-4a9e-b3ac-5602c2532eed,0e819a9c-851d-4a7a-9263-62cfc8ce1b48,328c55cf-0b57-4992-99b1-f6a21af62ade,7b415a2f-22b0-4735-a38e-567f1891acc5,c49ac752-3693-4be1-ab4d-91fbdd02f827,a079d928-7b78-4a85-a72b-4274d7fc2914,a20cbbf3-330b-46d9-a35b-92e89a022282,be6150a5-8060-4434-9724-afede307d495,3a896889-9029-4ef8-8684-ca378da4702b,7074bb7a-a595-434d-883e-95ef6f5f6667,9f7b555a-fc30-40c3-88d3-f2ece06454a2,a6204079-9e72-42ad-8f39-a65552fa2e68,c0d80b93-b2d5-4129-b998-b1312f98e2a8,3dc41419-2565-428e-b9b9-02e3fd93c470,a91a22f8-d2e6-4593-bcae-6d479ed38031,d0d3426f-8b7a-4501-adb4-86d8fa9099c8,a7dac6a9-eab5-4c01-882c-dbf6c774d7c3,c1240427-ae38-4335-9e80-492c13ef4893,b7ddde65-702e-4255-b1d7-09d01248e76d,00d34549-134c-48e2-b4bb-b54946da70cc,85c11e2a-628b-464f-a8bb-e97ab04a73eb,69a544f8-957d-4dda-a777-608ec7b95617,90789a49-ab1f-4229-bd89-e1c56ea8be3a,cd53e21a-e5ed-4d9b-a258-42a23e9715bb,8d409c7b-29d2-4df2-aa69-ab56c9339bc5,6566f489-f254-4bca-a86b-90f5ac7471d7,d9a44a36-6a3c-4d5d-9c66-b2c0afb479d1,9cf87956-37e6-4bed-9cea-a64b13ae0de5,a1a08e1d-1353-4d6f-94ce-f90872074c7c,59c68d10-55c1-4629-b05a-bedde7f343b5" + } + }, + "evaluation_timestamp": "1776696116.0", + "metric_config": { + "evaluation_description": "criteria_met_rate", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.2425, + "uncertainty": { + "standard_deviation": 0.029049527153467647, + "num_samples": 200 + }, + "details": { + "total_matched_trajectories": "200", + "stopping_reason_count_repetition_guard": "200", + "avg_total_tokens_target_model": "14050.67", + "total_total_tokens_target_model": "2810133.0", + "avg_total_tokens_other_models": "55615.32", + "total_total_tokens_other_models": "11123064.0", + "avg_total_tokens_all_models": "69665.99", + "total_total_tokens_all_models": "13933197.0", + "avg_turn_count": "30.11", + "total_turn_count": "6022.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "3600", + "attempt_timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "3600", + "max_connections": "10" + } + } + }, + { + "evaluation_name": "mean on healthbench/S-adaptive/+4ep/e182d015 for scorer healthbench_score", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "50", + "sample_ids": "c3eeecfc-de96-480f-929d-ac4a5e5a541c,269074dc-9495-484b-b2f9-0f0943c0f816,b97b1bae-b09f-4073-8e16-540bba766c60,8e88f21a-24c4-4f9c-b848-faf5e10e31b9,a49eb9af-926e-4f7c-94d0-133729059116,1049130c-e9c9-461d-b080-90027bc011c0,13bb02e4-844e-4972-9200-2befa5f911d8,2694a753-2f57-41eb-a7f9-8c4c7acdd2e8,b160df9f-be98-4a4d-9cc8-11a612ffaf87,5e0a84d7-b972-484c-a649-469b15f0a90f,c7606b1b-ca4d-4a71-8277-bcddff09a653,20cdebb7-9189-4f09-ad21-bfc8e8b75680,f24935ac-0530-4063-b2f1-ac45cd265037,1b8f338a-2e98-4a42-ba80-cba266f4b9fa,6f827e79-05c4-4383-8ed4-16d36260f29f,c835bebe-e998-4dcd-843a-2de03d4aabc6,70afd1ac-85ea-4455-a5a2-15b35378e10a,aa7a760a-5173-4a39-af83-48d6ae2230f0,f712eb09-82af-4fe8-800d-bc3af772a9ba,8ee6a990-8966-4b42-baa8-6c7e9f74ce74,267bda16-fd4b-4a9e-b3ac-5602c2532eed,0e819a9c-851d-4a7a-9263-62cfc8ce1b48,328c55cf-0b57-4992-99b1-f6a21af62ade,7b415a2f-22b0-4735-a38e-567f1891acc5,c49ac752-3693-4be1-ab4d-91fbdd02f827,a079d928-7b78-4a85-a72b-4274d7fc2914,a20cbbf3-330b-46d9-a35b-92e89a022282,be6150a5-8060-4434-9724-afede307d495,3a896889-9029-4ef8-8684-ca378da4702b,7074bb7a-a595-434d-883e-95ef6f5f6667,9f7b555a-fc30-40c3-88d3-f2ece06454a2,a6204079-9e72-42ad-8f39-a65552fa2e68,c0d80b93-b2d5-4129-b998-b1312f98e2a8,3dc41419-2565-428e-b9b9-02e3fd93c470,a91a22f8-d2e6-4593-bcae-6d479ed38031,d0d3426f-8b7a-4501-adb4-86d8fa9099c8,a7dac6a9-eab5-4c01-882c-dbf6c774d7c3,c1240427-ae38-4335-9e80-492c13ef4893,b7ddde65-702e-4255-b1d7-09d01248e76d,00d34549-134c-48e2-b4bb-b54946da70cc,85c11e2a-628b-464f-a8bb-e97ab04a73eb,69a544f8-957d-4dda-a777-608ec7b95617,90789a49-ab1f-4229-bd89-e1c56ea8be3a,cd53e21a-e5ed-4d9b-a258-42a23e9715bb,8d409c7b-29d2-4df2-aa69-ab56c9339bc5,6566f489-f254-4bca-a86b-90f5ac7471d7,d9a44a36-6a3c-4d5d-9c66-b2c0afb479d1,9cf87956-37e6-4bed-9cea-a64b13ae0de5,a1a08e1d-1353-4d6f-94ce-f90872074c7c,59c68d10-55c1-4629-b05a-bedde7f343b5" + } + }, + "evaluation_timestamp": "1776696116.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.11398601531230274, + "uncertainty": { + "num_samples": 200 + }, + "details": { + "total_matched_trajectories": "200", + "stopping_reason_count_repetition_guard": "200", + "avg_total_tokens_target_model": "14050.67", + "total_total_tokens_target_model": "2810133.0", + "avg_total_tokens_other_models": "55615.32", + "total_total_tokens_other_models": "11123064.0", + "avg_total_tokens_all_models": "69665.99", + "total_total_tokens_all_models": "13933197.0", + "avg_turn_count": "30.11", + "total_turn_count": "6022.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "3600", + "attempt_timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "3600", + "max_connections": "10" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/anthropic/claude-opus-4-5-20251101/9451b74f-abe7-4315-b10f-2f86ab2520b3_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "aa9d3c014d09d0ffd55ebe7ce9cd405c0c8becb778ec83a48079c8bb35434e44", + "total_rows": 200 + } +} \ No newline at end of file diff --git a/data/healthbench/anthropic/claude-opus-4-5-20251101/9451b74f-abe7-4315-b10f-2f86ab2520b3_samples.jsonl b/data/healthbench/anthropic/claude-opus-4-5-20251101/9451b74f-abe7-4315-b10f-2f86ab2520b3_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..ebcf82bd2d34ed5b898debbaede232d1a967513c --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-5-20251101/9451b74f-abe7-4315-b10f-2f86ab2520b3_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:aa9d3c014d09d0ffd55ebe7ce9cd405c0c8becb778ec83a48079c8bb35434e44 +size 13589312 diff --git a/data/healthbench/anthropic/claude-opus-4-5-20251101/b08f4d6a-6344-4cf1-a890-964819978804.json b/data/healthbench/anthropic/claude-opus-4-5-20251101/b08f4d6a-6344-4cf1-a890-964819978804.json new file mode 100644 index 0000000000000000000000000000000000000000..98d1ddbcbbc3e698d063d4a3cf55a0ac528a7e99 --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-5-20251101/b08f4d6a-6344-4cf1-a890-964819978804.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/anthropic_claude-opus-4-5-20251101/1786166636.278326", + "evaluation_timestamp": "1776165241.0", + "retrieved_timestamp": "1786166636.278326", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-5-20251101", + "id": "anthropic/claude-opus-4-5-20251101", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/anthropic/claude-opus-4-5-20251101/b08f4d6a-6344-4cf1-a890-964819978804_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "ab3078f67c4ed51e8e7404543c3a49dcee6b6af48ef0b2e346047c6fd17c8460", + "total_rows": 350 + } +} \ No newline at end of file diff --git a/data/healthbench/anthropic/claude-opus-4-5-20251101/b08f4d6a-6344-4cf1-a890-964819978804_samples.jsonl b/data/healthbench/anthropic/claude-opus-4-5-20251101/b08f4d6a-6344-4cf1-a890-964819978804_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..1d886026e67c3661f159af9b4f588213d086a3ed --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-5-20251101/b08f4d6a-6344-4cf1-a890-964819978804_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ab3078f67c4ed51e8e7404543c3a49dcee6b6af48ef0b2e346047c6fd17c8460 +size 23115374 diff --git a/data/healthbench/anthropic/claude-opus-4-5-20251101/c86f7afb-8c23-4756-a0ab-e2ad1732ca88.json b/data/healthbench/anthropic/claude-opus-4-5-20251101/c86f7afb-8c23-4756-a0ab-e2ad1732ca88.json new file mode 100644 index 0000000000000000000000000000000000000000..e5bfc8185c57795b8891b5c3b189729c33f44cf1 --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-5-20251101/c86f7afb-8c23-4756-a0ab-e2ad1732ca88.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/anthropic_claude-opus-4-5-20251101/1786168551.423578", + "evaluation_timestamp": "1776682987.0", + "retrieved_timestamp": "1786168551.423578", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-5-20251101", + "id": "anthropic/claude-opus-4-5-20251101", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/anthropic/claude-opus-4-5-20251101/c86f7afb-8c23-4756-a0ab-e2ad1732ca88_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "4a2b4aa9f4771c8822fd2bbdab68b5a8d4adb128b08e104206cacbc01f434e6b", + "total_rows": 50 + } +} \ No newline at end of file diff --git a/data/healthbench/anthropic/claude-opus-4-5-20251101/c86f7afb-8c23-4756-a0ab-e2ad1732ca88_samples.jsonl b/data/healthbench/anthropic/claude-opus-4-5-20251101/c86f7afb-8c23-4756-a0ab-e2ad1732ca88_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..98ea1fc487bd4a03bf5748ce17d3e47f6dd61fd5 --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-5-20251101/c86f7afb-8c23-4756-a0ab-e2ad1732ca88_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4a2b4aa9f4771c8822fd2bbdab68b5a8d4adb128b08e104206cacbc01f434e6b +size 3496251 diff --git a/data/healthbench/anthropic/claude-opus-4-5-20251101/c92fabc7-f5e0-4669-8dbe-3154c456a817.json b/data/healthbench/anthropic/claude-opus-4-5-20251101/c92fabc7-f5e0-4669-8dbe-3154c456a817.json new file mode 100644 index 0000000000000000000000000000000000000000..1b0222dd80628b74b0920c39b341578a446649c0 --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-5-20251101/c92fabc7-f5e0-4669-8dbe-3154c456a817.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/anthropic_claude-opus-4-5-20251101/1786168558.231424", + "evaluation_timestamp": "1776076027.0", + "retrieved_timestamp": "1786168558.231424", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-5-20251101", + "id": "anthropic/claude-opus-4-5-20251101", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/anthropic/claude-opus-4-5-20251101/c92fabc7-f5e0-4669-8dbe-3154c456a817_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "c6cb9efb0100d2288adadd5b70d06f8a3bafd221cbdd5b23ab281fb86b1f562f", + "total_rows": 80 + } +} \ No newline at end of file diff --git a/data/healthbench/anthropic/claude-opus-4-5-20251101/c92fabc7-f5e0-4669-8dbe-3154c456a817_samples.jsonl b/data/healthbench/anthropic/claude-opus-4-5-20251101/c92fabc7-f5e0-4669-8dbe-3154c456a817_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..2048894329dd1c9c6ff2d95587ae567379d00ada --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-5-20251101/c92fabc7-f5e0-4669-8dbe-3154c456a817_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c6cb9efb0100d2288adadd5b70d06f8a3bafd221cbdd5b23ab281fb86b1f562f +size 4806019 diff --git a/data/healthbench/anthropic/claude-opus-4-5-20251101/d4e50f98-4af9-4908-b19d-118d2daf8bbd.json b/data/healthbench/anthropic/claude-opus-4-5-20251101/d4e50f98-4af9-4908-b19d-118d2daf8bbd.json new file mode 100644 index 0000000000000000000000000000000000000000..3b293c7365fea8a1b166e94822ef8b64ed0e6042 --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-5-20251101/d4e50f98-4af9-4908-b19d-118d2daf8bbd.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/anthropic_claude-opus-4-5-20251101/1786168546.877655", + "evaluation_timestamp": "1776107987.0", + "retrieved_timestamp": "1786168546.877655", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-5-20251101", + "id": "anthropic/claude-opus-4-5-20251101", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/anthropic/claude-opus-4-5-20251101/d4e50f98-4af9-4908-b19d-118d2daf8bbd_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "4f70fe2827642e22706601326659b1bc7aefb38afa0765e8c710a173b928f41a", + "total_rows": 30 + } +} \ No newline at end of file diff --git a/data/healthbench/anthropic/claude-opus-4-5-20251101/d4e50f98-4af9-4908-b19d-118d2daf8bbd_samples.jsonl b/data/healthbench/anthropic/claude-opus-4-5-20251101/d4e50f98-4af9-4908-b19d-118d2daf8bbd_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..9a61f08a1fb4b1b4e1243fd11d9c8d9c6f20cab9 --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-5-20251101/d4e50f98-4af9-4908-b19d-118d2daf8bbd_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f70fe2827642e22706601326659b1bc7aefb38afa0765e8c710a173b928f41a +size 1856095 diff --git a/data/healthbench/anthropic/claude-opus-4-5-20251101/deeb16ae-6ac9-49cd-8791-fe832eb188aa.json b/data/healthbench/anthropic/claude-opus-4-5-20251101/deeb16ae-6ac9-49cd-8791-fe832eb188aa.json new file mode 100644 index 0000000000000000000000000000000000000000..571b3cbbd69b9b67431798498a3d4f8c5b550335 --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-5-20251101/deeb16ae-6ac9-49cd-8791-fe832eb188aa.json @@ -0,0 +1,292 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/anthropic_claude-opus-4-5-20251101/1786163938.536178", + "evaluation_timestamp": "1776165242.0", + "retrieved_timestamp": "1786163938.536178", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-5-20251101", + "id": "anthropic/claude-opus-4-5-20251101", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "bootstrap_score on healthbench/S-adaptive/+7ep for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "30", + "sample_ids": "1f548d5b-cd00-49a0-b327-283a2e00debd,0b8f1d60-2081-4562-98f7-b6a976fe1c6d,6f7a2ee9-e9c6-42d8-b79f-22dea966b8d2,19ec4833-86e9-4166-8b82-d1da09f31fd7,7ebc830a-8dbd-489b-9d61-4d8bacf0db8d,c971f9d1-5f6a-464e-b282-41c8f0d82f58,5a6e4a41-3ea6-4050-a971-93433fe34877,78510411-e03b-4f93-a369-c7e72d2e4908,bbd759b4-8e4d-4db7-bc36-f2e069396c1c,8f2a65de-dea7-48e8-8adb-6194eca26c08,6d5f483c-3e86-456d-bfd5-4e28de699ae5,f05491d8-d160-4b2b-bd5b-59d757585e39,85408180-238e-4c1b-92c1-55d8ce6c3717,ecd7453e-1dc1-4e52-bb95-67e8e5788573,0e073591-b3e7-4dc8-94d0-dc5aa93ca35c,7637b987-8fe2-49d0-8372-1ebf17284d70,5b294937-13e5-424e-8bb4-5d1904a2344a,c5bf7fc3-dfb0-4b47-9bbe-222e87b952e2,83cf8f2d-2857-4f01-a283-9595d8f4ae8e,7a4548e6-38b7-48dd-9088-d7d1219f7852,e0d1b955-1150-457e-96d7-91a9481999cd,24f9a6e7-b214-4011-94c4-6502f249a621,6a97773b-c33c-4f2e-9ac3-fdd53de0f175,f32ee0c1-ea9c-4cfb-8332-bd9579d23024,b6eb69cb-b911-4c81-a7a1-357d21d27109,c0dc053b-157b-4d13-9956-213b67ff6a36,3bb4a735-ace7-4342-b7a5-0bae083e6f82,66b448bf-2e49-488d-9d0b-f9912ab3f0f8,b12d2453-fd3b-462c-b9cc-554ff16bfd3f,0a548d04-0973-4343-a3bf-07f730bed2f6" + } + }, + "evaluation_timestamp": "1776165242.0", + "metric_config": { + "evaluation_description": "bootstrap_score", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.3849898941107965, + "uncertainty": { + "standard_deviation": 0.05250011500444398, + "num_samples": 210 + }, + "details": { + "total_matched_trajectories": "20", + "stopping_reason_count_repetition_guard": "20", + "avg_total_tokens_target_model": "11732.55", + "total_total_tokens_target_model": "234651.0", + "avg_total_tokens_other_models": "61776.40", + "total_total_tokens_other_models": "1235528.0", + "avg_total_tokens_all_models": "73508.95", + "total_total_tokens_all_models": "1470179.0", + "avg_turn_count": "32.70", + "total_turn_count": "654.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + } + } + }, + { + "evaluation_name": "std on healthbench/S-adaptive/+7ep for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "30", + "sample_ids": "1f548d5b-cd00-49a0-b327-283a2e00debd,0b8f1d60-2081-4562-98f7-b6a976fe1c6d,6f7a2ee9-e9c6-42d8-b79f-22dea966b8d2,19ec4833-86e9-4166-8b82-d1da09f31fd7,7ebc830a-8dbd-489b-9d61-4d8bacf0db8d,c971f9d1-5f6a-464e-b282-41c8f0d82f58,5a6e4a41-3ea6-4050-a971-93433fe34877,78510411-e03b-4f93-a369-c7e72d2e4908,bbd759b4-8e4d-4db7-bc36-f2e069396c1c,8f2a65de-dea7-48e8-8adb-6194eca26c08,6d5f483c-3e86-456d-bfd5-4e28de699ae5,f05491d8-d160-4b2b-bd5b-59d757585e39,85408180-238e-4c1b-92c1-55d8ce6c3717,ecd7453e-1dc1-4e52-bb95-67e8e5788573,0e073591-b3e7-4dc8-94d0-dc5aa93ca35c,7637b987-8fe2-49d0-8372-1ebf17284d70,5b294937-13e5-424e-8bb4-5d1904a2344a,c5bf7fc3-dfb0-4b47-9bbe-222e87b952e2,83cf8f2d-2857-4f01-a283-9595d8f4ae8e,7a4548e6-38b7-48dd-9088-d7d1219f7852,e0d1b955-1150-457e-96d7-91a9481999cd,24f9a6e7-b214-4011-94c4-6502f249a621,6a97773b-c33c-4f2e-9ac3-fdd53de0f175,f32ee0c1-ea9c-4cfb-8332-bd9579d23024,b6eb69cb-b911-4c81-a7a1-357d21d27109,c0dc053b-157b-4d13-9956-213b67ff6a36,3bb4a735-ace7-4342-b7a5-0bae083e6f82,66b448bf-2e49-488d-9d0b-f9912ab3f0f8,b12d2453-fd3b-462c-b9cc-554ff16bfd3f,0a548d04-0973-4343-a3bf-07f730bed2f6" + } + }, + "evaluation_timestamp": "1776165242.0", + "metric_config": { + "evaluation_description": "std", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.05250011500444398, + "uncertainty": { + "standard_deviation": 0.05250011500444398, + "num_samples": 210 + }, + "details": { + "total_matched_trajectories": "20", + "stopping_reason_count_repetition_guard": "20", + "avg_total_tokens_target_model": "11732.55", + "total_total_tokens_target_model": "234651.0", + "avg_total_tokens_other_models": "61776.40", + "total_total_tokens_other_models": "1235528.0", + "avg_total_tokens_all_models": "73508.95", + "total_total_tokens_all_models": "1470179.0", + "avg_turn_count": "32.70", + "total_turn_count": "654.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + } + } + }, + { + "evaluation_name": "criteria_met_rate on healthbench/S-adaptive/+7ep for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "30", + "sample_ids": "1f548d5b-cd00-49a0-b327-283a2e00debd,0b8f1d60-2081-4562-98f7-b6a976fe1c6d,6f7a2ee9-e9c6-42d8-b79f-22dea966b8d2,19ec4833-86e9-4166-8b82-d1da09f31fd7,7ebc830a-8dbd-489b-9d61-4d8bacf0db8d,c971f9d1-5f6a-464e-b282-41c8f0d82f58,5a6e4a41-3ea6-4050-a971-93433fe34877,78510411-e03b-4f93-a369-c7e72d2e4908,bbd759b4-8e4d-4db7-bc36-f2e069396c1c,8f2a65de-dea7-48e8-8adb-6194eca26c08,6d5f483c-3e86-456d-bfd5-4e28de699ae5,f05491d8-d160-4b2b-bd5b-59d757585e39,85408180-238e-4c1b-92c1-55d8ce6c3717,ecd7453e-1dc1-4e52-bb95-67e8e5788573,0e073591-b3e7-4dc8-94d0-dc5aa93ca35c,7637b987-8fe2-49d0-8372-1ebf17284d70,5b294937-13e5-424e-8bb4-5d1904a2344a,c5bf7fc3-dfb0-4b47-9bbe-222e87b952e2,83cf8f2d-2857-4f01-a283-9595d8f4ae8e,7a4548e6-38b7-48dd-9088-d7d1219f7852,e0d1b955-1150-457e-96d7-91a9481999cd,24f9a6e7-b214-4011-94c4-6502f249a621,6a97773b-c33c-4f2e-9ac3-fdd53de0f175,f32ee0c1-ea9c-4cfb-8332-bd9579d23024,b6eb69cb-b911-4c81-a7a1-357d21d27109,c0dc053b-157b-4d13-9956-213b67ff6a36,3bb4a735-ace7-4342-b7a5-0bae083e6f82,66b448bf-2e49-488d-9d0b-f9912ab3f0f8,b12d2453-fd3b-462c-b9cc-554ff16bfd3f,0a548d04-0973-4343-a3bf-07f730bed2f6" + } + }, + "evaluation_timestamp": "1776165242.0", + "metric_config": { + "evaluation_description": "criteria_met_rate", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.35186688311688313, + "uncertainty": { + "standard_deviation": 0.05250011500444398, + "num_samples": 210 + }, + "details": { + "total_matched_trajectories": "20", + "stopping_reason_count_repetition_guard": "20", + "avg_total_tokens_target_model": "11732.55", + "total_total_tokens_target_model": "234651.0", + "avg_total_tokens_other_models": "61776.40", + "total_total_tokens_other_models": "1235528.0", + "avg_total_tokens_all_models": "73508.95", + "total_total_tokens_all_models": "1470179.0", + "avg_turn_count": "32.70", + "total_turn_count": "654.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + } + } + }, + { + "evaluation_name": "mean on healthbench/S-adaptive/+7ep for scorer healthbench_score", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "30", + "sample_ids": "1f548d5b-cd00-49a0-b327-283a2e00debd,0b8f1d60-2081-4562-98f7-b6a976fe1c6d,6f7a2ee9-e9c6-42d8-b79f-22dea966b8d2,19ec4833-86e9-4166-8b82-d1da09f31fd7,7ebc830a-8dbd-489b-9d61-4d8bacf0db8d,c971f9d1-5f6a-464e-b282-41c8f0d82f58,5a6e4a41-3ea6-4050-a971-93433fe34877,78510411-e03b-4f93-a369-c7e72d2e4908,bbd759b4-8e4d-4db7-bc36-f2e069396c1c,8f2a65de-dea7-48e8-8adb-6194eca26c08,6d5f483c-3e86-456d-bfd5-4e28de699ae5,f05491d8-d160-4b2b-bd5b-59d757585e39,85408180-238e-4c1b-92c1-55d8ce6c3717,ecd7453e-1dc1-4e52-bb95-67e8e5788573,0e073591-b3e7-4dc8-94d0-dc5aa93ca35c,7637b987-8fe2-49d0-8372-1ebf17284d70,5b294937-13e5-424e-8bb4-5d1904a2344a,c5bf7fc3-dfb0-4b47-9bbe-222e87b952e2,83cf8f2d-2857-4f01-a283-9595d8f4ae8e,7a4548e6-38b7-48dd-9088-d7d1219f7852,e0d1b955-1150-457e-96d7-91a9481999cd,24f9a6e7-b214-4011-94c4-6502f249a621,6a97773b-c33c-4f2e-9ac3-fdd53de0f175,f32ee0c1-ea9c-4cfb-8332-bd9579d23024,b6eb69cb-b911-4c81-a7a1-357d21d27109,c0dc053b-157b-4d13-9956-213b67ff6a36,3bb4a735-ace7-4342-b7a5-0bae083e6f82,66b448bf-2e49-488d-9d0b-f9912ab3f0f8,b12d2453-fd3b-462c-b9cc-554ff16bfd3f,0a548d04-0973-4343-a3bf-07f730bed2f6" + } + }, + "evaluation_timestamp": "1776165242.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.3849898941107965, + "uncertainty": { + "num_samples": 210 + }, + "details": { + "total_matched_trajectories": "20", + "stopping_reason_count_repetition_guard": "20", + "avg_total_tokens_target_model": "11732.55", + "total_total_tokens_target_model": "234651.0", + "avg_total_tokens_other_models": "61776.40", + "total_total_tokens_other_models": "1235528.0", + "avg_total_tokens_all_models": "73508.95", + "total_total_tokens_all_models": "1470179.0", + "avg_turn_count": "32.70", + "total_turn_count": "654.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/anthropic/claude-opus-4-5-20251101/deeb16ae-6ac9-49cd-8791-fe832eb188aa_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "48c42c258812b3c885b882e2a3f671b3c2f563ee4e3198846a3308a9239972a3", + "total_rows": 210 + } +} \ No newline at end of file diff --git a/data/healthbench/anthropic/claude-opus-4-5-20251101/deeb16ae-6ac9-49cd-8791-fe832eb188aa_samples.jsonl b/data/healthbench/anthropic/claude-opus-4-5-20251101/deeb16ae-6ac9-49cd-8791-fe832eb188aa_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..0f4dfa2b19b7bf45d75aec5151f7976ee962b464 --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-5-20251101/deeb16ae-6ac9-49cd-8791-fe832eb188aa_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:48c42c258812b3c885b882e2a3f671b3c2f563ee4e3198846a3308a9239972a3 +size 15027935 diff --git a/data/healthbench/anthropic/claude-opus-4-5-20251101/df9136f0-5a24-4de5-b6a4-2b3fad75fdc8.json b/data/healthbench/anthropic/claude-opus-4-5-20251101/df9136f0-5a24-4de5-b6a4-2b3fad75fdc8.json new file mode 100644 index 0000000000000000000000000000000000000000..1e55553fb66c692e892b1389788824aaf874e732 --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-5-20251101/df9136f0-5a24-4de5-b6a4-2b3fad75fdc8.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/anthropic_claude-opus-4-5-20251101/1786166593.021462", + "evaluation_timestamp": "1775733995.0", + "retrieved_timestamp": "1786166593.021462", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-5-20251101", + "id": "anthropic/claude-opus-4-5-20251101", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/anthropic/claude-opus-4-5-20251101/df9136f0-5a24-4de5-b6a4-2b3fad75fdc8_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "6c5257779d61fedcad34b314f4636f6b7eec0da91fb8a22c015606a2590b7f11", + "total_rows": 100 + } +} \ No newline at end of file diff --git a/data/healthbench/anthropic/claude-opus-4-5-20251101/df9136f0-5a24-4de5-b6a4-2b3fad75fdc8_samples.jsonl b/data/healthbench/anthropic/claude-opus-4-5-20251101/df9136f0-5a24-4de5-b6a4-2b3fad75fdc8_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..6a3f07b366350bfd695ccf6caa5bde6b8012bb65 --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-5-20251101/df9136f0-5a24-4de5-b6a4-2b3fad75fdc8_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6c5257779d61fedcad34b314f4636f6b7eec0da91fb8a22c015606a2590b7f11 +size 6769716 diff --git a/data/healthbench/anthropic/claude-opus-4-5-20251101/e360d40a-764c-4e0c-ad0a-79a6e3a2779b.json b/data/healthbench/anthropic/claude-opus-4-5-20251101/e360d40a-764c-4e0c-ad0a-79a6e3a2779b.json new file mode 100644 index 0000000000000000000000000000000000000000..11891d55ec716bb47c902eab28762a361bd347c7 --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-5-20251101/e360d40a-764c-4e0c-ad0a-79a6e3a2779b.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/anthropic_claude-opus-4-5-20251101/1786168543.041853", + "evaluation_timestamp": "1775733995.0", + "retrieved_timestamp": "1786168543.041853", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-5-20251101", + "id": "anthropic/claude-opus-4-5-20251101", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/anthropic/claude-opus-4-5-20251101/e360d40a-764c-4e0c-ad0a-79a6e3a2779b_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "2a8628e3930fcf716414ea516d214dc1a505d57c14ce37fd225df76273c7fc69", + "total_rows": 100 + } +} \ No newline at end of file diff --git a/data/healthbench/anthropic/claude-opus-4-5-20251101/e360d40a-764c-4e0c-ad0a-79a6e3a2779b_samples.jsonl b/data/healthbench/anthropic/claude-opus-4-5-20251101/e360d40a-764c-4e0c-ad0a-79a6e3a2779b_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..92cb754dc2b0c93f67c6c4adf85f9efa86dc60a5 --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-5-20251101/e360d40a-764c-4e0c-ad0a-79a6e3a2779b_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2a8628e3930fcf716414ea516d214dc1a505d57c14ce37fd225df76273c7fc69 +size 6254786 diff --git a/data/healthbench/anthropic/claude-opus-4-5-20251101/e37c135b-6a0b-4f3e-88c3-7622cd613c2e.json b/data/healthbench/anthropic/claude-opus-4-5-20251101/e37c135b-6a0b-4f3e-88c3-7622cd613c2e.json new file mode 100644 index 0000000000000000000000000000000000000000..7329002e7eb9aaebb8f0ee3808c7dc7b66f8bbbd --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-5-20251101/e37c135b-6a0b-4f3e-88c3-7622cd613c2e.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/anthropic_claude-opus-4-5-20251101/1786166606.10988", + "evaluation_timestamp": "1776076027.0", + "retrieved_timestamp": "1786166606.10988", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-5-20251101", + "id": "anthropic/claude-opus-4-5-20251101", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/anthropic/claude-opus-4-5-20251101/e37c135b-6a0b-4f3e-88c3-7622cd613c2e_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "f932e184ef9b180b91b6a83652a53ab0f7340a4e03ae514c2b1ca36fde09b4b1", + "total_rows": 80 + } +} \ No newline at end of file diff --git a/data/healthbench/anthropic/claude-opus-4-5-20251101/e37c135b-6a0b-4f3e-88c3-7622cd613c2e_samples.jsonl b/data/healthbench/anthropic/claude-opus-4-5-20251101/e37c135b-6a0b-4f3e-88c3-7622cd613c2e_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..a619bd8a1299c629cb2626d8c4f5a76b00c56d5a --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-5-20251101/e37c135b-6a0b-4f3e-88c3-7622cd613c2e_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f932e184ef9b180b91b6a83652a53ab0f7340a4e03ae514c2b1ca36fde09b4b1 +size 5064249 diff --git a/data/healthbench/anthropic/claude-opus-4-5-20251101/f8583a62-c43f-4b65-a5f7-b8add3f30f83.json b/data/healthbench/anthropic/claude-opus-4-5-20251101/f8583a62-c43f-4b65-a5f7-b8add3f30f83.json new file mode 100644 index 0000000000000000000000000000000000000000..abbf3347a840d4eb8a4d3ba6e4c14680daf0cd42 --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-5-20251101/f8583a62-c43f-4b65-a5f7-b8add3f30f83.json @@ -0,0 +1,292 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/anthropic_claude-opus-4-5-20251101/1786163927.455855", + "evaluation_timestamp": "1776263015.0", + "retrieved_timestamp": "1786163927.455855", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-5-20251101", + "id": "anthropic/claude-opus-4-5-20251101", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "bootstrap_score on healthbench/S-adaptive/+5ep/6aedf2d4 for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "47", + "sample_ids": "e2cd9f23-e3dd-4b67-8984-a5412dd802e5,a94aad40-387e-4a37-b92f-61badd57388b,596ba714-29f7-447e-aa73-88d46ecd6e88,1233737e-2674-41d3-8519-b817aeb0b32e,c80a2a84-281d-41cc-a10b-32ee48c584f6,bf7718f2-bceb-4702-8d74-a7fc1bd7fb78,b7a82324-73ee-4996-8f71-8f455f789e90,22fe3eec-03b0-4cdf-a06a-87e71d236082,6c568ea6-86bd-4286-94ba-e3ccc8b67b21,905949d2-7a0a-4461-8f4b-257de6be6eed,c653b6eb-f9b9-4626-b702-ff73d71e8253,cf09b209-5434-4317-aa2a-7b07537b475e,77c0be2e-364a-40de-ae21-a277154e3d60,af750930-5fa5-4efe-8f85-b499641377d4,af62e0b8-bbd0-4dcc-b55c-cb4c7bbcfb2c,f3729e05-61ef-4306-9491-6f5659d4eef9,126236de-9e7a-4814-b5d7-98cbe89b33b2,9bb31d40-c155-4213-868d-103fa0cc7c56,97bec88b-6ef7-4f24-84d6-a2bcc592548b,779b5d14-1a6f-4d0a-8763-e0a25961a811,62b4d946-e447-4fb4-96ed-b6d36c8e2c62,d475913b-0b18-47e9-90b2-28449e1d6541,77dff7ec-2e02-4b93-86d6-5d7765b8dd3d,3a0eec54-fd86-47c1-883b-d8e828d2a327,b947fc6f-1d50-4e3f-b679-8484529b787b,e8e17de2-9b5f-4f20-a18e-fdd14a8b7192,b5b6d817-c524-4bef-badc-f93876657ea2,9ab66439-8090-4f6c-ba74-874591dfd1a6,8e447490-23cd-4d64-9e3a-b8ad46c76573,762d9eca-4835-4447-a649-f91740c74bd0,5c1cf475-6fff-42e2-8508-7e1a14ea9f65,06fb659f-a4aa-4a97-a831-ed2cd2aa7e69,608770a0-440d-4349-9a1c-863e9f4d3e24,a58987e9-8b59-461a-a68b-20efa1d37d51,2d878243-5d98-489a-b8f2-8809f813b6cf,452a9534-b4fc-4483-94e0-d34b06b8e299,e690c779-3cb8-4271-8541-a699bb1bd475,d984a2a8-1209-42bb-8f59-001421c96f8f,5caa2d1f-3ab9-4e25-97a2-af29c2dbbcfb,7b1d2983-ab41-48f2-bdcd-454cbf3f81b8,b267896f-cfc0-4bbe-9fda-fdfa287d1f43,ff514fc4-bd34-4a3f-be3f-840b0e1f8091,a131d5e7-2a27-426f-90f0-25eae3c4b490,620b9bf1-f0f3-4d68-aa74-1560fb406a68,9b4c14e9-c404-4646-87b7-d3badae8a68b,a57a4b5e-0f12-4103-8852-72bd2b7dba24,cf8040e2-39e3-4200-ab7f-b9b532310ac3" + } + }, + "evaluation_timestamp": "1776263015.0", + "metric_config": { + "evaluation_description": "bootstrap_score", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.3200653075259933, + "uncertainty": { + "standard_deviation": 0.036673666693607526, + "num_samples": 235 + }, + "details": { + "total_matched_trajectories": "30", + "stopping_reason_count_repetition_guard": "30", + "avg_total_tokens_target_model": "13773.63", + "total_total_tokens_target_model": "413209.0", + "avg_total_tokens_other_models": "36960.77", + "total_total_tokens_other_models": "1108823.0", + "avg_total_tokens_all_models": "50734.40", + "total_total_tokens_all_models": "1522032.0", + "avg_turn_count": "20.47", + "total_turn_count": "614.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + } + } + }, + { + "evaluation_name": "std on healthbench/S-adaptive/+5ep/6aedf2d4 for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "47", + "sample_ids": "e2cd9f23-e3dd-4b67-8984-a5412dd802e5,a94aad40-387e-4a37-b92f-61badd57388b,596ba714-29f7-447e-aa73-88d46ecd6e88,1233737e-2674-41d3-8519-b817aeb0b32e,c80a2a84-281d-41cc-a10b-32ee48c584f6,bf7718f2-bceb-4702-8d74-a7fc1bd7fb78,b7a82324-73ee-4996-8f71-8f455f789e90,22fe3eec-03b0-4cdf-a06a-87e71d236082,6c568ea6-86bd-4286-94ba-e3ccc8b67b21,905949d2-7a0a-4461-8f4b-257de6be6eed,c653b6eb-f9b9-4626-b702-ff73d71e8253,cf09b209-5434-4317-aa2a-7b07537b475e,77c0be2e-364a-40de-ae21-a277154e3d60,af750930-5fa5-4efe-8f85-b499641377d4,af62e0b8-bbd0-4dcc-b55c-cb4c7bbcfb2c,f3729e05-61ef-4306-9491-6f5659d4eef9,126236de-9e7a-4814-b5d7-98cbe89b33b2,9bb31d40-c155-4213-868d-103fa0cc7c56,97bec88b-6ef7-4f24-84d6-a2bcc592548b,779b5d14-1a6f-4d0a-8763-e0a25961a811,62b4d946-e447-4fb4-96ed-b6d36c8e2c62,d475913b-0b18-47e9-90b2-28449e1d6541,77dff7ec-2e02-4b93-86d6-5d7765b8dd3d,3a0eec54-fd86-47c1-883b-d8e828d2a327,b947fc6f-1d50-4e3f-b679-8484529b787b,e8e17de2-9b5f-4f20-a18e-fdd14a8b7192,b5b6d817-c524-4bef-badc-f93876657ea2,9ab66439-8090-4f6c-ba74-874591dfd1a6,8e447490-23cd-4d64-9e3a-b8ad46c76573,762d9eca-4835-4447-a649-f91740c74bd0,5c1cf475-6fff-42e2-8508-7e1a14ea9f65,06fb659f-a4aa-4a97-a831-ed2cd2aa7e69,608770a0-440d-4349-9a1c-863e9f4d3e24,a58987e9-8b59-461a-a68b-20efa1d37d51,2d878243-5d98-489a-b8f2-8809f813b6cf,452a9534-b4fc-4483-94e0-d34b06b8e299,e690c779-3cb8-4271-8541-a699bb1bd475,d984a2a8-1209-42bb-8f59-001421c96f8f,5caa2d1f-3ab9-4e25-97a2-af29c2dbbcfb,7b1d2983-ab41-48f2-bdcd-454cbf3f81b8,b267896f-cfc0-4bbe-9fda-fdfa287d1f43,ff514fc4-bd34-4a3f-be3f-840b0e1f8091,a131d5e7-2a27-426f-90f0-25eae3c4b490,620b9bf1-f0f3-4d68-aa74-1560fb406a68,9b4c14e9-c404-4646-87b7-d3badae8a68b,a57a4b5e-0f12-4103-8852-72bd2b7dba24,cf8040e2-39e3-4200-ab7f-b9b532310ac3" + } + }, + "evaluation_timestamp": "1776263015.0", + "metric_config": { + "evaluation_description": "std", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.036673666693607526, + "uncertainty": { + "standard_deviation": 0.036673666693607526, + "num_samples": 235 + }, + "details": { + "total_matched_trajectories": "30", + "stopping_reason_count_repetition_guard": "30", + "avg_total_tokens_target_model": "13773.63", + "total_total_tokens_target_model": "413209.0", + "avg_total_tokens_other_models": "36960.77", + "total_total_tokens_other_models": "1108823.0", + "avg_total_tokens_all_models": "50734.40", + "total_total_tokens_all_models": "1522032.0", + "avg_turn_count": "20.47", + "total_turn_count": "614.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + } + } + }, + { + "evaluation_name": "criteria_met_rate on healthbench/S-adaptive/+5ep/6aedf2d4 for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "47", + "sample_ids": "e2cd9f23-e3dd-4b67-8984-a5412dd802e5,a94aad40-387e-4a37-b92f-61badd57388b,596ba714-29f7-447e-aa73-88d46ecd6e88,1233737e-2674-41d3-8519-b817aeb0b32e,c80a2a84-281d-41cc-a10b-32ee48c584f6,bf7718f2-bceb-4702-8d74-a7fc1bd7fb78,b7a82324-73ee-4996-8f71-8f455f789e90,22fe3eec-03b0-4cdf-a06a-87e71d236082,6c568ea6-86bd-4286-94ba-e3ccc8b67b21,905949d2-7a0a-4461-8f4b-257de6be6eed,c653b6eb-f9b9-4626-b702-ff73d71e8253,cf09b209-5434-4317-aa2a-7b07537b475e,77c0be2e-364a-40de-ae21-a277154e3d60,af750930-5fa5-4efe-8f85-b499641377d4,af62e0b8-bbd0-4dcc-b55c-cb4c7bbcfb2c,f3729e05-61ef-4306-9491-6f5659d4eef9,126236de-9e7a-4814-b5d7-98cbe89b33b2,9bb31d40-c155-4213-868d-103fa0cc7c56,97bec88b-6ef7-4f24-84d6-a2bcc592548b,779b5d14-1a6f-4d0a-8763-e0a25961a811,62b4d946-e447-4fb4-96ed-b6d36c8e2c62,d475913b-0b18-47e9-90b2-28449e1d6541,77dff7ec-2e02-4b93-86d6-5d7765b8dd3d,3a0eec54-fd86-47c1-883b-d8e828d2a327,b947fc6f-1d50-4e3f-b679-8484529b787b,e8e17de2-9b5f-4f20-a18e-fdd14a8b7192,b5b6d817-c524-4bef-badc-f93876657ea2,9ab66439-8090-4f6c-ba74-874591dfd1a6,8e447490-23cd-4d64-9e3a-b8ad46c76573,762d9eca-4835-4447-a649-f91740c74bd0,5c1cf475-6fff-42e2-8508-7e1a14ea9f65,06fb659f-a4aa-4a97-a831-ed2cd2aa7e69,608770a0-440d-4349-9a1c-863e9f4d3e24,a58987e9-8b59-461a-a68b-20efa1d37d51,2d878243-5d98-489a-b8f2-8809f813b6cf,452a9534-b4fc-4483-94e0-d34b06b8e299,e690c779-3cb8-4271-8541-a699bb1bd475,d984a2a8-1209-42bb-8f59-001421c96f8f,5caa2d1f-3ab9-4e25-97a2-af29c2dbbcfb,7b1d2983-ab41-48f2-bdcd-454cbf3f81b8,b267896f-cfc0-4bbe-9fda-fdfa287d1f43,ff514fc4-bd34-4a3f-be3f-840b0e1f8091,a131d5e7-2a27-426f-90f0-25eae3c4b490,620b9bf1-f0f3-4d68-aa74-1560fb406a68,9b4c14e9-c404-4646-87b7-d3badae8a68b,a57a4b5e-0f12-4103-8852-72bd2b7dba24,cf8040e2-39e3-4200-ab7f-b9b532310ac3" + } + }, + "evaluation_timestamp": "1776263015.0", + "metric_config": { + "evaluation_description": "criteria_met_rate", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.3503759398496241, + "uncertainty": { + "standard_deviation": 0.036673666693607526, + "num_samples": 235 + }, + "details": { + "total_matched_trajectories": "30", + "stopping_reason_count_repetition_guard": "30", + "avg_total_tokens_target_model": "13773.63", + "total_total_tokens_target_model": "413209.0", + "avg_total_tokens_other_models": "36960.77", + "total_total_tokens_other_models": "1108823.0", + "avg_total_tokens_all_models": "50734.40", + "total_total_tokens_all_models": "1522032.0", + "avg_turn_count": "20.47", + "total_turn_count": "614.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + } + } + }, + { + "evaluation_name": "mean on healthbench/S-adaptive/+5ep/6aedf2d4 for scorer healthbench_score", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "47", + "sample_ids": "e2cd9f23-e3dd-4b67-8984-a5412dd802e5,a94aad40-387e-4a37-b92f-61badd57388b,596ba714-29f7-447e-aa73-88d46ecd6e88,1233737e-2674-41d3-8519-b817aeb0b32e,c80a2a84-281d-41cc-a10b-32ee48c584f6,bf7718f2-bceb-4702-8d74-a7fc1bd7fb78,b7a82324-73ee-4996-8f71-8f455f789e90,22fe3eec-03b0-4cdf-a06a-87e71d236082,6c568ea6-86bd-4286-94ba-e3ccc8b67b21,905949d2-7a0a-4461-8f4b-257de6be6eed,c653b6eb-f9b9-4626-b702-ff73d71e8253,cf09b209-5434-4317-aa2a-7b07537b475e,77c0be2e-364a-40de-ae21-a277154e3d60,af750930-5fa5-4efe-8f85-b499641377d4,af62e0b8-bbd0-4dcc-b55c-cb4c7bbcfb2c,f3729e05-61ef-4306-9491-6f5659d4eef9,126236de-9e7a-4814-b5d7-98cbe89b33b2,9bb31d40-c155-4213-868d-103fa0cc7c56,97bec88b-6ef7-4f24-84d6-a2bcc592548b,779b5d14-1a6f-4d0a-8763-e0a25961a811,62b4d946-e447-4fb4-96ed-b6d36c8e2c62,d475913b-0b18-47e9-90b2-28449e1d6541,77dff7ec-2e02-4b93-86d6-5d7765b8dd3d,3a0eec54-fd86-47c1-883b-d8e828d2a327,b947fc6f-1d50-4e3f-b679-8484529b787b,e8e17de2-9b5f-4f20-a18e-fdd14a8b7192,b5b6d817-c524-4bef-badc-f93876657ea2,9ab66439-8090-4f6c-ba74-874591dfd1a6,8e447490-23cd-4d64-9e3a-b8ad46c76573,762d9eca-4835-4447-a649-f91740c74bd0,5c1cf475-6fff-42e2-8508-7e1a14ea9f65,06fb659f-a4aa-4a97-a831-ed2cd2aa7e69,608770a0-440d-4349-9a1c-863e9f4d3e24,a58987e9-8b59-461a-a68b-20efa1d37d51,2d878243-5d98-489a-b8f2-8809f813b6cf,452a9534-b4fc-4483-94e0-d34b06b8e299,e690c779-3cb8-4271-8541-a699bb1bd475,d984a2a8-1209-42bb-8f59-001421c96f8f,5caa2d1f-3ab9-4e25-97a2-af29c2dbbcfb,7b1d2983-ab41-48f2-bdcd-454cbf3f81b8,b267896f-cfc0-4bbe-9fda-fdfa287d1f43,ff514fc4-bd34-4a3f-be3f-840b0e1f8091,a131d5e7-2a27-426f-90f0-25eae3c4b490,620b9bf1-f0f3-4d68-aa74-1560fb406a68,9b4c14e9-c404-4646-87b7-d3badae8a68b,a57a4b5e-0f12-4103-8852-72bd2b7dba24,cf8040e2-39e3-4200-ab7f-b9b532310ac3" + } + }, + "evaluation_timestamp": "1776263015.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.3200653075259933, + "uncertainty": { + "num_samples": 235 + }, + "details": { + "total_matched_trajectories": "30", + "stopping_reason_count_repetition_guard": "30", + "avg_total_tokens_target_model": "13773.63", + "total_total_tokens_target_model": "413209.0", + "avg_total_tokens_other_models": "36960.77", + "total_total_tokens_other_models": "1108823.0", + "avg_total_tokens_all_models": "50734.40", + "total_total_tokens_all_models": "1522032.0", + "avg_turn_count": "20.47", + "total_turn_count": "614.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/anthropic/claude-opus-4-5-20251101/f8583a62-c43f-4b65-a5f7-b8add3f30f83_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "280140cbb16a02626e623ff9577edde1af3a9592b50a072af13ce925afbc71cc", + "total_rows": 235 + } +} \ No newline at end of file diff --git a/data/healthbench/anthropic/claude-opus-4-5-20251101/f8583a62-c43f-4b65-a5f7-b8add3f30f83_samples.jsonl b/data/healthbench/anthropic/claude-opus-4-5-20251101/f8583a62-c43f-4b65-a5f7-b8add3f30f83_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..e03842086f2025febe671a71f12a2e43eb8cf364 --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-5-20251101/f8583a62-c43f-4b65-a5f7-b8add3f30f83_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:280140cbb16a02626e623ff9577edde1af3a9592b50a072af13ce925afbc71cc +size 13976295 diff --git a/data/healthbench/anthropic/claude-opus-4-6/0828a1a9-07d6-4a75-bb8d-d65bb6b1ee39.json b/data/healthbench/anthropic/claude-opus-4-6/0828a1a9-07d6-4a75-bb8d-d65bb6b1ee39.json new file mode 100644 index 0000000000000000000000000000000000000000..01208e344730680b8586a61687808659b3d84b04 --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-6/0828a1a9-07d6-4a75-bb8d-d65bb6b1ee39.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/anthropic_claude-opus-4-6/1786168615.228464", + "evaluation_timestamp": "1776682987.0", + "retrieved_timestamp": "1786168615.228464", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-6", + "id": "anthropic/claude-opus-4-6", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/anthropic/claude-opus-4-6/0828a1a9-07d6-4a75-bb8d-d65bb6b1ee39_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "e353fc466a760321534a4313e6ac334ab59eeda42c886db98890a429038138e6", + "total_rows": 40 + } +} \ No newline at end of file diff --git a/data/healthbench/anthropic/claude-opus-4-6/0828a1a9-07d6-4a75-bb8d-d65bb6b1ee39_samples.jsonl b/data/healthbench/anthropic/claude-opus-4-6/0828a1a9-07d6-4a75-bb8d-d65bb6b1ee39_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..73cf282139b91031894ab6914e8941582b79eef0 --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-6/0828a1a9-07d6-4a75-bb8d-d65bb6b1ee39_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e353fc466a760321534a4313e6ac334ab59eeda42c886db98890a429038138e6 +size 2730521 diff --git a/data/healthbench/anthropic/claude-opus-4-6/0b702b01-c47c-45be-99ed-f5d1b80894d5.json b/data/healthbench/anthropic/claude-opus-4-6/0b702b01-c47c-45be-99ed-f5d1b80894d5.json new file mode 100644 index 0000000000000000000000000000000000000000..52173f2bfc13859527ea86feb9d33f144dbf2a9c --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-6/0b702b01-c47c-45be-99ed-f5d1b80894d5.json @@ -0,0 +1,316 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/anthropic_claude-opus-4-6/1786163982.845012", + "evaluation_timestamp": "1776700114.0", + "retrieved_timestamp": "1786163982.845012", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-6", + "id": "anthropic/claude-opus-4-6", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "bootstrap_score on healthbench/S-adaptive/+4ep/8fc7b8eb for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "2", + "sample_ids": "1dc0b953-ef66-4a78-9989-7cf1053b5b41,3cf35bfe-d881-4881-a268-b5fa0607bf89" + } + }, + "evaluation_timestamp": "1776700114.0", + "metric_config": { + "evaluation_description": "bootstrap_score", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.2393162393162393, + "uncertainty": { + "standard_deviation": 0.007597318301841897, + "num_samples": 8 + }, + "details": { + "total_matched_trajectories": "8", + "stopping_reason_count_repetition_guard": "8", + "avg_total_tokens_target_model": "13494.75", + "total_total_tokens_target_model": "107958.0", + "avg_total_tokens_other_models": "94063.50", + "total_total_tokens_other_models": "752508.0", + "avg_total_tokens_all_models": "107558.25", + "total_total_tokens_all_models": "860466.0", + "avg_turn_count": "46.12", + "total_turn_count": "369.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "3600", + "attempt_timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "3600", + "max_connections": "10" + } + } + }, + { + "evaluation_name": "std on healthbench/S-adaptive/+4ep/8fc7b8eb for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "2", + "sample_ids": "1dc0b953-ef66-4a78-9989-7cf1053b5b41,3cf35bfe-d881-4881-a268-b5fa0607bf89" + } + }, + "evaluation_timestamp": "1776700114.0", + "metric_config": { + "evaluation_description": "std", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.007597318301841897, + "uncertainty": { + "standard_deviation": 0.007597318301841897, + "num_samples": 8 + }, + "details": { + "total_matched_trajectories": "8", + "stopping_reason_count_repetition_guard": "8", + "avg_total_tokens_target_model": "13494.75", + "total_total_tokens_target_model": "107958.0", + "avg_total_tokens_other_models": "94063.50", + "total_total_tokens_other_models": "752508.0", + "avg_total_tokens_all_models": "107558.25", + "total_total_tokens_all_models": "860466.0", + "avg_turn_count": "46.12", + "total_turn_count": "369.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "3600", + "attempt_timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "3600", + "max_connections": "10" + } + } + }, + { + "evaluation_name": "criteria_met_rate on healthbench/S-adaptive/+4ep/8fc7b8eb for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "2", + "sample_ids": "1dc0b953-ef66-4a78-9989-7cf1053b5b41,3cf35bfe-d881-4881-a268-b5fa0607bf89" + } + }, + "evaluation_timestamp": "1776700114.0", + "metric_config": { + "evaluation_description": "criteria_met_rate", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.36046511627906974, + "uncertainty": { + "standard_deviation": 0.007597318301841897, + "num_samples": 8 + }, + "details": { + "total_matched_trajectories": "8", + "stopping_reason_count_repetition_guard": "8", + "avg_total_tokens_target_model": "13494.75", + "total_total_tokens_target_model": "107958.0", + "avg_total_tokens_other_models": "94063.50", + "total_total_tokens_other_models": "752508.0", + "avg_total_tokens_all_models": "107558.25", + "total_total_tokens_all_models": "860466.0", + "avg_turn_count": "46.12", + "total_turn_count": "369.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "3600", + "attempt_timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "3600", + "max_connections": "10" + } + } + }, + { + "evaluation_name": "mean on healthbench/S-adaptive/+4ep/8fc7b8eb for scorer healthbench_score", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "2", + "sample_ids": "1dc0b953-ef66-4a78-9989-7cf1053b5b41,3cf35bfe-d881-4881-a268-b5fa0607bf89" + } + }, + "evaluation_timestamp": "1776700114.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.2393162393162393, + "uncertainty": { + "num_samples": 8 + }, + "details": { + "total_matched_trajectories": "8", + "stopping_reason_count_repetition_guard": "8", + "avg_total_tokens_target_model": "13494.75", + "total_total_tokens_target_model": "107958.0", + "avg_total_tokens_other_models": "94063.50", + "total_total_tokens_other_models": "752508.0", + "avg_total_tokens_all_models": "107558.25", + "total_total_tokens_all_models": "860466.0", + "avg_turn_count": "46.12", + "total_turn_count": "369.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "3600", + "attempt_timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "3600", + "max_connections": "10" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/anthropic/claude-opus-4-6/0b702b01-c47c-45be-99ed-f5d1b80894d5_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "ea10909e17da1b3c39c82bf9c7b10582c5294958dc10712062288be01dbb2de1", + "total_rows": 8 + } +} \ No newline at end of file diff --git a/data/healthbench/anthropic/claude-opus-4-6/0b702b01-c47c-45be-99ed-f5d1b80894d5_samples.jsonl b/data/healthbench/anthropic/claude-opus-4-6/0b702b01-c47c-45be-99ed-f5d1b80894d5_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..4c66b130d0bcca6dd1edc2d12accb8bc872a1a23 --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-6/0b702b01-c47c-45be-99ed-f5d1b80894d5_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ea10909e17da1b3c39c82bf9c7b10582c5294958dc10712062288be01dbb2de1 +size 842734 diff --git a/data/healthbench/anthropic/claude-opus-4-6/274620bc-ee25-4302-8b29-59d061225640.json b/data/healthbench/anthropic/claude-opus-4-6/274620bc-ee25-4302-8b29-59d061225640.json new file mode 100644 index 0000000000000000000000000000000000000000..79052bc26e5e93672ea6a4012cf82b54c1d3a631 --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-6/274620bc-ee25-4302-8b29-59d061225640.json @@ -0,0 +1,292 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/anthropic_claude-opus-4-6/1786163980.27949", + "evaluation_timestamp": "1776165242.0", + "retrieved_timestamp": "1786163980.27949", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-6", + "id": "anthropic/claude-opus-4-6", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "bootstrap_score on healthbench/S-adaptive/+3ep for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "60", + "sample_ids": "6c568ea6-86bd-4286-94ba-e3ccc8b67b21,905949d2-7a0a-4461-8f4b-257de6be6eed,c653b6eb-f9b9-4626-b702-ff73d71e8253,cf09b209-5434-4317-aa2a-7b07537b475e,77c0be2e-364a-40de-ae21-a277154e3d60,af750930-5fa5-4efe-8f85-b499641377d4,af62e0b8-bbd0-4dcc-b55c-cb4c7bbcfb2c,f3729e05-61ef-4306-9491-6f5659d4eef9,126236de-9e7a-4814-b5d7-98cbe89b33b2,9bb31d40-c155-4213-868d-103fa0cc7c56,97bec88b-6ef7-4f24-84d6-a2bcc592548b,779b5d14-1a6f-4d0a-8763-e0a25961a811,62b4d946-e447-4fb4-96ed-b6d36c8e2c62,d475913b-0b18-47e9-90b2-28449e1d6541,77dff7ec-2e02-4b93-86d6-5d7765b8dd3d,3a0eec54-fd86-47c1-883b-d8e828d2a327,b947fc6f-1d50-4e3f-b679-8484529b787b,e8e17de2-9b5f-4f20-a18e-fdd14a8b7192,b5b6d817-c524-4bef-badc-f93876657ea2,9ab66439-8090-4f6c-ba74-874591dfd1a6,8e447490-23cd-4d64-9e3a-b8ad46c76573,762d9eca-4835-4447-a649-f91740c74bd0,5c1cf475-6fff-42e2-8508-7e1a14ea9f65,06fb659f-a4aa-4a97-a831-ed2cd2aa7e69,608770a0-440d-4349-9a1c-863e9f4d3e24,a58987e9-8b59-461a-a68b-20efa1d37d51,2d878243-5d98-489a-b8f2-8809f813b6cf,452a9534-b4fc-4483-94e0-d34b06b8e299,e690c779-3cb8-4271-8541-a699bb1bd475,d984a2a8-1209-42bb-8f59-001421c96f8f,5caa2d1f-3ab9-4e25-97a2-af29c2dbbcfb,7b1d2983-ab41-48f2-bdcd-454cbf3f81b8,b267896f-cfc0-4bbe-9fda-fdfa287d1f43,ff514fc4-bd34-4a3f-be3f-840b0e1f8091,a131d5e7-2a27-426f-90f0-25eae3c4b490,620b9bf1-f0f3-4d68-aa74-1560fb406a68,9b4c14e9-c404-4646-87b7-d3badae8a68b,9385d26b-bf44-4ff7-9410-e181413540dd,a57a4b5e-0f12-4103-8852-72bd2b7dba24,cf8040e2-39e3-4200-ab7f-b9b532310ac3,4fc7b71f-9591-4ae1-bfca-2929c7225182,da05b57a-d776-45b6-96af-1750e0254eec,ef446d15-d019-4c5a-a052-ae8cc3ea3b81,0fe4ea94-633d-4f73-985b-cbcd9f5a0270,cbb0ca6e-aa5e-4683-b107-d7dfa4cb60ce,ae6ec2ce-5c23-4e65-802f-cef9f1624d71,cd3dbc9f-a39c-4d13-af60-cba7dbc3baad,02423ab9-f3ed-4096-a7df-bfa3ebb40c7c,084b0c27-79c5-42f6-9cbd-690d0e79167a,42465540-8af1-423b-9b35-a31de1e1d49e,74a50705-7db2-44be-84dc-92b2d960d344,3b79284e-1143-46b4-b6a0-78e5448ff773,afbfc79d-6455-40b8-b475-1bcde4ef0cc9,6d705f80-8259-4ca1-8645-163f42c85a3c,129057dd-4bf4-472d-a8cf-a063dbed60ad,2833344d-bff9-4bbe-8fab-6b13d64c6b2a,48669cdc-1a64-4c99-a6e8-9a7181341fa2,be70739f-234a-4a0f-9b14-d15b8800577b,4047884e-4629-457c-9d77-130e937fae60,eb97bae4-430e-45cd-a065-2df3ab5c600e" + } + }, + "evaluation_timestamp": "1776165242.0", + "metric_config": { + "evaluation_description": "bootstrap_score", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.4027922073540191, + "uncertainty": { + "standard_deviation": 0.035813685054733056, + "num_samples": 180 + }, + "details": { + "total_matched_trajectories": "36", + "stopping_reason_count_repetition_guard": "36", + "avg_total_tokens_target_model": "15419.17", + "total_total_tokens_target_model": "555090.0", + "avg_total_tokens_other_models": "57403.67", + "total_total_tokens_other_models": "2066532.0", + "avg_total_tokens_all_models": "72822.83", + "total_total_tokens_all_models": "2621622.0", + "avg_turn_count": "25.08", + "total_turn_count": "903.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + } + } + }, + { + "evaluation_name": "std on healthbench/S-adaptive/+3ep for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "60", + "sample_ids": "6c568ea6-86bd-4286-94ba-e3ccc8b67b21,905949d2-7a0a-4461-8f4b-257de6be6eed,c653b6eb-f9b9-4626-b702-ff73d71e8253,cf09b209-5434-4317-aa2a-7b07537b475e,77c0be2e-364a-40de-ae21-a277154e3d60,af750930-5fa5-4efe-8f85-b499641377d4,af62e0b8-bbd0-4dcc-b55c-cb4c7bbcfb2c,f3729e05-61ef-4306-9491-6f5659d4eef9,126236de-9e7a-4814-b5d7-98cbe89b33b2,9bb31d40-c155-4213-868d-103fa0cc7c56,97bec88b-6ef7-4f24-84d6-a2bcc592548b,779b5d14-1a6f-4d0a-8763-e0a25961a811,62b4d946-e447-4fb4-96ed-b6d36c8e2c62,d475913b-0b18-47e9-90b2-28449e1d6541,77dff7ec-2e02-4b93-86d6-5d7765b8dd3d,3a0eec54-fd86-47c1-883b-d8e828d2a327,b947fc6f-1d50-4e3f-b679-8484529b787b,e8e17de2-9b5f-4f20-a18e-fdd14a8b7192,b5b6d817-c524-4bef-badc-f93876657ea2,9ab66439-8090-4f6c-ba74-874591dfd1a6,8e447490-23cd-4d64-9e3a-b8ad46c76573,762d9eca-4835-4447-a649-f91740c74bd0,5c1cf475-6fff-42e2-8508-7e1a14ea9f65,06fb659f-a4aa-4a97-a831-ed2cd2aa7e69,608770a0-440d-4349-9a1c-863e9f4d3e24,a58987e9-8b59-461a-a68b-20efa1d37d51,2d878243-5d98-489a-b8f2-8809f813b6cf,452a9534-b4fc-4483-94e0-d34b06b8e299,e690c779-3cb8-4271-8541-a699bb1bd475,d984a2a8-1209-42bb-8f59-001421c96f8f,5caa2d1f-3ab9-4e25-97a2-af29c2dbbcfb,7b1d2983-ab41-48f2-bdcd-454cbf3f81b8,b267896f-cfc0-4bbe-9fda-fdfa287d1f43,ff514fc4-bd34-4a3f-be3f-840b0e1f8091,a131d5e7-2a27-426f-90f0-25eae3c4b490,620b9bf1-f0f3-4d68-aa74-1560fb406a68,9b4c14e9-c404-4646-87b7-d3badae8a68b,9385d26b-bf44-4ff7-9410-e181413540dd,a57a4b5e-0f12-4103-8852-72bd2b7dba24,cf8040e2-39e3-4200-ab7f-b9b532310ac3,4fc7b71f-9591-4ae1-bfca-2929c7225182,da05b57a-d776-45b6-96af-1750e0254eec,ef446d15-d019-4c5a-a052-ae8cc3ea3b81,0fe4ea94-633d-4f73-985b-cbcd9f5a0270,cbb0ca6e-aa5e-4683-b107-d7dfa4cb60ce,ae6ec2ce-5c23-4e65-802f-cef9f1624d71,cd3dbc9f-a39c-4d13-af60-cba7dbc3baad,02423ab9-f3ed-4096-a7df-bfa3ebb40c7c,084b0c27-79c5-42f6-9cbd-690d0e79167a,42465540-8af1-423b-9b35-a31de1e1d49e,74a50705-7db2-44be-84dc-92b2d960d344,3b79284e-1143-46b4-b6a0-78e5448ff773,afbfc79d-6455-40b8-b475-1bcde4ef0cc9,6d705f80-8259-4ca1-8645-163f42c85a3c,129057dd-4bf4-472d-a8cf-a063dbed60ad,2833344d-bff9-4bbe-8fab-6b13d64c6b2a,48669cdc-1a64-4c99-a6e8-9a7181341fa2,be70739f-234a-4a0f-9b14-d15b8800577b,4047884e-4629-457c-9d77-130e937fae60,eb97bae4-430e-45cd-a065-2df3ab5c600e" + } + }, + "evaluation_timestamp": "1776165242.0", + "metric_config": { + "evaluation_description": "std", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.035813685054733056, + "uncertainty": { + "standard_deviation": 0.035813685054733056, + "num_samples": 180 + }, + "details": { + "total_matched_trajectories": "36", + "stopping_reason_count_repetition_guard": "36", + "avg_total_tokens_target_model": "15419.17", + "total_total_tokens_target_model": "555090.0", + "avg_total_tokens_other_models": "57403.67", + "total_total_tokens_other_models": "2066532.0", + "avg_total_tokens_all_models": "72822.83", + "total_total_tokens_all_models": "2621622.0", + "avg_turn_count": "25.08", + "total_turn_count": "903.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + } + } + }, + { + "evaluation_name": "criteria_met_rate on healthbench/S-adaptive/+3ep for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "60", + "sample_ids": "6c568ea6-86bd-4286-94ba-e3ccc8b67b21,905949d2-7a0a-4461-8f4b-257de6be6eed,c653b6eb-f9b9-4626-b702-ff73d71e8253,cf09b209-5434-4317-aa2a-7b07537b475e,77c0be2e-364a-40de-ae21-a277154e3d60,af750930-5fa5-4efe-8f85-b499641377d4,af62e0b8-bbd0-4dcc-b55c-cb4c7bbcfb2c,f3729e05-61ef-4306-9491-6f5659d4eef9,126236de-9e7a-4814-b5d7-98cbe89b33b2,9bb31d40-c155-4213-868d-103fa0cc7c56,97bec88b-6ef7-4f24-84d6-a2bcc592548b,779b5d14-1a6f-4d0a-8763-e0a25961a811,62b4d946-e447-4fb4-96ed-b6d36c8e2c62,d475913b-0b18-47e9-90b2-28449e1d6541,77dff7ec-2e02-4b93-86d6-5d7765b8dd3d,3a0eec54-fd86-47c1-883b-d8e828d2a327,b947fc6f-1d50-4e3f-b679-8484529b787b,e8e17de2-9b5f-4f20-a18e-fdd14a8b7192,b5b6d817-c524-4bef-badc-f93876657ea2,9ab66439-8090-4f6c-ba74-874591dfd1a6,8e447490-23cd-4d64-9e3a-b8ad46c76573,762d9eca-4835-4447-a649-f91740c74bd0,5c1cf475-6fff-42e2-8508-7e1a14ea9f65,06fb659f-a4aa-4a97-a831-ed2cd2aa7e69,608770a0-440d-4349-9a1c-863e9f4d3e24,a58987e9-8b59-461a-a68b-20efa1d37d51,2d878243-5d98-489a-b8f2-8809f813b6cf,452a9534-b4fc-4483-94e0-d34b06b8e299,e690c779-3cb8-4271-8541-a699bb1bd475,d984a2a8-1209-42bb-8f59-001421c96f8f,5caa2d1f-3ab9-4e25-97a2-af29c2dbbcfb,7b1d2983-ab41-48f2-bdcd-454cbf3f81b8,b267896f-cfc0-4bbe-9fda-fdfa287d1f43,ff514fc4-bd34-4a3f-be3f-840b0e1f8091,a131d5e7-2a27-426f-90f0-25eae3c4b490,620b9bf1-f0f3-4d68-aa74-1560fb406a68,9b4c14e9-c404-4646-87b7-d3badae8a68b,9385d26b-bf44-4ff7-9410-e181413540dd,a57a4b5e-0f12-4103-8852-72bd2b7dba24,cf8040e2-39e3-4200-ab7f-b9b532310ac3,4fc7b71f-9591-4ae1-bfca-2929c7225182,da05b57a-d776-45b6-96af-1750e0254eec,ef446d15-d019-4c5a-a052-ae8cc3ea3b81,0fe4ea94-633d-4f73-985b-cbcd9f5a0270,cbb0ca6e-aa5e-4683-b107-d7dfa4cb60ce,ae6ec2ce-5c23-4e65-802f-cef9f1624d71,cd3dbc9f-a39c-4d13-af60-cba7dbc3baad,02423ab9-f3ed-4096-a7df-bfa3ebb40c7c,084b0c27-79c5-42f6-9cbd-690d0e79167a,42465540-8af1-423b-9b35-a31de1e1d49e,74a50705-7db2-44be-84dc-92b2d960d344,3b79284e-1143-46b4-b6a0-78e5448ff773,afbfc79d-6455-40b8-b475-1bcde4ef0cc9,6d705f80-8259-4ca1-8645-163f42c85a3c,129057dd-4bf4-472d-a8cf-a063dbed60ad,2833344d-bff9-4bbe-8fab-6b13d64c6b2a,48669cdc-1a64-4c99-a6e8-9a7181341fa2,be70739f-234a-4a0f-9b14-d15b8800577b,4047884e-4629-457c-9d77-130e937fae60,eb97bae4-430e-45cd-a065-2df3ab5c600e" + } + }, + "evaluation_timestamp": "1776165242.0", + "metric_config": { + "evaluation_description": "criteria_met_rate", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.441806331471136, + "uncertainty": { + "standard_deviation": 0.035813685054733056, + "num_samples": 180 + }, + "details": { + "total_matched_trajectories": "36", + "stopping_reason_count_repetition_guard": "36", + "avg_total_tokens_target_model": "15419.17", + "total_total_tokens_target_model": "555090.0", + "avg_total_tokens_other_models": "57403.67", + "total_total_tokens_other_models": "2066532.0", + "avg_total_tokens_all_models": "72822.83", + "total_total_tokens_all_models": "2621622.0", + "avg_turn_count": "25.08", + "total_turn_count": "903.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + } + } + }, + { + "evaluation_name": "mean on healthbench/S-adaptive/+3ep for scorer healthbench_score", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "60", + "sample_ids": "6c568ea6-86bd-4286-94ba-e3ccc8b67b21,905949d2-7a0a-4461-8f4b-257de6be6eed,c653b6eb-f9b9-4626-b702-ff73d71e8253,cf09b209-5434-4317-aa2a-7b07537b475e,77c0be2e-364a-40de-ae21-a277154e3d60,af750930-5fa5-4efe-8f85-b499641377d4,af62e0b8-bbd0-4dcc-b55c-cb4c7bbcfb2c,f3729e05-61ef-4306-9491-6f5659d4eef9,126236de-9e7a-4814-b5d7-98cbe89b33b2,9bb31d40-c155-4213-868d-103fa0cc7c56,97bec88b-6ef7-4f24-84d6-a2bcc592548b,779b5d14-1a6f-4d0a-8763-e0a25961a811,62b4d946-e447-4fb4-96ed-b6d36c8e2c62,d475913b-0b18-47e9-90b2-28449e1d6541,77dff7ec-2e02-4b93-86d6-5d7765b8dd3d,3a0eec54-fd86-47c1-883b-d8e828d2a327,b947fc6f-1d50-4e3f-b679-8484529b787b,e8e17de2-9b5f-4f20-a18e-fdd14a8b7192,b5b6d817-c524-4bef-badc-f93876657ea2,9ab66439-8090-4f6c-ba74-874591dfd1a6,8e447490-23cd-4d64-9e3a-b8ad46c76573,762d9eca-4835-4447-a649-f91740c74bd0,5c1cf475-6fff-42e2-8508-7e1a14ea9f65,06fb659f-a4aa-4a97-a831-ed2cd2aa7e69,608770a0-440d-4349-9a1c-863e9f4d3e24,a58987e9-8b59-461a-a68b-20efa1d37d51,2d878243-5d98-489a-b8f2-8809f813b6cf,452a9534-b4fc-4483-94e0-d34b06b8e299,e690c779-3cb8-4271-8541-a699bb1bd475,d984a2a8-1209-42bb-8f59-001421c96f8f,5caa2d1f-3ab9-4e25-97a2-af29c2dbbcfb,7b1d2983-ab41-48f2-bdcd-454cbf3f81b8,b267896f-cfc0-4bbe-9fda-fdfa287d1f43,ff514fc4-bd34-4a3f-be3f-840b0e1f8091,a131d5e7-2a27-426f-90f0-25eae3c4b490,620b9bf1-f0f3-4d68-aa74-1560fb406a68,9b4c14e9-c404-4646-87b7-d3badae8a68b,9385d26b-bf44-4ff7-9410-e181413540dd,a57a4b5e-0f12-4103-8852-72bd2b7dba24,cf8040e2-39e3-4200-ab7f-b9b532310ac3,4fc7b71f-9591-4ae1-bfca-2929c7225182,da05b57a-d776-45b6-96af-1750e0254eec,ef446d15-d019-4c5a-a052-ae8cc3ea3b81,0fe4ea94-633d-4f73-985b-cbcd9f5a0270,cbb0ca6e-aa5e-4683-b107-d7dfa4cb60ce,ae6ec2ce-5c23-4e65-802f-cef9f1624d71,cd3dbc9f-a39c-4d13-af60-cba7dbc3baad,02423ab9-f3ed-4096-a7df-bfa3ebb40c7c,084b0c27-79c5-42f6-9cbd-690d0e79167a,42465540-8af1-423b-9b35-a31de1e1d49e,74a50705-7db2-44be-84dc-92b2d960d344,3b79284e-1143-46b4-b6a0-78e5448ff773,afbfc79d-6455-40b8-b475-1bcde4ef0cc9,6d705f80-8259-4ca1-8645-163f42c85a3c,129057dd-4bf4-472d-a8cf-a063dbed60ad,2833344d-bff9-4bbe-8fab-6b13d64c6b2a,48669cdc-1a64-4c99-a6e8-9a7181341fa2,be70739f-234a-4a0f-9b14-d15b8800577b,4047884e-4629-457c-9d77-130e937fae60,eb97bae4-430e-45cd-a065-2df3ab5c600e" + } + }, + "evaluation_timestamp": "1776165242.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.4027922073540191, + "uncertainty": { + "num_samples": 180 + }, + "details": { + "total_matched_trajectories": "36", + "stopping_reason_count_repetition_guard": "36", + "avg_total_tokens_target_model": "15419.17", + "total_total_tokens_target_model": "555090.0", + "avg_total_tokens_other_models": "57403.67", + "total_total_tokens_other_models": "2066532.0", + "avg_total_tokens_all_models": "72822.83", + "total_total_tokens_all_models": "2621622.0", + "avg_turn_count": "25.08", + "total_turn_count": "903.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/anthropic/claude-opus-4-6/274620bc-ee25-4302-8b29-59d061225640_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "909c70b5cc72a26cc9362404c3ca8b9e9c7dcf48628f5f087b4d86640fc447d5", + "total_rows": 180 + } +} \ No newline at end of file diff --git a/data/healthbench/anthropic/claude-opus-4-6/274620bc-ee25-4302-8b29-59d061225640_samples.jsonl b/data/healthbench/anthropic/claude-opus-4-6/274620bc-ee25-4302-8b29-59d061225640_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..b12092b2056d10003e31422e89738e8876bc08c7 --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-6/274620bc-ee25-4302-8b29-59d061225640_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:909c70b5cc72a26cc9362404c3ca8b9e9c7dcf48628f5f087b4d86640fc447d5 +size 11748819 diff --git a/data/healthbench/anthropic/claude-opus-4-6/2909f1ed-5042-45a8-a9b0-b5b8391e4ca7.json b/data/healthbench/anthropic/claude-opus-4-6/2909f1ed-5042-45a8-a9b0-b5b8391e4ca7.json new file mode 100644 index 0000000000000000000000000000000000000000..aebef328165d02726809752e4c590fba9ad3fcd5 --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-6/2909f1ed-5042-45a8-a9b0-b5b8391e4ca7.json @@ -0,0 +1,316 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/anthropic_claude-opus-4-6/1786163960.389354", + "evaluation_timestamp": "1776682987.0", + "retrieved_timestamp": "1786163960.389354", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-6", + "id": "anthropic/claude-opus-4-6", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "bootstrap_score on healthbench/S-adaptive/+3ep/bdd03026 for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "1", + "sample_ids": "aa7a760a-5173-4a39-af83-48d6ae2230f0" + } + }, + "evaluation_timestamp": "1776682987.0", + "metric_config": { + "evaluation_description": "bootstrap_score", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.437984496124031, + "uncertainty": { + "standard_deviation": 5.551115123125783e-17, + "num_samples": 3 + }, + "details": { + "total_matched_trajectories": "3", + "stopping_reason_count_repetition_guard": "3", + "avg_total_tokens_target_model": "7385.33", + "total_total_tokens_target_model": "22156.0", + "avg_total_tokens_other_models": "54714.00", + "total_total_tokens_other_models": "164142.0", + "avg_total_tokens_all_models": "62099.33", + "total_total_tokens_all_models": "186298.0", + "avg_turn_count": "37.00", + "total_turn_count": "111.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "3600", + "attempt_timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "3600", + "max_connections": "10" + } + } + }, + { + "evaluation_name": "std on healthbench/S-adaptive/+3ep/bdd03026 for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "1", + "sample_ids": "aa7a760a-5173-4a39-af83-48d6ae2230f0" + } + }, + "evaluation_timestamp": "1776682987.0", + "metric_config": { + "evaluation_description": "std", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 5.551115123125783e-17, + "uncertainty": { + "standard_deviation": 5.551115123125783e-17, + "num_samples": 3 + }, + "details": { + "total_matched_trajectories": "3", + "stopping_reason_count_repetition_guard": "3", + "avg_total_tokens_target_model": "7385.33", + "total_total_tokens_target_model": "22156.0", + "avg_total_tokens_other_models": "54714.00", + "total_total_tokens_other_models": "164142.0", + "avg_total_tokens_all_models": "62099.33", + "total_total_tokens_all_models": "186298.0", + "avg_turn_count": "37.00", + "total_turn_count": "111.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "3600", + "attempt_timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "3600", + "max_connections": "10" + } + } + }, + { + "evaluation_name": "criteria_met_rate on healthbench/S-adaptive/+3ep/bdd03026 for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "1", + "sample_ids": "aa7a760a-5173-4a39-af83-48d6ae2230f0" + } + }, + "evaluation_timestamp": "1776682987.0", + "metric_config": { + "evaluation_description": "criteria_met_rate", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.3333333333333333, + "uncertainty": { + "standard_deviation": 5.551115123125783e-17, + "num_samples": 3 + }, + "details": { + "total_matched_trajectories": "3", + "stopping_reason_count_repetition_guard": "3", + "avg_total_tokens_target_model": "7385.33", + "total_total_tokens_target_model": "22156.0", + "avg_total_tokens_other_models": "54714.00", + "total_total_tokens_other_models": "164142.0", + "avg_total_tokens_all_models": "62099.33", + "total_total_tokens_all_models": "186298.0", + "avg_turn_count": "37.00", + "total_turn_count": "111.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "3600", + "attempt_timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "3600", + "max_connections": "10" + } + } + }, + { + "evaluation_name": "mean on healthbench/S-adaptive/+3ep/bdd03026 for scorer healthbench_score", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "1", + "sample_ids": "aa7a760a-5173-4a39-af83-48d6ae2230f0" + } + }, + "evaluation_timestamp": "1776682987.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.437984496124031, + "uncertainty": { + "num_samples": 3 + }, + "details": { + "total_matched_trajectories": "3", + "stopping_reason_count_repetition_guard": "3", + "avg_total_tokens_target_model": "7385.33", + "total_total_tokens_target_model": "22156.0", + "avg_total_tokens_other_models": "54714.00", + "total_total_tokens_other_models": "164142.0", + "avg_total_tokens_all_models": "62099.33", + "total_total_tokens_all_models": "186298.0", + "avg_turn_count": "37.00", + "total_turn_count": "111.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "3600", + "attempt_timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "3600", + "max_connections": "10" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/anthropic/claude-opus-4-6/2909f1ed-5042-45a8-a9b0-b5b8391e4ca7_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "079a8309720aef35d61ef864696edc8d301e8532ac3ccb6b144028588a0dffa3", + "total_rows": 3 + } +} \ No newline at end of file diff --git a/data/healthbench/anthropic/claude-opus-4-6/2909f1ed-5042-45a8-a9b0-b5b8391e4ca7_samples.jsonl b/data/healthbench/anthropic/claude-opus-4-6/2909f1ed-5042-45a8-a9b0-b5b8391e4ca7_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..ee458e7b6212df55c6dfc22854191947370905df --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-6/2909f1ed-5042-45a8-a9b0-b5b8391e4ca7_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:079a8309720aef35d61ef864696edc8d301e8532ac3ccb6b144028588a0dffa3 +size 233923 diff --git a/data/healthbench/anthropic/claude-opus-4-6/2d8bbf1f-98ca-48ac-875d-874b9b68680c.json b/data/healthbench/anthropic/claude-opus-4-6/2d8bbf1f-98ca-48ac-875d-874b9b68680c.json new file mode 100644 index 0000000000000000000000000000000000000000..17c39617df001f82aa8fdbf77774000d10111bc9 --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-6/2d8bbf1f-98ca-48ac-875d-874b9b68680c.json @@ -0,0 +1,316 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/anthropic_claude-opus-4-6/1786163950.314015", + "evaluation_timestamp": "1776682987.0", + "retrieved_timestamp": "1786163950.314015", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-6", + "id": "anthropic/claude-opus-4-6", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "bootstrap_score on healthbench/S-adaptive/+2ep/bfd1d095 for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "17", + "sample_ids": "c3eeecfc-de96-480f-929d-ac4a5e5a541c,269074dc-9495-484b-b2f9-0f0943c0f816,b97b1bae-b09f-4073-8e16-540bba766c60,8e88f21a-24c4-4f9c-b848-faf5e10e31b9,a49eb9af-926e-4f7c-94d0-133729059116,1049130c-e9c9-461d-b080-90027bc011c0,13bb02e4-844e-4972-9200-2befa5f911d8,2694a753-2f57-41eb-a7f9-8c4c7acdd2e8,b160df9f-be98-4a4d-9cc8-11a612ffaf87,5e0a84d7-b972-484c-a649-469b15f0a90f,c7606b1b-ca4d-4a71-8277-bcddff09a653,20cdebb7-9189-4f09-ad21-bfc8e8b75680,f24935ac-0530-4063-b2f1-ac45cd265037,1b8f338a-2e98-4a42-ba80-cba266f4b9fa,6f827e79-05c4-4383-8ed4-16d36260f29f,c835bebe-e998-4dcd-843a-2de03d4aabc6,70afd1ac-85ea-4455-a5a2-15b35378e10a" + } + }, + "evaluation_timestamp": "1776682987.0", + "metric_config": { + "evaluation_description": "bootstrap_score", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.20948923056100074, + "uncertainty": { + "standard_deviation": 0.04807021582461973, + "num_samples": 34 + }, + "details": { + "total_matched_trajectories": "34", + "stopping_reason_count_repetition_guard": "34", + "avg_total_tokens_target_model": "13876.35", + "total_total_tokens_target_model": "471796.0", + "avg_total_tokens_other_models": "66503.15", + "total_total_tokens_other_models": "2261107.0", + "avg_total_tokens_all_models": "80379.50", + "total_total_tokens_all_models": "2732903.0", + "avg_turn_count": "32.88", + "total_turn_count": "1118.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "3600", + "attempt_timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "3600", + "max_connections": "10" + } + } + }, + { + "evaluation_name": "std on healthbench/S-adaptive/+2ep/bfd1d095 for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "17", + "sample_ids": "c3eeecfc-de96-480f-929d-ac4a5e5a541c,269074dc-9495-484b-b2f9-0f0943c0f816,b97b1bae-b09f-4073-8e16-540bba766c60,8e88f21a-24c4-4f9c-b848-faf5e10e31b9,a49eb9af-926e-4f7c-94d0-133729059116,1049130c-e9c9-461d-b080-90027bc011c0,13bb02e4-844e-4972-9200-2befa5f911d8,2694a753-2f57-41eb-a7f9-8c4c7acdd2e8,b160df9f-be98-4a4d-9cc8-11a612ffaf87,5e0a84d7-b972-484c-a649-469b15f0a90f,c7606b1b-ca4d-4a71-8277-bcddff09a653,20cdebb7-9189-4f09-ad21-bfc8e8b75680,f24935ac-0530-4063-b2f1-ac45cd265037,1b8f338a-2e98-4a42-ba80-cba266f4b9fa,6f827e79-05c4-4383-8ed4-16d36260f29f,c835bebe-e998-4dcd-843a-2de03d4aabc6,70afd1ac-85ea-4455-a5a2-15b35378e10a" + } + }, + "evaluation_timestamp": "1776682987.0", + "metric_config": { + "evaluation_description": "std", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.04807021582461973, + "uncertainty": { + "standard_deviation": 0.04807021582461973, + "num_samples": 34 + }, + "details": { + "total_matched_trajectories": "34", + "stopping_reason_count_repetition_guard": "34", + "avg_total_tokens_target_model": "13876.35", + "total_total_tokens_target_model": "471796.0", + "avg_total_tokens_other_models": "66503.15", + "total_total_tokens_other_models": "2261107.0", + "avg_total_tokens_all_models": "80379.50", + "total_total_tokens_all_models": "2732903.0", + "avg_turn_count": "32.88", + "total_turn_count": "1118.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "3600", + "attempt_timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "3600", + "max_connections": "10" + } + } + }, + { + "evaluation_name": "criteria_met_rate on healthbench/S-adaptive/+2ep/bfd1d095 for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "17", + "sample_ids": "c3eeecfc-de96-480f-929d-ac4a5e5a541c,269074dc-9495-484b-b2f9-0f0943c0f816,b97b1bae-b09f-4073-8e16-540bba766c60,8e88f21a-24c4-4f9c-b848-faf5e10e31b9,a49eb9af-926e-4f7c-94d0-133729059116,1049130c-e9c9-461d-b080-90027bc011c0,13bb02e4-844e-4972-9200-2befa5f911d8,2694a753-2f57-41eb-a7f9-8c4c7acdd2e8,b160df9f-be98-4a4d-9cc8-11a612ffaf87,5e0a84d7-b972-484c-a649-469b15f0a90f,c7606b1b-ca4d-4a71-8277-bcddff09a653,20cdebb7-9189-4f09-ad21-bfc8e8b75680,f24935ac-0530-4063-b2f1-ac45cd265037,1b8f338a-2e98-4a42-ba80-cba266f4b9fa,6f827e79-05c4-4383-8ed4-16d36260f29f,c835bebe-e998-4dcd-843a-2de03d4aabc6,70afd1ac-85ea-4455-a5a2-15b35378e10a" + } + }, + "evaluation_timestamp": "1776682987.0", + "metric_config": { + "evaluation_description": "criteria_met_rate", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.33555555555555555, + "uncertainty": { + "standard_deviation": 0.04807021582461973, + "num_samples": 34 + }, + "details": { + "total_matched_trajectories": "34", + "stopping_reason_count_repetition_guard": "34", + "avg_total_tokens_target_model": "13876.35", + "total_total_tokens_target_model": "471796.0", + "avg_total_tokens_other_models": "66503.15", + "total_total_tokens_other_models": "2261107.0", + "avg_total_tokens_all_models": "80379.50", + "total_total_tokens_all_models": "2732903.0", + "avg_turn_count": "32.88", + "total_turn_count": "1118.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "3600", + "attempt_timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "3600", + "max_connections": "10" + } + } + }, + { + "evaluation_name": "mean on healthbench/S-adaptive/+2ep/bfd1d095 for scorer healthbench_score", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "17", + "sample_ids": "c3eeecfc-de96-480f-929d-ac4a5e5a541c,269074dc-9495-484b-b2f9-0f0943c0f816,b97b1bae-b09f-4073-8e16-540bba766c60,8e88f21a-24c4-4f9c-b848-faf5e10e31b9,a49eb9af-926e-4f7c-94d0-133729059116,1049130c-e9c9-461d-b080-90027bc011c0,13bb02e4-844e-4972-9200-2befa5f911d8,2694a753-2f57-41eb-a7f9-8c4c7acdd2e8,b160df9f-be98-4a4d-9cc8-11a612ffaf87,5e0a84d7-b972-484c-a649-469b15f0a90f,c7606b1b-ca4d-4a71-8277-bcddff09a653,20cdebb7-9189-4f09-ad21-bfc8e8b75680,f24935ac-0530-4063-b2f1-ac45cd265037,1b8f338a-2e98-4a42-ba80-cba266f4b9fa,6f827e79-05c4-4383-8ed4-16d36260f29f,c835bebe-e998-4dcd-843a-2de03d4aabc6,70afd1ac-85ea-4455-a5a2-15b35378e10a" + } + }, + "evaluation_timestamp": "1776682987.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.20948923056100074, + "uncertainty": { + "num_samples": 34 + }, + "details": { + "total_matched_trajectories": "34", + "stopping_reason_count_repetition_guard": "34", + "avg_total_tokens_target_model": "13876.35", + "total_total_tokens_target_model": "471796.0", + "avg_total_tokens_other_models": "66503.15", + "total_total_tokens_other_models": "2261107.0", + "avg_total_tokens_all_models": "80379.50", + "total_total_tokens_all_models": "2732903.0", + "avg_turn_count": "32.88", + "total_turn_count": "1118.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "3600", + "attempt_timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "3600", + "max_connections": "10" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/anthropic/claude-opus-4-6/2d8bbf1f-98ca-48ac-875d-874b9b68680c_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "cc58a432bd9ff5b66482c1fb6bd6570d6e718a491d6c06925aeba5778fb07c4e", + "total_rows": 34 + } +} \ No newline at end of file diff --git a/data/healthbench/anthropic/claude-opus-4-6/2d8bbf1f-98ca-48ac-875d-874b9b68680c_samples.jsonl b/data/healthbench/anthropic/claude-opus-4-6/2d8bbf1f-98ca-48ac-875d-874b9b68680c_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..e51179ad1ebf949490c273dcabe251b4cdbd612e --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-6/2d8bbf1f-98ca-48ac-875d-874b9b68680c_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cc58a432bd9ff5b66482c1fb6bd6570d6e718a491d6c06925aeba5778fb07c4e +size 2565581 diff --git a/data/healthbench/anthropic/claude-opus-4-6/368648a5-1a6b-434c-8dd7-4ab6cf2454b6.json b/data/healthbench/anthropic/claude-opus-4-6/368648a5-1a6b-434c-8dd7-4ab6cf2454b6.json new file mode 100644 index 0000000000000000000000000000000000000000..c924b31277a10a1593ec41ba13d6af303eb788fd --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-6/368648a5-1a6b-434c-8dd7-4ab6cf2454b6.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/anthropic_claude-opus-4-6/1786168597.177467", + "evaluation_timestamp": "1776076027.0", + "retrieved_timestamp": "1786168597.177467", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-6", + "id": "anthropic/claude-opus-4-6", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/anthropic/claude-opus-4-6/368648a5-1a6b-434c-8dd7-4ab6cf2454b6_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "2b554c4e21a24758a7e03dfd4ff0a0af9c57ca12e50d98a96a216d2248c8d62f", + "total_rows": 60 + } +} \ No newline at end of file diff --git a/data/healthbench/anthropic/claude-opus-4-6/368648a5-1a6b-434c-8dd7-4ab6cf2454b6_samples.jsonl b/data/healthbench/anthropic/claude-opus-4-6/368648a5-1a6b-434c-8dd7-4ab6cf2454b6_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..367ee047b1fa38dc7e8fc919bd186b3eb3e7e0f9 --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-6/368648a5-1a6b-434c-8dd7-4ab6cf2454b6_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2b554c4e21a24758a7e03dfd4ff0a0af9c57ca12e50d98a96a216d2248c8d62f +size 3845628 diff --git a/data/healthbench/anthropic/claude-opus-4-6/449331cd-ef7b-4f2d-8ba9-55a1fe1ba82e.json b/data/healthbench/anthropic/claude-opus-4-6/449331cd-ef7b-4f2d-8ba9-55a1fe1ba82e.json new file mode 100644 index 0000000000000000000000000000000000000000..a9efaa530e1241d36eb5796dc9329aadd1815bad --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-6/449331cd-ef7b-4f2d-8ba9-55a1fe1ba82e.json @@ -0,0 +1,244 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/anthropic_claude-opus-4-6/1786163973.499856", + "evaluation_timestamp": "1776076028.0", + "retrieved_timestamp": "1786163973.499856", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-6", + "id": "anthropic/claude-opus-4-6", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "bootstrap_score on healthbench/S-adaptive/+3ep for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "6", + "sample_ids": "596ba714-29f7-447e-aa73-88d46ecd6e88,1233737e-2674-41d3-8519-b817aeb0b32e,c80a2a84-281d-41cc-a10b-32ee48c584f6,bf7718f2-bceb-4702-8d74-a7fc1bd7fb78,b7a82324-73ee-4996-8f71-8f455f789e90,22fe3eec-03b0-4cdf-a06a-87e71d236082" + } + }, + "evaluation_timestamp": "1776076028.0", + "metric_config": { + "evaluation_description": "bootstrap_score", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.4280355778418447, + "uncertainty": { + "standard_deviation": 0.10334277745897112, + "num_samples": 18 + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + } + } + }, + { + "evaluation_name": "std on healthbench/S-adaptive/+3ep for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "6", + "sample_ids": "596ba714-29f7-447e-aa73-88d46ecd6e88,1233737e-2674-41d3-8519-b817aeb0b32e,c80a2a84-281d-41cc-a10b-32ee48c584f6,bf7718f2-bceb-4702-8d74-a7fc1bd7fb78,b7a82324-73ee-4996-8f71-8f455f789e90,22fe3eec-03b0-4cdf-a06a-87e71d236082" + } + }, + "evaluation_timestamp": "1776076028.0", + "metric_config": { + "evaluation_description": "std", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.10334277745897112, + "uncertainty": { + "standard_deviation": 0.10334277745897112, + "num_samples": 18 + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + } + } + }, + { + "evaluation_name": "criteria_met_rate on healthbench/S-adaptive/+3ep for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "6", + "sample_ids": "596ba714-29f7-447e-aa73-88d46ecd6e88,1233737e-2674-41d3-8519-b817aeb0b32e,c80a2a84-281d-41cc-a10b-32ee48c584f6,bf7718f2-bceb-4702-8d74-a7fc1bd7fb78,b7a82324-73ee-4996-8f71-8f455f789e90,22fe3eec-03b0-4cdf-a06a-87e71d236082" + } + }, + "evaluation_timestamp": "1776076028.0", + "metric_config": { + "evaluation_description": "criteria_met_rate", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.4222222222222222, + "uncertainty": { + "standard_deviation": 0.10334277745897112, + "num_samples": 18 + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + } + } + }, + { + "evaluation_name": "mean on healthbench/S-adaptive/+3ep for scorer healthbench_score", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "6", + "sample_ids": "596ba714-29f7-447e-aa73-88d46ecd6e88,1233737e-2674-41d3-8519-b817aeb0b32e,c80a2a84-281d-41cc-a10b-32ee48c584f6,bf7718f2-bceb-4702-8d74-a7fc1bd7fb78,b7a82324-73ee-4996-8f71-8f455f789e90,22fe3eec-03b0-4cdf-a06a-87e71d236082" + } + }, + "evaluation_timestamp": "1776076028.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.4280355778418447, + "uncertainty": { + "num_samples": 18 + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/anthropic/claude-opus-4-6/449331cd-ef7b-4f2d-8ba9-55a1fe1ba82e_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "4747c9da704c5c3cba30aebddf4f772f68d4cf40e34ea4d2e88ae18454798879", + "total_rows": 18 + } +} \ No newline at end of file diff --git a/data/healthbench/anthropic/claude-opus-4-6/449331cd-ef7b-4f2d-8ba9-55a1fe1ba82e_samples.jsonl b/data/healthbench/anthropic/claude-opus-4-6/449331cd-ef7b-4f2d-8ba9-55a1fe1ba82e_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..cb62bbb611a2c5310a91054104dffecea3cf5697 --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-6/449331cd-ef7b-4f2d-8ba9-55a1fe1ba82e_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4747c9da704c5c3cba30aebddf4f772f68d4cf40e34ea4d2e88ae18454798879 +size 1001756 diff --git a/data/healthbench/anthropic/claude-opus-4-6/456333c1-2a84-4e62-af38-85548154a24f.json b/data/healthbench/anthropic/claude-opus-4-6/456333c1-2a84-4e62-af38-85548154a24f.json new file mode 100644 index 0000000000000000000000000000000000000000..b24043dcffa5d89ceee0aeb9c8e6c3239e43ca9d --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-6/456333c1-2a84-4e62-af38-85548154a24f.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/anthropic_claude-opus-4-6/1786168610.369274", + "evaluation_timestamp": "1776259832.0", + "retrieved_timestamp": "1786168610.369274", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-6", + "id": "anthropic/claude-opus-4-6", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/anthropic/claude-opus-4-6/456333c1-2a84-4e62-af38-85548154a24f_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "ee2831c3648978d3fe6837dc053a0f1b59af5bfa0d768b982daabfe80bffafee", + "total_rows": 190 + } +} \ No newline at end of file diff --git a/data/healthbench/anthropic/claude-opus-4-6/456333c1-2a84-4e62-af38-85548154a24f_samples.jsonl b/data/healthbench/anthropic/claude-opus-4-6/456333c1-2a84-4e62-af38-85548154a24f_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..9cd8b225f31b3e1a8a65f20a2e50fa3a5b1834d9 --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-6/456333c1-2a84-4e62-af38-85548154a24f_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ee2831c3648978d3fe6837dc053a0f1b59af5bfa0d768b982daabfe80bffafee +size 12168320 diff --git a/data/healthbench/anthropic/claude-opus-4-6/52e6fcce-341d-4302-b975-56f03605799b.json b/data/healthbench/anthropic/claude-opus-4-6/52e6fcce-341d-4302-b975-56f03605799b.json new file mode 100644 index 0000000000000000000000000000000000000000..d0b78ec1b39ab5dafd27750a8a89a88a4607e7d5 --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-6/52e6fcce-341d-4302-b975-56f03605799b.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/anthropic_claude-opus-4-6/1786168591.471425", + "evaluation_timestamp": "1776165242.0", + "retrieved_timestamp": "1786168591.471425", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-6", + "id": "anthropic/claude-opus-4-6", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/anthropic/claude-opus-4-6/52e6fcce-341d-4302-b975-56f03605799b_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "b2126d88e36ca990e0c02d170eef39fc03e539cfd8e497c75d5885839c18fc10", + "total_rows": 100 + } +} \ No newline at end of file diff --git a/data/healthbench/anthropic/claude-opus-4-6/52e6fcce-341d-4302-b975-56f03605799b_samples.jsonl b/data/healthbench/anthropic/claude-opus-4-6/52e6fcce-341d-4302-b975-56f03605799b_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..8760b642b0490f2be0803f20b501f9d3ff9d1ef9 --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-6/52e6fcce-341d-4302-b975-56f03605799b_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b2126d88e36ca990e0c02d170eef39fc03e539cfd8e497c75d5885839c18fc10 +size 6653836 diff --git a/data/healthbench/anthropic/claude-opus-4-6/660bc367-cc17-4410-b38f-5177122cf7fb.json b/data/healthbench/anthropic/claude-opus-4-6/660bc367-cc17-4410-b38f-5177122cf7fb.json new file mode 100644 index 0000000000000000000000000000000000000000..61f62651bc72ce6d5703797636c673c61562d76b --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-6/660bc367-cc17-4410-b38f-5177122cf7fb.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/anthropic_claude-opus-4-6/1786168637.919266", + "evaluation_timestamp": "1775733995.0", + "retrieved_timestamp": "1786168637.919266", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-6", + "id": "anthropic/claude-opus-4-6", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/anthropic/claude-opus-4-6/660bc367-cc17-4410-b38f-5177122cf7fb_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "d1c436d45beaf37fe818732721510628560e656de9da4e3e73c9ed329123f25d", + "total_rows": 270 + } +} \ No newline at end of file diff --git a/data/healthbench/anthropic/claude-opus-4-6/660bc367-cc17-4410-b38f-5177122cf7fb_samples.jsonl b/data/healthbench/anthropic/claude-opus-4-6/660bc367-cc17-4410-b38f-5177122cf7fb_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..de0d00577282ac255d61dae566013128e80d3ee0 --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-6/660bc367-cc17-4410-b38f-5177122cf7fb_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d1c436d45beaf37fe818732721510628560e656de9da4e3e73c9ed329123f25d +size 17174352 diff --git a/data/healthbench/anthropic/claude-opus-4-6/6b523743-f8de-460d-818d-d3dad33d619b.json b/data/healthbench/anthropic/claude-opus-4-6/6b523743-f8de-460d-818d-d3dad33d619b.json new file mode 100644 index 0000000000000000000000000000000000000000..e12b0c9aadf1d72acd4f4cfbb1cccafca186f6ff --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-6/6b523743-f8de-460d-818d-d3dad33d619b.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/anthropic_claude-opus-4-6/1786166710.331372", + "evaluation_timestamp": "1776165241.0", + "retrieved_timestamp": "1786166710.331372", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-6", + "id": "anthropic/claude-opus-4-6", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/anthropic/claude-opus-4-6/6b523743-f8de-460d-818d-d3dad33d619b_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "13e742fdf623308c65a904eee2bcc551c455ba2a0843dc375fa8c0b40b107e7e", + "total_rows": 350 + } +} \ No newline at end of file diff --git a/data/healthbench/anthropic/claude-opus-4-6/6b523743-f8de-460d-818d-d3dad33d619b_samples.jsonl b/data/healthbench/anthropic/claude-opus-4-6/6b523743-f8de-460d-818d-d3dad33d619b_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..f8067f7908ff3f223814c5e1381e61f591984744 --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-6/6b523743-f8de-460d-818d-d3dad33d619b_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:13e742fdf623308c65a904eee2bcc551c455ba2a0843dc375fa8c0b40b107e7e +size 26212178 diff --git a/data/healthbench/anthropic/claude-opus-4-6/6e6ecc70-e588-4e67-be9a-bef81ae6a136.json b/data/healthbench/anthropic/claude-opus-4-6/6e6ecc70-e588-4e67-be9a-bef81ae6a136.json new file mode 100644 index 0000000000000000000000000000000000000000..0574bc607bdac02a8674cd6722865e627527af8a --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-6/6e6ecc70-e588-4e67-be9a-bef81ae6a136.json @@ -0,0 +1,292 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/anthropic_claude-opus-4-6/1786163955.689341", + "evaluation_timestamp": "1776259833.0", + "retrieved_timestamp": "1786163955.689341", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-6", + "id": "anthropic/claude-opus-4-6", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "bootstrap_score on healthbench/S-adaptive/+2ep/84c63c00 for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "66", + "sample_ids": "596ba714-29f7-447e-aa73-88d46ecd6e88,1233737e-2674-41d3-8519-b817aeb0b32e,c80a2a84-281d-41cc-a10b-32ee48c584f6,bf7718f2-bceb-4702-8d74-a7fc1bd7fb78,b7a82324-73ee-4996-8f71-8f455f789e90,22fe3eec-03b0-4cdf-a06a-87e71d236082,6c568ea6-86bd-4286-94ba-e3ccc8b67b21,905949d2-7a0a-4461-8f4b-257de6be6eed,c653b6eb-f9b9-4626-b702-ff73d71e8253,cf09b209-5434-4317-aa2a-7b07537b475e,77c0be2e-364a-40de-ae21-a277154e3d60,af750930-5fa5-4efe-8f85-b499641377d4,af62e0b8-bbd0-4dcc-b55c-cb4c7bbcfb2c,f3729e05-61ef-4306-9491-6f5659d4eef9,126236de-9e7a-4814-b5d7-98cbe89b33b2,9bb31d40-c155-4213-868d-103fa0cc7c56,97bec88b-6ef7-4f24-84d6-a2bcc592548b,779b5d14-1a6f-4d0a-8763-e0a25961a811,62b4d946-e447-4fb4-96ed-b6d36c8e2c62,d475913b-0b18-47e9-90b2-28449e1d6541,77dff7ec-2e02-4b93-86d6-5d7765b8dd3d,3a0eec54-fd86-47c1-883b-d8e828d2a327,b947fc6f-1d50-4e3f-b679-8484529b787b,e8e17de2-9b5f-4f20-a18e-fdd14a8b7192,b5b6d817-c524-4bef-badc-f93876657ea2,9ab66439-8090-4f6c-ba74-874591dfd1a6,8e447490-23cd-4d64-9e3a-b8ad46c76573,762d9eca-4835-4447-a649-f91740c74bd0,5c1cf475-6fff-42e2-8508-7e1a14ea9f65,06fb659f-a4aa-4a97-a831-ed2cd2aa7e69,608770a0-440d-4349-9a1c-863e9f4d3e24,a58987e9-8b59-461a-a68b-20efa1d37d51,2d878243-5d98-489a-b8f2-8809f813b6cf,452a9534-b4fc-4483-94e0-d34b06b8e299,e690c779-3cb8-4271-8541-a699bb1bd475,d984a2a8-1209-42bb-8f59-001421c96f8f,5caa2d1f-3ab9-4e25-97a2-af29c2dbbcfb,7b1d2983-ab41-48f2-bdcd-454cbf3f81b8,b267896f-cfc0-4bbe-9fda-fdfa287d1f43,ff514fc4-bd34-4a3f-be3f-840b0e1f8091,a131d5e7-2a27-426f-90f0-25eae3c4b490,620b9bf1-f0f3-4d68-aa74-1560fb406a68,9b4c14e9-c404-4646-87b7-d3badae8a68b,9385d26b-bf44-4ff7-9410-e181413540dd,a57a4b5e-0f12-4103-8852-72bd2b7dba24,cf8040e2-39e3-4200-ab7f-b9b532310ac3,4fc7b71f-9591-4ae1-bfca-2929c7225182,da05b57a-d776-45b6-96af-1750e0254eec,ef446d15-d019-4c5a-a052-ae8cc3ea3b81,0fe4ea94-633d-4f73-985b-cbcd9f5a0270,cbb0ca6e-aa5e-4683-b107-d7dfa4cb60ce,ae6ec2ce-5c23-4e65-802f-cef9f1624d71,cd3dbc9f-a39c-4d13-af60-cba7dbc3baad,02423ab9-f3ed-4096-a7df-bfa3ebb40c7c,084b0c27-79c5-42f6-9cbd-690d0e79167a,42465540-8af1-423b-9b35-a31de1e1d49e,74a50705-7db2-44be-84dc-92b2d960d344,3b79284e-1143-46b4-b6a0-78e5448ff773,afbfc79d-6455-40b8-b475-1bcde4ef0cc9,6d705f80-8259-4ca1-8645-163f42c85a3c,129057dd-4bf4-472d-a8cf-a063dbed60ad,2833344d-bff9-4bbe-8fab-6b13d64c6b2a,48669cdc-1a64-4c99-a6e8-9a7181341fa2,be70739f-234a-4a0f-9b14-d15b8800577b,4047884e-4629-457c-9d77-130e937fae60,eb97bae4-430e-45cd-a065-2df3ab5c600e" + } + }, + "evaluation_timestamp": "1776259833.0", + "metric_config": { + "evaluation_description": "bootstrap_score", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.42316034231123917, + "uncertainty": { + "standard_deviation": 0.03685554153687695, + "num_samples": 132 + }, + "details": { + "total_matched_trajectories": "24", + "stopping_reason_count_repetition_guard": "24", + "avg_total_tokens_target_model": "14457.92", + "total_total_tokens_target_model": "346990.0", + "avg_total_tokens_other_models": "55445.62", + "total_total_tokens_other_models": "1330695.0", + "avg_total_tokens_all_models": "69903.54", + "total_total_tokens_all_models": "1677685.0", + "avg_turn_count": "24.92", + "total_turn_count": "598.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + } + } + }, + { + "evaluation_name": "std on healthbench/S-adaptive/+2ep/84c63c00 for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "66", + "sample_ids": "596ba714-29f7-447e-aa73-88d46ecd6e88,1233737e-2674-41d3-8519-b817aeb0b32e,c80a2a84-281d-41cc-a10b-32ee48c584f6,bf7718f2-bceb-4702-8d74-a7fc1bd7fb78,b7a82324-73ee-4996-8f71-8f455f789e90,22fe3eec-03b0-4cdf-a06a-87e71d236082,6c568ea6-86bd-4286-94ba-e3ccc8b67b21,905949d2-7a0a-4461-8f4b-257de6be6eed,c653b6eb-f9b9-4626-b702-ff73d71e8253,cf09b209-5434-4317-aa2a-7b07537b475e,77c0be2e-364a-40de-ae21-a277154e3d60,af750930-5fa5-4efe-8f85-b499641377d4,af62e0b8-bbd0-4dcc-b55c-cb4c7bbcfb2c,f3729e05-61ef-4306-9491-6f5659d4eef9,126236de-9e7a-4814-b5d7-98cbe89b33b2,9bb31d40-c155-4213-868d-103fa0cc7c56,97bec88b-6ef7-4f24-84d6-a2bcc592548b,779b5d14-1a6f-4d0a-8763-e0a25961a811,62b4d946-e447-4fb4-96ed-b6d36c8e2c62,d475913b-0b18-47e9-90b2-28449e1d6541,77dff7ec-2e02-4b93-86d6-5d7765b8dd3d,3a0eec54-fd86-47c1-883b-d8e828d2a327,b947fc6f-1d50-4e3f-b679-8484529b787b,e8e17de2-9b5f-4f20-a18e-fdd14a8b7192,b5b6d817-c524-4bef-badc-f93876657ea2,9ab66439-8090-4f6c-ba74-874591dfd1a6,8e447490-23cd-4d64-9e3a-b8ad46c76573,762d9eca-4835-4447-a649-f91740c74bd0,5c1cf475-6fff-42e2-8508-7e1a14ea9f65,06fb659f-a4aa-4a97-a831-ed2cd2aa7e69,608770a0-440d-4349-9a1c-863e9f4d3e24,a58987e9-8b59-461a-a68b-20efa1d37d51,2d878243-5d98-489a-b8f2-8809f813b6cf,452a9534-b4fc-4483-94e0-d34b06b8e299,e690c779-3cb8-4271-8541-a699bb1bd475,d984a2a8-1209-42bb-8f59-001421c96f8f,5caa2d1f-3ab9-4e25-97a2-af29c2dbbcfb,7b1d2983-ab41-48f2-bdcd-454cbf3f81b8,b267896f-cfc0-4bbe-9fda-fdfa287d1f43,ff514fc4-bd34-4a3f-be3f-840b0e1f8091,a131d5e7-2a27-426f-90f0-25eae3c4b490,620b9bf1-f0f3-4d68-aa74-1560fb406a68,9b4c14e9-c404-4646-87b7-d3badae8a68b,9385d26b-bf44-4ff7-9410-e181413540dd,a57a4b5e-0f12-4103-8852-72bd2b7dba24,cf8040e2-39e3-4200-ab7f-b9b532310ac3,4fc7b71f-9591-4ae1-bfca-2929c7225182,da05b57a-d776-45b6-96af-1750e0254eec,ef446d15-d019-4c5a-a052-ae8cc3ea3b81,0fe4ea94-633d-4f73-985b-cbcd9f5a0270,cbb0ca6e-aa5e-4683-b107-d7dfa4cb60ce,ae6ec2ce-5c23-4e65-802f-cef9f1624d71,cd3dbc9f-a39c-4d13-af60-cba7dbc3baad,02423ab9-f3ed-4096-a7df-bfa3ebb40c7c,084b0c27-79c5-42f6-9cbd-690d0e79167a,42465540-8af1-423b-9b35-a31de1e1d49e,74a50705-7db2-44be-84dc-92b2d960d344,3b79284e-1143-46b4-b6a0-78e5448ff773,afbfc79d-6455-40b8-b475-1bcde4ef0cc9,6d705f80-8259-4ca1-8645-163f42c85a3c,129057dd-4bf4-472d-a8cf-a063dbed60ad,2833344d-bff9-4bbe-8fab-6b13d64c6b2a,48669cdc-1a64-4c99-a6e8-9a7181341fa2,be70739f-234a-4a0f-9b14-d15b8800577b,4047884e-4629-457c-9d77-130e937fae60,eb97bae4-430e-45cd-a065-2df3ab5c600e" + } + }, + "evaluation_timestamp": "1776259833.0", + "metric_config": { + "evaluation_description": "std", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.03685554153687695, + "uncertainty": { + "standard_deviation": 0.03685554153687695, + "num_samples": 132 + }, + "details": { + "total_matched_trajectories": "24", + "stopping_reason_count_repetition_guard": "24", + "avg_total_tokens_target_model": "14457.92", + "total_total_tokens_target_model": "346990.0", + "avg_total_tokens_other_models": "55445.62", + "total_total_tokens_other_models": "1330695.0", + "avg_total_tokens_all_models": "69903.54", + "total_total_tokens_all_models": "1677685.0", + "avg_turn_count": "24.92", + "total_turn_count": "598.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + } + } + }, + { + "evaluation_name": "criteria_met_rate on healthbench/S-adaptive/+2ep/84c63c00 for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "66", + "sample_ids": "596ba714-29f7-447e-aa73-88d46ecd6e88,1233737e-2674-41d3-8519-b817aeb0b32e,c80a2a84-281d-41cc-a10b-32ee48c584f6,bf7718f2-bceb-4702-8d74-a7fc1bd7fb78,b7a82324-73ee-4996-8f71-8f455f789e90,22fe3eec-03b0-4cdf-a06a-87e71d236082,6c568ea6-86bd-4286-94ba-e3ccc8b67b21,905949d2-7a0a-4461-8f4b-257de6be6eed,c653b6eb-f9b9-4626-b702-ff73d71e8253,cf09b209-5434-4317-aa2a-7b07537b475e,77c0be2e-364a-40de-ae21-a277154e3d60,af750930-5fa5-4efe-8f85-b499641377d4,af62e0b8-bbd0-4dcc-b55c-cb4c7bbcfb2c,f3729e05-61ef-4306-9491-6f5659d4eef9,126236de-9e7a-4814-b5d7-98cbe89b33b2,9bb31d40-c155-4213-868d-103fa0cc7c56,97bec88b-6ef7-4f24-84d6-a2bcc592548b,779b5d14-1a6f-4d0a-8763-e0a25961a811,62b4d946-e447-4fb4-96ed-b6d36c8e2c62,d475913b-0b18-47e9-90b2-28449e1d6541,77dff7ec-2e02-4b93-86d6-5d7765b8dd3d,3a0eec54-fd86-47c1-883b-d8e828d2a327,b947fc6f-1d50-4e3f-b679-8484529b787b,e8e17de2-9b5f-4f20-a18e-fdd14a8b7192,b5b6d817-c524-4bef-badc-f93876657ea2,9ab66439-8090-4f6c-ba74-874591dfd1a6,8e447490-23cd-4d64-9e3a-b8ad46c76573,762d9eca-4835-4447-a649-f91740c74bd0,5c1cf475-6fff-42e2-8508-7e1a14ea9f65,06fb659f-a4aa-4a97-a831-ed2cd2aa7e69,608770a0-440d-4349-9a1c-863e9f4d3e24,a58987e9-8b59-461a-a68b-20efa1d37d51,2d878243-5d98-489a-b8f2-8809f813b6cf,452a9534-b4fc-4483-94e0-d34b06b8e299,e690c779-3cb8-4271-8541-a699bb1bd475,d984a2a8-1209-42bb-8f59-001421c96f8f,5caa2d1f-3ab9-4e25-97a2-af29c2dbbcfb,7b1d2983-ab41-48f2-bdcd-454cbf3f81b8,b267896f-cfc0-4bbe-9fda-fdfa287d1f43,ff514fc4-bd34-4a3f-be3f-840b0e1f8091,a131d5e7-2a27-426f-90f0-25eae3c4b490,620b9bf1-f0f3-4d68-aa74-1560fb406a68,9b4c14e9-c404-4646-87b7-d3badae8a68b,9385d26b-bf44-4ff7-9410-e181413540dd,a57a4b5e-0f12-4103-8852-72bd2b7dba24,cf8040e2-39e3-4200-ab7f-b9b532310ac3,4fc7b71f-9591-4ae1-bfca-2929c7225182,da05b57a-d776-45b6-96af-1750e0254eec,ef446d15-d019-4c5a-a052-ae8cc3ea3b81,0fe4ea94-633d-4f73-985b-cbcd9f5a0270,cbb0ca6e-aa5e-4683-b107-d7dfa4cb60ce,ae6ec2ce-5c23-4e65-802f-cef9f1624d71,cd3dbc9f-a39c-4d13-af60-cba7dbc3baad,02423ab9-f3ed-4096-a7df-bfa3ebb40c7c,084b0c27-79c5-42f6-9cbd-690d0e79167a,42465540-8af1-423b-9b35-a31de1e1d49e,74a50705-7db2-44be-84dc-92b2d960d344,3b79284e-1143-46b4-b6a0-78e5448ff773,afbfc79d-6455-40b8-b475-1bcde4ef0cc9,6d705f80-8259-4ca1-8645-163f42c85a3c,129057dd-4bf4-472d-a8cf-a063dbed60ad,2833344d-bff9-4bbe-8fab-6b13d64c6b2a,48669cdc-1a64-4c99-a6e8-9a7181341fa2,be70739f-234a-4a0f-9b14-d15b8800577b,4047884e-4629-457c-9d77-130e937fae60,eb97bae4-430e-45cd-a065-2df3ab5c600e" + } + }, + "evaluation_timestamp": "1776259833.0", + "metric_config": { + "evaluation_description": "criteria_met_rate", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.4536082474226804, + "uncertainty": { + "standard_deviation": 0.03685554153687695, + "num_samples": 132 + }, + "details": { + "total_matched_trajectories": "24", + "stopping_reason_count_repetition_guard": "24", + "avg_total_tokens_target_model": "14457.92", + "total_total_tokens_target_model": "346990.0", + "avg_total_tokens_other_models": "55445.62", + "total_total_tokens_other_models": "1330695.0", + "avg_total_tokens_all_models": "69903.54", + "total_total_tokens_all_models": "1677685.0", + "avg_turn_count": "24.92", + "total_turn_count": "598.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + } + } + }, + { + "evaluation_name": "mean on healthbench/S-adaptive/+2ep/84c63c00 for scorer healthbench_score", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "66", + "sample_ids": "596ba714-29f7-447e-aa73-88d46ecd6e88,1233737e-2674-41d3-8519-b817aeb0b32e,c80a2a84-281d-41cc-a10b-32ee48c584f6,bf7718f2-bceb-4702-8d74-a7fc1bd7fb78,b7a82324-73ee-4996-8f71-8f455f789e90,22fe3eec-03b0-4cdf-a06a-87e71d236082,6c568ea6-86bd-4286-94ba-e3ccc8b67b21,905949d2-7a0a-4461-8f4b-257de6be6eed,c653b6eb-f9b9-4626-b702-ff73d71e8253,cf09b209-5434-4317-aa2a-7b07537b475e,77c0be2e-364a-40de-ae21-a277154e3d60,af750930-5fa5-4efe-8f85-b499641377d4,af62e0b8-bbd0-4dcc-b55c-cb4c7bbcfb2c,f3729e05-61ef-4306-9491-6f5659d4eef9,126236de-9e7a-4814-b5d7-98cbe89b33b2,9bb31d40-c155-4213-868d-103fa0cc7c56,97bec88b-6ef7-4f24-84d6-a2bcc592548b,779b5d14-1a6f-4d0a-8763-e0a25961a811,62b4d946-e447-4fb4-96ed-b6d36c8e2c62,d475913b-0b18-47e9-90b2-28449e1d6541,77dff7ec-2e02-4b93-86d6-5d7765b8dd3d,3a0eec54-fd86-47c1-883b-d8e828d2a327,b947fc6f-1d50-4e3f-b679-8484529b787b,e8e17de2-9b5f-4f20-a18e-fdd14a8b7192,b5b6d817-c524-4bef-badc-f93876657ea2,9ab66439-8090-4f6c-ba74-874591dfd1a6,8e447490-23cd-4d64-9e3a-b8ad46c76573,762d9eca-4835-4447-a649-f91740c74bd0,5c1cf475-6fff-42e2-8508-7e1a14ea9f65,06fb659f-a4aa-4a97-a831-ed2cd2aa7e69,608770a0-440d-4349-9a1c-863e9f4d3e24,a58987e9-8b59-461a-a68b-20efa1d37d51,2d878243-5d98-489a-b8f2-8809f813b6cf,452a9534-b4fc-4483-94e0-d34b06b8e299,e690c779-3cb8-4271-8541-a699bb1bd475,d984a2a8-1209-42bb-8f59-001421c96f8f,5caa2d1f-3ab9-4e25-97a2-af29c2dbbcfb,7b1d2983-ab41-48f2-bdcd-454cbf3f81b8,b267896f-cfc0-4bbe-9fda-fdfa287d1f43,ff514fc4-bd34-4a3f-be3f-840b0e1f8091,a131d5e7-2a27-426f-90f0-25eae3c4b490,620b9bf1-f0f3-4d68-aa74-1560fb406a68,9b4c14e9-c404-4646-87b7-d3badae8a68b,9385d26b-bf44-4ff7-9410-e181413540dd,a57a4b5e-0f12-4103-8852-72bd2b7dba24,cf8040e2-39e3-4200-ab7f-b9b532310ac3,4fc7b71f-9591-4ae1-bfca-2929c7225182,da05b57a-d776-45b6-96af-1750e0254eec,ef446d15-d019-4c5a-a052-ae8cc3ea3b81,0fe4ea94-633d-4f73-985b-cbcd9f5a0270,cbb0ca6e-aa5e-4683-b107-d7dfa4cb60ce,ae6ec2ce-5c23-4e65-802f-cef9f1624d71,cd3dbc9f-a39c-4d13-af60-cba7dbc3baad,02423ab9-f3ed-4096-a7df-bfa3ebb40c7c,084b0c27-79c5-42f6-9cbd-690d0e79167a,42465540-8af1-423b-9b35-a31de1e1d49e,74a50705-7db2-44be-84dc-92b2d960d344,3b79284e-1143-46b4-b6a0-78e5448ff773,afbfc79d-6455-40b8-b475-1bcde4ef0cc9,6d705f80-8259-4ca1-8645-163f42c85a3c,129057dd-4bf4-472d-a8cf-a063dbed60ad,2833344d-bff9-4bbe-8fab-6b13d64c6b2a,48669cdc-1a64-4c99-a6e8-9a7181341fa2,be70739f-234a-4a0f-9b14-d15b8800577b,4047884e-4629-457c-9d77-130e937fae60,eb97bae4-430e-45cd-a065-2df3ab5c600e" + } + }, + "evaluation_timestamp": "1776259833.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.42316034231123917, + "uncertainty": { + "num_samples": 132 + }, + "details": { + "total_matched_trajectories": "24", + "stopping_reason_count_repetition_guard": "24", + "avg_total_tokens_target_model": "14457.92", + "total_total_tokens_target_model": "346990.0", + "avg_total_tokens_other_models": "55445.62", + "total_total_tokens_other_models": "1330695.0", + "avg_total_tokens_all_models": "69903.54", + "total_total_tokens_all_models": "1677685.0", + "avg_turn_count": "24.92", + "total_turn_count": "598.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/anthropic/claude-opus-4-6/6e6ecc70-e588-4e67-be9a-bef81ae6a136_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "748ad64750dfd5fe08d5c490d08dc586a50c072a51595230d5d15eada76d7590", + "total_rows": 132 + } +} \ No newline at end of file diff --git a/data/healthbench/anthropic/claude-opus-4-6/6e6ecc70-e588-4e67-be9a-bef81ae6a136_samples.jsonl b/data/healthbench/anthropic/claude-opus-4-6/6e6ecc70-e588-4e67-be9a-bef81ae6a136_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..19959a18f7394f0e729315bebfe98619019b7264 --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-6/6e6ecc70-e588-4e67-be9a-bef81ae6a136_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:748ad64750dfd5fe08d5c490d08dc586a50c072a51595230d5d15eada76d7590 +size 8429221 diff --git a/data/healthbench/anthropic/claude-opus-4-6/73edb3bd-7eb0-4d45-be43-3306603909c4.json b/data/healthbench/anthropic/claude-opus-4-6/73edb3bd-7eb0-4d45-be43-3306603909c4.json new file mode 100644 index 0000000000000000000000000000000000000000..3568dbcc260287cbe0dad879faccb4692754b431 --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-6/73edb3bd-7eb0-4d45-be43-3306603909c4.json @@ -0,0 +1,316 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/anthropic_claude-opus-4-6/1786163989.207867", + "evaluation_timestamp": "1776703489.0", + "retrieved_timestamp": "1786163989.207867", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-6", + "id": "anthropic/claude-opus-4-6", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "bootstrap_score on healthbench/S-adaptive/40c8ecf6 for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "32", + "sample_ids": "bc8d3e4a-8ba0-4cab-be1c-b20ae3af91e4,99306c29-bcef-49d5-bcd1-78d4e541e197,d0fcfaf0-7c6d-4ead-915b-6d9d9130e747,59c68d10-55c1-4629-b05a-bedde7f343b5,d86ee090-3a2e-44d0-abed-6f0f40a38c08,25cf07b8-606f-4448-99ef-ea87ca3097ba,1ca222ed-cc79-4e78-9ca3-3547e7b37e3a,63e4e0d4-0cb1-442e-8316-8c12a6e3ec14,0eb6fd6d-eaaa-46db-ab16-02e610f238a9,99b22b39-30d5-40dc-82d9-c1cd3c793a44,8c342edc-619b-4079-a28c-b7129f3670a6,e8991792-c875-4f40-99a0-54c902d42a52,54349c88-e80a-4728-b181-305f97905d59,f93b0850-2c60-479d-831b-b8c8f9493d61,c63a4306-4003-45a6-8be1-d669457473a9,d3cbb579-70a5-46e8-95f6-72cc9afa4ba3,3852d050-c009-4e33-b953-fd734a5b5e73,5dc23430-dfbb-4794-9f77-0d4dd6093033,67e76447-f16a-4a65-aba7-fc7166fd9b63,a783e7a3-cecc-4d15-b18c-3f2ddfc2dc6c,62e9d35f-153d-4744-b97c-9abfb2a90b1e,04b6e0c8-2835-415e-865d-aca9f1cd3c57,7a283823-00ee-489d-8dbc-356e2857526f,e89e00c9-29e9-420c-813d-e06ddc39d585,51433193-b0c3-4725-a3a8-9dd15e969b0c,48c89cd1-0ecf-4738-8fcc-392269585bd3,92ea478e-6bf8-4ffb-b88a-f57ccf5f2346,36c6e780-d4a8-47b2-8a36-90ee4b97607f,5fb6812f-9816-4d5d-b7e0-4ec3e9df5dc1,c29c95d1-5b5e-4f53-9ec5-83e1f72b5aa5,b81c059a-1137-47cf-bab8-ab5eb8ff34d5,886864a6-a0e4-468f-aaf4-26cf8dc5e541" + } + }, + "evaluation_timestamp": "1776703489.0", + "metric_config": { + "evaluation_description": "bootstrap_score", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.18029613245601867, + "uncertainty": { + "standard_deviation": 0.04770576098335035, + "num_samples": 160 + }, + "details": { + "total_matched_trajectories": "160", + "stopping_reason_count_repetition_guard": "160", + "avg_total_tokens_target_model": "22475.67", + "total_total_tokens_target_model": "3596108.0", + "avg_total_tokens_other_models": "55352.01", + "total_total_tokens_other_models": "8856322.0", + "avg_total_tokens_all_models": "77827.69", + "total_total_tokens_all_models": "12452430.0", + "avg_turn_count": "28.11", + "total_turn_count": "4497.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "3600", + "attempt_timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "3600", + "max_connections": "10" + } + } + }, + { + "evaluation_name": "std on healthbench/S-adaptive/40c8ecf6 for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "32", + "sample_ids": "bc8d3e4a-8ba0-4cab-be1c-b20ae3af91e4,99306c29-bcef-49d5-bcd1-78d4e541e197,d0fcfaf0-7c6d-4ead-915b-6d9d9130e747,59c68d10-55c1-4629-b05a-bedde7f343b5,d86ee090-3a2e-44d0-abed-6f0f40a38c08,25cf07b8-606f-4448-99ef-ea87ca3097ba,1ca222ed-cc79-4e78-9ca3-3547e7b37e3a,63e4e0d4-0cb1-442e-8316-8c12a6e3ec14,0eb6fd6d-eaaa-46db-ab16-02e610f238a9,99b22b39-30d5-40dc-82d9-c1cd3c793a44,8c342edc-619b-4079-a28c-b7129f3670a6,e8991792-c875-4f40-99a0-54c902d42a52,54349c88-e80a-4728-b181-305f97905d59,f93b0850-2c60-479d-831b-b8c8f9493d61,c63a4306-4003-45a6-8be1-d669457473a9,d3cbb579-70a5-46e8-95f6-72cc9afa4ba3,3852d050-c009-4e33-b953-fd734a5b5e73,5dc23430-dfbb-4794-9f77-0d4dd6093033,67e76447-f16a-4a65-aba7-fc7166fd9b63,a783e7a3-cecc-4d15-b18c-3f2ddfc2dc6c,62e9d35f-153d-4744-b97c-9abfb2a90b1e,04b6e0c8-2835-415e-865d-aca9f1cd3c57,7a283823-00ee-489d-8dbc-356e2857526f,e89e00c9-29e9-420c-813d-e06ddc39d585,51433193-b0c3-4725-a3a8-9dd15e969b0c,48c89cd1-0ecf-4738-8fcc-392269585bd3,92ea478e-6bf8-4ffb-b88a-f57ccf5f2346,36c6e780-d4a8-47b2-8a36-90ee4b97607f,5fb6812f-9816-4d5d-b7e0-4ec3e9df5dc1,c29c95d1-5b5e-4f53-9ec5-83e1f72b5aa5,b81c059a-1137-47cf-bab8-ab5eb8ff34d5,886864a6-a0e4-468f-aaf4-26cf8dc5e541" + } + }, + "evaluation_timestamp": "1776703489.0", + "metric_config": { + "evaluation_description": "std", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.04770576098335035, + "uncertainty": { + "standard_deviation": 0.04770576098335035, + "num_samples": 160 + }, + "details": { + "total_matched_trajectories": "160", + "stopping_reason_count_repetition_guard": "160", + "avg_total_tokens_target_model": "22475.67", + "total_total_tokens_target_model": "3596108.0", + "avg_total_tokens_other_models": "55352.01", + "total_total_tokens_other_models": "8856322.0", + "avg_total_tokens_all_models": "77827.69", + "total_total_tokens_all_models": "12452430.0", + "avg_turn_count": "28.11", + "total_turn_count": "4497.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "3600", + "attempt_timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "3600", + "max_connections": "10" + } + } + }, + { + "evaluation_name": "criteria_met_rate on healthbench/S-adaptive/40c8ecf6 for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "32", + "sample_ids": "bc8d3e4a-8ba0-4cab-be1c-b20ae3af91e4,99306c29-bcef-49d5-bcd1-78d4e541e197,d0fcfaf0-7c6d-4ead-915b-6d9d9130e747,59c68d10-55c1-4629-b05a-bedde7f343b5,d86ee090-3a2e-44d0-abed-6f0f40a38c08,25cf07b8-606f-4448-99ef-ea87ca3097ba,1ca222ed-cc79-4e78-9ca3-3547e7b37e3a,63e4e0d4-0cb1-442e-8316-8c12a6e3ec14,0eb6fd6d-eaaa-46db-ab16-02e610f238a9,99b22b39-30d5-40dc-82d9-c1cd3c793a44,8c342edc-619b-4079-a28c-b7129f3670a6,e8991792-c875-4f40-99a0-54c902d42a52,54349c88-e80a-4728-b181-305f97905d59,f93b0850-2c60-479d-831b-b8c8f9493d61,c63a4306-4003-45a6-8be1-d669457473a9,d3cbb579-70a5-46e8-95f6-72cc9afa4ba3,3852d050-c009-4e33-b953-fd734a5b5e73,5dc23430-dfbb-4794-9f77-0d4dd6093033,67e76447-f16a-4a65-aba7-fc7166fd9b63,a783e7a3-cecc-4d15-b18c-3f2ddfc2dc6c,62e9d35f-153d-4744-b97c-9abfb2a90b1e,04b6e0c8-2835-415e-865d-aca9f1cd3c57,7a283823-00ee-489d-8dbc-356e2857526f,e89e00c9-29e9-420c-813d-e06ddc39d585,51433193-b0c3-4725-a3a8-9dd15e969b0c,48c89cd1-0ecf-4738-8fcc-392269585bd3,92ea478e-6bf8-4ffb-b88a-f57ccf5f2346,36c6e780-d4a8-47b2-8a36-90ee4b97607f,5fb6812f-9816-4d5d-b7e0-4ec3e9df5dc1,c29c95d1-5b5e-4f53-9ec5-83e1f72b5aa5,b81c059a-1137-47cf-bab8-ab5eb8ff34d5,886864a6-a0e4-468f-aaf4-26cf8dc5e541" + } + }, + "evaluation_timestamp": "1776703489.0", + "metric_config": { + "evaluation_description": "criteria_met_rate", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.3208333333333333, + "uncertainty": { + "standard_deviation": 0.04770576098335035, + "num_samples": 160 + }, + "details": { + "total_matched_trajectories": "160", + "stopping_reason_count_repetition_guard": "160", + "avg_total_tokens_target_model": "22475.67", + "total_total_tokens_target_model": "3596108.0", + "avg_total_tokens_other_models": "55352.01", + "total_total_tokens_other_models": "8856322.0", + "avg_total_tokens_all_models": "77827.69", + "total_total_tokens_all_models": "12452430.0", + "avg_turn_count": "28.11", + "total_turn_count": "4497.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "3600", + "attempt_timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "3600", + "max_connections": "10" + } + } + }, + { + "evaluation_name": "mean on healthbench/S-adaptive/40c8ecf6 for scorer healthbench_score", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "32", + "sample_ids": "bc8d3e4a-8ba0-4cab-be1c-b20ae3af91e4,99306c29-bcef-49d5-bcd1-78d4e541e197,d0fcfaf0-7c6d-4ead-915b-6d9d9130e747,59c68d10-55c1-4629-b05a-bedde7f343b5,d86ee090-3a2e-44d0-abed-6f0f40a38c08,25cf07b8-606f-4448-99ef-ea87ca3097ba,1ca222ed-cc79-4e78-9ca3-3547e7b37e3a,63e4e0d4-0cb1-442e-8316-8c12a6e3ec14,0eb6fd6d-eaaa-46db-ab16-02e610f238a9,99b22b39-30d5-40dc-82d9-c1cd3c793a44,8c342edc-619b-4079-a28c-b7129f3670a6,e8991792-c875-4f40-99a0-54c902d42a52,54349c88-e80a-4728-b181-305f97905d59,f93b0850-2c60-479d-831b-b8c8f9493d61,c63a4306-4003-45a6-8be1-d669457473a9,d3cbb579-70a5-46e8-95f6-72cc9afa4ba3,3852d050-c009-4e33-b953-fd734a5b5e73,5dc23430-dfbb-4794-9f77-0d4dd6093033,67e76447-f16a-4a65-aba7-fc7166fd9b63,a783e7a3-cecc-4d15-b18c-3f2ddfc2dc6c,62e9d35f-153d-4744-b97c-9abfb2a90b1e,04b6e0c8-2835-415e-865d-aca9f1cd3c57,7a283823-00ee-489d-8dbc-356e2857526f,e89e00c9-29e9-420c-813d-e06ddc39d585,51433193-b0c3-4725-a3a8-9dd15e969b0c,48c89cd1-0ecf-4738-8fcc-392269585bd3,92ea478e-6bf8-4ffb-b88a-f57ccf5f2346,36c6e780-d4a8-47b2-8a36-90ee4b97607f,5fb6812f-9816-4d5d-b7e0-4ec3e9df5dc1,c29c95d1-5b5e-4f53-9ec5-83e1f72b5aa5,b81c059a-1137-47cf-bab8-ab5eb8ff34d5,886864a6-a0e4-468f-aaf4-26cf8dc5e541" + } + }, + "evaluation_timestamp": "1776703489.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.18029613245601867, + "uncertainty": { + "num_samples": 160 + }, + "details": { + "total_matched_trajectories": "160", + "stopping_reason_count_repetition_guard": "160", + "avg_total_tokens_target_model": "22475.67", + "total_total_tokens_target_model": "3596108.0", + "avg_total_tokens_other_models": "55352.01", + "total_total_tokens_other_models": "8856322.0", + "avg_total_tokens_all_models": "77827.69", + "total_total_tokens_all_models": "12452430.0", + "avg_turn_count": "28.11", + "total_turn_count": "4497.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "3600", + "attempt_timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "3600", + "max_connections": "10" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/anthropic/claude-opus-4-6/73edb3bd-7eb0-4d45-be43-3306603909c4_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "8ed028275522758cb79f82db4d509b14912cf4d216ebb795ccd4e94d93ecaad9", + "total_rows": 160 + } +} \ No newline at end of file diff --git a/data/healthbench/anthropic/claude-opus-4-6/73edb3bd-7eb0-4d45-be43-3306603909c4_samples.jsonl b/data/healthbench/anthropic/claude-opus-4-6/73edb3bd-7eb0-4d45-be43-3306603909c4_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..268a0614afa847724301cfb096fea1216f005ec4 --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-6/73edb3bd-7eb0-4d45-be43-3306603909c4_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8ed028275522758cb79f82db4d509b14912cf4d216ebb795ccd4e94d93ecaad9 +size 11015801 diff --git a/data/healthbench/anthropic/claude-opus-4-6/79115736-85ae-4a80-86ce-a9b27a51a38f.json b/data/healthbench/anthropic/claude-opus-4-6/79115736-85ae-4a80-86ce-a9b27a51a38f.json new file mode 100644 index 0000000000000000000000000000000000000000..532254893d8ae098c5aaf88563fef60b3efe290b --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-6/79115736-85ae-4a80-86ce-a9b27a51a38f.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/anthropic_claude-opus-4-6/1786166651.296378", + "evaluation_timestamp": "1775733995.0", + "retrieved_timestamp": "1786166651.296378", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-6", + "id": "anthropic/claude-opus-4-6", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/anthropic/claude-opus-4-6/79115736-85ae-4a80-86ce-a9b27a51a38f_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "7e12b8c3aa979819e29216bbb3120433b28a3d1e343815d6334fa0b9014bb070", + "total_rows": 50 + } +} \ No newline at end of file diff --git a/data/healthbench/anthropic/claude-opus-4-6/79115736-85ae-4a80-86ce-a9b27a51a38f_samples.jsonl b/data/healthbench/anthropic/claude-opus-4-6/79115736-85ae-4a80-86ce-a9b27a51a38f_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..791efaf7c543ee6a98434c125ffffa15572b13f7 --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-6/79115736-85ae-4a80-86ce-a9b27a51a38f_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7e12b8c3aa979819e29216bbb3120433b28a3d1e343815d6334fa0b9014bb070 +size 3349688 diff --git a/data/healthbench/anthropic/claude-opus-4-6/98f6ebbd-456b-4b09-8de3-0fff0181eff1.json b/data/healthbench/anthropic/claude-opus-4-6/98f6ebbd-456b-4b09-8de3-0fff0181eff1.json new file mode 100644 index 0000000000000000000000000000000000000000..42148ab06c1874bb0875d7b74fd8ccd22ef943da --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-6/98f6ebbd-456b-4b09-8de3-0fff0181eff1.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/anthropic_claude-opus-4-6/1786168622.748769", + "evaluation_timestamp": "1776259832.0", + "retrieved_timestamp": "1786168622.748769", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-6", + "id": "anthropic/claude-opus-4-6", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/anthropic/claude-opus-4-6/98f6ebbd-456b-4b09-8de3-0fff0181eff1_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "cae1239ddb53c7d33f174d2b072eca7defec29cfb3b87994870bd383b28c2bcf", + "total_rows": 100 + } +} \ No newline at end of file diff --git a/data/healthbench/anthropic/claude-opus-4-6/98f6ebbd-456b-4b09-8de3-0fff0181eff1_samples.jsonl b/data/healthbench/anthropic/claude-opus-4-6/98f6ebbd-456b-4b09-8de3-0fff0181eff1_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..283d78f3502736dd6708bcb4bd5d29c233b7b880 --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-6/98f6ebbd-456b-4b09-8de3-0fff0181eff1_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cae1239ddb53c7d33f174d2b072eca7defec29cfb3b87994870bd383b28c2bcf +size 6422256 diff --git a/data/healthbench/anthropic/claude-opus-4-6/9fe9f794-0385-4ec6-ba4e-3a1831a78e18.json b/data/healthbench/anthropic/claude-opus-4-6/9fe9f794-0385-4ec6-ba4e-3a1831a78e18.json new file mode 100644 index 0000000000000000000000000000000000000000..9a089c90042fcd9d9a70f9f5879dfeaf4354443a --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-6/9fe9f794-0385-4ec6-ba4e-3a1831a78e18.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/anthropic_claude-opus-4-6/1786166669.747331", + "evaluation_timestamp": "1775665083.0", + "retrieved_timestamp": "1786166669.747331", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-6", + "id": "anthropic/claude-opus-4-6", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/anthropic/claude-opus-4-6/9fe9f794-0385-4ec6-ba4e-3a1831a78e18_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "77d237e320e1c9b8f1aac2c2be32ca12a333293e873dcead4c83b4368a26d566", + "total_rows": 202 + } +} \ No newline at end of file diff --git a/data/healthbench/anthropic/claude-opus-4-6/9fe9f794-0385-4ec6-ba4e-3a1831a78e18_samples.jsonl b/data/healthbench/anthropic/claude-opus-4-6/9fe9f794-0385-4ec6-ba4e-3a1831a78e18_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..9b33c56d4eac8fc39a5acfced049feeb57cd57bd --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-6/9fe9f794-0385-4ec6-ba4e-3a1831a78e18_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:77d237e320e1c9b8f1aac2c2be32ca12a333293e873dcead4c83b4368a26d566 +size 15784232 diff --git a/data/healthbench/anthropic/claude-opus-4-6/a672cd75-3afd-4e97-9719-4800a886c8cf.json b/data/healthbench/anthropic/claude-opus-4-6/a672cd75-3afd-4e97-9719-4800a886c8cf.json new file mode 100644 index 0000000000000000000000000000000000000000..0390fc1e8e17af8dfea27de5eb20769dacb89118 --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-6/a672cd75-3afd-4e97-9719-4800a886c8cf.json @@ -0,0 +1,316 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/anthropic_claude-opus-4-6/1786163958.604383", + "evaluation_timestamp": "1776696116.0", + "retrieved_timestamp": "1786163958.604383", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-6", + "id": "anthropic/claude-opus-4-6", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "bootstrap_score on healthbench/S-adaptive/+2ep/4e133fba for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "8", + "sample_ids": "f712eb09-82af-4fe8-800d-bc3af772a9ba,8ee6a990-8966-4b42-baa8-6c7e9f74ce74,267bda16-fd4b-4a9e-b3ac-5602c2532eed,328c55cf-0b57-4992-99b1-f6a21af62ade,7b415a2f-22b0-4735-a38e-567f1891acc5,c49ac752-3693-4be1-ab4d-91fbdd02f827,a20cbbf3-330b-46d9-a35b-92e89a022282,9f7b555a-fc30-40c3-88d3-f2ece06454a2" + } + }, + "evaluation_timestamp": "1776696116.0", + "metric_config": { + "evaluation_description": "bootstrap_score", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.31051265895015895, + "uncertainty": { + "standard_deviation": 0.17686610600764807, + "num_samples": 16 + }, + "details": { + "total_matched_trajectories": "16", + "stopping_reason_count_repetition_guard": "16", + "avg_total_tokens_target_model": "17574.94", + "total_total_tokens_target_model": "281199.0", + "avg_total_tokens_other_models": "41398.25", + "total_total_tokens_other_models": "662372.0", + "avg_total_tokens_all_models": "58973.19", + "total_total_tokens_all_models": "943571.0", + "avg_turn_count": "19.81", + "total_turn_count": "317.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "3600", + "attempt_timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "3600", + "max_connections": "10" + } + } + }, + { + "evaluation_name": "std on healthbench/S-adaptive/+2ep/4e133fba for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "8", + "sample_ids": "f712eb09-82af-4fe8-800d-bc3af772a9ba,8ee6a990-8966-4b42-baa8-6c7e9f74ce74,267bda16-fd4b-4a9e-b3ac-5602c2532eed,328c55cf-0b57-4992-99b1-f6a21af62ade,7b415a2f-22b0-4735-a38e-567f1891acc5,c49ac752-3693-4be1-ab4d-91fbdd02f827,a20cbbf3-330b-46d9-a35b-92e89a022282,9f7b555a-fc30-40c3-88d3-f2ece06454a2" + } + }, + "evaluation_timestamp": "1776696116.0", + "metric_config": { + "evaluation_description": "std", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.17686610600764807, + "uncertainty": { + "standard_deviation": 0.17686610600764807, + "num_samples": 16 + }, + "details": { + "total_matched_trajectories": "16", + "stopping_reason_count_repetition_guard": "16", + "avg_total_tokens_target_model": "17574.94", + "total_total_tokens_target_model": "281199.0", + "avg_total_tokens_other_models": "41398.25", + "total_total_tokens_other_models": "662372.0", + "avg_total_tokens_all_models": "58973.19", + "total_total_tokens_all_models": "943571.0", + "avg_turn_count": "19.81", + "total_turn_count": "317.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "3600", + "attempt_timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "3600", + "max_connections": "10" + } + } + }, + { + "evaluation_name": "criteria_met_rate on healthbench/S-adaptive/+2ep/4e133fba for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "8", + "sample_ids": "f712eb09-82af-4fe8-800d-bc3af772a9ba,8ee6a990-8966-4b42-baa8-6c7e9f74ce74,267bda16-fd4b-4a9e-b3ac-5602c2532eed,328c55cf-0b57-4992-99b1-f6a21af62ade,7b415a2f-22b0-4735-a38e-567f1891acc5,c49ac752-3693-4be1-ab4d-91fbdd02f827,a20cbbf3-330b-46d9-a35b-92e89a022282,9f7b555a-fc30-40c3-88d3-f2ece06454a2" + } + }, + "evaluation_timestamp": "1776696116.0", + "metric_config": { + "evaluation_description": "criteria_met_rate", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.42857142857142855, + "uncertainty": { + "standard_deviation": 0.17686610600764807, + "num_samples": 16 + }, + "details": { + "total_matched_trajectories": "16", + "stopping_reason_count_repetition_guard": "16", + "avg_total_tokens_target_model": "17574.94", + "total_total_tokens_target_model": "281199.0", + "avg_total_tokens_other_models": "41398.25", + "total_total_tokens_other_models": "662372.0", + "avg_total_tokens_all_models": "58973.19", + "total_total_tokens_all_models": "943571.0", + "avg_turn_count": "19.81", + "total_turn_count": "317.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "3600", + "attempt_timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "3600", + "max_connections": "10" + } + } + }, + { + "evaluation_name": "mean on healthbench/S-adaptive/+2ep/4e133fba for scorer healthbench_score", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "8", + "sample_ids": "f712eb09-82af-4fe8-800d-bc3af772a9ba,8ee6a990-8966-4b42-baa8-6c7e9f74ce74,267bda16-fd4b-4a9e-b3ac-5602c2532eed,328c55cf-0b57-4992-99b1-f6a21af62ade,7b415a2f-22b0-4735-a38e-567f1891acc5,c49ac752-3693-4be1-ab4d-91fbdd02f827,a20cbbf3-330b-46d9-a35b-92e89a022282,9f7b555a-fc30-40c3-88d3-f2ece06454a2" + } + }, + "evaluation_timestamp": "1776696116.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.31051265895015895, + "uncertainty": { + "num_samples": 16 + }, + "details": { + "total_matched_trajectories": "16", + "stopping_reason_count_repetition_guard": "16", + "avg_total_tokens_target_model": "17574.94", + "total_total_tokens_target_model": "281199.0", + "avg_total_tokens_other_models": "41398.25", + "total_total_tokens_other_models": "662372.0", + "avg_total_tokens_all_models": "58973.19", + "total_total_tokens_all_models": "943571.0", + "avg_turn_count": "19.81", + "total_turn_count": "317.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "3600", + "attempt_timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "3600", + "max_connections": "10" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/anthropic/claude-opus-4-6/a672cd75-3afd-4e97-9719-4800a886c8cf_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "2f0ac8a50d78fd16a35a4cbb95afddde6edf5d42352515d830d525e58fda9ab3", + "total_rows": 16 + } +} \ No newline at end of file diff --git a/data/healthbench/anthropic/claude-opus-4-6/a672cd75-3afd-4e97-9719-4800a886c8cf_samples.jsonl b/data/healthbench/anthropic/claude-opus-4-6/a672cd75-3afd-4e97-9719-4800a886c8cf_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..428db0993ecead39ebb05652a7729241e2595f2c --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-6/a672cd75-3afd-4e97-9719-4800a886c8cf_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2f0ac8a50d78fd16a35a4cbb95afddde6edf5d42352515d830d525e58fda9ab3 +size 881338 diff --git a/data/healthbench/anthropic/claude-opus-4-6/b4b41bb4-5e5f-415b-a8eb-3cafdad45dc4.json b/data/healthbench/anthropic/claude-opus-4-6/b4b41bb4-5e5f-415b-a8eb-3cafdad45dc4.json new file mode 100644 index 0000000000000000000000000000000000000000..fe294b810122a9f8909ef46c1a5d09a40b600c00 --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-6/b4b41bb4-5e5f-415b-a8eb-3cafdad45dc4.json @@ -0,0 +1,316 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/anthropic_claude-opus-4-6/1786163964.593079", + "evaluation_timestamp": "1776696116.0", + "retrieved_timestamp": "1786163964.593079", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-6", + "id": "anthropic/claude-opus-4-6", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "bootstrap_score on healthbench/S-adaptive/+3ep/dd74d450 for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "24", + "sample_ids": "0e819a9c-851d-4a7a-9263-62cfc8ce1b48,a079d928-7b78-4a85-a72b-4274d7fc2914,be6150a5-8060-4434-9724-afede307d495,3a896889-9029-4ef8-8684-ca378da4702b,7074bb7a-a595-434d-883e-95ef6f5f6667,a6204079-9e72-42ad-8f39-a65552fa2e68,c0d80b93-b2d5-4129-b998-b1312f98e2a8,3dc41419-2565-428e-b9b9-02e3fd93c470,a91a22f8-d2e6-4593-bcae-6d479ed38031,d0d3426f-8b7a-4501-adb4-86d8fa9099c8,a7dac6a9-eab5-4c01-882c-dbf6c774d7c3,c1240427-ae38-4335-9e80-492c13ef4893,b7ddde65-702e-4255-b1d7-09d01248e76d,00d34549-134c-48e2-b4bb-b54946da70cc,85c11e2a-628b-464f-a8bb-e97ab04a73eb,69a544f8-957d-4dda-a777-608ec7b95617,90789a49-ab1f-4229-bd89-e1c56ea8be3a,cd53e21a-e5ed-4d9b-a258-42a23e9715bb,8d409c7b-29d2-4df2-aa69-ab56c9339bc5,6566f489-f254-4bca-a86b-90f5ac7471d7,d9a44a36-6a3c-4d5d-9c66-b2c0afb479d1,9cf87956-37e6-4bed-9cea-a64b13ae0de5,a1a08e1d-1353-4d6f-94ce-f90872074c7c,51d9c72b-ba60-47c5-9cc9-3ed6212a3982" + } + }, + "evaluation_timestamp": "1776696116.0", + "metric_config": { + "evaluation_description": "bootstrap_score", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.25736396776595305, + "uncertainty": { + "standard_deviation": 0.042749277836137675, + "num_samples": 72 + }, + "details": { + "total_matched_trajectories": "72", + "stopping_reason_count_repetition_guard": "72", + "avg_total_tokens_target_model": "16720.56", + "total_total_tokens_target_model": "1203880.0", + "avg_total_tokens_other_models": "77280.19", + "total_total_tokens_other_models": "5564174.0", + "avg_total_tokens_all_models": "94000.75", + "total_total_tokens_all_models": "6768054.0", + "avg_turn_count": "29.07", + "total_turn_count": "2093.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "3600", + "attempt_timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "3600", + "max_connections": "10" + } + } + }, + { + "evaluation_name": "std on healthbench/S-adaptive/+3ep/dd74d450 for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "24", + "sample_ids": "0e819a9c-851d-4a7a-9263-62cfc8ce1b48,a079d928-7b78-4a85-a72b-4274d7fc2914,be6150a5-8060-4434-9724-afede307d495,3a896889-9029-4ef8-8684-ca378da4702b,7074bb7a-a595-434d-883e-95ef6f5f6667,a6204079-9e72-42ad-8f39-a65552fa2e68,c0d80b93-b2d5-4129-b998-b1312f98e2a8,3dc41419-2565-428e-b9b9-02e3fd93c470,a91a22f8-d2e6-4593-bcae-6d479ed38031,d0d3426f-8b7a-4501-adb4-86d8fa9099c8,a7dac6a9-eab5-4c01-882c-dbf6c774d7c3,c1240427-ae38-4335-9e80-492c13ef4893,b7ddde65-702e-4255-b1d7-09d01248e76d,00d34549-134c-48e2-b4bb-b54946da70cc,85c11e2a-628b-464f-a8bb-e97ab04a73eb,69a544f8-957d-4dda-a777-608ec7b95617,90789a49-ab1f-4229-bd89-e1c56ea8be3a,cd53e21a-e5ed-4d9b-a258-42a23e9715bb,8d409c7b-29d2-4df2-aa69-ab56c9339bc5,6566f489-f254-4bca-a86b-90f5ac7471d7,d9a44a36-6a3c-4d5d-9c66-b2c0afb479d1,9cf87956-37e6-4bed-9cea-a64b13ae0de5,a1a08e1d-1353-4d6f-94ce-f90872074c7c,51d9c72b-ba60-47c5-9cc9-3ed6212a3982" + } + }, + "evaluation_timestamp": "1776696116.0", + "metric_config": { + "evaluation_description": "std", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.042749277836137675, + "uncertainty": { + "standard_deviation": 0.042749277836137675, + "num_samples": 72 + }, + "details": { + "total_matched_trajectories": "72", + "stopping_reason_count_repetition_guard": "72", + "avg_total_tokens_target_model": "16720.56", + "total_total_tokens_target_model": "1203880.0", + "avg_total_tokens_other_models": "77280.19", + "total_total_tokens_other_models": "5564174.0", + "avg_total_tokens_all_models": "94000.75", + "total_total_tokens_all_models": "6768054.0", + "avg_turn_count": "29.07", + "total_turn_count": "2093.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "3600", + "attempt_timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "3600", + "max_connections": "10" + } + } + }, + { + "evaluation_name": "criteria_met_rate on healthbench/S-adaptive/+3ep/dd74d450 for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "24", + "sample_ids": "0e819a9c-851d-4a7a-9263-62cfc8ce1b48,a079d928-7b78-4a85-a72b-4274d7fc2914,be6150a5-8060-4434-9724-afede307d495,3a896889-9029-4ef8-8684-ca378da4702b,7074bb7a-a595-434d-883e-95ef6f5f6667,a6204079-9e72-42ad-8f39-a65552fa2e68,c0d80b93-b2d5-4129-b998-b1312f98e2a8,3dc41419-2565-428e-b9b9-02e3fd93c470,a91a22f8-d2e6-4593-bcae-6d479ed38031,d0d3426f-8b7a-4501-adb4-86d8fa9099c8,a7dac6a9-eab5-4c01-882c-dbf6c774d7c3,c1240427-ae38-4335-9e80-492c13ef4893,b7ddde65-702e-4255-b1d7-09d01248e76d,00d34549-134c-48e2-b4bb-b54946da70cc,85c11e2a-628b-464f-a8bb-e97ab04a73eb,69a544f8-957d-4dda-a777-608ec7b95617,90789a49-ab1f-4229-bd89-e1c56ea8be3a,cd53e21a-e5ed-4d9b-a258-42a23e9715bb,8d409c7b-29d2-4df2-aa69-ab56c9339bc5,6566f489-f254-4bca-a86b-90f5ac7471d7,d9a44a36-6a3c-4d5d-9c66-b2c0afb479d1,9cf87956-37e6-4bed-9cea-a64b13ae0de5,a1a08e1d-1353-4d6f-94ce-f90872074c7c,51d9c72b-ba60-47c5-9cc9-3ed6212a3982" + } + }, + "evaluation_timestamp": "1776696116.0", + "metric_config": { + "evaluation_description": "criteria_met_rate", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.32792207792207795, + "uncertainty": { + "standard_deviation": 0.042749277836137675, + "num_samples": 72 + }, + "details": { + "total_matched_trajectories": "72", + "stopping_reason_count_repetition_guard": "72", + "avg_total_tokens_target_model": "16720.56", + "total_total_tokens_target_model": "1203880.0", + "avg_total_tokens_other_models": "77280.19", + "total_total_tokens_other_models": "5564174.0", + "avg_total_tokens_all_models": "94000.75", + "total_total_tokens_all_models": "6768054.0", + "avg_turn_count": "29.07", + "total_turn_count": "2093.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "3600", + "attempt_timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "3600", + "max_connections": "10" + } + } + }, + { + "evaluation_name": "mean on healthbench/S-adaptive/+3ep/dd74d450 for scorer healthbench_score", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "24", + "sample_ids": "0e819a9c-851d-4a7a-9263-62cfc8ce1b48,a079d928-7b78-4a85-a72b-4274d7fc2914,be6150a5-8060-4434-9724-afede307d495,3a896889-9029-4ef8-8684-ca378da4702b,7074bb7a-a595-434d-883e-95ef6f5f6667,a6204079-9e72-42ad-8f39-a65552fa2e68,c0d80b93-b2d5-4129-b998-b1312f98e2a8,3dc41419-2565-428e-b9b9-02e3fd93c470,a91a22f8-d2e6-4593-bcae-6d479ed38031,d0d3426f-8b7a-4501-adb4-86d8fa9099c8,a7dac6a9-eab5-4c01-882c-dbf6c774d7c3,c1240427-ae38-4335-9e80-492c13ef4893,b7ddde65-702e-4255-b1d7-09d01248e76d,00d34549-134c-48e2-b4bb-b54946da70cc,85c11e2a-628b-464f-a8bb-e97ab04a73eb,69a544f8-957d-4dda-a777-608ec7b95617,90789a49-ab1f-4229-bd89-e1c56ea8be3a,cd53e21a-e5ed-4d9b-a258-42a23e9715bb,8d409c7b-29d2-4df2-aa69-ab56c9339bc5,6566f489-f254-4bca-a86b-90f5ac7471d7,d9a44a36-6a3c-4d5d-9c66-b2c0afb479d1,9cf87956-37e6-4bed-9cea-a64b13ae0de5,a1a08e1d-1353-4d6f-94ce-f90872074c7c,51d9c72b-ba60-47c5-9cc9-3ed6212a3982" + } + }, + "evaluation_timestamp": "1776696116.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.25736396776595305, + "uncertainty": { + "num_samples": 72 + }, + "details": { + "total_matched_trajectories": "72", + "stopping_reason_count_repetition_guard": "72", + "avg_total_tokens_target_model": "16720.56", + "total_total_tokens_target_model": "1203880.0", + "avg_total_tokens_other_models": "77280.19", + "total_total_tokens_other_models": "5564174.0", + "avg_total_tokens_all_models": "94000.75", + "total_total_tokens_all_models": "6768054.0", + "avg_turn_count": "29.07", + "total_turn_count": "2093.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "3600", + "attempt_timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "3600", + "max_connections": "10" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/anthropic/claude-opus-4-6/b4b41bb4-5e5f-415b-a8eb-3cafdad45dc4_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "78dff831146672037aa8a1b6e1948a7996364100ffca8faa222a84d0e1852304", + "total_rows": 72 + } +} \ No newline at end of file diff --git a/data/healthbench/anthropic/claude-opus-4-6/b4b41bb4-5e5f-415b-a8eb-3cafdad45dc4_samples.jsonl b/data/healthbench/anthropic/claude-opus-4-6/b4b41bb4-5e5f-415b-a8eb-3cafdad45dc4_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..7048a059b67a0f5d29f9bf897f9e49a16a0d3082 --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-6/b4b41bb4-5e5f-415b-a8eb-3cafdad45dc4_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:78dff831146672037aa8a1b6e1948a7996364100ffca8faa222a84d0e1852304 +size 5267083 diff --git a/data/healthbench/anthropic/claude-opus-4-6/b6a2c7d5-7116-4d6b-a8f4-7b00b375c4ba.json b/data/healthbench/anthropic/claude-opus-4-6/b6a2c7d5-7116-4d6b-a8f4-7b00b375c4ba.json new file mode 100644 index 0000000000000000000000000000000000000000..db9966b4e71d71fac16ef17d4229b21c97cde9ae --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-6/b6a2c7d5-7116-4d6b-a8f4-7b00b375c4ba.json @@ -0,0 +1,292 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/anthropic_claude-opus-4-6/1786163999.909884", + "evaluation_timestamp": "1775733996.0", + "retrieved_timestamp": "1786163999.909884", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-6", + "id": "anthropic/claude-opus-4-6", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "bootstrap_score on healthbench/S-adaptive/+7ep for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "34", + "sample_ids": "1f548d5b-cd00-49a0-b327-283a2e00debd,0b8f1d60-2081-4562-98f7-b6a976fe1c6d,6f7a2ee9-e9c6-42d8-b79f-22dea966b8d2,19ec4833-86e9-4166-8b82-d1da09f31fd7,7ebc830a-8dbd-489b-9d61-4d8bacf0db8d,c971f9d1-5f6a-464e-b282-41c8f0d82f58,5a6e4a41-3ea6-4050-a971-93433fe34877,78510411-e03b-4f93-a369-c7e72d2e4908,bbd759b4-8e4d-4db7-bc36-f2e069396c1c,8f2a65de-dea7-48e8-8adb-6194eca26c08,6d5f483c-3e86-456d-bfd5-4e28de699ae5,f05491d8-d160-4b2b-bd5b-59d757585e39,85408180-238e-4c1b-92c1-55d8ce6c3717,ecd7453e-1dc1-4e52-bb95-67e8e5788573,0e073591-b3e7-4dc8-94d0-dc5aa93ca35c,7637b987-8fe2-49d0-8372-1ebf17284d70,5b294937-13e5-424e-8bb4-5d1904a2344a,c5bf7fc3-dfb0-4b47-9bbe-222e87b952e2,83cf8f2d-2857-4f01-a283-9595d8f4ae8e,7a4548e6-38b7-48dd-9088-d7d1219f7852,e0d1b955-1150-457e-96d7-91a9481999cd,24f9a6e7-b214-4011-94c4-6502f249a621,6a97773b-c33c-4f2e-9ac3-fdd53de0f175,f32ee0c1-ea9c-4cfb-8332-bd9579d23024,b6eb69cb-b911-4c81-a7a1-357d21d27109,c0dc053b-157b-4d13-9956-213b67ff6a36,3bb4a735-ace7-4342-b7a5-0bae083e6f82,66b448bf-2e49-488d-9d0b-f9912ab3f0f8,b12d2453-fd3b-462c-b9cc-554ff16bfd3f,0a548d04-0973-4343-a3bf-07f730bed2f6,e2cd9f23-e3dd-4b67-8984-a5412dd802e5,1fb6ab91-a999-437d-8c93-c5937c0a89a3,42af0c9f-a715-45c4-ab24-c2948f6f943d,a94aad40-387e-4a37-b92f-61badd57388b" + } + }, + "evaluation_timestamp": "1775733996.0", + "metric_config": { + "evaluation_description": "bootstrap_score", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.4315032560349089, + "uncertainty": { + "standard_deviation": 0.04863449898413763, + "num_samples": 238 + }, + "details": { + "total_matched_trajectories": "8", + "stopping_reason_count_repetition_guard": "8", + "avg_total_tokens_target_model": "20033.50", + "total_total_tokens_target_model": "160268.0", + "avg_total_tokens_other_models": "84150.00", + "total_total_tokens_other_models": "673200.0", + "avg_total_tokens_all_models": "104183.50", + "total_total_tokens_all_models": "833468.0", + "avg_turn_count": "33.12", + "total_turn_count": "265.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + } + } + }, + { + "evaluation_name": "std on healthbench/S-adaptive/+7ep for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "34", + "sample_ids": "1f548d5b-cd00-49a0-b327-283a2e00debd,0b8f1d60-2081-4562-98f7-b6a976fe1c6d,6f7a2ee9-e9c6-42d8-b79f-22dea966b8d2,19ec4833-86e9-4166-8b82-d1da09f31fd7,7ebc830a-8dbd-489b-9d61-4d8bacf0db8d,c971f9d1-5f6a-464e-b282-41c8f0d82f58,5a6e4a41-3ea6-4050-a971-93433fe34877,78510411-e03b-4f93-a369-c7e72d2e4908,bbd759b4-8e4d-4db7-bc36-f2e069396c1c,8f2a65de-dea7-48e8-8adb-6194eca26c08,6d5f483c-3e86-456d-bfd5-4e28de699ae5,f05491d8-d160-4b2b-bd5b-59d757585e39,85408180-238e-4c1b-92c1-55d8ce6c3717,ecd7453e-1dc1-4e52-bb95-67e8e5788573,0e073591-b3e7-4dc8-94d0-dc5aa93ca35c,7637b987-8fe2-49d0-8372-1ebf17284d70,5b294937-13e5-424e-8bb4-5d1904a2344a,c5bf7fc3-dfb0-4b47-9bbe-222e87b952e2,83cf8f2d-2857-4f01-a283-9595d8f4ae8e,7a4548e6-38b7-48dd-9088-d7d1219f7852,e0d1b955-1150-457e-96d7-91a9481999cd,24f9a6e7-b214-4011-94c4-6502f249a621,6a97773b-c33c-4f2e-9ac3-fdd53de0f175,f32ee0c1-ea9c-4cfb-8332-bd9579d23024,b6eb69cb-b911-4c81-a7a1-357d21d27109,c0dc053b-157b-4d13-9956-213b67ff6a36,3bb4a735-ace7-4342-b7a5-0bae083e6f82,66b448bf-2e49-488d-9d0b-f9912ab3f0f8,b12d2453-fd3b-462c-b9cc-554ff16bfd3f,0a548d04-0973-4343-a3bf-07f730bed2f6,e2cd9f23-e3dd-4b67-8984-a5412dd802e5,1fb6ab91-a999-437d-8c93-c5937c0a89a3,42af0c9f-a715-45c4-ab24-c2948f6f943d,a94aad40-387e-4a37-b92f-61badd57388b" + } + }, + "evaluation_timestamp": "1775733996.0", + "metric_config": { + "evaluation_description": "std", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.04863449898413763, + "uncertainty": { + "standard_deviation": 0.04863449898413763, + "num_samples": 238 + }, + "details": { + "total_matched_trajectories": "8", + "stopping_reason_count_repetition_guard": "8", + "avg_total_tokens_target_model": "20033.50", + "total_total_tokens_target_model": "160268.0", + "avg_total_tokens_other_models": "84150.00", + "total_total_tokens_other_models": "673200.0", + "avg_total_tokens_all_models": "104183.50", + "total_total_tokens_all_models": "833468.0", + "avg_turn_count": "33.12", + "total_turn_count": "265.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + } + } + }, + { + "evaluation_name": "criteria_met_rate on healthbench/S-adaptive/+7ep for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "34", + "sample_ids": "1f548d5b-cd00-49a0-b327-283a2e00debd,0b8f1d60-2081-4562-98f7-b6a976fe1c6d,6f7a2ee9-e9c6-42d8-b79f-22dea966b8d2,19ec4833-86e9-4166-8b82-d1da09f31fd7,7ebc830a-8dbd-489b-9d61-4d8bacf0db8d,c971f9d1-5f6a-464e-b282-41c8f0d82f58,5a6e4a41-3ea6-4050-a971-93433fe34877,78510411-e03b-4f93-a369-c7e72d2e4908,bbd759b4-8e4d-4db7-bc36-f2e069396c1c,8f2a65de-dea7-48e8-8adb-6194eca26c08,6d5f483c-3e86-456d-bfd5-4e28de699ae5,f05491d8-d160-4b2b-bd5b-59d757585e39,85408180-238e-4c1b-92c1-55d8ce6c3717,ecd7453e-1dc1-4e52-bb95-67e8e5788573,0e073591-b3e7-4dc8-94d0-dc5aa93ca35c,7637b987-8fe2-49d0-8372-1ebf17284d70,5b294937-13e5-424e-8bb4-5d1904a2344a,c5bf7fc3-dfb0-4b47-9bbe-222e87b952e2,83cf8f2d-2857-4f01-a283-9595d8f4ae8e,7a4548e6-38b7-48dd-9088-d7d1219f7852,e0d1b955-1150-457e-96d7-91a9481999cd,24f9a6e7-b214-4011-94c4-6502f249a621,6a97773b-c33c-4f2e-9ac3-fdd53de0f175,f32ee0c1-ea9c-4cfb-8332-bd9579d23024,b6eb69cb-b911-4c81-a7a1-357d21d27109,c0dc053b-157b-4d13-9956-213b67ff6a36,3bb4a735-ace7-4342-b7a5-0bae083e6f82,66b448bf-2e49-488d-9d0b-f9912ab3f0f8,b12d2453-fd3b-462c-b9cc-554ff16bfd3f,0a548d04-0973-4343-a3bf-07f730bed2f6,e2cd9f23-e3dd-4b67-8984-a5412dd802e5,1fb6ab91-a999-437d-8c93-c5937c0a89a3,42af0c9f-a715-45c4-ab24-c2948f6f943d,a94aad40-387e-4a37-b92f-61badd57388b" + } + }, + "evaluation_timestamp": "1775733996.0", + "metric_config": { + "evaluation_description": "criteria_met_rate", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.4221972253468316, + "uncertainty": { + "standard_deviation": 0.04863449898413763, + "num_samples": 238 + }, + "details": { + "total_matched_trajectories": "8", + "stopping_reason_count_repetition_guard": "8", + "avg_total_tokens_target_model": "20033.50", + "total_total_tokens_target_model": "160268.0", + "avg_total_tokens_other_models": "84150.00", + "total_total_tokens_other_models": "673200.0", + "avg_total_tokens_all_models": "104183.50", + "total_total_tokens_all_models": "833468.0", + "avg_turn_count": "33.12", + "total_turn_count": "265.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + } + } + }, + { + "evaluation_name": "mean on healthbench/S-adaptive/+7ep for scorer healthbench_score", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "34", + "sample_ids": "1f548d5b-cd00-49a0-b327-283a2e00debd,0b8f1d60-2081-4562-98f7-b6a976fe1c6d,6f7a2ee9-e9c6-42d8-b79f-22dea966b8d2,19ec4833-86e9-4166-8b82-d1da09f31fd7,7ebc830a-8dbd-489b-9d61-4d8bacf0db8d,c971f9d1-5f6a-464e-b282-41c8f0d82f58,5a6e4a41-3ea6-4050-a971-93433fe34877,78510411-e03b-4f93-a369-c7e72d2e4908,bbd759b4-8e4d-4db7-bc36-f2e069396c1c,8f2a65de-dea7-48e8-8adb-6194eca26c08,6d5f483c-3e86-456d-bfd5-4e28de699ae5,f05491d8-d160-4b2b-bd5b-59d757585e39,85408180-238e-4c1b-92c1-55d8ce6c3717,ecd7453e-1dc1-4e52-bb95-67e8e5788573,0e073591-b3e7-4dc8-94d0-dc5aa93ca35c,7637b987-8fe2-49d0-8372-1ebf17284d70,5b294937-13e5-424e-8bb4-5d1904a2344a,c5bf7fc3-dfb0-4b47-9bbe-222e87b952e2,83cf8f2d-2857-4f01-a283-9595d8f4ae8e,7a4548e6-38b7-48dd-9088-d7d1219f7852,e0d1b955-1150-457e-96d7-91a9481999cd,24f9a6e7-b214-4011-94c4-6502f249a621,6a97773b-c33c-4f2e-9ac3-fdd53de0f175,f32ee0c1-ea9c-4cfb-8332-bd9579d23024,b6eb69cb-b911-4c81-a7a1-357d21d27109,c0dc053b-157b-4d13-9956-213b67ff6a36,3bb4a735-ace7-4342-b7a5-0bae083e6f82,66b448bf-2e49-488d-9d0b-f9912ab3f0f8,b12d2453-fd3b-462c-b9cc-554ff16bfd3f,0a548d04-0973-4343-a3bf-07f730bed2f6,e2cd9f23-e3dd-4b67-8984-a5412dd802e5,1fb6ab91-a999-437d-8c93-c5937c0a89a3,42af0c9f-a715-45c4-ab24-c2948f6f943d,a94aad40-387e-4a37-b92f-61badd57388b" + } + }, + "evaluation_timestamp": "1775733996.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.4315032560349089, + "uncertainty": { + "num_samples": 238 + }, + "details": { + "total_matched_trajectories": "8", + "stopping_reason_count_repetition_guard": "8", + "avg_total_tokens_target_model": "20033.50", + "total_total_tokens_target_model": "160268.0", + "avg_total_tokens_other_models": "84150.00", + "total_total_tokens_other_models": "673200.0", + "avg_total_tokens_all_models": "104183.50", + "total_total_tokens_all_models": "833468.0", + "avg_turn_count": "33.12", + "total_turn_count": "265.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/anthropic/claude-opus-4-6/b6a2c7d5-7116-4d6b-a8f4-7b00b375c4ba_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "98b36232af52b1ea4b6f237d524dfd9ebe1f0a2b0a9f2b777b86b6b0028c063d", + "total_rows": 238 + } +} \ No newline at end of file diff --git a/data/healthbench/anthropic/claude-opus-4-6/b6a2c7d5-7116-4d6b-a8f4-7b00b375c4ba_samples.jsonl b/data/healthbench/anthropic/claude-opus-4-6/b6a2c7d5-7116-4d6b-a8f4-7b00b375c4ba_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..a77b1f12093f36269c4cedd8b9a247e401d68be2 --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-6/b6a2c7d5-7116-4d6b-a8f4-7b00b375c4ba_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:98b36232af52b1ea4b6f237d524dfd9ebe1f0a2b0a9f2b777b86b6b0028c063d +size 16205448 diff --git a/data/healthbench/anthropic/claude-opus-4-6/b6f3dac7-e5d5-45c1-80b4-5b58bd7a42ab.json b/data/healthbench/anthropic/claude-opus-4-6/b6f3dac7-e5d5-45c1-80b4-5b58bd7a42ab.json new file mode 100644 index 0000000000000000000000000000000000000000..d34084f4e6099b05658f863047e1dbf79c52f493 --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-6/b6f3dac7-e5d5-45c1-80b4-5b58bd7a42ab.json @@ -0,0 +1,292 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/anthropic_claude-opus-4-6/1786163970.740956", + "evaluation_timestamp": "1776263110.0", + "retrieved_timestamp": "1786163970.740956", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-6", + "id": "anthropic/claude-opus-4-6", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "bootstrap_score on healthbench/S-adaptive/+3ep/adc441a2 for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "34", + "sample_ids": "1f548d5b-cd00-49a0-b327-283a2e00debd,0b8f1d60-2081-4562-98f7-b6a976fe1c6d,6f7a2ee9-e9c6-42d8-b79f-22dea966b8d2,19ec4833-86e9-4166-8b82-d1da09f31fd7,7ebc830a-8dbd-489b-9d61-4d8bacf0db8d,c971f9d1-5f6a-464e-b282-41c8f0d82f58,5a6e4a41-3ea6-4050-a971-93433fe34877,78510411-e03b-4f93-a369-c7e72d2e4908,bbd759b4-8e4d-4db7-bc36-f2e069396c1c,8f2a65de-dea7-48e8-8adb-6194eca26c08,6d5f483c-3e86-456d-bfd5-4e28de699ae5,f05491d8-d160-4b2b-bd5b-59d757585e39,85408180-238e-4c1b-92c1-55d8ce6c3717,ecd7453e-1dc1-4e52-bb95-67e8e5788573,0e073591-b3e7-4dc8-94d0-dc5aa93ca35c,7637b987-8fe2-49d0-8372-1ebf17284d70,5b294937-13e5-424e-8bb4-5d1904a2344a,c5bf7fc3-dfb0-4b47-9bbe-222e87b952e2,83cf8f2d-2857-4f01-a283-9595d8f4ae8e,7a4548e6-38b7-48dd-9088-d7d1219f7852,e0d1b955-1150-457e-96d7-91a9481999cd,24f9a6e7-b214-4011-94c4-6502f249a621,6a97773b-c33c-4f2e-9ac3-fdd53de0f175,f32ee0c1-ea9c-4cfb-8332-bd9579d23024,b6eb69cb-b911-4c81-a7a1-357d21d27109,c0dc053b-157b-4d13-9956-213b67ff6a36,3bb4a735-ace7-4342-b7a5-0bae083e6f82,66b448bf-2e49-488d-9d0b-f9912ab3f0f8,b12d2453-fd3b-462c-b9cc-554ff16bfd3f,0a548d04-0973-4343-a3bf-07f730bed2f6,e2cd9f23-e3dd-4b67-8984-a5412dd802e5,1fb6ab91-a999-437d-8c93-c5937c0a89a3,42af0c9f-a715-45c4-ab24-c2948f6f943d,a94aad40-387e-4a37-b92f-61badd57388b" + } + }, + "evaluation_timestamp": "1776263110.0", + "metric_config": { + "evaluation_description": "bootstrap_score", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.43851846663291094, + "uncertainty": { + "standard_deviation": 0.04972561288243523, + "num_samples": 102 + }, + "details": { + "total_matched_trajectories": "12", + "stopping_reason_count_repetition_guard": "12", + "avg_total_tokens_target_model": "21967.25", + "total_total_tokens_target_model": "263607.0", + "avg_total_tokens_other_models": "90428.75", + "total_total_tokens_other_models": "1085145.0", + "avg_total_tokens_all_models": "112396.00", + "total_total_tokens_all_models": "1348752.0", + "avg_turn_count": "33.33", + "total_turn_count": "400.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + } + } + }, + { + "evaluation_name": "std on healthbench/S-adaptive/+3ep/adc441a2 for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "34", + "sample_ids": "1f548d5b-cd00-49a0-b327-283a2e00debd,0b8f1d60-2081-4562-98f7-b6a976fe1c6d,6f7a2ee9-e9c6-42d8-b79f-22dea966b8d2,19ec4833-86e9-4166-8b82-d1da09f31fd7,7ebc830a-8dbd-489b-9d61-4d8bacf0db8d,c971f9d1-5f6a-464e-b282-41c8f0d82f58,5a6e4a41-3ea6-4050-a971-93433fe34877,78510411-e03b-4f93-a369-c7e72d2e4908,bbd759b4-8e4d-4db7-bc36-f2e069396c1c,8f2a65de-dea7-48e8-8adb-6194eca26c08,6d5f483c-3e86-456d-bfd5-4e28de699ae5,f05491d8-d160-4b2b-bd5b-59d757585e39,85408180-238e-4c1b-92c1-55d8ce6c3717,ecd7453e-1dc1-4e52-bb95-67e8e5788573,0e073591-b3e7-4dc8-94d0-dc5aa93ca35c,7637b987-8fe2-49d0-8372-1ebf17284d70,5b294937-13e5-424e-8bb4-5d1904a2344a,c5bf7fc3-dfb0-4b47-9bbe-222e87b952e2,83cf8f2d-2857-4f01-a283-9595d8f4ae8e,7a4548e6-38b7-48dd-9088-d7d1219f7852,e0d1b955-1150-457e-96d7-91a9481999cd,24f9a6e7-b214-4011-94c4-6502f249a621,6a97773b-c33c-4f2e-9ac3-fdd53de0f175,f32ee0c1-ea9c-4cfb-8332-bd9579d23024,b6eb69cb-b911-4c81-a7a1-357d21d27109,c0dc053b-157b-4d13-9956-213b67ff6a36,3bb4a735-ace7-4342-b7a5-0bae083e6f82,66b448bf-2e49-488d-9d0b-f9912ab3f0f8,b12d2453-fd3b-462c-b9cc-554ff16bfd3f,0a548d04-0973-4343-a3bf-07f730bed2f6,e2cd9f23-e3dd-4b67-8984-a5412dd802e5,1fb6ab91-a999-437d-8c93-c5937c0a89a3,42af0c9f-a715-45c4-ab24-c2948f6f943d,a94aad40-387e-4a37-b92f-61badd57388b" + } + }, + "evaluation_timestamp": "1776263110.0", + "metric_config": { + "evaluation_description": "std", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.04972561288243523, + "uncertainty": { + "standard_deviation": 0.04972561288243523, + "num_samples": 102 + }, + "details": { + "total_matched_trajectories": "12", + "stopping_reason_count_repetition_guard": "12", + "avg_total_tokens_target_model": "21967.25", + "total_total_tokens_target_model": "263607.0", + "avg_total_tokens_other_models": "90428.75", + "total_total_tokens_other_models": "1085145.0", + "avg_total_tokens_all_models": "112396.00", + "total_total_tokens_all_models": "1348752.0", + "avg_turn_count": "33.33", + "total_turn_count": "400.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + } + } + }, + { + "evaluation_name": "criteria_met_rate on healthbench/S-adaptive/+3ep/adc441a2 for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "34", + "sample_ids": "1f548d5b-cd00-49a0-b327-283a2e00debd,0b8f1d60-2081-4562-98f7-b6a976fe1c6d,6f7a2ee9-e9c6-42d8-b79f-22dea966b8d2,19ec4833-86e9-4166-8b82-d1da09f31fd7,7ebc830a-8dbd-489b-9d61-4d8bacf0db8d,c971f9d1-5f6a-464e-b282-41c8f0d82f58,5a6e4a41-3ea6-4050-a971-93433fe34877,78510411-e03b-4f93-a369-c7e72d2e4908,bbd759b4-8e4d-4db7-bc36-f2e069396c1c,8f2a65de-dea7-48e8-8adb-6194eca26c08,6d5f483c-3e86-456d-bfd5-4e28de699ae5,f05491d8-d160-4b2b-bd5b-59d757585e39,85408180-238e-4c1b-92c1-55d8ce6c3717,ecd7453e-1dc1-4e52-bb95-67e8e5788573,0e073591-b3e7-4dc8-94d0-dc5aa93ca35c,7637b987-8fe2-49d0-8372-1ebf17284d70,5b294937-13e5-424e-8bb4-5d1904a2344a,c5bf7fc3-dfb0-4b47-9bbe-222e87b952e2,83cf8f2d-2857-4f01-a283-9595d8f4ae8e,7a4548e6-38b7-48dd-9088-d7d1219f7852,e0d1b955-1150-457e-96d7-91a9481999cd,24f9a6e7-b214-4011-94c4-6502f249a621,6a97773b-c33c-4f2e-9ac3-fdd53de0f175,f32ee0c1-ea9c-4cfb-8332-bd9579d23024,b6eb69cb-b911-4c81-a7a1-357d21d27109,c0dc053b-157b-4d13-9956-213b67ff6a36,3bb4a735-ace7-4342-b7a5-0bae083e6f82,66b448bf-2e49-488d-9d0b-f9912ab3f0f8,b12d2453-fd3b-462c-b9cc-554ff16bfd3f,0a548d04-0973-4343-a3bf-07f730bed2f6,e2cd9f23-e3dd-4b67-8984-a5412dd802e5,1fb6ab91-a999-437d-8c93-c5937c0a89a3,42af0c9f-a715-45c4-ab24-c2948f6f943d,a94aad40-387e-4a37-b92f-61badd57388b" + } + }, + "evaluation_timestamp": "1776263110.0", + "metric_config": { + "evaluation_description": "criteria_met_rate", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.41907261592300954, + "uncertainty": { + "standard_deviation": 0.04972561288243523, + "num_samples": 102 + }, + "details": { + "total_matched_trajectories": "12", + "stopping_reason_count_repetition_guard": "12", + "avg_total_tokens_target_model": "21967.25", + "total_total_tokens_target_model": "263607.0", + "avg_total_tokens_other_models": "90428.75", + "total_total_tokens_other_models": "1085145.0", + "avg_total_tokens_all_models": "112396.00", + "total_total_tokens_all_models": "1348752.0", + "avg_turn_count": "33.33", + "total_turn_count": "400.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + } + } + }, + { + "evaluation_name": "mean on healthbench/S-adaptive/+3ep/adc441a2 for scorer healthbench_score", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "34", + "sample_ids": "1f548d5b-cd00-49a0-b327-283a2e00debd,0b8f1d60-2081-4562-98f7-b6a976fe1c6d,6f7a2ee9-e9c6-42d8-b79f-22dea966b8d2,19ec4833-86e9-4166-8b82-d1da09f31fd7,7ebc830a-8dbd-489b-9d61-4d8bacf0db8d,c971f9d1-5f6a-464e-b282-41c8f0d82f58,5a6e4a41-3ea6-4050-a971-93433fe34877,78510411-e03b-4f93-a369-c7e72d2e4908,bbd759b4-8e4d-4db7-bc36-f2e069396c1c,8f2a65de-dea7-48e8-8adb-6194eca26c08,6d5f483c-3e86-456d-bfd5-4e28de699ae5,f05491d8-d160-4b2b-bd5b-59d757585e39,85408180-238e-4c1b-92c1-55d8ce6c3717,ecd7453e-1dc1-4e52-bb95-67e8e5788573,0e073591-b3e7-4dc8-94d0-dc5aa93ca35c,7637b987-8fe2-49d0-8372-1ebf17284d70,5b294937-13e5-424e-8bb4-5d1904a2344a,c5bf7fc3-dfb0-4b47-9bbe-222e87b952e2,83cf8f2d-2857-4f01-a283-9595d8f4ae8e,7a4548e6-38b7-48dd-9088-d7d1219f7852,e0d1b955-1150-457e-96d7-91a9481999cd,24f9a6e7-b214-4011-94c4-6502f249a621,6a97773b-c33c-4f2e-9ac3-fdd53de0f175,f32ee0c1-ea9c-4cfb-8332-bd9579d23024,b6eb69cb-b911-4c81-a7a1-357d21d27109,c0dc053b-157b-4d13-9956-213b67ff6a36,3bb4a735-ace7-4342-b7a5-0bae083e6f82,66b448bf-2e49-488d-9d0b-f9912ab3f0f8,b12d2453-fd3b-462c-b9cc-554ff16bfd3f,0a548d04-0973-4343-a3bf-07f730bed2f6,e2cd9f23-e3dd-4b67-8984-a5412dd802e5,1fb6ab91-a999-437d-8c93-c5937c0a89a3,42af0c9f-a715-45c4-ab24-c2948f6f943d,a94aad40-387e-4a37-b92f-61badd57388b" + } + }, + "evaluation_timestamp": "1776263110.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.43851846663291094, + "uncertainty": { + "num_samples": 102 + }, + "details": { + "total_matched_trajectories": "12", + "stopping_reason_count_repetition_guard": "12", + "avg_total_tokens_target_model": "21967.25", + "total_total_tokens_target_model": "263607.0", + "avg_total_tokens_other_models": "90428.75", + "total_total_tokens_other_models": "1085145.0", + "avg_total_tokens_all_models": "112396.00", + "total_total_tokens_all_models": "1348752.0", + "avg_turn_count": "33.33", + "total_turn_count": "400.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/anthropic/claude-opus-4-6/b6f3dac7-e5d5-45c1-80b4-5b58bd7a42ab_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "fdf7aeffb72b5d609d24cd332c1d2a712b8fba9c895f54d2408522adc8d8eb0e", + "total_rows": 102 + } +} \ No newline at end of file diff --git a/data/healthbench/anthropic/claude-opus-4-6/b6f3dac7-e5d5-45c1-80b4-5b58bd7a42ab_samples.jsonl b/data/healthbench/anthropic/claude-opus-4-6/b6f3dac7-e5d5-45c1-80b4-5b58bd7a42ab_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..249ad16af80c3efb1717762834304f41dcc76776 --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-6/b6f3dac7-e5d5-45c1-80b4-5b58bd7a42ab_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fdf7aeffb72b5d609d24cd332c1d2a712b8fba9c895f54d2408522adc8d8eb0e +size 12738225 diff --git a/data/healthbench/anthropic/claude-opus-4-6/f09f04ac-7006-433d-8ff4-fbb45d1b1a40.json b/data/healthbench/anthropic/claude-opus-4-6/f09f04ac-7006-433d-8ff4-fbb45d1b1a40.json new file mode 100644 index 0000000000000000000000000000000000000000..89c84312c511124cb3b5910405fd6f5ddef43f6d --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-6/f09f04ac-7006-433d-8ff4-fbb45d1b1a40.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/anthropic_claude-opus-4-6/1786166689.459087", + "evaluation_timestamp": "1776696116.0", + "retrieved_timestamp": "1786166689.459087", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-6", + "id": "anthropic/claude-opus-4-6", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/anthropic/claude-opus-4-6/f09f04ac-7006-433d-8ff4-fbb45d1b1a40_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "a019158f2a03992c046396a2b624fe78855aa512d45f41cd93a2937866896367", + "total_rows": 279 + } +} \ No newline at end of file diff --git a/data/healthbench/anthropic/claude-opus-4-6/f09f04ac-7006-433d-8ff4-fbb45d1b1a40_samples.jsonl b/data/healthbench/anthropic/claude-opus-4-6/f09f04ac-7006-433d-8ff4-fbb45d1b1a40_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..bc6b453cf707a844aa7c6e3233c34f95a9ee01b6 --- /dev/null +++ b/data/healthbench/anthropic/claude-opus-4-6/f09f04ac-7006-433d-8ff4-fbb45d1b1a40_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a019158f2a03992c046396a2b624fe78855aa512d45f41cd93a2937866896367 +size 22032695 diff --git a/data/healthbench/openai/gpt-5-2025-08-07/0d42f6c2-58b2-41d9-8d97-d989c7ee1fbd.json b/data/healthbench/openai/gpt-5-2025-08-07/0d42f6c2-58b2-41d9-8d97-d989c7ee1fbd.json new file mode 100644 index 0000000000000000000000000000000000000000..16ed1cea5d0566510968a31bd461d25d64d07af8 --- /dev/null +++ b/data/healthbench/openai/gpt-5-2025-08-07/0d42f6c2-58b2-41d9-8d97-d989c7ee1fbd.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/openai_gpt-5-2025-08-07/1786166536.570412", + "evaluation_timestamp": "1776335705.0", + "retrieved_timestamp": "1786166536.570412", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5-2025-08-07", + "id": "openai/gpt-5-2025-08-07", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/openai/gpt-5-2025-08-07/0d42f6c2-58b2-41d9-8d97-d989c7ee1fbd_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "b3ef937e09821b8faead1945c4c7114a583a276288b07bb33ad30a3eeb6a1c6f", + "total_rows": 129 + } +} \ No newline at end of file diff --git a/data/healthbench/openai/gpt-5-2025-08-07/0d42f6c2-58b2-41d9-8d97-d989c7ee1fbd_samples.jsonl b/data/healthbench/openai/gpt-5-2025-08-07/0d42f6c2-58b2-41d9-8d97-d989c7ee1fbd_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..8e1cdb1074a30944c837a92af59eda7a4416c627 --- /dev/null +++ b/data/healthbench/openai/gpt-5-2025-08-07/0d42f6c2-58b2-41d9-8d97-d989c7ee1fbd_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b3ef937e09821b8faead1945c4c7114a583a276288b07bb33ad30a3eeb6a1c6f +size 8978006 diff --git a/data/healthbench/openai/gpt-5-2025-08-07/2525b612-f151-45d4-9dec-2fb13f80632f.json b/data/healthbench/openai/gpt-5-2025-08-07/2525b612-f151-45d4-9dec-2fb13f80632f.json new file mode 100644 index 0000000000000000000000000000000000000000..7750df47dc06db7e7a023f76cfc2dc91afcc369a --- /dev/null +++ b/data/healthbench/openai/gpt-5-2025-08-07/2525b612-f151-45d4-9dec-2fb13f80632f.json @@ -0,0 +1,304 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/openai_gpt-5-2025-08-07/1786163869.411049", + "evaluation_timestamp": "1776259833.0", + "retrieved_timestamp": "1786163869.411049", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5-2025-08-07", + "id": "openai/gpt-5-2025-08-07", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "bootstrap_score on healthbench/S-adaptive/+2ep/4e1c875a for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "5", + "sample_ids": "2833344d-bff9-4bbe-8fab-6b13d64c6b2a,48669cdc-1a64-4c99-a6e8-9a7181341fa2,be70739f-234a-4a0f-9b14-d15b8800577b,4047884e-4629-457c-9d77-130e937fae60,eb97bae4-430e-45cd-a065-2df3ab5c600e" + } + }, + "evaluation_timestamp": "1776259833.0", + "metric_config": { + "evaluation_description": "bootstrap_score", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.5576095343797898, + "uncertainty": { + "standard_deviation": 0.12610701515654624, + "num_samples": 10 + }, + "details": { + "total_matched_trajectories": "2", + "stopping_reason_count_repetition_guard": "2", + "avg_total_tokens_target_model": "21370.00", + "total_total_tokens_target_model": "42740.0", + "avg_total_tokens_other_models": "33608.00", + "total_total_tokens_other_models": "67216.0", + "avg_total_tokens_all_models": "54978.00", + "total_total_tokens_all_models": "109956.0", + "avg_turn_count": "15.00", + "total_turn_count": "30.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "reasoning_effort": "\"high\"" + } + } + }, + { + "evaluation_name": "std on healthbench/S-adaptive/+2ep/4e1c875a for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "5", + "sample_ids": "2833344d-bff9-4bbe-8fab-6b13d64c6b2a,48669cdc-1a64-4c99-a6e8-9a7181341fa2,be70739f-234a-4a0f-9b14-d15b8800577b,4047884e-4629-457c-9d77-130e937fae60,eb97bae4-430e-45cd-a065-2df3ab5c600e" + } + }, + "evaluation_timestamp": "1776259833.0", + "metric_config": { + "evaluation_description": "std", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.12610701515654624, + "uncertainty": { + "standard_deviation": 0.12610701515654624, + "num_samples": 10 + }, + "details": { + "total_matched_trajectories": "2", + "stopping_reason_count_repetition_guard": "2", + "avg_total_tokens_target_model": "21370.00", + "total_total_tokens_target_model": "42740.0", + "avg_total_tokens_other_models": "33608.00", + "total_total_tokens_other_models": "67216.0", + "avg_total_tokens_all_models": "54978.00", + "total_total_tokens_all_models": "109956.0", + "avg_turn_count": "15.00", + "total_turn_count": "30.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "reasoning_effort": "\"high\"" + } + } + }, + { + "evaluation_name": "criteria_met_rate on healthbench/S-adaptive/+2ep/4e1c875a for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "5", + "sample_ids": "2833344d-bff9-4bbe-8fab-6b13d64c6b2a,48669cdc-1a64-4c99-a6e8-9a7181341fa2,be70739f-234a-4a0f-9b14-d15b8800577b,4047884e-4629-457c-9d77-130e937fae60,eb97bae4-430e-45cd-a065-2df3ab5c600e" + } + }, + "evaluation_timestamp": "1776259833.0", + "metric_config": { + "evaluation_description": "criteria_met_rate", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.5, + "uncertainty": { + "standard_deviation": 0.12610701515654624, + "num_samples": 10 + }, + "details": { + "total_matched_trajectories": "2", + "stopping_reason_count_repetition_guard": "2", + "avg_total_tokens_target_model": "21370.00", + "total_total_tokens_target_model": "42740.0", + "avg_total_tokens_other_models": "33608.00", + "total_total_tokens_other_models": "67216.0", + "avg_total_tokens_all_models": "54978.00", + "total_total_tokens_all_models": "109956.0", + "avg_turn_count": "15.00", + "total_turn_count": "30.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "reasoning_effort": "\"high\"" + } + } + }, + { + "evaluation_name": "mean on healthbench/S-adaptive/+2ep/4e1c875a for scorer healthbench_score", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "5", + "sample_ids": "2833344d-bff9-4bbe-8fab-6b13d64c6b2a,48669cdc-1a64-4c99-a6e8-9a7181341fa2,be70739f-234a-4a0f-9b14-d15b8800577b,4047884e-4629-457c-9d77-130e937fae60,eb97bae4-430e-45cd-a065-2df3ab5c600e" + } + }, + "evaluation_timestamp": "1776259833.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.5576095343797898, + "uncertainty": { + "num_samples": 10 + }, + "details": { + "total_matched_trajectories": "2", + "stopping_reason_count_repetition_guard": "2", + "avg_total_tokens_target_model": "21370.00", + "total_total_tokens_target_model": "42740.0", + "avg_total_tokens_other_models": "33608.00", + "total_total_tokens_other_models": "67216.0", + "avg_total_tokens_all_models": "54978.00", + "total_total_tokens_all_models": "109956.0", + "avg_turn_count": "15.00", + "total_turn_count": "30.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/openai/gpt-5-2025-08-07/2525b612-f151-45d4-9dec-2fb13f80632f_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "d9712737ecf24209e0abd80c6fa31a81dc7ab0b7167f16c5f37cdecde6cff70c", + "total_rows": 10 + } +} \ No newline at end of file diff --git a/data/healthbench/openai/gpt-5-2025-08-07/2525b612-f151-45d4-9dec-2fb13f80632f_samples.jsonl b/data/healthbench/openai/gpt-5-2025-08-07/2525b612-f151-45d4-9dec-2fb13f80632f_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..304c64d944a596af57b49e2ce8361efdc7f81db9 --- /dev/null +++ b/data/healthbench/openai/gpt-5-2025-08-07/2525b612-f151-45d4-9dec-2fb13f80632f_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d9712737ecf24209e0abd80c6fa31a81dc7ab0b7167f16c5f37cdecde6cff70c +size 768833 diff --git a/data/healthbench/openai/gpt-5-2025-08-07/56b35e77-4918-4e61-8ab6-40d7f9a7e382.json b/data/healthbench/openai/gpt-5-2025-08-07/56b35e77-4918-4e61-8ab6-40d7f9a7e382.json new file mode 100644 index 0000000000000000000000000000000000000000..f82038dc3e559100b1831fcb75df3a17ad45953b --- /dev/null +++ b/data/healthbench/openai/gpt-5-2025-08-07/56b35e77-4918-4e61-8ab6-40d7f9a7e382.json @@ -0,0 +1,256 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/openai_gpt-5-2025-08-07/1786163865.413531", + "evaluation_timestamp": "1776076028.0", + "retrieved_timestamp": "1786163865.413531", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5-2025-08-07", + "id": "openai/gpt-5-2025-08-07", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "bootstrap_score on healthbench/S-adaptive/+1ep for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "5", + "sample_ids": "e0d1b955-1150-457e-96d7-91a9481999cd,24f9a6e7-b214-4011-94c4-6502f249a621,6a97773b-c33c-4f2e-9ac3-fdd53de0f175,f32ee0c1-ea9c-4cfb-8332-bd9579d23024,b6eb69cb-b911-4c81-a7a1-357d21d27109" + } + }, + "evaluation_timestamp": "1776076028.0", + "metric_config": { + "evaluation_description": "bootstrap_score", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.6448979591836734, + "uncertainty": { + "standard_deviation": 0.15805518039589583, + "num_samples": 5 + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "reasoning_effort": "\"high\"" + } + } + }, + { + "evaluation_name": "std on healthbench/S-adaptive/+1ep for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "5", + "sample_ids": "e0d1b955-1150-457e-96d7-91a9481999cd,24f9a6e7-b214-4011-94c4-6502f249a621,6a97773b-c33c-4f2e-9ac3-fdd53de0f175,f32ee0c1-ea9c-4cfb-8332-bd9579d23024,b6eb69cb-b911-4c81-a7a1-357d21d27109" + } + }, + "evaluation_timestamp": "1776076028.0", + "metric_config": { + "evaluation_description": "std", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.15805518039589583, + "uncertainty": { + "standard_deviation": 0.15805518039589583, + "num_samples": 5 + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "reasoning_effort": "\"high\"" + } + } + }, + { + "evaluation_name": "criteria_met_rate on healthbench/S-adaptive/+1ep for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "5", + "sample_ids": "e0d1b955-1150-457e-96d7-91a9481999cd,24f9a6e7-b214-4011-94c4-6502f249a621,6a97773b-c33c-4f2e-9ac3-fdd53de0f175,f32ee0c1-ea9c-4cfb-8332-bd9579d23024,b6eb69cb-b911-4c81-a7a1-357d21d27109" + } + }, + "evaluation_timestamp": "1776076028.0", + "metric_config": { + "evaluation_description": "criteria_met_rate", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.358974358974359, + "uncertainty": { + "standard_deviation": 0.15805518039589583, + "num_samples": 5 + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "reasoning_effort": "\"high\"" + } + } + }, + { + "evaluation_name": "mean on healthbench/S-adaptive/+1ep for scorer healthbench_score", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "5", + "sample_ids": "e0d1b955-1150-457e-96d7-91a9481999cd,24f9a6e7-b214-4011-94c4-6502f249a621,6a97773b-c33c-4f2e-9ac3-fdd53de0f175,f32ee0c1-ea9c-4cfb-8332-bd9579d23024,b6eb69cb-b911-4c81-a7a1-357d21d27109" + } + }, + "evaluation_timestamp": "1776076028.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.6448979591836734, + "uncertainty": { + "num_samples": 5 + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/openai/gpt-5-2025-08-07/56b35e77-4918-4e61-8ab6-40d7f9a7e382_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "348167e4d8b8b901872c86557adf70da1189b0c08d9d265d5fc14b8b16179296", + "total_rows": 5 + } +} \ No newline at end of file diff --git a/data/healthbench/openai/gpt-5-2025-08-07/56b35e77-4918-4e61-8ab6-40d7f9a7e382_samples.jsonl b/data/healthbench/openai/gpt-5-2025-08-07/56b35e77-4918-4e61-8ab6-40d7f9a7e382_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..85deba4ef9ddd8831e30f6a5492ffc586753517a --- /dev/null +++ b/data/healthbench/openai/gpt-5-2025-08-07/56b35e77-4918-4e61-8ab6-40d7f9a7e382_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:348167e4d8b8b901872c86557adf70da1189b0c08d9d265d5fc14b8b16179296 +size 287241 diff --git a/data/healthbench/openai/gpt-5-2025-08-07/5fe89f68-7f48-4698-b112-9802619d30fd.json b/data/healthbench/openai/gpt-5-2025-08-07/5fe89f68-7f48-4698-b112-9802619d30fd.json new file mode 100644 index 0000000000000000000000000000000000000000..1c02c4922f67bc15fa9012102125cd0ad7998e87 --- /dev/null +++ b/data/healthbench/openai/gpt-5-2025-08-07/5fe89f68-7f48-4698-b112-9802619d30fd.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/openai_gpt-5-2025-08-07/1786168473.277448", + "evaluation_timestamp": "1776259832.0", + "retrieved_timestamp": "1786168473.277448", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5-2025-08-07", + "id": "openai/gpt-5-2025-08-07", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/openai/gpt-5-2025-08-07/5fe89f68-7f48-4698-b112-9802619d30fd_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "7930b304e25727cbb6c19edf1ddc80e955645f3ed834b8fd9e74f4e2d155de62", + "total_rows": 200 + } +} \ No newline at end of file diff --git a/data/healthbench/openai/gpt-5-2025-08-07/5fe89f68-7f48-4698-b112-9802619d30fd_samples.jsonl b/data/healthbench/openai/gpt-5-2025-08-07/5fe89f68-7f48-4698-b112-9802619d30fd_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..a95570d3272d72b90739133b9d23f0376f20b6de --- /dev/null +++ b/data/healthbench/openai/gpt-5-2025-08-07/5fe89f68-7f48-4698-b112-9802619d30fd_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7930b304e25727cbb6c19edf1ddc80e955645f3ed834b8fd9e74f4e2d155de62 +size 13149907 diff --git a/data/healthbench/openai/gpt-5-2025-08-07/672125ff-810f-4364-a8ce-be06146e18a7.json b/data/healthbench/openai/gpt-5-2025-08-07/672125ff-810f-4364-a8ce-be06146e18a7.json new file mode 100644 index 0000000000000000000000000000000000000000..ce7e18b1dfca7ab7a2631c344c9112e6e1bab3d2 --- /dev/null +++ b/data/healthbench/openai/gpt-5-2025-08-07/672125ff-810f-4364-a8ce-be06146e18a7.json @@ -0,0 +1,320 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/openai_gpt-5-2025-08-07/1786163867.328717", + "evaluation_timestamp": "1776779722.0", + "retrieved_timestamp": "1786163867.328717", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5-2025-08-07", + "id": "openai/gpt-5-2025-08-07", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "bootstrap_score on healthbench/S-adaptive/+1ep/ea2c2278 for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "7", + "sample_ids": "8e88f21a-24c4-4f9c-b848-faf5e10e31b9,a49eb9af-926e-4f7c-94d0-133729059116,2694a753-2f57-41eb-a7f9-8c4c7acdd2e8,5e0a84d7-b972-484c-a649-469b15f0a90f,c7606b1b-ca4d-4a71-8277-bcddff09a653,f24935ac-0530-4063-b2f1-ac45cd265037,6f827e79-05c4-4383-8ed4-16d36260f29f" + } + }, + "evaluation_timestamp": "1776779722.0", + "metric_config": { + "evaluation_description": "bootstrap_score", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.23399935022742038, + "uncertainty": { + "standard_deviation": 0.0762152570969854, + "num_samples": 6 + }, + "details": { + "total_matched_trajectories": "6", + "stopping_reason_count_repetition_guard": "6", + "avg_total_tokens_target_model": "16363.00", + "total_total_tokens_target_model": "98178.0", + "avg_total_tokens_other_models": "60395.00", + "total_total_tokens_other_models": "362370.0", + "avg_total_tokens_all_models": "76758.00", + "total_total_tokens_all_models": "460548.0", + "avg_turn_count": "26.00", + "total_turn_count": "156.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + }, + { + "evaluation_name": "std on healthbench/S-adaptive/+1ep/ea2c2278 for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "7", + "sample_ids": "8e88f21a-24c4-4f9c-b848-faf5e10e31b9,a49eb9af-926e-4f7c-94d0-133729059116,2694a753-2f57-41eb-a7f9-8c4c7acdd2e8,5e0a84d7-b972-484c-a649-469b15f0a90f,c7606b1b-ca4d-4a71-8277-bcddff09a653,f24935ac-0530-4063-b2f1-ac45cd265037,6f827e79-05c4-4383-8ed4-16d36260f29f" + } + }, + "evaluation_timestamp": "1776779722.0", + "metric_config": { + "evaluation_description": "std", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.0762152570969854, + "uncertainty": { + "standard_deviation": 0.0762152570969854, + "num_samples": 6 + }, + "details": { + "total_matched_trajectories": "6", + "stopping_reason_count_repetition_guard": "6", + "avg_total_tokens_target_model": "16363.00", + "total_total_tokens_target_model": "98178.0", + "avg_total_tokens_other_models": "60395.00", + "total_total_tokens_other_models": "362370.0", + "avg_total_tokens_all_models": "76758.00", + "total_total_tokens_all_models": "460548.0", + "avg_turn_count": "26.00", + "total_turn_count": "156.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + }, + { + "evaluation_name": "criteria_met_rate on healthbench/S-adaptive/+1ep/ea2c2278 for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "7", + "sample_ids": "8e88f21a-24c4-4f9c-b848-faf5e10e31b9,a49eb9af-926e-4f7c-94d0-133729059116,2694a753-2f57-41eb-a7f9-8c4c7acdd2e8,5e0a84d7-b972-484c-a649-469b15f0a90f,c7606b1b-ca4d-4a71-8277-bcddff09a653,f24935ac-0530-4063-b2f1-ac45cd265037,6f827e79-05c4-4383-8ed4-16d36260f29f" + } + }, + "evaluation_timestamp": "1776779722.0", + "metric_config": { + "evaluation_description": "criteria_met_rate", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.3333333333333333, + "uncertainty": { + "standard_deviation": 0.0762152570969854, + "num_samples": 6 + }, + "details": { + "total_matched_trajectories": "6", + "stopping_reason_count_repetition_guard": "6", + "avg_total_tokens_target_model": "16363.00", + "total_total_tokens_target_model": "98178.0", + "avg_total_tokens_other_models": "60395.00", + "total_total_tokens_other_models": "362370.0", + "avg_total_tokens_all_models": "76758.00", + "total_total_tokens_all_models": "460548.0", + "avg_turn_count": "26.00", + "total_turn_count": "156.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + }, + { + "evaluation_name": "mean on healthbench/S-adaptive/+1ep/ea2c2278 for scorer healthbench_score", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "7", + "sample_ids": "8e88f21a-24c4-4f9c-b848-faf5e10e31b9,a49eb9af-926e-4f7c-94d0-133729059116,2694a753-2f57-41eb-a7f9-8c4c7acdd2e8,5e0a84d7-b972-484c-a649-469b15f0a90f,c7606b1b-ca4d-4a71-8277-bcddff09a653,f24935ac-0530-4063-b2f1-ac45cd265037,6f827e79-05c4-4383-8ed4-16d36260f29f" + } + }, + "evaluation_timestamp": "1776779722.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.23399935022742038, + "uncertainty": { + "num_samples": 6 + }, + "details": { + "total_matched_trajectories": "6", + "stopping_reason_count_repetition_guard": "6", + "avg_total_tokens_target_model": "16363.00", + "total_total_tokens_target_model": "98178.0", + "avg_total_tokens_other_models": "60395.00", + "total_total_tokens_other_models": "362370.0", + "avg_total_tokens_all_models": "76758.00", + "total_total_tokens_all_models": "460548.0", + "avg_turn_count": "26.00", + "total_turn_count": "156.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/openai/gpt-5-2025-08-07/672125ff-810f-4364-a8ce-be06146e18a7_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "81e563e63b7b4ebc95a0b595f6c1b9436c6a5db019269f5cd941f3a1401e56cf", + "total_rows": 6 + } +} \ No newline at end of file diff --git a/data/healthbench/openai/gpt-5-2025-08-07/672125ff-810f-4364-a8ce-be06146e18a7_samples.jsonl b/data/healthbench/openai/gpt-5-2025-08-07/672125ff-810f-4364-a8ce-be06146e18a7_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..69a44f54b3dd5b263fbb1de78ddc05ba0dc311f8 --- /dev/null +++ b/data/healthbench/openai/gpt-5-2025-08-07/672125ff-810f-4364-a8ce-be06146e18a7_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:81e563e63b7b4ebc95a0b595f6c1b9436c6a5db019269f5cd941f3a1401e56cf +size 394337 diff --git a/data/healthbench/openai/gpt-5-2025-08-07/746031e4-3329-4a92-9c7c-5307a54c8d7c.json b/data/healthbench/openai/gpt-5-2025-08-07/746031e4-3329-4a92-9c7c-5307a54c8d7c.json new file mode 100644 index 0000000000000000000000000000000000000000..bec5cd98a8f616cbf69df390199c800b868b74e5 --- /dev/null +++ b/data/healthbench/openai/gpt-5-2025-08-07/746031e4-3329-4a92-9c7c-5307a54c8d7c.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/openai_gpt-5-2025-08-07/1786166527.973413", + "evaluation_timestamp": "1776335705.0", + "retrieved_timestamp": "1786166527.973413", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5-2025-08-07", + "id": "openai/gpt-5-2025-08-07", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/openai/gpt-5-2025-08-07/746031e4-3329-4a92-9c7c-5307a54c8d7c_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "19623797e369e50273f4f5cf31a6c0f23868f999ddec0d9665c5530d25229e94", + "total_rows": 120 + } +} \ No newline at end of file diff --git a/data/healthbench/openai/gpt-5-2025-08-07/746031e4-3329-4a92-9c7c-5307a54c8d7c_samples.jsonl b/data/healthbench/openai/gpt-5-2025-08-07/746031e4-3329-4a92-9c7c-5307a54c8d7c_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..c893ed74723b6ea026a9b9c897be78450aecef3a --- /dev/null +++ b/data/healthbench/openai/gpt-5-2025-08-07/746031e4-3329-4a92-9c7c-5307a54c8d7c_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:19623797e369e50273f4f5cf31a6c0f23868f999ddec0d9665c5530d25229e94 +size 8346131 diff --git a/data/healthbench/openai/gpt-5-2025-08-07/794ccffc-437d-4bd1-aee2-123b8220e504.json b/data/healthbench/openai/gpt-5-2025-08-07/794ccffc-437d-4bd1-aee2-123b8220e504.json new file mode 100644 index 0000000000000000000000000000000000000000..f7026e55c525ed8f3b16bf6e8c5a75a19a047259 --- /dev/null +++ b/data/healthbench/openai/gpt-5-2025-08-07/794ccffc-437d-4bd1-aee2-123b8220e504.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/openai_gpt-5-2025-08-07/1786166552.739253", + "evaluation_timestamp": "1776259832.0", + "retrieved_timestamp": "1786166552.739253", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5-2025-08-07", + "id": "openai/gpt-5-2025-08-07", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/openai/gpt-5-2025-08-07/794ccffc-437d-4bd1-aee2-123b8220e504_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "a9b81d2a359a0ee6c643aaa915c23cd3d6f5f7199dadd4608ec403bb694817e7", + "total_rows": 250 + } +} \ No newline at end of file diff --git a/data/healthbench/openai/gpt-5-2025-08-07/794ccffc-437d-4bd1-aee2-123b8220e504_samples.jsonl b/data/healthbench/openai/gpt-5-2025-08-07/794ccffc-437d-4bd1-aee2-123b8220e504_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..db0ab1be036aba6d2ef94edc6608911044d39fa5 --- /dev/null +++ b/data/healthbench/openai/gpt-5-2025-08-07/794ccffc-437d-4bd1-aee2-123b8220e504_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a9b81d2a359a0ee6c643aaa915c23cd3d6f5f7199dadd4608ec403bb694817e7 +size 17150485 diff --git a/data/healthbench/openai/gpt-5-2025-08-07/7c7b4029-2ae1-438b-a3e4-79fa985cf826.json b/data/healthbench/openai/gpt-5-2025-08-07/7c7b4029-2ae1-438b-a3e4-79fa985cf826.json new file mode 100644 index 0000000000000000000000000000000000000000..6a10a5fd17a4497113bdcb8ef8879a41498b451d --- /dev/null +++ b/data/healthbench/openai/gpt-5-2025-08-07/7c7b4029-2ae1-438b-a3e4-79fa985cf826.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/openai_gpt-5-2025-08-07/1786168460.207794", + "evaluation_timestamp": "1776076027.0", + "retrieved_timestamp": "1786168460.207794", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5-2025-08-07", + "id": "openai/gpt-5-2025-08-07", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/openai/gpt-5-2025-08-07/7c7b4029-2ae1-438b-a3e4-79fa985cf826_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "451c412bfbe60d5b3c0dfd84cffa0f4e8b42d252cdafb1fe9197aec4f2409e9f", + "total_rows": 70 + } +} \ No newline at end of file diff --git a/data/healthbench/openai/gpt-5-2025-08-07/7c7b4029-2ae1-438b-a3e4-79fa985cf826_samples.jsonl b/data/healthbench/openai/gpt-5-2025-08-07/7c7b4029-2ae1-438b-a3e4-79fa985cf826_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..e2f5303ef7547a261f75e186962377bacd292e0a --- /dev/null +++ b/data/healthbench/openai/gpt-5-2025-08-07/7c7b4029-2ae1-438b-a3e4-79fa985cf826_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:451c412bfbe60d5b3c0dfd84cffa0f4e8b42d252cdafb1fe9197aec4f2409e9f +size 4488958 diff --git a/data/healthbench/openai/gpt-5-2025-08-07/81c14e0b-fcfb-4cd6-8cf4-2d9a4ddbf854.json b/data/healthbench/openai/gpt-5-2025-08-07/81c14e0b-fcfb-4cd6-8cf4-2d9a4ddbf854.json new file mode 100644 index 0000000000000000000000000000000000000000..5abd44469aefbc7af0ffc800f560668d2f93f373 --- /dev/null +++ b/data/healthbench/openai/gpt-5-2025-08-07/81c14e0b-fcfb-4cd6-8cf4-2d9a4ddbf854.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/openai_gpt-5-2025-08-07/1786168516.889656", + "evaluation_timestamp": "1776165241.0", + "retrieved_timestamp": "1786168516.889656", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5-2025-08-07", + "id": "openai/gpt-5-2025-08-07", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/openai/gpt-5-2025-08-07/81c14e0b-fcfb-4cd6-8cf4-2d9a4ddbf854_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "3f3e096fb701e775d61b1e57e35ed8eb03ba6dcc55e20a26623df2e89c7ef9fa", + "total_rows": 351 + } +} \ No newline at end of file diff --git a/data/healthbench/openai/gpt-5-2025-08-07/81c14e0b-fcfb-4cd6-8cf4-2d9a4ddbf854_samples.jsonl b/data/healthbench/openai/gpt-5-2025-08-07/81c14e0b-fcfb-4cd6-8cf4-2d9a4ddbf854_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..9bdf25be2c3408d0c21bd8062e703aa25dc3dc1a --- /dev/null +++ b/data/healthbench/openai/gpt-5-2025-08-07/81c14e0b-fcfb-4cd6-8cf4-2d9a4ddbf854_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3f3e096fb701e775d61b1e57e35ed8eb03ba6dcc55e20a26623df2e89c7ef9fa +size 22318921 diff --git a/data/healthbench/openai/gpt-5-2025-08-07/8ed1e6da-d3cf-4285-98da-3740d56f3fc9.json b/data/healthbench/openai/gpt-5-2025-08-07/8ed1e6da-d3cf-4285-98da-3740d56f3fc9.json new file mode 100644 index 0000000000000000000000000000000000000000..2e322eeea13746167382532a457efa79813a98c5 --- /dev/null +++ b/data/healthbench/openai/gpt-5-2025-08-07/8ed1e6da-d3cf-4285-98da-3740d56f3fc9.json @@ -0,0 +1,304 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/openai_gpt-5-2025-08-07/1786163873.224801", + "evaluation_timestamp": "1776261896.0", + "retrieved_timestamp": "1786163873.224801", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5-2025-08-07", + "id": "openai/gpt-5-2025-08-07", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "bootstrap_score on healthbench/S-adaptive/+3ep/c59db474 for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "26", + "sample_ids": "1f548d5b-cd00-49a0-b327-283a2e00debd,0b8f1d60-2081-4562-98f7-b6a976fe1c6d,6f7a2ee9-e9c6-42d8-b79f-22dea966b8d2,19ec4833-86e9-4166-8b82-d1da09f31fd7,7ebc830a-8dbd-489b-9d61-4d8bacf0db8d,c971f9d1-5f6a-464e-b282-41c8f0d82f58,5a6e4a41-3ea6-4050-a971-93433fe34877,78510411-e03b-4f93-a369-c7e72d2e4908,bbd759b4-8e4d-4db7-bc36-f2e069396c1c,8f2a65de-dea7-48e8-8adb-6194eca26c08,6d5f483c-3e86-456d-bfd5-4e28de699ae5,f05491d8-d160-4b2b-bd5b-59d757585e39,85408180-238e-4c1b-92c1-55d8ce6c3717,ecd7453e-1dc1-4e52-bb95-67e8e5788573,0e073591-b3e7-4dc8-94d0-dc5aa93ca35c,7637b987-8fe2-49d0-8372-1ebf17284d70,5b294937-13e5-424e-8bb4-5d1904a2344a,c5bf7fc3-dfb0-4b47-9bbe-222e87b952e2,83cf8f2d-2857-4f01-a283-9595d8f4ae8e,7a4548e6-38b7-48dd-9088-d7d1219f7852,e0d1b955-1150-457e-96d7-91a9481999cd,24f9a6e7-b214-4011-94c4-6502f249a621,6a97773b-c33c-4f2e-9ac3-fdd53de0f175,f32ee0c1-ea9c-4cfb-8332-bd9579d23024,b6eb69cb-b911-4c81-a7a1-357d21d27109,c0dc053b-157b-4d13-9956-213b67ff6a36" + } + }, + "evaluation_timestamp": "1776261896.0", + "metric_config": { + "evaluation_description": "bootstrap_score", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.47675577569810834, + "uncertainty": { + "standard_deviation": 0.04955664771026938, + "num_samples": 78 + }, + "details": { + "total_matched_trajectories": "12", + "stopping_reason_count_repetition_guard": "12", + "avg_total_tokens_target_model": "20832.08", + "total_total_tokens_target_model": "249985.0", + "avg_total_tokens_other_models": "85459.17", + "total_total_tokens_other_models": "1025510.0", + "avg_total_tokens_all_models": "106291.25", + "total_total_tokens_all_models": "1275495.0", + "avg_turn_count": "32.50", + "total_turn_count": "390.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "reasoning_effort": "\"high\"" + } + } + }, + { + "evaluation_name": "std on healthbench/S-adaptive/+3ep/c59db474 for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "26", + "sample_ids": "1f548d5b-cd00-49a0-b327-283a2e00debd,0b8f1d60-2081-4562-98f7-b6a976fe1c6d,6f7a2ee9-e9c6-42d8-b79f-22dea966b8d2,19ec4833-86e9-4166-8b82-d1da09f31fd7,7ebc830a-8dbd-489b-9d61-4d8bacf0db8d,c971f9d1-5f6a-464e-b282-41c8f0d82f58,5a6e4a41-3ea6-4050-a971-93433fe34877,78510411-e03b-4f93-a369-c7e72d2e4908,bbd759b4-8e4d-4db7-bc36-f2e069396c1c,8f2a65de-dea7-48e8-8adb-6194eca26c08,6d5f483c-3e86-456d-bfd5-4e28de699ae5,f05491d8-d160-4b2b-bd5b-59d757585e39,85408180-238e-4c1b-92c1-55d8ce6c3717,ecd7453e-1dc1-4e52-bb95-67e8e5788573,0e073591-b3e7-4dc8-94d0-dc5aa93ca35c,7637b987-8fe2-49d0-8372-1ebf17284d70,5b294937-13e5-424e-8bb4-5d1904a2344a,c5bf7fc3-dfb0-4b47-9bbe-222e87b952e2,83cf8f2d-2857-4f01-a283-9595d8f4ae8e,7a4548e6-38b7-48dd-9088-d7d1219f7852,e0d1b955-1150-457e-96d7-91a9481999cd,24f9a6e7-b214-4011-94c4-6502f249a621,6a97773b-c33c-4f2e-9ac3-fdd53de0f175,f32ee0c1-ea9c-4cfb-8332-bd9579d23024,b6eb69cb-b911-4c81-a7a1-357d21d27109,c0dc053b-157b-4d13-9956-213b67ff6a36" + } + }, + "evaluation_timestamp": "1776261896.0", + "metric_config": { + "evaluation_description": "std", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.04955664771026938, + "uncertainty": { + "standard_deviation": 0.04955664771026938, + "num_samples": 78 + }, + "details": { + "total_matched_trajectories": "12", + "stopping_reason_count_repetition_guard": "12", + "avg_total_tokens_target_model": "20832.08", + "total_total_tokens_target_model": "249985.0", + "avg_total_tokens_other_models": "85459.17", + "total_total_tokens_other_models": "1025510.0", + "avg_total_tokens_all_models": "106291.25", + "total_total_tokens_all_models": "1275495.0", + "avg_turn_count": "32.50", + "total_turn_count": "390.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "reasoning_effort": "\"high\"" + } + } + }, + { + "evaluation_name": "criteria_met_rate on healthbench/S-adaptive/+3ep/c59db474 for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "26", + "sample_ids": "1f548d5b-cd00-49a0-b327-283a2e00debd,0b8f1d60-2081-4562-98f7-b6a976fe1c6d,6f7a2ee9-e9c6-42d8-b79f-22dea966b8d2,19ec4833-86e9-4166-8b82-d1da09f31fd7,7ebc830a-8dbd-489b-9d61-4d8bacf0db8d,c971f9d1-5f6a-464e-b282-41c8f0d82f58,5a6e4a41-3ea6-4050-a971-93433fe34877,78510411-e03b-4f93-a369-c7e72d2e4908,bbd759b4-8e4d-4db7-bc36-f2e069396c1c,8f2a65de-dea7-48e8-8adb-6194eca26c08,6d5f483c-3e86-456d-bfd5-4e28de699ae5,f05491d8-d160-4b2b-bd5b-59d757585e39,85408180-238e-4c1b-92c1-55d8ce6c3717,ecd7453e-1dc1-4e52-bb95-67e8e5788573,0e073591-b3e7-4dc8-94d0-dc5aa93ca35c,7637b987-8fe2-49d0-8372-1ebf17284d70,5b294937-13e5-424e-8bb4-5d1904a2344a,c5bf7fc3-dfb0-4b47-9bbe-222e87b952e2,83cf8f2d-2857-4f01-a283-9595d8f4ae8e,7a4548e6-38b7-48dd-9088-d7d1219f7852,e0d1b955-1150-457e-96d7-91a9481999cd,24f9a6e7-b214-4011-94c4-6502f249a621,6a97773b-c33c-4f2e-9ac3-fdd53de0f175,f32ee0c1-ea9c-4cfb-8332-bd9579d23024,b6eb69cb-b911-4c81-a7a1-357d21d27109,c0dc053b-157b-4d13-9956-213b67ff6a36" + } + }, + "evaluation_timestamp": "1776261896.0", + "metric_config": { + "evaluation_description": "criteria_met_rate", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.47712418300653603, + "uncertainty": { + "standard_deviation": 0.04955664771026938, + "num_samples": 78 + }, + "details": { + "total_matched_trajectories": "12", + "stopping_reason_count_repetition_guard": "12", + "avg_total_tokens_target_model": "20832.08", + "total_total_tokens_target_model": "249985.0", + "avg_total_tokens_other_models": "85459.17", + "total_total_tokens_other_models": "1025510.0", + "avg_total_tokens_all_models": "106291.25", + "total_total_tokens_all_models": "1275495.0", + "avg_turn_count": "32.50", + "total_turn_count": "390.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "reasoning_effort": "\"high\"" + } + } + }, + { + "evaluation_name": "mean on healthbench/S-adaptive/+3ep/c59db474 for scorer healthbench_score", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "26", + "sample_ids": "1f548d5b-cd00-49a0-b327-283a2e00debd,0b8f1d60-2081-4562-98f7-b6a976fe1c6d,6f7a2ee9-e9c6-42d8-b79f-22dea966b8d2,19ec4833-86e9-4166-8b82-d1da09f31fd7,7ebc830a-8dbd-489b-9d61-4d8bacf0db8d,c971f9d1-5f6a-464e-b282-41c8f0d82f58,5a6e4a41-3ea6-4050-a971-93433fe34877,78510411-e03b-4f93-a369-c7e72d2e4908,bbd759b4-8e4d-4db7-bc36-f2e069396c1c,8f2a65de-dea7-48e8-8adb-6194eca26c08,6d5f483c-3e86-456d-bfd5-4e28de699ae5,f05491d8-d160-4b2b-bd5b-59d757585e39,85408180-238e-4c1b-92c1-55d8ce6c3717,ecd7453e-1dc1-4e52-bb95-67e8e5788573,0e073591-b3e7-4dc8-94d0-dc5aa93ca35c,7637b987-8fe2-49d0-8372-1ebf17284d70,5b294937-13e5-424e-8bb4-5d1904a2344a,c5bf7fc3-dfb0-4b47-9bbe-222e87b952e2,83cf8f2d-2857-4f01-a283-9595d8f4ae8e,7a4548e6-38b7-48dd-9088-d7d1219f7852,e0d1b955-1150-457e-96d7-91a9481999cd,24f9a6e7-b214-4011-94c4-6502f249a621,6a97773b-c33c-4f2e-9ac3-fdd53de0f175,f32ee0c1-ea9c-4cfb-8332-bd9579d23024,b6eb69cb-b911-4c81-a7a1-357d21d27109,c0dc053b-157b-4d13-9956-213b67ff6a36" + } + }, + "evaluation_timestamp": "1776261896.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.47675577569810834, + "uncertainty": { + "num_samples": 78 + }, + "details": { + "total_matched_trajectories": "12", + "stopping_reason_count_repetition_guard": "12", + "avg_total_tokens_target_model": "20832.08", + "total_total_tokens_target_model": "249985.0", + "avg_total_tokens_other_models": "85459.17", + "total_total_tokens_other_models": "1025510.0", + "avg_total_tokens_all_models": "106291.25", + "total_total_tokens_all_models": "1275495.0", + "avg_turn_count": "32.50", + "total_turn_count": "390.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/openai/gpt-5-2025-08-07/8ed1e6da-d3cf-4285-98da-3740d56f3fc9_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "e93d7681c2f67e1028611940c4fd6dd82d273c6c9fb98f6f510497d6b4d1e22e", + "total_rows": 78 + } +} \ No newline at end of file diff --git a/data/healthbench/openai/gpt-5-2025-08-07/8ed1e6da-d3cf-4285-98da-3740d56f3fc9_samples.jsonl b/data/healthbench/openai/gpt-5-2025-08-07/8ed1e6da-d3cf-4285-98da-3740d56f3fc9_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..34db5aa8e3eb84efff8cd33f7be352fd5823adac --- /dev/null +++ b/data/healthbench/openai/gpt-5-2025-08-07/8ed1e6da-d3cf-4285-98da-3740d56f3fc9_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e93d7681c2f67e1028611940c4fd6dd82d273c6c9fb98f6f510497d6b4d1e22e +size 5224153 diff --git a/data/healthbench/openai/gpt-5-2025-08-07/918416be-dff3-4a91-921b-676c24034076.json b/data/healthbench/openai/gpt-5-2025-08-07/918416be-dff3-4a91-921b-676c24034076.json new file mode 100644 index 0000000000000000000000000000000000000000..b5f892a12a8f63edeef3fbeb2050532aa489af5d --- /dev/null +++ b/data/healthbench/openai/gpt-5-2025-08-07/918416be-dff3-4a91-921b-676c24034076.json @@ -0,0 +1,320 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/openai_gpt-5-2025-08-07/1786163900.528634", + "evaluation_timestamp": "1776696116.0", + "retrieved_timestamp": "1786163900.528634", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5-2025-08-07", + "id": "openai/gpt-5-2025-08-07", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "bootstrap_score on healthbench/S-adaptive/6ce2ac11 for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "66", + "sample_ids": "f712eb09-82af-4fe8-800d-bc3af772a9ba,8ee6a990-8966-4b42-baa8-6c7e9f74ce74,267bda16-fd4b-4a9e-b3ac-5602c2532eed,0e819a9c-851d-4a7a-9263-62cfc8ce1b48,328c55cf-0b57-4992-99b1-f6a21af62ade,7b415a2f-22b0-4735-a38e-567f1891acc5,c49ac752-3693-4be1-ab4d-91fbdd02f827,a079d928-7b78-4a85-a72b-4274d7fc2914,a20cbbf3-330b-46d9-a35b-92e89a022282,be6150a5-8060-4434-9724-afede307d495,3a896889-9029-4ef8-8684-ca378da4702b,7074bb7a-a595-434d-883e-95ef6f5f6667,9f7b555a-fc30-40c3-88d3-f2ece06454a2,a6204079-9e72-42ad-8f39-a65552fa2e68,c0d80b93-b2d5-4129-b998-b1312f98e2a8,3dc41419-2565-428e-b9b9-02e3fd93c470,a91a22f8-d2e6-4593-bcae-6d479ed38031,d0d3426f-8b7a-4501-adb4-86d8fa9099c8,a7dac6a9-eab5-4c01-882c-dbf6c774d7c3,c1240427-ae38-4335-9e80-492c13ef4893,b7ddde65-702e-4255-b1d7-09d01248e76d,00d34549-134c-48e2-b4bb-b54946da70cc,85c11e2a-628b-464f-a8bb-e97ab04a73eb,69a544f8-957d-4dda-a777-608ec7b95617,90789a49-ab1f-4229-bd89-e1c56ea8be3a,cd53e21a-e5ed-4d9b-a258-42a23e9715bb,8d409c7b-29d2-4df2-aa69-ab56c9339bc5,6566f489-f254-4bca-a86b-90f5ac7471d7,d9a44a36-6a3c-4d5d-9c66-b2c0afb479d1,9cf87956-37e6-4bed-9cea-a64b13ae0de5,1dc0b953-ef66-4a78-9989-7cf1053b5b41,a1a08e1d-1353-4d6f-94ce-f90872074c7c,51d9c72b-ba60-47c5-9cc9-3ed6212a3982,3cf35bfe-d881-4881-a268-b5fa0607bf89,bc8d3e4a-8ba0-4cab-be1c-b20ae3af91e4,99306c29-bcef-49d5-bcd1-78d4e541e197,d0fcfaf0-7c6d-4ead-915b-6d9d9130e747,59c68d10-55c1-4629-b05a-bedde7f343b5,d86ee090-3a2e-44d0-abed-6f0f40a38c08,25cf07b8-606f-4448-99ef-ea87ca3097ba,1ca222ed-cc79-4e78-9ca3-3547e7b37e3a,63e4e0d4-0cb1-442e-8316-8c12a6e3ec14,0eb6fd6d-eaaa-46db-ab16-02e610f238a9,99b22b39-30d5-40dc-82d9-c1cd3c793a44,8c342edc-619b-4079-a28c-b7129f3670a6,e8991792-c875-4f40-99a0-54c902d42a52,54349c88-e80a-4728-b181-305f97905d59,f93b0850-2c60-479d-831b-b8c8f9493d61,c63a4306-4003-45a6-8be1-d669457473a9,d3cbb579-70a5-46e8-95f6-72cc9afa4ba3,3852d050-c009-4e33-b953-fd734a5b5e73,5dc23430-dfbb-4794-9f77-0d4dd6093033,67e76447-f16a-4a65-aba7-fc7166fd9b63,a783e7a3-cecc-4d15-b18c-3f2ddfc2dc6c,62e9d35f-153d-4744-b97c-9abfb2a90b1e,04b6e0c8-2835-415e-865d-aca9f1cd3c57,7a283823-00ee-489d-8dbc-356e2857526f,e89e00c9-29e9-420c-813d-e06ddc39d585,51433193-b0c3-4725-a3a8-9dd15e969b0c,48c89cd1-0ecf-4738-8fcc-392269585bd3,92ea478e-6bf8-4ffb-b88a-f57ccf5f2346,36c6e780-d4a8-47b2-8a36-90ee4b97607f,5fb6812f-9816-4d5d-b7e0-4ec3e9df5dc1,c29c95d1-5b5e-4f53-9ec5-83e1f72b5aa5,b81c059a-1137-47cf-bab8-ab5eb8ff34d5,886864a6-a0e4-468f-aaf4-26cf8dc5e541" + } + }, + "evaluation_timestamp": "1776696116.0", + "metric_config": { + "evaluation_description": "bootstrap_score", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.2416969742523121, + "uncertainty": { + "standard_deviation": 0.034674948869492205, + "num_samples": 318 + }, + "details": { + "total_matched_trajectories": "318", + "stopping_reason_count_repetition_guard": "318", + "avg_total_tokens_target_model": "21254.46", + "total_total_tokens_target_model": "6758919.0", + "avg_total_tokens_other_models": "68935.17", + "total_total_tokens_other_models": "21921383.0", + "avg_total_tokens_all_models": "90189.63", + "total_total_tokens_all_models": "28680302.0", + "avg_turn_count": "25.64", + "total_turn_count": "8155.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "3600", + "attempt_timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + }, + { + "evaluation_name": "std on healthbench/S-adaptive/6ce2ac11 for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "66", + "sample_ids": "f712eb09-82af-4fe8-800d-bc3af772a9ba,8ee6a990-8966-4b42-baa8-6c7e9f74ce74,267bda16-fd4b-4a9e-b3ac-5602c2532eed,0e819a9c-851d-4a7a-9263-62cfc8ce1b48,328c55cf-0b57-4992-99b1-f6a21af62ade,7b415a2f-22b0-4735-a38e-567f1891acc5,c49ac752-3693-4be1-ab4d-91fbdd02f827,a079d928-7b78-4a85-a72b-4274d7fc2914,a20cbbf3-330b-46d9-a35b-92e89a022282,be6150a5-8060-4434-9724-afede307d495,3a896889-9029-4ef8-8684-ca378da4702b,7074bb7a-a595-434d-883e-95ef6f5f6667,9f7b555a-fc30-40c3-88d3-f2ece06454a2,a6204079-9e72-42ad-8f39-a65552fa2e68,c0d80b93-b2d5-4129-b998-b1312f98e2a8,3dc41419-2565-428e-b9b9-02e3fd93c470,a91a22f8-d2e6-4593-bcae-6d479ed38031,d0d3426f-8b7a-4501-adb4-86d8fa9099c8,a7dac6a9-eab5-4c01-882c-dbf6c774d7c3,c1240427-ae38-4335-9e80-492c13ef4893,b7ddde65-702e-4255-b1d7-09d01248e76d,00d34549-134c-48e2-b4bb-b54946da70cc,85c11e2a-628b-464f-a8bb-e97ab04a73eb,69a544f8-957d-4dda-a777-608ec7b95617,90789a49-ab1f-4229-bd89-e1c56ea8be3a,cd53e21a-e5ed-4d9b-a258-42a23e9715bb,8d409c7b-29d2-4df2-aa69-ab56c9339bc5,6566f489-f254-4bca-a86b-90f5ac7471d7,d9a44a36-6a3c-4d5d-9c66-b2c0afb479d1,9cf87956-37e6-4bed-9cea-a64b13ae0de5,1dc0b953-ef66-4a78-9989-7cf1053b5b41,a1a08e1d-1353-4d6f-94ce-f90872074c7c,51d9c72b-ba60-47c5-9cc9-3ed6212a3982,3cf35bfe-d881-4881-a268-b5fa0607bf89,bc8d3e4a-8ba0-4cab-be1c-b20ae3af91e4,99306c29-bcef-49d5-bcd1-78d4e541e197,d0fcfaf0-7c6d-4ead-915b-6d9d9130e747,59c68d10-55c1-4629-b05a-bedde7f343b5,d86ee090-3a2e-44d0-abed-6f0f40a38c08,25cf07b8-606f-4448-99ef-ea87ca3097ba,1ca222ed-cc79-4e78-9ca3-3547e7b37e3a,63e4e0d4-0cb1-442e-8316-8c12a6e3ec14,0eb6fd6d-eaaa-46db-ab16-02e610f238a9,99b22b39-30d5-40dc-82d9-c1cd3c793a44,8c342edc-619b-4079-a28c-b7129f3670a6,e8991792-c875-4f40-99a0-54c902d42a52,54349c88-e80a-4728-b181-305f97905d59,f93b0850-2c60-479d-831b-b8c8f9493d61,c63a4306-4003-45a6-8be1-d669457473a9,d3cbb579-70a5-46e8-95f6-72cc9afa4ba3,3852d050-c009-4e33-b953-fd734a5b5e73,5dc23430-dfbb-4794-9f77-0d4dd6093033,67e76447-f16a-4a65-aba7-fc7166fd9b63,a783e7a3-cecc-4d15-b18c-3f2ddfc2dc6c,62e9d35f-153d-4744-b97c-9abfb2a90b1e,04b6e0c8-2835-415e-865d-aca9f1cd3c57,7a283823-00ee-489d-8dbc-356e2857526f,e89e00c9-29e9-420c-813d-e06ddc39d585,51433193-b0c3-4725-a3a8-9dd15e969b0c,48c89cd1-0ecf-4738-8fcc-392269585bd3,92ea478e-6bf8-4ffb-b88a-f57ccf5f2346,36c6e780-d4a8-47b2-8a36-90ee4b97607f,5fb6812f-9816-4d5d-b7e0-4ec3e9df5dc1,c29c95d1-5b5e-4f53-9ec5-83e1f72b5aa5,b81c059a-1137-47cf-bab8-ab5eb8ff34d5,886864a6-a0e4-468f-aaf4-26cf8dc5e541" + } + }, + "evaluation_timestamp": "1776696116.0", + "metric_config": { + "evaluation_description": "std", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.034674948869492205, + "uncertainty": { + "standard_deviation": 0.034674948869492205, + "num_samples": 318 + }, + "details": { + "total_matched_trajectories": "318", + "stopping_reason_count_repetition_guard": "318", + "avg_total_tokens_target_model": "21254.46", + "total_total_tokens_target_model": "6758919.0", + "avg_total_tokens_other_models": "68935.17", + "total_total_tokens_other_models": "21921383.0", + "avg_total_tokens_all_models": "90189.63", + "total_total_tokens_all_models": "28680302.0", + "avg_turn_count": "25.64", + "total_turn_count": "8155.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "3600", + "attempt_timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + }, + { + "evaluation_name": "criteria_met_rate on healthbench/S-adaptive/6ce2ac11 for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "66", + "sample_ids": "f712eb09-82af-4fe8-800d-bc3af772a9ba,8ee6a990-8966-4b42-baa8-6c7e9f74ce74,267bda16-fd4b-4a9e-b3ac-5602c2532eed,0e819a9c-851d-4a7a-9263-62cfc8ce1b48,328c55cf-0b57-4992-99b1-f6a21af62ade,7b415a2f-22b0-4735-a38e-567f1891acc5,c49ac752-3693-4be1-ab4d-91fbdd02f827,a079d928-7b78-4a85-a72b-4274d7fc2914,a20cbbf3-330b-46d9-a35b-92e89a022282,be6150a5-8060-4434-9724-afede307d495,3a896889-9029-4ef8-8684-ca378da4702b,7074bb7a-a595-434d-883e-95ef6f5f6667,9f7b555a-fc30-40c3-88d3-f2ece06454a2,a6204079-9e72-42ad-8f39-a65552fa2e68,c0d80b93-b2d5-4129-b998-b1312f98e2a8,3dc41419-2565-428e-b9b9-02e3fd93c470,a91a22f8-d2e6-4593-bcae-6d479ed38031,d0d3426f-8b7a-4501-adb4-86d8fa9099c8,a7dac6a9-eab5-4c01-882c-dbf6c774d7c3,c1240427-ae38-4335-9e80-492c13ef4893,b7ddde65-702e-4255-b1d7-09d01248e76d,00d34549-134c-48e2-b4bb-b54946da70cc,85c11e2a-628b-464f-a8bb-e97ab04a73eb,69a544f8-957d-4dda-a777-608ec7b95617,90789a49-ab1f-4229-bd89-e1c56ea8be3a,cd53e21a-e5ed-4d9b-a258-42a23e9715bb,8d409c7b-29d2-4df2-aa69-ab56c9339bc5,6566f489-f254-4bca-a86b-90f5ac7471d7,d9a44a36-6a3c-4d5d-9c66-b2c0afb479d1,9cf87956-37e6-4bed-9cea-a64b13ae0de5,1dc0b953-ef66-4a78-9989-7cf1053b5b41,a1a08e1d-1353-4d6f-94ce-f90872074c7c,51d9c72b-ba60-47c5-9cc9-3ed6212a3982,3cf35bfe-d881-4881-a268-b5fa0607bf89,bc8d3e4a-8ba0-4cab-be1c-b20ae3af91e4,99306c29-bcef-49d5-bcd1-78d4e541e197,d0fcfaf0-7c6d-4ead-915b-6d9d9130e747,59c68d10-55c1-4629-b05a-bedde7f343b5,d86ee090-3a2e-44d0-abed-6f0f40a38c08,25cf07b8-606f-4448-99ef-ea87ca3097ba,1ca222ed-cc79-4e78-9ca3-3547e7b37e3a,63e4e0d4-0cb1-442e-8316-8c12a6e3ec14,0eb6fd6d-eaaa-46db-ab16-02e610f238a9,99b22b39-30d5-40dc-82d9-c1cd3c793a44,8c342edc-619b-4079-a28c-b7129f3670a6,e8991792-c875-4f40-99a0-54c902d42a52,54349c88-e80a-4728-b181-305f97905d59,f93b0850-2c60-479d-831b-b8c8f9493d61,c63a4306-4003-45a6-8be1-d669457473a9,d3cbb579-70a5-46e8-95f6-72cc9afa4ba3,3852d050-c009-4e33-b953-fd734a5b5e73,5dc23430-dfbb-4794-9f77-0d4dd6093033,67e76447-f16a-4a65-aba7-fc7166fd9b63,a783e7a3-cecc-4d15-b18c-3f2ddfc2dc6c,62e9d35f-153d-4744-b97c-9abfb2a90b1e,04b6e0c8-2835-415e-865d-aca9f1cd3c57,7a283823-00ee-489d-8dbc-356e2857526f,e89e00c9-29e9-420c-813d-e06ddc39d585,51433193-b0c3-4725-a3a8-9dd15e969b0c,48c89cd1-0ecf-4738-8fcc-392269585bd3,92ea478e-6bf8-4ffb-b88a-f57ccf5f2346,36c6e780-d4a8-47b2-8a36-90ee4b97607f,5fb6812f-9816-4d5d-b7e0-4ec3e9df5dc1,c29c95d1-5b5e-4f53-9ec5-83e1f72b5aa5,b81c059a-1137-47cf-bab8-ab5eb8ff34d5,886864a6-a0e4-468f-aaf4-26cf8dc5e541" + } + }, + "evaluation_timestamp": "1776696116.0", + "metric_config": { + "evaluation_description": "criteria_met_rate", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.3943696724988784, + "uncertainty": { + "standard_deviation": 0.034674948869492205, + "num_samples": 318 + }, + "details": { + "total_matched_trajectories": "318", + "stopping_reason_count_repetition_guard": "318", + "avg_total_tokens_target_model": "21254.46", + "total_total_tokens_target_model": "6758919.0", + "avg_total_tokens_other_models": "68935.17", + "total_total_tokens_other_models": "21921383.0", + "avg_total_tokens_all_models": "90189.63", + "total_total_tokens_all_models": "28680302.0", + "avg_turn_count": "25.64", + "total_turn_count": "8155.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "3600", + "attempt_timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + }, + { + "evaluation_name": "mean on healthbench/S-adaptive/6ce2ac11 for scorer healthbench_score", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "66", + "sample_ids": "f712eb09-82af-4fe8-800d-bc3af772a9ba,8ee6a990-8966-4b42-baa8-6c7e9f74ce74,267bda16-fd4b-4a9e-b3ac-5602c2532eed,0e819a9c-851d-4a7a-9263-62cfc8ce1b48,328c55cf-0b57-4992-99b1-f6a21af62ade,7b415a2f-22b0-4735-a38e-567f1891acc5,c49ac752-3693-4be1-ab4d-91fbdd02f827,a079d928-7b78-4a85-a72b-4274d7fc2914,a20cbbf3-330b-46d9-a35b-92e89a022282,be6150a5-8060-4434-9724-afede307d495,3a896889-9029-4ef8-8684-ca378da4702b,7074bb7a-a595-434d-883e-95ef6f5f6667,9f7b555a-fc30-40c3-88d3-f2ece06454a2,a6204079-9e72-42ad-8f39-a65552fa2e68,c0d80b93-b2d5-4129-b998-b1312f98e2a8,3dc41419-2565-428e-b9b9-02e3fd93c470,a91a22f8-d2e6-4593-bcae-6d479ed38031,d0d3426f-8b7a-4501-adb4-86d8fa9099c8,a7dac6a9-eab5-4c01-882c-dbf6c774d7c3,c1240427-ae38-4335-9e80-492c13ef4893,b7ddde65-702e-4255-b1d7-09d01248e76d,00d34549-134c-48e2-b4bb-b54946da70cc,85c11e2a-628b-464f-a8bb-e97ab04a73eb,69a544f8-957d-4dda-a777-608ec7b95617,90789a49-ab1f-4229-bd89-e1c56ea8be3a,cd53e21a-e5ed-4d9b-a258-42a23e9715bb,8d409c7b-29d2-4df2-aa69-ab56c9339bc5,6566f489-f254-4bca-a86b-90f5ac7471d7,d9a44a36-6a3c-4d5d-9c66-b2c0afb479d1,9cf87956-37e6-4bed-9cea-a64b13ae0de5,1dc0b953-ef66-4a78-9989-7cf1053b5b41,a1a08e1d-1353-4d6f-94ce-f90872074c7c,51d9c72b-ba60-47c5-9cc9-3ed6212a3982,3cf35bfe-d881-4881-a268-b5fa0607bf89,bc8d3e4a-8ba0-4cab-be1c-b20ae3af91e4,99306c29-bcef-49d5-bcd1-78d4e541e197,d0fcfaf0-7c6d-4ead-915b-6d9d9130e747,59c68d10-55c1-4629-b05a-bedde7f343b5,d86ee090-3a2e-44d0-abed-6f0f40a38c08,25cf07b8-606f-4448-99ef-ea87ca3097ba,1ca222ed-cc79-4e78-9ca3-3547e7b37e3a,63e4e0d4-0cb1-442e-8316-8c12a6e3ec14,0eb6fd6d-eaaa-46db-ab16-02e610f238a9,99b22b39-30d5-40dc-82d9-c1cd3c793a44,8c342edc-619b-4079-a28c-b7129f3670a6,e8991792-c875-4f40-99a0-54c902d42a52,54349c88-e80a-4728-b181-305f97905d59,f93b0850-2c60-479d-831b-b8c8f9493d61,c63a4306-4003-45a6-8be1-d669457473a9,d3cbb579-70a5-46e8-95f6-72cc9afa4ba3,3852d050-c009-4e33-b953-fd734a5b5e73,5dc23430-dfbb-4794-9f77-0d4dd6093033,67e76447-f16a-4a65-aba7-fc7166fd9b63,a783e7a3-cecc-4d15-b18c-3f2ddfc2dc6c,62e9d35f-153d-4744-b97c-9abfb2a90b1e,04b6e0c8-2835-415e-865d-aca9f1cd3c57,7a283823-00ee-489d-8dbc-356e2857526f,e89e00c9-29e9-420c-813d-e06ddc39d585,51433193-b0c3-4725-a3a8-9dd15e969b0c,48c89cd1-0ecf-4738-8fcc-392269585bd3,92ea478e-6bf8-4ffb-b88a-f57ccf5f2346,36c6e780-d4a8-47b2-8a36-90ee4b97607f,5fb6812f-9816-4d5d-b7e0-4ec3e9df5dc1,c29c95d1-5b5e-4f53-9ec5-83e1f72b5aa5,b81c059a-1137-47cf-bab8-ab5eb8ff34d5,886864a6-a0e4-468f-aaf4-26cf8dc5e541" + } + }, + "evaluation_timestamp": "1776696116.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.2416969742523121, + "uncertainty": { + "num_samples": 318 + }, + "details": { + "total_matched_trajectories": "318", + "stopping_reason_count_repetition_guard": "318", + "avg_total_tokens_target_model": "21254.46", + "total_total_tokens_target_model": "6758919.0", + "avg_total_tokens_other_models": "68935.17", + "total_total_tokens_other_models": "21921383.0", + "avg_total_tokens_all_models": "90189.63", + "total_total_tokens_all_models": "28680302.0", + "avg_turn_count": "25.64", + "total_turn_count": "8155.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "3600", + "attempt_timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/openai/gpt-5-2025-08-07/918416be-dff3-4a91-921b-676c24034076_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "871fe4821c71ec17d156d2ed3ca00f7adcb05b0751f8fbc4bee1439ac476ffe1", + "total_rows": 318 + } +} \ No newline at end of file diff --git a/data/healthbench/openai/gpt-5-2025-08-07/918416be-dff3-4a91-921b-676c24034076_samples.jsonl b/data/healthbench/openai/gpt-5-2025-08-07/918416be-dff3-4a91-921b-676c24034076_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..a4c37324ca4a9c1c85fb697e5ba82460730fdb4f --- /dev/null +++ b/data/healthbench/openai/gpt-5-2025-08-07/918416be-dff3-4a91-921b-676c24034076_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:871fe4821c71ec17d156d2ed3ca00f7adcb05b0751f8fbc4bee1439ac476ffe1 +size 22190058 diff --git a/data/healthbench/openai/gpt-5-2025-08-07/9992ce29-5674-46bb-8ae8-5efcc3deb208.json b/data/healthbench/openai/gpt-5-2025-08-07/9992ce29-5674-46bb-8ae8-5efcc3deb208.json new file mode 100644 index 0000000000000000000000000000000000000000..15c0c9d6fbf1c8af7bc5d2c29f281d4d0f251997 --- /dev/null +++ b/data/healthbench/openai/gpt-5-2025-08-07/9992ce29-5674-46bb-8ae8-5efcc3deb208.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/openai_gpt-5-2025-08-07/1786166583.552322", + "evaluation_timestamp": "1776696116.0", + "retrieved_timestamp": "1786166583.552322", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5-2025-08-07", + "id": "openai/gpt-5-2025-08-07", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/openai/gpt-5-2025-08-07/9992ce29-5674-46bb-8ae8-5efcc3deb208_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "b77f1d2b975c073007aac7e14f63183de8f8081620b83bc53cece14e09fad732", + "total_rows": 302 + } +} \ No newline at end of file diff --git a/data/healthbench/openai/gpt-5-2025-08-07/9992ce29-5674-46bb-8ae8-5efcc3deb208_samples.jsonl b/data/healthbench/openai/gpt-5-2025-08-07/9992ce29-5674-46bb-8ae8-5efcc3deb208_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..a311db007dfb6a5a6b5873449a5c3f0c6203c931 --- /dev/null +++ b/data/healthbench/openai/gpt-5-2025-08-07/9992ce29-5674-46bb-8ae8-5efcc3deb208_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b77f1d2b975c073007aac7e14f63183de8f8081620b83bc53cece14e09fad732 +size 22238991 diff --git a/data/healthbench/openai/gpt-5-2025-08-07/9daea19a-2cd6-4ac9-ab41-782c5cd48a48.json b/data/healthbench/openai/gpt-5-2025-08-07/9daea19a-2cd6-4ac9-ab41-782c5cd48a48.json new file mode 100644 index 0000000000000000000000000000000000000000..60d57f4ffad3b0cd57fe05bb13a5e601a4956e85 --- /dev/null +++ b/data/healthbench/openai/gpt-5-2025-08-07/9daea19a-2cd6-4ac9-ab41-782c5cd48a48.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/openai_gpt-5-2025-08-07/1786168532.989595", + "evaluation_timestamp": "1775733995.0", + "retrieved_timestamp": "1786168532.989595", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5-2025-08-07", + "id": "openai/gpt-5-2025-08-07", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/openai/gpt-5-2025-08-07/9daea19a-2cd6-4ac9-ab41-782c5cd48a48_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "4803bd4d131dae180db002195e5f4c0bdaf3f88e3bf583f4051acd86ae96e71b", + "total_rows": 170 + } +} \ No newline at end of file diff --git a/data/healthbench/openai/gpt-5-2025-08-07/9daea19a-2cd6-4ac9-ab41-782c5cd48a48_samples.jsonl b/data/healthbench/openai/gpt-5-2025-08-07/9daea19a-2cd6-4ac9-ab41-782c5cd48a48_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..cceeace08743fed29b909749927118dbb9586c06 --- /dev/null +++ b/data/healthbench/openai/gpt-5-2025-08-07/9daea19a-2cd6-4ac9-ab41-782c5cd48a48_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4803bd4d131dae180db002195e5f4c0bdaf3f88e3bf583f4051acd86ae96e71b +size 11794730 diff --git a/data/healthbench/openai/gpt-5-2025-08-07/a2070f7b-f111-40ab-8a96-7981769954ac.json b/data/healthbench/openai/gpt-5-2025-08-07/a2070f7b-f111-40ab-8a96-7981769954ac.json new file mode 100644 index 0000000000000000000000000000000000000000..681760d57983c75b6444d46362e13fa6a603cce2 --- /dev/null +++ b/data/healthbench/openai/gpt-5-2025-08-07/a2070f7b-f111-40ab-8a96-7981769954ac.json @@ -0,0 +1,320 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/openai_gpt-5-2025-08-07/1786163863.563298", + "evaluation_timestamp": "1776696116.0", + "retrieved_timestamp": "1786163863.563298", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5-2025-08-07", + "id": "openai/gpt-5-2025-08-07", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "bootstrap_score on healthbench/S-adaptive/+1ep/93232a3f for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "14", + "sample_ids": "b97b1bae-b09f-4073-8e16-540bba766c60,a49eb9af-926e-4f7c-94d0-133729059116,1049130c-e9c9-461d-b080-90027bc011c0,13bb02e4-844e-4972-9200-2befa5f911d8,2694a753-2f57-41eb-a7f9-8c4c7acdd2e8,5e0a84d7-b972-484c-a649-469b15f0a90f,c7606b1b-ca4d-4a71-8277-bcddff09a653,20cdebb7-9189-4f09-ad21-bfc8e8b75680,f24935ac-0530-4063-b2f1-ac45cd265037,1b8f338a-2e98-4a42-ba80-cba266f4b9fa,6f827e79-05c4-4383-8ed4-16d36260f29f,c835bebe-e998-4dcd-843a-2de03d4aabc6,70afd1ac-85ea-4455-a5a2-15b35378e10a,aa7a760a-5173-4a39-af83-48d6ae2230f0" + } + }, + "evaluation_timestamp": "1776696116.0", + "metric_config": { + "evaluation_description": "bootstrap_score", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.3764351460224393, + "uncertainty": { + "standard_deviation": 0.0700006890420607, + "num_samples": 12 + }, + "details": { + "total_matched_trajectories": "12", + "stopping_reason_count_repetition_guard": "12", + "avg_total_tokens_target_model": "19781.75", + "total_total_tokens_target_model": "237381.0", + "avg_total_tokens_other_models": "71631.08", + "total_total_tokens_other_models": "859573.0", + "avg_total_tokens_all_models": "91412.83", + "total_total_tokens_all_models": "1096954.0", + "avg_turn_count": "29.33", + "total_turn_count": "352.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "3600", + "attempt_timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + }, + { + "evaluation_name": "std on healthbench/S-adaptive/+1ep/93232a3f for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "14", + "sample_ids": "b97b1bae-b09f-4073-8e16-540bba766c60,a49eb9af-926e-4f7c-94d0-133729059116,1049130c-e9c9-461d-b080-90027bc011c0,13bb02e4-844e-4972-9200-2befa5f911d8,2694a753-2f57-41eb-a7f9-8c4c7acdd2e8,5e0a84d7-b972-484c-a649-469b15f0a90f,c7606b1b-ca4d-4a71-8277-bcddff09a653,20cdebb7-9189-4f09-ad21-bfc8e8b75680,f24935ac-0530-4063-b2f1-ac45cd265037,1b8f338a-2e98-4a42-ba80-cba266f4b9fa,6f827e79-05c4-4383-8ed4-16d36260f29f,c835bebe-e998-4dcd-843a-2de03d4aabc6,70afd1ac-85ea-4455-a5a2-15b35378e10a,aa7a760a-5173-4a39-af83-48d6ae2230f0" + } + }, + "evaluation_timestamp": "1776696116.0", + "metric_config": { + "evaluation_description": "std", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.0700006890420607, + "uncertainty": { + "standard_deviation": 0.0700006890420607, + "num_samples": 12 + }, + "details": { + "total_matched_trajectories": "12", + "stopping_reason_count_repetition_guard": "12", + "avg_total_tokens_target_model": "19781.75", + "total_total_tokens_target_model": "237381.0", + "avg_total_tokens_other_models": "71631.08", + "total_total_tokens_other_models": "859573.0", + "avg_total_tokens_all_models": "91412.83", + "total_total_tokens_all_models": "1096954.0", + "avg_turn_count": "29.33", + "total_turn_count": "352.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "3600", + "attempt_timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + }, + { + "evaluation_name": "criteria_met_rate on healthbench/S-adaptive/+1ep/93232a3f for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "14", + "sample_ids": "b97b1bae-b09f-4073-8e16-540bba766c60,a49eb9af-926e-4f7c-94d0-133729059116,1049130c-e9c9-461d-b080-90027bc011c0,13bb02e4-844e-4972-9200-2befa5f911d8,2694a753-2f57-41eb-a7f9-8c4c7acdd2e8,5e0a84d7-b972-484c-a649-469b15f0a90f,c7606b1b-ca4d-4a71-8277-bcddff09a653,20cdebb7-9189-4f09-ad21-bfc8e8b75680,f24935ac-0530-4063-b2f1-ac45cd265037,1b8f338a-2e98-4a42-ba80-cba266f4b9fa,6f827e79-05c4-4383-8ed4-16d36260f29f,c835bebe-e998-4dcd-843a-2de03d4aabc6,70afd1ac-85ea-4455-a5a2-15b35378e10a,aa7a760a-5173-4a39-af83-48d6ae2230f0" + } + }, + "evaluation_timestamp": "1776696116.0", + "metric_config": { + "evaluation_description": "criteria_met_rate", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.3924050632911392, + "uncertainty": { + "standard_deviation": 0.0700006890420607, + "num_samples": 12 + }, + "details": { + "total_matched_trajectories": "12", + "stopping_reason_count_repetition_guard": "12", + "avg_total_tokens_target_model": "19781.75", + "total_total_tokens_target_model": "237381.0", + "avg_total_tokens_other_models": "71631.08", + "total_total_tokens_other_models": "859573.0", + "avg_total_tokens_all_models": "91412.83", + "total_total_tokens_all_models": "1096954.0", + "avg_turn_count": "29.33", + "total_turn_count": "352.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "3600", + "attempt_timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + }, + { + "evaluation_name": "mean on healthbench/S-adaptive/+1ep/93232a3f for scorer healthbench_score", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "14", + "sample_ids": "b97b1bae-b09f-4073-8e16-540bba766c60,a49eb9af-926e-4f7c-94d0-133729059116,1049130c-e9c9-461d-b080-90027bc011c0,13bb02e4-844e-4972-9200-2befa5f911d8,2694a753-2f57-41eb-a7f9-8c4c7acdd2e8,5e0a84d7-b972-484c-a649-469b15f0a90f,c7606b1b-ca4d-4a71-8277-bcddff09a653,20cdebb7-9189-4f09-ad21-bfc8e8b75680,f24935ac-0530-4063-b2f1-ac45cd265037,1b8f338a-2e98-4a42-ba80-cba266f4b9fa,6f827e79-05c4-4383-8ed4-16d36260f29f,c835bebe-e998-4dcd-843a-2de03d4aabc6,70afd1ac-85ea-4455-a5a2-15b35378e10a,aa7a760a-5173-4a39-af83-48d6ae2230f0" + } + }, + "evaluation_timestamp": "1776696116.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.3764351460224393, + "uncertainty": { + "num_samples": 12 + }, + "details": { + "total_matched_trajectories": "12", + "stopping_reason_count_repetition_guard": "12", + "avg_total_tokens_target_model": "19781.75", + "total_total_tokens_target_model": "237381.0", + "avg_total_tokens_other_models": "71631.08", + "total_total_tokens_other_models": "859573.0", + "avg_total_tokens_all_models": "91412.83", + "total_total_tokens_all_models": "1096954.0", + "avg_turn_count": "29.33", + "total_turn_count": "352.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "3600", + "attempt_timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/openai/gpt-5-2025-08-07/a2070f7b-f111-40ab-8a96-7981769954ac_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "026b83c18c8e258889c57b0a9d280e7f777752562be5fe96ddd6cd3ca988c44a", + "total_rows": 12 + } +} \ No newline at end of file diff --git a/data/healthbench/openai/gpt-5-2025-08-07/a2070f7b-f111-40ab-8a96-7981769954ac_samples.jsonl b/data/healthbench/openai/gpt-5-2025-08-07/a2070f7b-f111-40ab-8a96-7981769954ac_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..c94ea4e8c1a4a691ab5b28948262f10d81388d8f --- /dev/null +++ b/data/healthbench/openai/gpt-5-2025-08-07/a2070f7b-f111-40ab-8a96-7981769954ac_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:026b83c18c8e258889c57b0a9d280e7f777752562be5fe96ddd6cd3ca988c44a +size 871478 diff --git a/data/healthbench/openai/gpt-5-2025-08-07/aaee40de-ecb9-478f-ac2b-128af6513d0d.json b/data/healthbench/openai/gpt-5-2025-08-07/aaee40de-ecb9-478f-ac2b-128af6513d0d.json new file mode 100644 index 0000000000000000000000000000000000000000..b0ef123b9186ce7f2375f8a4abd4c7e8d5a8690e --- /dev/null +++ b/data/healthbench/openai/gpt-5-2025-08-07/aaee40de-ecb9-478f-ac2b-128af6513d0d.json @@ -0,0 +1,304 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/openai_gpt-5-2025-08-07/1786163886.946565", + "evaluation_timestamp": "1776263448.0", + "retrieved_timestamp": "1786163886.946565", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5-2025-08-07", + "id": "openai/gpt-5-2025-08-07", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "bootstrap_score on healthbench/S-adaptive/+4ep/a56f7cfb for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "69", + "sample_ids": "3bb4a735-ace7-4342-b7a5-0bae083e6f82,66b448bf-2e49-488d-9d0b-f9912ab3f0f8,b12d2453-fd3b-462c-b9cc-554ff16bfd3f,0a548d04-0973-4343-a3bf-07f730bed2f6,e2cd9f23-e3dd-4b67-8984-a5412dd802e5,1fb6ab91-a999-437d-8c93-c5937c0a89a3,42af0c9f-a715-45c4-ab24-c2948f6f943d,a94aad40-387e-4a37-b92f-61badd57388b,596ba714-29f7-447e-aa73-88d46ecd6e88,1233737e-2674-41d3-8519-b817aeb0b32e,c80a2a84-281d-41cc-a10b-32ee48c584f6,bf7718f2-bceb-4702-8d74-a7fc1bd7fb78,b7a82324-73ee-4996-8f71-8f455f789e90,22fe3eec-03b0-4cdf-a06a-87e71d236082,6c568ea6-86bd-4286-94ba-e3ccc8b67b21,905949d2-7a0a-4461-8f4b-257de6be6eed,c653b6eb-f9b9-4626-b702-ff73d71e8253,cf09b209-5434-4317-aa2a-7b07537b475e,77c0be2e-364a-40de-ae21-a277154e3d60,af750930-5fa5-4efe-8f85-b499641377d4,af62e0b8-bbd0-4dcc-b55c-cb4c7bbcfb2c,f3729e05-61ef-4306-9491-6f5659d4eef9,126236de-9e7a-4814-b5d7-98cbe89b33b2,9bb31d40-c155-4213-868d-103fa0cc7c56,97bec88b-6ef7-4f24-84d6-a2bcc592548b,779b5d14-1a6f-4d0a-8763-e0a25961a811,62b4d946-e447-4fb4-96ed-b6d36c8e2c62,d475913b-0b18-47e9-90b2-28449e1d6541,77dff7ec-2e02-4b93-86d6-5d7765b8dd3d,3a0eec54-fd86-47c1-883b-d8e828d2a327,b947fc6f-1d50-4e3f-b679-8484529b787b,e8e17de2-9b5f-4f20-a18e-fdd14a8b7192,b5b6d817-c524-4bef-badc-f93876657ea2,9ab66439-8090-4f6c-ba74-874591dfd1a6,8e447490-23cd-4d64-9e3a-b8ad46c76573,762d9eca-4835-4447-a649-f91740c74bd0,5c1cf475-6fff-42e2-8508-7e1a14ea9f65,06fb659f-a4aa-4a97-a831-ed2cd2aa7e69,608770a0-440d-4349-9a1c-863e9f4d3e24,a58987e9-8b59-461a-a68b-20efa1d37d51,2d878243-5d98-489a-b8f2-8809f813b6cf,452a9534-b4fc-4483-94e0-d34b06b8e299,e690c779-3cb8-4271-8541-a699bb1bd475,d984a2a8-1209-42bb-8f59-001421c96f8f,5caa2d1f-3ab9-4e25-97a2-af29c2dbbcfb,7b1d2983-ab41-48f2-bdcd-454cbf3f81b8,b267896f-cfc0-4bbe-9fda-fdfa287d1f43,ff514fc4-bd34-4a3f-be3f-840b0e1f8091,a131d5e7-2a27-426f-90f0-25eae3c4b490,620b9bf1-f0f3-4d68-aa74-1560fb406a68,9b4c14e9-c404-4646-87b7-d3badae8a68b,9385d26b-bf44-4ff7-9410-e181413540dd,a57a4b5e-0f12-4103-8852-72bd2b7dba24,cf8040e2-39e3-4200-ab7f-b9b532310ac3,4fc7b71f-9591-4ae1-bfca-2929c7225182,da05b57a-d776-45b6-96af-1750e0254eec,ef446d15-d019-4c5a-a052-ae8cc3ea3b81,0fe4ea94-633d-4f73-985b-cbcd9f5a0270,cbb0ca6e-aa5e-4683-b107-d7dfa4cb60ce,ae6ec2ce-5c23-4e65-802f-cef9f1624d71,cd3dbc9f-a39c-4d13-af60-cba7dbc3baad,02423ab9-f3ed-4096-a7df-bfa3ebb40c7c,084b0c27-79c5-42f6-9cbd-690d0e79167a,42465540-8af1-423b-9b35-a31de1e1d49e,74a50705-7db2-44be-84dc-92b2d960d344,3b79284e-1143-46b4-b6a0-78e5448ff773,afbfc79d-6455-40b8-b475-1bcde4ef0cc9,6d705f80-8259-4ca1-8645-163f42c85a3c,129057dd-4bf4-472d-a8cf-a063dbed60ad" + } + }, + "evaluation_timestamp": "1776263448.0", + "metric_config": { + "evaluation_description": "bootstrap_score", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.4248689122403643, + "uncertainty": { + "standard_deviation": 0.036781397484136745, + "num_samples": 276 + }, + "details": { + "total_matched_trajectories": "44", + "stopping_reason_count_repetition_guard": "44", + "avg_total_tokens_target_model": "23119.23", + "total_total_tokens_target_model": "1017246.0", + "avg_total_tokens_other_models": "66597.77", + "total_total_tokens_other_models": "2930302.0", + "avg_total_tokens_all_models": "89717.00", + "total_total_tokens_all_models": "3947548.0", + "avg_turn_count": "25.55", + "total_turn_count": "1124.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "reasoning_effort": "\"high\"" + } + } + }, + { + "evaluation_name": "std on healthbench/S-adaptive/+4ep/a56f7cfb for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "69", + "sample_ids": "3bb4a735-ace7-4342-b7a5-0bae083e6f82,66b448bf-2e49-488d-9d0b-f9912ab3f0f8,b12d2453-fd3b-462c-b9cc-554ff16bfd3f,0a548d04-0973-4343-a3bf-07f730bed2f6,e2cd9f23-e3dd-4b67-8984-a5412dd802e5,1fb6ab91-a999-437d-8c93-c5937c0a89a3,42af0c9f-a715-45c4-ab24-c2948f6f943d,a94aad40-387e-4a37-b92f-61badd57388b,596ba714-29f7-447e-aa73-88d46ecd6e88,1233737e-2674-41d3-8519-b817aeb0b32e,c80a2a84-281d-41cc-a10b-32ee48c584f6,bf7718f2-bceb-4702-8d74-a7fc1bd7fb78,b7a82324-73ee-4996-8f71-8f455f789e90,22fe3eec-03b0-4cdf-a06a-87e71d236082,6c568ea6-86bd-4286-94ba-e3ccc8b67b21,905949d2-7a0a-4461-8f4b-257de6be6eed,c653b6eb-f9b9-4626-b702-ff73d71e8253,cf09b209-5434-4317-aa2a-7b07537b475e,77c0be2e-364a-40de-ae21-a277154e3d60,af750930-5fa5-4efe-8f85-b499641377d4,af62e0b8-bbd0-4dcc-b55c-cb4c7bbcfb2c,f3729e05-61ef-4306-9491-6f5659d4eef9,126236de-9e7a-4814-b5d7-98cbe89b33b2,9bb31d40-c155-4213-868d-103fa0cc7c56,97bec88b-6ef7-4f24-84d6-a2bcc592548b,779b5d14-1a6f-4d0a-8763-e0a25961a811,62b4d946-e447-4fb4-96ed-b6d36c8e2c62,d475913b-0b18-47e9-90b2-28449e1d6541,77dff7ec-2e02-4b93-86d6-5d7765b8dd3d,3a0eec54-fd86-47c1-883b-d8e828d2a327,b947fc6f-1d50-4e3f-b679-8484529b787b,e8e17de2-9b5f-4f20-a18e-fdd14a8b7192,b5b6d817-c524-4bef-badc-f93876657ea2,9ab66439-8090-4f6c-ba74-874591dfd1a6,8e447490-23cd-4d64-9e3a-b8ad46c76573,762d9eca-4835-4447-a649-f91740c74bd0,5c1cf475-6fff-42e2-8508-7e1a14ea9f65,06fb659f-a4aa-4a97-a831-ed2cd2aa7e69,608770a0-440d-4349-9a1c-863e9f4d3e24,a58987e9-8b59-461a-a68b-20efa1d37d51,2d878243-5d98-489a-b8f2-8809f813b6cf,452a9534-b4fc-4483-94e0-d34b06b8e299,e690c779-3cb8-4271-8541-a699bb1bd475,d984a2a8-1209-42bb-8f59-001421c96f8f,5caa2d1f-3ab9-4e25-97a2-af29c2dbbcfb,7b1d2983-ab41-48f2-bdcd-454cbf3f81b8,b267896f-cfc0-4bbe-9fda-fdfa287d1f43,ff514fc4-bd34-4a3f-be3f-840b0e1f8091,a131d5e7-2a27-426f-90f0-25eae3c4b490,620b9bf1-f0f3-4d68-aa74-1560fb406a68,9b4c14e9-c404-4646-87b7-d3badae8a68b,9385d26b-bf44-4ff7-9410-e181413540dd,a57a4b5e-0f12-4103-8852-72bd2b7dba24,cf8040e2-39e3-4200-ab7f-b9b532310ac3,4fc7b71f-9591-4ae1-bfca-2929c7225182,da05b57a-d776-45b6-96af-1750e0254eec,ef446d15-d019-4c5a-a052-ae8cc3ea3b81,0fe4ea94-633d-4f73-985b-cbcd9f5a0270,cbb0ca6e-aa5e-4683-b107-d7dfa4cb60ce,ae6ec2ce-5c23-4e65-802f-cef9f1624d71,cd3dbc9f-a39c-4d13-af60-cba7dbc3baad,02423ab9-f3ed-4096-a7df-bfa3ebb40c7c,084b0c27-79c5-42f6-9cbd-690d0e79167a,42465540-8af1-423b-9b35-a31de1e1d49e,74a50705-7db2-44be-84dc-92b2d960d344,3b79284e-1143-46b4-b6a0-78e5448ff773,afbfc79d-6455-40b8-b475-1bcde4ef0cc9,6d705f80-8259-4ca1-8645-163f42c85a3c,129057dd-4bf4-472d-a8cf-a063dbed60ad" + } + }, + "evaluation_timestamp": "1776263448.0", + "metric_config": { + "evaluation_description": "std", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.036781397484136745, + "uncertainty": { + "standard_deviation": 0.036781397484136745, + "num_samples": 276 + }, + "details": { + "total_matched_trajectories": "44", + "stopping_reason_count_repetition_guard": "44", + "avg_total_tokens_target_model": "23119.23", + "total_total_tokens_target_model": "1017246.0", + "avg_total_tokens_other_models": "66597.77", + "total_total_tokens_other_models": "2930302.0", + "avg_total_tokens_all_models": "89717.00", + "total_total_tokens_all_models": "3947548.0", + "avg_turn_count": "25.55", + "total_turn_count": "1124.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "reasoning_effort": "\"high\"" + } + } + }, + { + "evaluation_name": "criteria_met_rate on healthbench/S-adaptive/+4ep/a56f7cfb for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "69", + "sample_ids": "3bb4a735-ace7-4342-b7a5-0bae083e6f82,66b448bf-2e49-488d-9d0b-f9912ab3f0f8,b12d2453-fd3b-462c-b9cc-554ff16bfd3f,0a548d04-0973-4343-a3bf-07f730bed2f6,e2cd9f23-e3dd-4b67-8984-a5412dd802e5,1fb6ab91-a999-437d-8c93-c5937c0a89a3,42af0c9f-a715-45c4-ab24-c2948f6f943d,a94aad40-387e-4a37-b92f-61badd57388b,596ba714-29f7-447e-aa73-88d46ecd6e88,1233737e-2674-41d3-8519-b817aeb0b32e,c80a2a84-281d-41cc-a10b-32ee48c584f6,bf7718f2-bceb-4702-8d74-a7fc1bd7fb78,b7a82324-73ee-4996-8f71-8f455f789e90,22fe3eec-03b0-4cdf-a06a-87e71d236082,6c568ea6-86bd-4286-94ba-e3ccc8b67b21,905949d2-7a0a-4461-8f4b-257de6be6eed,c653b6eb-f9b9-4626-b702-ff73d71e8253,cf09b209-5434-4317-aa2a-7b07537b475e,77c0be2e-364a-40de-ae21-a277154e3d60,af750930-5fa5-4efe-8f85-b499641377d4,af62e0b8-bbd0-4dcc-b55c-cb4c7bbcfb2c,f3729e05-61ef-4306-9491-6f5659d4eef9,126236de-9e7a-4814-b5d7-98cbe89b33b2,9bb31d40-c155-4213-868d-103fa0cc7c56,97bec88b-6ef7-4f24-84d6-a2bcc592548b,779b5d14-1a6f-4d0a-8763-e0a25961a811,62b4d946-e447-4fb4-96ed-b6d36c8e2c62,d475913b-0b18-47e9-90b2-28449e1d6541,77dff7ec-2e02-4b93-86d6-5d7765b8dd3d,3a0eec54-fd86-47c1-883b-d8e828d2a327,b947fc6f-1d50-4e3f-b679-8484529b787b,e8e17de2-9b5f-4f20-a18e-fdd14a8b7192,b5b6d817-c524-4bef-badc-f93876657ea2,9ab66439-8090-4f6c-ba74-874591dfd1a6,8e447490-23cd-4d64-9e3a-b8ad46c76573,762d9eca-4835-4447-a649-f91740c74bd0,5c1cf475-6fff-42e2-8508-7e1a14ea9f65,06fb659f-a4aa-4a97-a831-ed2cd2aa7e69,608770a0-440d-4349-9a1c-863e9f4d3e24,a58987e9-8b59-461a-a68b-20efa1d37d51,2d878243-5d98-489a-b8f2-8809f813b6cf,452a9534-b4fc-4483-94e0-d34b06b8e299,e690c779-3cb8-4271-8541-a699bb1bd475,d984a2a8-1209-42bb-8f59-001421c96f8f,5caa2d1f-3ab9-4e25-97a2-af29c2dbbcfb,7b1d2983-ab41-48f2-bdcd-454cbf3f81b8,b267896f-cfc0-4bbe-9fda-fdfa287d1f43,ff514fc4-bd34-4a3f-be3f-840b0e1f8091,a131d5e7-2a27-426f-90f0-25eae3c4b490,620b9bf1-f0f3-4d68-aa74-1560fb406a68,9b4c14e9-c404-4646-87b7-d3badae8a68b,9385d26b-bf44-4ff7-9410-e181413540dd,a57a4b5e-0f12-4103-8852-72bd2b7dba24,cf8040e2-39e3-4200-ab7f-b9b532310ac3,4fc7b71f-9591-4ae1-bfca-2929c7225182,da05b57a-d776-45b6-96af-1750e0254eec,ef446d15-d019-4c5a-a052-ae8cc3ea3b81,0fe4ea94-633d-4f73-985b-cbcd9f5a0270,cbb0ca6e-aa5e-4683-b107-d7dfa4cb60ce,ae6ec2ce-5c23-4e65-802f-cef9f1624d71,cd3dbc9f-a39c-4d13-af60-cba7dbc3baad,02423ab9-f3ed-4096-a7df-bfa3ebb40c7c,084b0c27-79c5-42f6-9cbd-690d0e79167a,42465540-8af1-423b-9b35-a31de1e1d49e,74a50705-7db2-44be-84dc-92b2d960d344,3b79284e-1143-46b4-b6a0-78e5448ff773,afbfc79d-6455-40b8-b475-1bcde4ef0cc9,6d705f80-8259-4ca1-8645-163f42c85a3c,129057dd-4bf4-472d-a8cf-a063dbed60ad" + } + }, + "evaluation_timestamp": "1776263448.0", + "metric_config": { + "evaluation_description": "criteria_met_rate", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.48166023166023164, + "uncertainty": { + "standard_deviation": 0.036781397484136745, + "num_samples": 276 + }, + "details": { + "total_matched_trajectories": "44", + "stopping_reason_count_repetition_guard": "44", + "avg_total_tokens_target_model": "23119.23", + "total_total_tokens_target_model": "1017246.0", + "avg_total_tokens_other_models": "66597.77", + "total_total_tokens_other_models": "2930302.0", + "avg_total_tokens_all_models": "89717.00", + "total_total_tokens_all_models": "3947548.0", + "avg_turn_count": "25.55", + "total_turn_count": "1124.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "reasoning_effort": "\"high\"" + } + } + }, + { + "evaluation_name": "mean on healthbench/S-adaptive/+4ep/a56f7cfb for scorer healthbench_score", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "69", + "sample_ids": "3bb4a735-ace7-4342-b7a5-0bae083e6f82,66b448bf-2e49-488d-9d0b-f9912ab3f0f8,b12d2453-fd3b-462c-b9cc-554ff16bfd3f,0a548d04-0973-4343-a3bf-07f730bed2f6,e2cd9f23-e3dd-4b67-8984-a5412dd802e5,1fb6ab91-a999-437d-8c93-c5937c0a89a3,42af0c9f-a715-45c4-ab24-c2948f6f943d,a94aad40-387e-4a37-b92f-61badd57388b,596ba714-29f7-447e-aa73-88d46ecd6e88,1233737e-2674-41d3-8519-b817aeb0b32e,c80a2a84-281d-41cc-a10b-32ee48c584f6,bf7718f2-bceb-4702-8d74-a7fc1bd7fb78,b7a82324-73ee-4996-8f71-8f455f789e90,22fe3eec-03b0-4cdf-a06a-87e71d236082,6c568ea6-86bd-4286-94ba-e3ccc8b67b21,905949d2-7a0a-4461-8f4b-257de6be6eed,c653b6eb-f9b9-4626-b702-ff73d71e8253,cf09b209-5434-4317-aa2a-7b07537b475e,77c0be2e-364a-40de-ae21-a277154e3d60,af750930-5fa5-4efe-8f85-b499641377d4,af62e0b8-bbd0-4dcc-b55c-cb4c7bbcfb2c,f3729e05-61ef-4306-9491-6f5659d4eef9,126236de-9e7a-4814-b5d7-98cbe89b33b2,9bb31d40-c155-4213-868d-103fa0cc7c56,97bec88b-6ef7-4f24-84d6-a2bcc592548b,779b5d14-1a6f-4d0a-8763-e0a25961a811,62b4d946-e447-4fb4-96ed-b6d36c8e2c62,d475913b-0b18-47e9-90b2-28449e1d6541,77dff7ec-2e02-4b93-86d6-5d7765b8dd3d,3a0eec54-fd86-47c1-883b-d8e828d2a327,b947fc6f-1d50-4e3f-b679-8484529b787b,e8e17de2-9b5f-4f20-a18e-fdd14a8b7192,b5b6d817-c524-4bef-badc-f93876657ea2,9ab66439-8090-4f6c-ba74-874591dfd1a6,8e447490-23cd-4d64-9e3a-b8ad46c76573,762d9eca-4835-4447-a649-f91740c74bd0,5c1cf475-6fff-42e2-8508-7e1a14ea9f65,06fb659f-a4aa-4a97-a831-ed2cd2aa7e69,608770a0-440d-4349-9a1c-863e9f4d3e24,a58987e9-8b59-461a-a68b-20efa1d37d51,2d878243-5d98-489a-b8f2-8809f813b6cf,452a9534-b4fc-4483-94e0-d34b06b8e299,e690c779-3cb8-4271-8541-a699bb1bd475,d984a2a8-1209-42bb-8f59-001421c96f8f,5caa2d1f-3ab9-4e25-97a2-af29c2dbbcfb,7b1d2983-ab41-48f2-bdcd-454cbf3f81b8,b267896f-cfc0-4bbe-9fda-fdfa287d1f43,ff514fc4-bd34-4a3f-be3f-840b0e1f8091,a131d5e7-2a27-426f-90f0-25eae3c4b490,620b9bf1-f0f3-4d68-aa74-1560fb406a68,9b4c14e9-c404-4646-87b7-d3badae8a68b,9385d26b-bf44-4ff7-9410-e181413540dd,a57a4b5e-0f12-4103-8852-72bd2b7dba24,cf8040e2-39e3-4200-ab7f-b9b532310ac3,4fc7b71f-9591-4ae1-bfca-2929c7225182,da05b57a-d776-45b6-96af-1750e0254eec,ef446d15-d019-4c5a-a052-ae8cc3ea3b81,0fe4ea94-633d-4f73-985b-cbcd9f5a0270,cbb0ca6e-aa5e-4683-b107-d7dfa4cb60ce,ae6ec2ce-5c23-4e65-802f-cef9f1624d71,cd3dbc9f-a39c-4d13-af60-cba7dbc3baad,02423ab9-f3ed-4096-a7df-bfa3ebb40c7c,084b0c27-79c5-42f6-9cbd-690d0e79167a,42465540-8af1-423b-9b35-a31de1e1d49e,74a50705-7db2-44be-84dc-92b2d960d344,3b79284e-1143-46b4-b6a0-78e5448ff773,afbfc79d-6455-40b8-b475-1bcde4ef0cc9,6d705f80-8259-4ca1-8645-163f42c85a3c,129057dd-4bf4-472d-a8cf-a063dbed60ad" + } + }, + "evaluation_timestamp": "1776263448.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.4248689122403643, + "uncertainty": { + "num_samples": 276 + }, + "details": { + "total_matched_trajectories": "44", + "stopping_reason_count_repetition_guard": "44", + "avg_total_tokens_target_model": "23119.23", + "total_total_tokens_target_model": "1017246.0", + "avg_total_tokens_other_models": "66597.77", + "total_total_tokens_other_models": "2930302.0", + "avg_total_tokens_all_models": "89717.00", + "total_total_tokens_all_models": "3947548.0", + "avg_turn_count": "25.55", + "total_turn_count": "1124.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/openai/gpt-5-2025-08-07/aaee40de-ecb9-478f-ac2b-128af6513d0d_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "dc5f7e706b43f4e53ef2a749ef49bb88d1ea3aa3731609967def71cb30506208", + "total_rows": 276 + } +} \ No newline at end of file diff --git a/data/healthbench/openai/gpt-5-2025-08-07/aaee40de-ecb9-478f-ac2b-128af6513d0d_samples.jsonl b/data/healthbench/openai/gpt-5-2025-08-07/aaee40de-ecb9-478f-ac2b-128af6513d0d_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..b62fdb41ff4ec43891bd8fee7ed61bea300eb9cb --- /dev/null +++ b/data/healthbench/openai/gpt-5-2025-08-07/aaee40de-ecb9-478f-ac2b-128af6513d0d_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dc5f7e706b43f4e53ef2a749ef49bb88d1ea3aa3731609967def71cb30506208 +size 17818123 diff --git a/data/healthbench/openai/gpt-5-2025-08-07/b5e9b430-43ea-460f-9cdd-2814fa0fb13f.json b/data/healthbench/openai/gpt-5-2025-08-07/b5e9b430-43ea-460f-9cdd-2814fa0fb13f.json new file mode 100644 index 0000000000000000000000000000000000000000..4f9b0918ba0aeda62e84e5cbe0bf51b1e00c2882 --- /dev/null +++ b/data/healthbench/openai/gpt-5-2025-08-07/b5e9b430-43ea-460f-9cdd-2814fa0fb13f.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/openai_gpt-5-2025-08-07/1786168495.280465", + "evaluation_timestamp": "1776682987.0", + "retrieved_timestamp": "1786168495.280465", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5-2025-08-07", + "id": "openai/gpt-5-2025-08-07", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/openai/gpt-5-2025-08-07/b5e9b430-43ea-460f-9cdd-2814fa0fb13f_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "3b443138454e9371eaf0e1ebe6b4da4e53f49b3a8a9c4eefb4c3a79b3d1a9693", + "total_rows": 33 + } +} \ No newline at end of file diff --git a/data/healthbench/openai/gpt-5-2025-08-07/b5e9b430-43ea-460f-9cdd-2814fa0fb13f_samples.jsonl b/data/healthbench/openai/gpt-5-2025-08-07/b5e9b430-43ea-460f-9cdd-2814fa0fb13f_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..a3816674685645e35421394dc0a26d0d12b53ee7 --- /dev/null +++ b/data/healthbench/openai/gpt-5-2025-08-07/b5e9b430-43ea-460f-9cdd-2814fa0fb13f_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3b443138454e9371eaf0e1ebe6b4da4e53f49b3a8a9c4eefb4c3a79b3d1a9693 +size 2404125 diff --git a/data/healthbench/openai/gpt-5-2025-08-07/ea59a3dc-1b63-4102-974c-bdd402b06f57.json b/data/healthbench/openai/gpt-5-2025-08-07/ea59a3dc-1b63-4102-974c-bdd402b06f57.json new file mode 100644 index 0000000000000000000000000000000000000000..7b86a237b3a022784372f35849fa24d6a1f95c7d --- /dev/null +++ b/data/healthbench/openai/gpt-5-2025-08-07/ea59a3dc-1b63-4102-974c-bdd402b06f57.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/openai_gpt-5-2025-08-07/1786166476.813519", + "evaluation_timestamp": "1776165242.0", + "retrieved_timestamp": "1786166476.813519", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5-2025-08-07", + "id": "openai/gpt-5-2025-08-07", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/openai/gpt-5-2025-08-07/ea59a3dc-1b63-4102-974c-bdd402b06f57_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "1e0542538207b026369acd2e91b6dcf68a24e73ec0ed3ad3f97263b2ab0938f5", + "total_rows": 100 + } +} \ No newline at end of file diff --git a/data/healthbench/openai/gpt-5-2025-08-07/ea59a3dc-1b63-4102-974c-bdd402b06f57_samples.jsonl b/data/healthbench/openai/gpt-5-2025-08-07/ea59a3dc-1b63-4102-974c-bdd402b06f57_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..d50c714b77eed81f6ac00828e8835ce73bf0c466 --- /dev/null +++ b/data/healthbench/openai/gpt-5-2025-08-07/ea59a3dc-1b63-4102-974c-bdd402b06f57_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1e0542538207b026369acd2e91b6dcf68a24e73ec0ed3ad3f97263b2ab0938f5 +size 6837156 diff --git a/data/healthbench/openai/gpt-5-2025-08-07/f01ed4c5-c9cf-4ae7-9814-835dbd5c1a89.json b/data/healthbench/openai/gpt-5-2025-08-07/f01ed4c5-c9cf-4ae7-9814-835dbd5c1a89.json new file mode 100644 index 0000000000000000000000000000000000000000..b1cecc5dab41fe48b3dd9f72fa0e06c87c7367c5 --- /dev/null +++ b/data/healthbench/openai/gpt-5-2025-08-07/f01ed4c5-c9cf-4ae7-9814-835dbd5c1a89.json @@ -0,0 +1,304 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/openai_gpt-5-2025-08-07/1786163905.452497", + "evaluation_timestamp": "1776165244.0", + "retrieved_timestamp": "1786163905.452497", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5-2025-08-07", + "id": "openai/gpt-5-2025-08-07", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "bootstrap_score on healthbench/S-adaptive/+8ep for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "5", + "sample_ids": "2833344d-bff9-4bbe-8fab-6b13d64c6b2a,48669cdc-1a64-4c99-a6e8-9a7181341fa2,be70739f-234a-4a0f-9b14-d15b8800577b,4047884e-4629-457c-9d77-130e937fae60,eb97bae4-430e-45cd-a065-2df3ab5c600e" + } + }, + "evaluation_timestamp": "1776165244.0", + "metric_config": { + "evaluation_description": "bootstrap_score", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.5447735981184926, + "uncertainty": { + "standard_deviation": 0.13199006663904367, + "num_samples": 40 + }, + "details": { + "total_matched_trajectories": "3", + "stopping_reason_count_repetition_guard": "3", + "avg_total_tokens_target_model": "23176.67", + "total_total_tokens_target_model": "69530.0", + "avg_total_tokens_other_models": "33958.00", + "total_total_tokens_other_models": "101874.0", + "avg_total_tokens_all_models": "57134.67", + "total_total_tokens_all_models": "171404.0", + "avg_turn_count": "15.00", + "total_turn_count": "45.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "reasoning_effort": "\"high\"" + } + } + }, + { + "evaluation_name": "std on healthbench/S-adaptive/+8ep for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "5", + "sample_ids": "2833344d-bff9-4bbe-8fab-6b13d64c6b2a,48669cdc-1a64-4c99-a6e8-9a7181341fa2,be70739f-234a-4a0f-9b14-d15b8800577b,4047884e-4629-457c-9d77-130e937fae60,eb97bae4-430e-45cd-a065-2df3ab5c600e" + } + }, + "evaluation_timestamp": "1776165244.0", + "metric_config": { + "evaluation_description": "std", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.13199006663904367, + "uncertainty": { + "standard_deviation": 0.13199006663904367, + "num_samples": 40 + }, + "details": { + "total_matched_trajectories": "3", + "stopping_reason_count_repetition_guard": "3", + "avg_total_tokens_target_model": "23176.67", + "total_total_tokens_target_model": "69530.0", + "avg_total_tokens_other_models": "33958.00", + "total_total_tokens_other_models": "101874.0", + "avg_total_tokens_all_models": "57134.67", + "total_total_tokens_all_models": "171404.0", + "avg_turn_count": "15.00", + "total_turn_count": "45.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "reasoning_effort": "\"high\"" + } + } + }, + { + "evaluation_name": "criteria_met_rate on healthbench/S-adaptive/+8ep for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "5", + "sample_ids": "2833344d-bff9-4bbe-8fab-6b13d64c6b2a,48669cdc-1a64-4c99-a6e8-9a7181341fa2,be70739f-234a-4a0f-9b14-d15b8800577b,4047884e-4629-457c-9d77-130e937fae60,eb97bae4-430e-45cd-a065-2df3ab5c600e" + } + }, + "evaluation_timestamp": "1776165244.0", + "metric_config": { + "evaluation_description": "criteria_met_rate", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.4847972972972973, + "uncertainty": { + "standard_deviation": 0.13199006663904367, + "num_samples": 40 + }, + "details": { + "total_matched_trajectories": "3", + "stopping_reason_count_repetition_guard": "3", + "avg_total_tokens_target_model": "23176.67", + "total_total_tokens_target_model": "69530.0", + "avg_total_tokens_other_models": "33958.00", + "total_total_tokens_other_models": "101874.0", + "avg_total_tokens_all_models": "57134.67", + "total_total_tokens_all_models": "171404.0", + "avg_turn_count": "15.00", + "total_turn_count": "45.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "reasoning_effort": "\"high\"" + } + } + }, + { + "evaluation_name": "mean on healthbench/S-adaptive/+8ep for scorer healthbench_score", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "5", + "sample_ids": "2833344d-bff9-4bbe-8fab-6b13d64c6b2a,48669cdc-1a64-4c99-a6e8-9a7181341fa2,be70739f-234a-4a0f-9b14-d15b8800577b,4047884e-4629-457c-9d77-130e937fae60,eb97bae4-430e-45cd-a065-2df3ab5c600e" + } + }, + "evaluation_timestamp": "1776165244.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.5447735981184926, + "uncertainty": { + "num_samples": 40 + }, + "details": { + "total_matched_trajectories": "3", + "stopping_reason_count_repetition_guard": "3", + "avg_total_tokens_target_model": "23176.67", + "total_total_tokens_target_model": "69530.0", + "avg_total_tokens_other_models": "33958.00", + "total_total_tokens_other_models": "101874.0", + "avg_total_tokens_all_models": "57134.67", + "total_total_tokens_all_models": "171404.0", + "avg_turn_count": "15.00", + "total_turn_count": "45.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/openai/gpt-5-2025-08-07/f01ed4c5-c9cf-4ae7-9814-835dbd5c1a89_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "4efc21a9c59f04e33fa89dede76ff578ae0a4bf690f6c6027631d896d6bf7466", + "total_rows": 40 + } +} \ No newline at end of file diff --git a/data/healthbench/openai/gpt-5-2025-08-07/f01ed4c5-c9cf-4ae7-9814-835dbd5c1a89_samples.jsonl b/data/healthbench/openai/gpt-5-2025-08-07/f01ed4c5-c9cf-4ae7-9814-835dbd5c1a89_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..1a737b9f614dd5a37a1cf15fd8d8bcdfc95651db --- /dev/null +++ b/data/healthbench/openai/gpt-5-2025-08-07/f01ed4c5-c9cf-4ae7-9814-835dbd5c1a89_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4efc21a9c59f04e33fa89dede76ff578ae0a4bf690f6c6027631d896d6bf7466 +size 3032103 diff --git a/data/healthbench/openai/gpt-5-2025-08-07/f2999644-bfd3-4cb1-8311-87bae3cf0def.json b/data/healthbench/openai/gpt-5-2025-08-07/f2999644-bfd3-4cb1-8311-87bae3cf0def.json new file mode 100644 index 0000000000000000000000000000000000000000..9086c089d5f06407dd19674649244759401d4251 --- /dev/null +++ b/data/healthbench/openai/gpt-5-2025-08-07/f2999644-bfd3-4cb1-8311-87bae3cf0def.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/openai_gpt-5-2025-08-07/1786168487.280189", + "evaluation_timestamp": "1776165242.0", + "retrieved_timestamp": "1786168487.280189", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5-2025-08-07", + "id": "openai/gpt-5-2025-08-07", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/openai/gpt-5-2025-08-07/f2999644-bfd3-4cb1-8311-87bae3cf0def_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "ada5820af7379098f58f007d49422bfc43836fb8f1798bb49867fc115652a999", + "total_rows": 200 + } +} \ No newline at end of file diff --git a/data/healthbench/openai/gpt-5-2025-08-07/f2999644-bfd3-4cb1-8311-87bae3cf0def_samples.jsonl b/data/healthbench/openai/gpt-5-2025-08-07/f2999644-bfd3-4cb1-8311-87bae3cf0def_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..21fab60761274706b6afd0acc5419271c3a6d63b --- /dev/null +++ b/data/healthbench/openai/gpt-5-2025-08-07/f2999644-bfd3-4cb1-8311-87bae3cf0def_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ada5820af7379098f58f007d49422bfc43836fb8f1798bb49867fc115652a999 +size 13119105 diff --git a/data/healthbench/openai/gpt-5.2-2025-12-11/155a40d5-fca2-42c3-85d7-6f3206475606.json b/data/healthbench/openai/gpt-5.2-2025-12-11/155a40d5-fca2-42c3-85d7-6f3206475606.json new file mode 100644 index 0000000000000000000000000000000000000000..c74bc99e492999162efefc4f1bcc3b3e7ab9c05d --- /dev/null +++ b/data/healthbench/openai/gpt-5.2-2025-12-11/155a40d5-fca2-42c3-85d7-6f3206475606.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/openai_gpt-5.2-2025-12-11/1786166360.193905", + "evaluation_timestamp": "1776682987.0", + "retrieved_timestamp": "1786166360.193905", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/openai/gpt-5.2-2025-12-11/155a40d5-fca2-42c3-85d7-6f3206475606_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "cd253170753cf25f62cad78783582ba5cc3fdf26c79b8b8a2d321e4e6d24bbe7", + "total_rows": 19 + } +} \ No newline at end of file diff --git a/data/healthbench/openai/gpt-5.2-2025-12-11/155a40d5-fca2-42c3-85d7-6f3206475606_samples.jsonl b/data/healthbench/openai/gpt-5.2-2025-12-11/155a40d5-fca2-42c3-85d7-6f3206475606_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..09832ea0c5d3156925377b7b19401bdd8cea3868 --- /dev/null +++ b/data/healthbench/openai/gpt-5.2-2025-12-11/155a40d5-fca2-42c3-85d7-6f3206475606_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cd253170753cf25f62cad78783582ba5cc3fdf26c79b8b8a2d321e4e6d24bbe7 +size 1263358 diff --git a/data/healthbench/openai/gpt-5.2-2025-12-11/2da88d86-166e-4f24-9eac-4718ad4249aa.json b/data/healthbench/openai/gpt-5.2-2025-12-11/2da88d86-166e-4f24-9eac-4718ad4249aa.json new file mode 100644 index 0000000000000000000000000000000000000000..402ab78d918e3706b6f39bcb0171cfb50bbf238f --- /dev/null +++ b/data/healthbench/openai/gpt-5.2-2025-12-11/2da88d86-166e-4f24-9eac-4718ad4249aa.json @@ -0,0 +1,304 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/openai_gpt-5.2-2025-12-11/1786163768.070545", + "evaluation_timestamp": "1776335705.0", + "retrieved_timestamp": "1786163768.070545", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "bootstrap_score on healthbench/S-adaptive/+1ep/be90264c for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "25", + "sample_ids": "a131d5e7-2a27-426f-90f0-25eae3c4b490,9b4c14e9-c404-4646-87b7-d3badae8a68b,9385d26b-bf44-4ff7-9410-e181413540dd,a57a4b5e-0f12-4103-8852-72bd2b7dba24,cf8040e2-39e3-4200-ab7f-b9b532310ac3,4fc7b71f-9591-4ae1-bfca-2929c7225182,da05b57a-d776-45b6-96af-1750e0254eec,ef446d15-d019-4c5a-a052-ae8cc3ea3b81,0fe4ea94-633d-4f73-985b-cbcd9f5a0270,cbb0ca6e-aa5e-4683-b107-d7dfa4cb60ce,ae6ec2ce-5c23-4e65-802f-cef9f1624d71,cd3dbc9f-a39c-4d13-af60-cba7dbc3baad,02423ab9-f3ed-4096-a7df-bfa3ebb40c7c,084b0c27-79c5-42f6-9cbd-690d0e79167a,42465540-8af1-423b-9b35-a31de1e1d49e,74a50705-7db2-44be-84dc-92b2d960d344,3b79284e-1143-46b4-b6a0-78e5448ff773,afbfc79d-6455-40b8-b475-1bcde4ef0cc9,6d705f80-8259-4ca1-8645-163f42c85a3c,129057dd-4bf4-472d-a8cf-a063dbed60ad,2833344d-bff9-4bbe-8fab-6b13d64c6b2a,48669cdc-1a64-4c99-a6e8-9a7181341fa2,be70739f-234a-4a0f-9b14-d15b8800577b,4047884e-4629-457c-9d77-130e937fae60,eb97bae4-430e-45cd-a065-2df3ab5c600e" + } + }, + "evaluation_timestamp": "1776335705.0", + "metric_config": { + "evaluation_description": "bootstrap_score", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.4559743657753237, + "uncertainty": { + "standard_deviation": 0.05295834041642024, + "num_samples": 25 + }, + "details": { + "total_matched_trajectories": "6", + "stopping_reason_count_repetition_guard": "6", + "avg_total_tokens_target_model": "9663.83", + "total_total_tokens_target_model": "57983.0", + "avg_total_tokens_other_models": "64424.17", + "total_total_tokens_other_models": "386545.0", + "avg_total_tokens_all_models": "74088.00", + "total_total_tokens_all_models": "444528.0", + "avg_turn_count": "29.50", + "total_turn_count": "177.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "reasoning_effort": "\"high\"" + } + } + }, + { + "evaluation_name": "std on healthbench/S-adaptive/+1ep/be90264c for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "25", + "sample_ids": "a131d5e7-2a27-426f-90f0-25eae3c4b490,9b4c14e9-c404-4646-87b7-d3badae8a68b,9385d26b-bf44-4ff7-9410-e181413540dd,a57a4b5e-0f12-4103-8852-72bd2b7dba24,cf8040e2-39e3-4200-ab7f-b9b532310ac3,4fc7b71f-9591-4ae1-bfca-2929c7225182,da05b57a-d776-45b6-96af-1750e0254eec,ef446d15-d019-4c5a-a052-ae8cc3ea3b81,0fe4ea94-633d-4f73-985b-cbcd9f5a0270,cbb0ca6e-aa5e-4683-b107-d7dfa4cb60ce,ae6ec2ce-5c23-4e65-802f-cef9f1624d71,cd3dbc9f-a39c-4d13-af60-cba7dbc3baad,02423ab9-f3ed-4096-a7df-bfa3ebb40c7c,084b0c27-79c5-42f6-9cbd-690d0e79167a,42465540-8af1-423b-9b35-a31de1e1d49e,74a50705-7db2-44be-84dc-92b2d960d344,3b79284e-1143-46b4-b6a0-78e5448ff773,afbfc79d-6455-40b8-b475-1bcde4ef0cc9,6d705f80-8259-4ca1-8645-163f42c85a3c,129057dd-4bf4-472d-a8cf-a063dbed60ad,2833344d-bff9-4bbe-8fab-6b13d64c6b2a,48669cdc-1a64-4c99-a6e8-9a7181341fa2,be70739f-234a-4a0f-9b14-d15b8800577b,4047884e-4629-457c-9d77-130e937fae60,eb97bae4-430e-45cd-a065-2df3ab5c600e" + } + }, + "evaluation_timestamp": "1776335705.0", + "metric_config": { + "evaluation_description": "std", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.05295834041642024, + "uncertainty": { + "standard_deviation": 0.05295834041642024, + "num_samples": 25 + }, + "details": { + "total_matched_trajectories": "6", + "stopping_reason_count_repetition_guard": "6", + "avg_total_tokens_target_model": "9663.83", + "total_total_tokens_target_model": "57983.0", + "avg_total_tokens_other_models": "64424.17", + "total_total_tokens_other_models": "386545.0", + "avg_total_tokens_all_models": "74088.00", + "total_total_tokens_all_models": "444528.0", + "avg_turn_count": "29.50", + "total_turn_count": "177.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "reasoning_effort": "\"high\"" + } + } + }, + { + "evaluation_name": "criteria_met_rate on healthbench/S-adaptive/+1ep/be90264c for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "25", + "sample_ids": "a131d5e7-2a27-426f-90f0-25eae3c4b490,9b4c14e9-c404-4646-87b7-d3badae8a68b,9385d26b-bf44-4ff7-9410-e181413540dd,a57a4b5e-0f12-4103-8852-72bd2b7dba24,cf8040e2-39e3-4200-ab7f-b9b532310ac3,4fc7b71f-9591-4ae1-bfca-2929c7225182,da05b57a-d776-45b6-96af-1750e0254eec,ef446d15-d019-4c5a-a052-ae8cc3ea3b81,0fe4ea94-633d-4f73-985b-cbcd9f5a0270,cbb0ca6e-aa5e-4683-b107-d7dfa4cb60ce,ae6ec2ce-5c23-4e65-802f-cef9f1624d71,cd3dbc9f-a39c-4d13-af60-cba7dbc3baad,02423ab9-f3ed-4096-a7df-bfa3ebb40c7c,084b0c27-79c5-42f6-9cbd-690d0e79167a,42465540-8af1-423b-9b35-a31de1e1d49e,74a50705-7db2-44be-84dc-92b2d960d344,3b79284e-1143-46b4-b6a0-78e5448ff773,afbfc79d-6455-40b8-b475-1bcde4ef0cc9,6d705f80-8259-4ca1-8645-163f42c85a3c,129057dd-4bf4-472d-a8cf-a063dbed60ad,2833344d-bff9-4bbe-8fab-6b13d64c6b2a,48669cdc-1a64-4c99-a6e8-9a7181341fa2,be70739f-234a-4a0f-9b14-d15b8800577b,4047884e-4629-457c-9d77-130e937fae60,eb97bae4-430e-45cd-a065-2df3ab5c600e" + } + }, + "evaluation_timestamp": "1776335705.0", + "metric_config": { + "evaluation_description": "criteria_met_rate", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.4402730375426621, + "uncertainty": { + "standard_deviation": 0.05295834041642024, + "num_samples": 25 + }, + "details": { + "total_matched_trajectories": "6", + "stopping_reason_count_repetition_guard": "6", + "avg_total_tokens_target_model": "9663.83", + "total_total_tokens_target_model": "57983.0", + "avg_total_tokens_other_models": "64424.17", + "total_total_tokens_other_models": "386545.0", + "avg_total_tokens_all_models": "74088.00", + "total_total_tokens_all_models": "444528.0", + "avg_turn_count": "29.50", + "total_turn_count": "177.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "reasoning_effort": "\"high\"" + } + } + }, + { + "evaluation_name": "mean on healthbench/S-adaptive/+1ep/be90264c for scorer healthbench_score", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "25", + "sample_ids": "a131d5e7-2a27-426f-90f0-25eae3c4b490,9b4c14e9-c404-4646-87b7-d3badae8a68b,9385d26b-bf44-4ff7-9410-e181413540dd,a57a4b5e-0f12-4103-8852-72bd2b7dba24,cf8040e2-39e3-4200-ab7f-b9b532310ac3,4fc7b71f-9591-4ae1-bfca-2929c7225182,da05b57a-d776-45b6-96af-1750e0254eec,ef446d15-d019-4c5a-a052-ae8cc3ea3b81,0fe4ea94-633d-4f73-985b-cbcd9f5a0270,cbb0ca6e-aa5e-4683-b107-d7dfa4cb60ce,ae6ec2ce-5c23-4e65-802f-cef9f1624d71,cd3dbc9f-a39c-4d13-af60-cba7dbc3baad,02423ab9-f3ed-4096-a7df-bfa3ebb40c7c,084b0c27-79c5-42f6-9cbd-690d0e79167a,42465540-8af1-423b-9b35-a31de1e1d49e,74a50705-7db2-44be-84dc-92b2d960d344,3b79284e-1143-46b4-b6a0-78e5448ff773,afbfc79d-6455-40b8-b475-1bcde4ef0cc9,6d705f80-8259-4ca1-8645-163f42c85a3c,129057dd-4bf4-472d-a8cf-a063dbed60ad,2833344d-bff9-4bbe-8fab-6b13d64c6b2a,48669cdc-1a64-4c99-a6e8-9a7181341fa2,be70739f-234a-4a0f-9b14-d15b8800577b,4047884e-4629-457c-9d77-130e937fae60,eb97bae4-430e-45cd-a065-2df3ab5c600e" + } + }, + "evaluation_timestamp": "1776335705.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.4559743657753237, + "uncertainty": { + "num_samples": 25 + }, + "details": { + "total_matched_trajectories": "6", + "stopping_reason_count_repetition_guard": "6", + "avg_total_tokens_target_model": "9663.83", + "total_total_tokens_target_model": "57983.0", + "avg_total_tokens_other_models": "64424.17", + "total_total_tokens_other_models": "386545.0", + "avg_total_tokens_all_models": "74088.00", + "total_total_tokens_all_models": "444528.0", + "avg_turn_count": "29.50", + "total_turn_count": "177.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/openai/gpt-5.2-2025-12-11/2da88d86-166e-4f24-9eac-4718ad4249aa_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "64f5773eca0bc2486dcd64a5b68138ff9dd3ec53f1c8fc5a98b80011492d4e04", + "total_rows": 25 + } +} \ No newline at end of file diff --git a/data/healthbench/openai/gpt-5.2-2025-12-11/2da88d86-166e-4f24-9eac-4718ad4249aa_samples.jsonl b/data/healthbench/openai/gpt-5.2-2025-12-11/2da88d86-166e-4f24-9eac-4718ad4249aa_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..e6277930cb0f9759fdd46bbf09a5495181c884a0 --- /dev/null +++ b/data/healthbench/openai/gpt-5.2-2025-12-11/2da88d86-166e-4f24-9eac-4718ad4249aa_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:64f5773eca0bc2486dcd64a5b68138ff9dd3ec53f1c8fc5a98b80011492d4e04 +size 1590425 diff --git a/data/healthbench/openai/gpt-5.2-2025-12-11/32461749-fecc-477c-9540-622dcfa8bb5b.json b/data/healthbench/openai/gpt-5.2-2025-12-11/32461749-fecc-477c-9540-622dcfa8bb5b.json new file mode 100644 index 0000000000000000000000000000000000000000..d996fdc4830b0fb389a5ab398a98437c71cd6ee9 --- /dev/null +++ b/data/healthbench/openai/gpt-5.2-2025-12-11/32461749-fecc-477c-9540-622dcfa8bb5b.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/openai_gpt-5.2-2025-12-11/1786166357.537446", + "evaluation_timestamp": "1776136554.0", + "retrieved_timestamp": "1786166357.537446", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/openai/gpt-5.2-2025-12-11/32461749-fecc-477c-9540-622dcfa8bb5b_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "491e159ddcaef45480331ba2a36f4fbe77e1e505927c9d55651ede1ca2f62617", + "total_rows": 10 + } +} \ No newline at end of file diff --git a/data/healthbench/openai/gpt-5.2-2025-12-11/32461749-fecc-477c-9540-622dcfa8bb5b_samples.jsonl b/data/healthbench/openai/gpt-5.2-2025-12-11/32461749-fecc-477c-9540-622dcfa8bb5b_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..f38f8e48dfd59a02076d29e6f2db2d1f4b71eec3 --- /dev/null +++ b/data/healthbench/openai/gpt-5.2-2025-12-11/32461749-fecc-477c-9540-622dcfa8bb5b_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:491e159ddcaef45480331ba2a36f4fbe77e1e505927c9d55651ede1ca2f62617 +size 680944 diff --git a/data/healthbench/openai/gpt-5.2-2025-12-11/35bf8f23-d2df-4897-9a13-95d149b31a25.json b/data/healthbench/openai/gpt-5.2-2025-12-11/35bf8f23-d2df-4897-9a13-95d149b31a25.json new file mode 100644 index 0000000000000000000000000000000000000000..d43d48fb9e278560195ce69e79683cbf926f84df --- /dev/null +++ b/data/healthbench/openai/gpt-5.2-2025-12-11/35bf8f23-d2df-4897-9a13-95d149b31a25.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/openai_gpt-5.2-2025-12-11/1786166399.433704", + "evaluation_timestamp": "1775733995.0", + "retrieved_timestamp": "1786166399.433704", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/openai/gpt-5.2-2025-12-11/35bf8f23-d2df-4897-9a13-95d149b31a25_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "0bac97971fb83738545af5bf87d5fc398eab4b76f878be7a82ff6971b716d322", + "total_rows": 160 + } +} \ No newline at end of file diff --git a/data/healthbench/openai/gpt-5.2-2025-12-11/35bf8f23-d2df-4897-9a13-95d149b31a25_samples.jsonl b/data/healthbench/openai/gpt-5.2-2025-12-11/35bf8f23-d2df-4897-9a13-95d149b31a25_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..bb384e0e05a45d8c1186063a8aad15b7e94a0e95 --- /dev/null +++ b/data/healthbench/openai/gpt-5.2-2025-12-11/35bf8f23-d2df-4897-9a13-95d149b31a25_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0bac97971fb83738545af5bf87d5fc398eab4b76f878be7a82ff6971b716d322 +size 10499438 diff --git a/data/healthbench/openai/gpt-5.2-2025-12-11/51f587e9-5842-41e0-95c2-47a24925a6d4.json b/data/healthbench/openai/gpt-5.2-2025-12-11/51f587e9-5842-41e0-95c2-47a24925a6d4.json new file mode 100644 index 0000000000000000000000000000000000000000..c5fd979a5919dd332ffcbfeda81927bf0f7bef79 --- /dev/null +++ b/data/healthbench/openai/gpt-5.2-2025-12-11/51f587e9-5842-41e0-95c2-47a24925a6d4.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/openai_gpt-5.2-2025-12-11/1786166384.63572", + "evaluation_timestamp": "1776076027.0", + "retrieved_timestamp": "1786166384.63572", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/openai/gpt-5.2-2025-12-11/51f587e9-5842-41e0-95c2-47a24925a6d4_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "7db21c8e9b2db027a3b6ea261eceb94bd97ed44baa451538df0dd254487d43ef", + "total_rows": 80 + } +} \ No newline at end of file diff --git a/data/healthbench/openai/gpt-5.2-2025-12-11/51f587e9-5842-41e0-95c2-47a24925a6d4_samples.jsonl b/data/healthbench/openai/gpt-5.2-2025-12-11/51f587e9-5842-41e0-95c2-47a24925a6d4_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..c360e23051787c35793c476b67c8d3ea3dfec8b6 --- /dev/null +++ b/data/healthbench/openai/gpt-5.2-2025-12-11/51f587e9-5842-41e0-95c2-47a24925a6d4_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7db21c8e9b2db027a3b6ea261eceb94bd97ed44baa451538df0dd254487d43ef +size 5178858 diff --git a/data/healthbench/openai/gpt-5.2-2025-12-11/54598837-51c8-4aeb-a516-b3ed22077325.json b/data/healthbench/openai/gpt-5.2-2025-12-11/54598837-51c8-4aeb-a516-b3ed22077325.json new file mode 100644 index 0000000000000000000000000000000000000000..17e029a41086b160270d19114fe56d75b7c5a1d1 --- /dev/null +++ b/data/healthbench/openai/gpt-5.2-2025-12-11/54598837-51c8-4aeb-a516-b3ed22077325.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/openai_gpt-5.2-2025-12-11/1786166377.533516", + "evaluation_timestamp": "1776165242.0", + "retrieved_timestamp": "1786166377.533516", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/openai/gpt-5.2-2025-12-11/54598837-51c8-4aeb-a516-b3ed22077325_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "3c196c70fcc939d4ed8847fbc79ce84bd6c8a0cd18526932bc9a14d986c16381", + "total_rows": 200 + } +} \ No newline at end of file diff --git a/data/healthbench/openai/gpt-5.2-2025-12-11/54598837-51c8-4aeb-a516-b3ed22077325_samples.jsonl b/data/healthbench/openai/gpt-5.2-2025-12-11/54598837-51c8-4aeb-a516-b3ed22077325_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..2407a92a7a57d95bb3728c273dbad91a6d9d86de --- /dev/null +++ b/data/healthbench/openai/gpt-5.2-2025-12-11/54598837-51c8-4aeb-a516-b3ed22077325_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3c196c70fcc939d4ed8847fbc79ce84bd6c8a0cd18526932bc9a14d986c16381 +size 12766796 diff --git a/data/healthbench/openai/gpt-5.2-2025-12-11/76a71e91-1590-4310-8a41-c27b6b9a89dc.json b/data/healthbench/openai/gpt-5.2-2025-12-11/76a71e91-1590-4310-8a41-c27b6b9a89dc.json new file mode 100644 index 0000000000000000000000000000000000000000..3f267c629237e41be7603944834f6a8eb498c949 --- /dev/null +++ b/data/healthbench/openai/gpt-5.2-2025-12-11/76a71e91-1590-4310-8a41-c27b6b9a89dc.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/openai_gpt-5.2-2025-12-11/1786168305.842927", + "evaluation_timestamp": "1776259832.0", + "retrieved_timestamp": "1786168305.842927", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/openai/gpt-5.2-2025-12-11/76a71e91-1590-4310-8a41-c27b6b9a89dc_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "c64cc973ccfa7660385ebd2d837f3cc19a62e851402021a5443ef8e40747eacb", + "total_rows": 160 + } +} \ No newline at end of file diff --git a/data/healthbench/openai/gpt-5.2-2025-12-11/76a71e91-1590-4310-8a41-c27b6b9a89dc_samples.jsonl b/data/healthbench/openai/gpt-5.2-2025-12-11/76a71e91-1590-4310-8a41-c27b6b9a89dc_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..feb81b471d8fbd32d36ce38d73c4d39cc2227543 --- /dev/null +++ b/data/healthbench/openai/gpt-5.2-2025-12-11/76a71e91-1590-4310-8a41-c27b6b9a89dc_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c64cc973ccfa7660385ebd2d837f3cc19a62e851402021a5443ef8e40747eacb +size 10337210 diff --git a/data/healthbench/openai/gpt-5.2-2025-12-11/8a2762de-132e-44e8-a265-c860196c1d48.json b/data/healthbench/openai/gpt-5.2-2025-12-11/8a2762de-132e-44e8-a265-c860196c1d48.json new file mode 100644 index 0000000000000000000000000000000000000000..ca6a2605baa8781bb1265d1d5f893f327de39e1b --- /dev/null +++ b/data/healthbench/openai/gpt-5.2-2025-12-11/8a2762de-132e-44e8-a265-c860196c1d48.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/openai_gpt-5.2-2025-12-11/1786168362.636885", + "evaluation_timestamp": "1776259832.0", + "retrieved_timestamp": "1786168362.636885", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/openai/gpt-5.2-2025-12-11/8a2762de-132e-44e8-a265-c860196c1d48_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "0aae5fa0365bec10cb3dbe54493615c6b239546628215d66c2d6ff6a088fd014", + "total_rows": 220 + } +} \ No newline at end of file diff --git a/data/healthbench/openai/gpt-5.2-2025-12-11/8a2762de-132e-44e8-a265-c860196c1d48_samples.jsonl b/data/healthbench/openai/gpt-5.2-2025-12-11/8a2762de-132e-44e8-a265-c860196c1d48_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..1b3bff85b899e8c3da59aa367bdeb70b4e6a8ce6 --- /dev/null +++ b/data/healthbench/openai/gpt-5.2-2025-12-11/8a2762de-132e-44e8-a265-c860196c1d48_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0aae5fa0365bec10cb3dbe54493615c6b239546628215d66c2d6ff6a088fd014 +size 14834319 diff --git a/data/healthbench/openai/gpt-5.2-2025-12-11/a3065968-e6e0-40cc-9822-26f5742aad2d.json b/data/healthbench/openai/gpt-5.2-2025-12-11/a3065968-e6e0-40cc-9822-26f5742aad2d.json new file mode 100644 index 0000000000000000000000000000000000000000..eeaade45791a8102009a56e0c2a1e2dcdab51689 --- /dev/null +++ b/data/healthbench/openai/gpt-5.2-2025-12-11/a3065968-e6e0-40cc-9822-26f5742aad2d.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/openai_gpt-5.2-2025-12-11/1786168280.609335", + "evaluation_timestamp": "1775604415.0", + "retrieved_timestamp": "1786168280.609335", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.202,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/openai/gpt-5.2-2025-12-11/a3065968-e6e0-40cc-9822-26f5742aad2d_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "cd0af953ad08d60c97797c79798c70be76fc386ca432ff666755bfd9879d13db", + "total_rows": 100 + } +} \ No newline at end of file diff --git a/data/healthbench/openai/gpt-5.2-2025-12-11/a3065968-e6e0-40cc-9822-26f5742aad2d_samples.jsonl b/data/healthbench/openai/gpt-5.2-2025-12-11/a3065968-e6e0-40cc-9822-26f5742aad2d_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..9c6a6802640e48ba1ff088bc12bc50c871bded2c --- /dev/null +++ b/data/healthbench/openai/gpt-5.2-2025-12-11/a3065968-e6e0-40cc-9822-26f5742aad2d_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cd0af953ad08d60c97797c79798c70be76fc386ca432ff666755bfd9879d13db +size 6478702 diff --git a/data/healthbench/openai/gpt-5.2-2025-12-11/a7afcfc6-1c74-4c24-b082-e123b7b5ea6c.json b/data/healthbench/openai/gpt-5.2-2025-12-11/a7afcfc6-1c74-4c24-b082-e123b7b5ea6c.json new file mode 100644 index 0000000000000000000000000000000000000000..a0375dac4d29710d167fbdde0081b9dea7a1cb80 --- /dev/null +++ b/data/healthbench/openai/gpt-5.2-2025-12-11/a7afcfc6-1c74-4c24-b082-e123b7b5ea6c.json @@ -0,0 +1,320 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/openai_gpt-5.2-2025-12-11/1786163775.284779", + "evaluation_timestamp": "1776696116.0", + "retrieved_timestamp": "1786163775.284779", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "bootstrap_score on healthbench/S-adaptive/+4ep/1bea1e9b for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "40", + "sample_ids": "f712eb09-82af-4fe8-800d-bc3af772a9ba,8ee6a990-8966-4b42-baa8-6c7e9f74ce74,267bda16-fd4b-4a9e-b3ac-5602c2532eed,0e819a9c-851d-4a7a-9263-62cfc8ce1b48,328c55cf-0b57-4992-99b1-f6a21af62ade,7b415a2f-22b0-4735-a38e-567f1891acc5,c49ac752-3693-4be1-ab4d-91fbdd02f827,a079d928-7b78-4a85-a72b-4274d7fc2914,a20cbbf3-330b-46d9-a35b-92e89a022282,be6150a5-8060-4434-9724-afede307d495,3a896889-9029-4ef8-8684-ca378da4702b,7074bb7a-a595-434d-883e-95ef6f5f6667,9f7b555a-fc30-40c3-88d3-f2ece06454a2,a6204079-9e72-42ad-8f39-a65552fa2e68,c0d80b93-b2d5-4129-b998-b1312f98e2a8,3dc41419-2565-428e-b9b9-02e3fd93c470,a91a22f8-d2e6-4593-bcae-6d479ed38031,d0d3426f-8b7a-4501-adb4-86d8fa9099c8,a7dac6a9-eab5-4c01-882c-dbf6c774d7c3,c1240427-ae38-4335-9e80-492c13ef4893,b7ddde65-702e-4255-b1d7-09d01248e76d,00d34549-134c-48e2-b4bb-b54946da70cc,85c11e2a-628b-464f-a8bb-e97ab04a73eb,69a544f8-957d-4dda-a777-608ec7b95617,90789a49-ab1f-4229-bd89-e1c56ea8be3a,cd53e21a-e5ed-4d9b-a258-42a23e9715bb,8d409c7b-29d2-4df2-aa69-ab56c9339bc5,6566f489-f254-4bca-a86b-90f5ac7471d7,d9a44a36-6a3c-4d5d-9c66-b2c0afb479d1,9cf87956-37e6-4bed-9cea-a64b13ae0de5,a1a08e1d-1353-4d6f-94ce-f90872074c7c,51d9c72b-ba60-47c5-9cc9-3ed6212a3982,bc8d3e4a-8ba0-4cab-be1c-b20ae3af91e4,99306c29-bcef-49d5-bcd1-78d4e541e197,d0fcfaf0-7c6d-4ead-915b-6d9d9130e747,59c68d10-55c1-4629-b05a-bedde7f343b5,d86ee090-3a2e-44d0-abed-6f0f40a38c08,25cf07b8-606f-4448-99ef-ea87ca3097ba,1ca222ed-cc79-4e78-9ca3-3547e7b37e3a,0eb6fd6d-eaaa-46db-ab16-02e610f238a9" + } + }, + "evaluation_timestamp": "1776696116.0", + "metric_config": { + "evaluation_description": "bootstrap_score", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.30537771766452904, + "uncertainty": { + "standard_deviation": 0.05033294822633026, + "num_samples": 157 + }, + "details": { + "total_matched_trajectories": "157", + "stopping_reason_count_repetition_guard": "157", + "avg_total_tokens_target_model": "12464.96", + "total_total_tokens_target_model": "1956998.0", + "avg_total_tokens_other_models": "60266.62", + "total_total_tokens_other_models": "9461859.0", + "avg_total_tokens_all_models": "72731.57", + "total_total_tokens_all_models": "11418857.0", + "avg_turn_count": "24.89", + "total_turn_count": "3908.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "3600", + "attempt_timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + }, + { + "evaluation_name": "std on healthbench/S-adaptive/+4ep/1bea1e9b for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "40", + "sample_ids": "f712eb09-82af-4fe8-800d-bc3af772a9ba,8ee6a990-8966-4b42-baa8-6c7e9f74ce74,267bda16-fd4b-4a9e-b3ac-5602c2532eed,0e819a9c-851d-4a7a-9263-62cfc8ce1b48,328c55cf-0b57-4992-99b1-f6a21af62ade,7b415a2f-22b0-4735-a38e-567f1891acc5,c49ac752-3693-4be1-ab4d-91fbdd02f827,a079d928-7b78-4a85-a72b-4274d7fc2914,a20cbbf3-330b-46d9-a35b-92e89a022282,be6150a5-8060-4434-9724-afede307d495,3a896889-9029-4ef8-8684-ca378da4702b,7074bb7a-a595-434d-883e-95ef6f5f6667,9f7b555a-fc30-40c3-88d3-f2ece06454a2,a6204079-9e72-42ad-8f39-a65552fa2e68,c0d80b93-b2d5-4129-b998-b1312f98e2a8,3dc41419-2565-428e-b9b9-02e3fd93c470,a91a22f8-d2e6-4593-bcae-6d479ed38031,d0d3426f-8b7a-4501-adb4-86d8fa9099c8,a7dac6a9-eab5-4c01-882c-dbf6c774d7c3,c1240427-ae38-4335-9e80-492c13ef4893,b7ddde65-702e-4255-b1d7-09d01248e76d,00d34549-134c-48e2-b4bb-b54946da70cc,85c11e2a-628b-464f-a8bb-e97ab04a73eb,69a544f8-957d-4dda-a777-608ec7b95617,90789a49-ab1f-4229-bd89-e1c56ea8be3a,cd53e21a-e5ed-4d9b-a258-42a23e9715bb,8d409c7b-29d2-4df2-aa69-ab56c9339bc5,6566f489-f254-4bca-a86b-90f5ac7471d7,d9a44a36-6a3c-4d5d-9c66-b2c0afb479d1,9cf87956-37e6-4bed-9cea-a64b13ae0de5,a1a08e1d-1353-4d6f-94ce-f90872074c7c,51d9c72b-ba60-47c5-9cc9-3ed6212a3982,bc8d3e4a-8ba0-4cab-be1c-b20ae3af91e4,99306c29-bcef-49d5-bcd1-78d4e541e197,d0fcfaf0-7c6d-4ead-915b-6d9d9130e747,59c68d10-55c1-4629-b05a-bedde7f343b5,d86ee090-3a2e-44d0-abed-6f0f40a38c08,25cf07b8-606f-4448-99ef-ea87ca3097ba,1ca222ed-cc79-4e78-9ca3-3547e7b37e3a,0eb6fd6d-eaaa-46db-ab16-02e610f238a9" + } + }, + "evaluation_timestamp": "1776696116.0", + "metric_config": { + "evaluation_description": "std", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.05033294822633026, + "uncertainty": { + "standard_deviation": 0.05033294822633026, + "num_samples": 157 + }, + "details": { + "total_matched_trajectories": "157", + "stopping_reason_count_repetition_guard": "157", + "avg_total_tokens_target_model": "12464.96", + "total_total_tokens_target_model": "1956998.0", + "avg_total_tokens_other_models": "60266.62", + "total_total_tokens_other_models": "9461859.0", + "avg_total_tokens_all_models": "72731.57", + "total_total_tokens_all_models": "11418857.0", + "avg_turn_count": "24.89", + "total_turn_count": "3908.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "3600", + "attempt_timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + }, + { + "evaluation_name": "criteria_met_rate on healthbench/S-adaptive/+4ep/1bea1e9b for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "40", + "sample_ids": "f712eb09-82af-4fe8-800d-bc3af772a9ba,8ee6a990-8966-4b42-baa8-6c7e9f74ce74,267bda16-fd4b-4a9e-b3ac-5602c2532eed,0e819a9c-851d-4a7a-9263-62cfc8ce1b48,328c55cf-0b57-4992-99b1-f6a21af62ade,7b415a2f-22b0-4735-a38e-567f1891acc5,c49ac752-3693-4be1-ab4d-91fbdd02f827,a079d928-7b78-4a85-a72b-4274d7fc2914,a20cbbf3-330b-46d9-a35b-92e89a022282,be6150a5-8060-4434-9724-afede307d495,3a896889-9029-4ef8-8684-ca378da4702b,7074bb7a-a595-434d-883e-95ef6f5f6667,9f7b555a-fc30-40c3-88d3-f2ece06454a2,a6204079-9e72-42ad-8f39-a65552fa2e68,c0d80b93-b2d5-4129-b998-b1312f98e2a8,3dc41419-2565-428e-b9b9-02e3fd93c470,a91a22f8-d2e6-4593-bcae-6d479ed38031,d0d3426f-8b7a-4501-adb4-86d8fa9099c8,a7dac6a9-eab5-4c01-882c-dbf6c774d7c3,c1240427-ae38-4335-9e80-492c13ef4893,b7ddde65-702e-4255-b1d7-09d01248e76d,00d34549-134c-48e2-b4bb-b54946da70cc,85c11e2a-628b-464f-a8bb-e97ab04a73eb,69a544f8-957d-4dda-a777-608ec7b95617,90789a49-ab1f-4229-bd89-e1c56ea8be3a,cd53e21a-e5ed-4d9b-a258-42a23e9715bb,8d409c7b-29d2-4df2-aa69-ab56c9339bc5,6566f489-f254-4bca-a86b-90f5ac7471d7,d9a44a36-6a3c-4d5d-9c66-b2c0afb479d1,9cf87956-37e6-4bed-9cea-a64b13ae0de5,a1a08e1d-1353-4d6f-94ce-f90872074c7c,51d9c72b-ba60-47c5-9cc9-3ed6212a3982,bc8d3e4a-8ba0-4cab-be1c-b20ae3af91e4,99306c29-bcef-49d5-bcd1-78d4e541e197,d0fcfaf0-7c6d-4ead-915b-6d9d9130e747,59c68d10-55c1-4629-b05a-bedde7f343b5,d86ee090-3a2e-44d0-abed-6f0f40a38c08,25cf07b8-606f-4448-99ef-ea87ca3097ba,1ca222ed-cc79-4e78-9ca3-3547e7b37e3a,0eb6fd6d-eaaa-46db-ab16-02e610f238a9" + } + }, + "evaluation_timestamp": "1776696116.0", + "metric_config": { + "evaluation_description": "criteria_met_rate", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.38921113689095127, + "uncertainty": { + "standard_deviation": 0.05033294822633026, + "num_samples": 157 + }, + "details": { + "total_matched_trajectories": "157", + "stopping_reason_count_repetition_guard": "157", + "avg_total_tokens_target_model": "12464.96", + "total_total_tokens_target_model": "1956998.0", + "avg_total_tokens_other_models": "60266.62", + "total_total_tokens_other_models": "9461859.0", + "avg_total_tokens_all_models": "72731.57", + "total_total_tokens_all_models": "11418857.0", + "avg_turn_count": "24.89", + "total_turn_count": "3908.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "3600", + "attempt_timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + }, + { + "evaluation_name": "mean on healthbench/S-adaptive/+4ep/1bea1e9b for scorer healthbench_score", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "40", + "sample_ids": "f712eb09-82af-4fe8-800d-bc3af772a9ba,8ee6a990-8966-4b42-baa8-6c7e9f74ce74,267bda16-fd4b-4a9e-b3ac-5602c2532eed,0e819a9c-851d-4a7a-9263-62cfc8ce1b48,328c55cf-0b57-4992-99b1-f6a21af62ade,7b415a2f-22b0-4735-a38e-567f1891acc5,c49ac752-3693-4be1-ab4d-91fbdd02f827,a079d928-7b78-4a85-a72b-4274d7fc2914,a20cbbf3-330b-46d9-a35b-92e89a022282,be6150a5-8060-4434-9724-afede307d495,3a896889-9029-4ef8-8684-ca378da4702b,7074bb7a-a595-434d-883e-95ef6f5f6667,9f7b555a-fc30-40c3-88d3-f2ece06454a2,a6204079-9e72-42ad-8f39-a65552fa2e68,c0d80b93-b2d5-4129-b998-b1312f98e2a8,3dc41419-2565-428e-b9b9-02e3fd93c470,a91a22f8-d2e6-4593-bcae-6d479ed38031,d0d3426f-8b7a-4501-adb4-86d8fa9099c8,a7dac6a9-eab5-4c01-882c-dbf6c774d7c3,c1240427-ae38-4335-9e80-492c13ef4893,b7ddde65-702e-4255-b1d7-09d01248e76d,00d34549-134c-48e2-b4bb-b54946da70cc,85c11e2a-628b-464f-a8bb-e97ab04a73eb,69a544f8-957d-4dda-a777-608ec7b95617,90789a49-ab1f-4229-bd89-e1c56ea8be3a,cd53e21a-e5ed-4d9b-a258-42a23e9715bb,8d409c7b-29d2-4df2-aa69-ab56c9339bc5,6566f489-f254-4bca-a86b-90f5ac7471d7,d9a44a36-6a3c-4d5d-9c66-b2c0afb479d1,9cf87956-37e6-4bed-9cea-a64b13ae0de5,a1a08e1d-1353-4d6f-94ce-f90872074c7c,51d9c72b-ba60-47c5-9cc9-3ed6212a3982,bc8d3e4a-8ba0-4cab-be1c-b20ae3af91e4,99306c29-bcef-49d5-bcd1-78d4e541e197,d0fcfaf0-7c6d-4ead-915b-6d9d9130e747,59c68d10-55c1-4629-b05a-bedde7f343b5,d86ee090-3a2e-44d0-abed-6f0f40a38c08,25cf07b8-606f-4448-99ef-ea87ca3097ba,1ca222ed-cc79-4e78-9ca3-3547e7b37e3a,0eb6fd6d-eaaa-46db-ab16-02e610f238a9" + } + }, + "evaluation_timestamp": "1776696116.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.30537771766452904, + "uncertainty": { + "num_samples": 157 + }, + "details": { + "total_matched_trajectories": "157", + "stopping_reason_count_repetition_guard": "157", + "avg_total_tokens_target_model": "12464.96", + "total_total_tokens_target_model": "1956998.0", + "avg_total_tokens_other_models": "60266.62", + "total_total_tokens_other_models": "9461859.0", + "avg_total_tokens_all_models": "72731.57", + "total_total_tokens_all_models": "11418857.0", + "avg_turn_count": "24.89", + "total_turn_count": "3908.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "3600", + "attempt_timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/openai/gpt-5.2-2025-12-11/a7afcfc6-1c74-4c24-b082-e123b7b5ea6c_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "38a8c8e28ac45ea63699094ee32d44ea617afaea506e6f50321db3a9b3b94cfd", + "total_rows": 157 + } +} \ No newline at end of file diff --git a/data/healthbench/openai/gpt-5.2-2025-12-11/a7afcfc6-1c74-4c24-b082-e123b7b5ea6c_samples.jsonl b/data/healthbench/openai/gpt-5.2-2025-12-11/a7afcfc6-1c74-4c24-b082-e123b7b5ea6c_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..a47230b489a7b5ac581c32ad24d070a91a05bb6d --- /dev/null +++ b/data/healthbench/openai/gpt-5.2-2025-12-11/a7afcfc6-1c74-4c24-b082-e123b7b5ea6c_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:38a8c8e28ac45ea63699094ee32d44ea617afaea506e6f50321db3a9b3b94cfd +size 10489498 diff --git a/data/healthbench/openai/gpt-5.2-2025-12-11/b09b771a-8f40-496f-bd07-9f2bd6a4b582.json b/data/healthbench/openai/gpt-5.2-2025-12-11/b09b771a-8f40-496f-bd07-9f2bd6a4b582.json new file mode 100644 index 0000000000000000000000000000000000000000..1842009d29b989fc205a8a7c0b4dca6a32c17314 --- /dev/null +++ b/data/healthbench/openai/gpt-5.2-2025-12-11/b09b771a-8f40-496f-bd07-9f2bd6a4b582.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/openai_gpt-5.2-2025-12-11/1786166389.407278", + "evaluation_timestamp": "1776076027.0", + "retrieved_timestamp": "1786166389.407278", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/openai/gpt-5.2-2025-12-11/b09b771a-8f40-496f-bd07-9f2bd6a4b582_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "588b6abe9bb90f914dad12765a49bdd09f59db2d64496d3247cd988c4b1213ab", + "total_rows": 80 + } +} \ No newline at end of file diff --git a/data/healthbench/openai/gpt-5.2-2025-12-11/b09b771a-8f40-496f-bd07-9f2bd6a4b582_samples.jsonl b/data/healthbench/openai/gpt-5.2-2025-12-11/b09b771a-8f40-496f-bd07-9f2bd6a4b582_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..ff6abcc79d29aab39d63fd644bd420df86d9cb6d --- /dev/null +++ b/data/healthbench/openai/gpt-5.2-2025-12-11/b09b771a-8f40-496f-bd07-9f2bd6a4b582_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:588b6abe9bb90f914dad12765a49bdd09f59db2d64496d3247cd988c4b1213ab +size 5118032 diff --git a/data/healthbench/openai/gpt-5.2-2025-12-11/b6f0360d-51ae-4906-948a-9c4f94eee24d.json b/data/healthbench/openai/gpt-5.2-2025-12-11/b6f0360d-51ae-4906-948a-9c4f94eee24d.json new file mode 100644 index 0000000000000000000000000000000000000000..e44e833244d78d77bc442b687acd2c69c47451eb --- /dev/null +++ b/data/healthbench/openai/gpt-5.2-2025-12-11/b6f0360d-51ae-4906-948a-9c4f94eee24d.json @@ -0,0 +1,320 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/openai_gpt-5.2-2025-12-11/1786163770.133652", + "evaluation_timestamp": "1776682987.0", + "retrieved_timestamp": "1786163770.133652", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "bootstrap_score on healthbench/S-adaptive/+1ep/1cf893a8 for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "4", + "sample_ids": "6f827e79-05c4-4383-8ed4-16d36260f29f,c835bebe-e998-4dcd-843a-2de03d4aabc6,70afd1ac-85ea-4455-a5a2-15b35378e10a,aa7a760a-5173-4a39-af83-48d6ae2230f0" + } + }, + "evaluation_timestamp": "1776682987.0", + "metric_config": { + "evaluation_description": "bootstrap_score", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.5658980409974902, + "uncertainty": { + "standard_deviation": 0.04926275183504485, + "num_samples": 4 + }, + "details": { + "total_matched_trajectories": "4", + "stopping_reason_count_repetition_guard": "4", + "avg_total_tokens_target_model": "8555.25", + "total_total_tokens_target_model": "34221.0", + "avg_total_tokens_other_models": "70742.00", + "total_total_tokens_other_models": "282968.0", + "avg_total_tokens_all_models": "79297.25", + "total_total_tokens_all_models": "317189.0", + "avg_turn_count": "35.25", + "total_turn_count": "141.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "3600", + "attempt_timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + }, + { + "evaluation_name": "std on healthbench/S-adaptive/+1ep/1cf893a8 for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "4", + "sample_ids": "6f827e79-05c4-4383-8ed4-16d36260f29f,c835bebe-e998-4dcd-843a-2de03d4aabc6,70afd1ac-85ea-4455-a5a2-15b35378e10a,aa7a760a-5173-4a39-af83-48d6ae2230f0" + } + }, + "evaluation_timestamp": "1776682987.0", + "metric_config": { + "evaluation_description": "std", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.04926275183504485, + "uncertainty": { + "standard_deviation": 0.04926275183504485, + "num_samples": 4 + }, + "details": { + "total_matched_trajectories": "4", + "stopping_reason_count_repetition_guard": "4", + "avg_total_tokens_target_model": "8555.25", + "total_total_tokens_target_model": "34221.0", + "avg_total_tokens_other_models": "70742.00", + "total_total_tokens_other_models": "282968.0", + "avg_total_tokens_all_models": "79297.25", + "total_total_tokens_all_models": "317189.0", + "avg_turn_count": "35.25", + "total_turn_count": "141.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "3600", + "attempt_timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + }, + { + "evaluation_name": "criteria_met_rate on healthbench/S-adaptive/+1ep/1cf893a8 for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "4", + "sample_ids": "6f827e79-05c4-4383-8ed4-16d36260f29f,c835bebe-e998-4dcd-843a-2de03d4aabc6,70afd1ac-85ea-4455-a5a2-15b35378e10a,aa7a760a-5173-4a39-af83-48d6ae2230f0" + } + }, + "evaluation_timestamp": "1776682987.0", + "metric_config": { + "evaluation_description": "criteria_met_rate", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.4375, + "uncertainty": { + "standard_deviation": 0.04926275183504485, + "num_samples": 4 + }, + "details": { + "total_matched_trajectories": "4", + "stopping_reason_count_repetition_guard": "4", + "avg_total_tokens_target_model": "8555.25", + "total_total_tokens_target_model": "34221.0", + "avg_total_tokens_other_models": "70742.00", + "total_total_tokens_other_models": "282968.0", + "avg_total_tokens_all_models": "79297.25", + "total_total_tokens_all_models": "317189.0", + "avg_turn_count": "35.25", + "total_turn_count": "141.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "3600", + "attempt_timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + }, + { + "evaluation_name": "mean on healthbench/S-adaptive/+1ep/1cf893a8 for scorer healthbench_score", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "4", + "sample_ids": "6f827e79-05c4-4383-8ed4-16d36260f29f,c835bebe-e998-4dcd-843a-2de03d4aabc6,70afd1ac-85ea-4455-a5a2-15b35378e10a,aa7a760a-5173-4a39-af83-48d6ae2230f0" + } + }, + "evaluation_timestamp": "1776682987.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.5658980409974902, + "uncertainty": { + "num_samples": 4 + }, + "details": { + "total_matched_trajectories": "4", + "stopping_reason_count_repetition_guard": "4", + "avg_total_tokens_target_model": "8555.25", + "total_total_tokens_target_model": "34221.0", + "avg_total_tokens_other_models": "70742.00", + "total_total_tokens_other_models": "282968.0", + "avg_total_tokens_all_models": "79297.25", + "total_total_tokens_all_models": "317189.0", + "avg_turn_count": "35.25", + "total_turn_count": "141.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "3600", + "attempt_timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/openai/gpt-5.2-2025-12-11/b6f0360d-51ae-4906-948a-9c4f94eee24d_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "cdf2d94905060e737b89dd6f85a77b5461de0e2a1354a3c7bdb7aa5a1feb71c4", + "total_rows": 4 + } +} \ No newline at end of file diff --git a/data/healthbench/openai/gpt-5.2-2025-12-11/b6f0360d-51ae-4906-948a-9c4f94eee24d_samples.jsonl b/data/healthbench/openai/gpt-5.2-2025-12-11/b6f0360d-51ae-4906-948a-9c4f94eee24d_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..977636563165dc60173623e5f314f8b57493ce68 --- /dev/null +++ b/data/healthbench/openai/gpt-5.2-2025-12-11/b6f0360d-51ae-4906-948a-9c4f94eee24d_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cdf2d94905060e737b89dd6f85a77b5461de0e2a1354a3c7bdb7aa5a1feb71c4 +size 335376 diff --git a/data/healthbench/openai/gpt-5.2-2025-12-11/ba8ad251-c25c-42df-9496-5026e690eb4c.json b/data/healthbench/openai/gpt-5.2-2025-12-11/ba8ad251-c25c-42df-9496-5026e690eb4c.json new file mode 100644 index 0000000000000000000000000000000000000000..5ab074a84d757da19d28df2adc8afe22a8de1cfc --- /dev/null +++ b/data/healthbench/openai/gpt-5.2-2025-12-11/ba8ad251-c25c-42df-9496-5026e690eb4c.json @@ -0,0 +1,304 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/openai_gpt-5.2-2025-12-11/1786163783.819131", + "evaluation_timestamp": "1776259833.0", + "retrieved_timestamp": "1786163783.819131", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "bootstrap_score on healthbench/S-adaptive/+4ep/1582ffe1 for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "51", + "sample_ids": "1f548d5b-cd00-49a0-b327-283a2e00debd,0b8f1d60-2081-4562-98f7-b6a976fe1c6d,6f7a2ee9-e9c6-42d8-b79f-22dea966b8d2,19ec4833-86e9-4166-8b82-d1da09f31fd7,7ebc830a-8dbd-489b-9d61-4d8bacf0db8d,c971f9d1-5f6a-464e-b282-41c8f0d82f58,5a6e4a41-3ea6-4050-a971-93433fe34877,78510411-e03b-4f93-a369-c7e72d2e4908,bbd759b4-8e4d-4db7-bc36-f2e069396c1c,8f2a65de-dea7-48e8-8adb-6194eca26c08,6d5f483c-3e86-456d-bfd5-4e28de699ae5,f05491d8-d160-4b2b-bd5b-59d757585e39,85408180-238e-4c1b-92c1-55d8ce6c3717,ecd7453e-1dc1-4e52-bb95-67e8e5788573,0e073591-b3e7-4dc8-94d0-dc5aa93ca35c,7637b987-8fe2-49d0-8372-1ebf17284d70,5b294937-13e5-424e-8bb4-5d1904a2344a,c5bf7fc3-dfb0-4b47-9bbe-222e87b952e2,83cf8f2d-2857-4f01-a283-9595d8f4ae8e,7a4548e6-38b7-48dd-9088-d7d1219f7852,e0d1b955-1150-457e-96d7-91a9481999cd,24f9a6e7-b214-4011-94c4-6502f249a621,6a97773b-c33c-4f2e-9ac3-fdd53de0f175,f32ee0c1-ea9c-4cfb-8332-bd9579d23024,b6eb69cb-b911-4c81-a7a1-357d21d27109,c0dc053b-157b-4d13-9956-213b67ff6a36,3bb4a735-ace7-4342-b7a5-0bae083e6f82,66b448bf-2e49-488d-9d0b-f9912ab3f0f8,b12d2453-fd3b-462c-b9cc-554ff16bfd3f,0a548d04-0973-4343-a3bf-07f730bed2f6,e2cd9f23-e3dd-4b67-8984-a5412dd802e5,1fb6ab91-a999-437d-8c93-c5937c0a89a3,42af0c9f-a715-45c4-ab24-c2948f6f943d,a94aad40-387e-4a37-b92f-61badd57388b,596ba714-29f7-447e-aa73-88d46ecd6e88,1233737e-2674-41d3-8519-b817aeb0b32e,c80a2a84-281d-41cc-a10b-32ee48c584f6,bf7718f2-bceb-4702-8d74-a7fc1bd7fb78,b7a82324-73ee-4996-8f71-8f455f789e90,22fe3eec-03b0-4cdf-a06a-87e71d236082,6c568ea6-86bd-4286-94ba-e3ccc8b67b21,905949d2-7a0a-4461-8f4b-257de6be6eed,c653b6eb-f9b9-4626-b702-ff73d71e8253,cf09b209-5434-4317-aa2a-7b07537b475e,77c0be2e-364a-40de-ae21-a277154e3d60,af750930-5fa5-4efe-8f85-b499641377d4,af62e0b8-bbd0-4dcc-b55c-cb4c7bbcfb2c,f3729e05-61ef-4306-9491-6f5659d4eef9,126236de-9e7a-4814-b5d7-98cbe89b33b2,9bb31d40-c155-4213-868d-103fa0cc7c56,97bec88b-6ef7-4f24-84d6-a2bcc592548b" + } + }, + "evaluation_timestamp": "1776259833.0", + "metric_config": { + "evaluation_description": "bootstrap_score", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.5276888876033371, + "uncertainty": { + "standard_deviation": 0.04085821178066376, + "num_samples": 204 + }, + "details": { + "total_matched_trajectories": "28", + "stopping_reason_count_repetition_guard": "28", + "avg_total_tokens_target_model": "13010.82", + "total_total_tokens_target_model": "364303.0", + "avg_total_tokens_other_models": "64499.39", + "total_total_tokens_other_models": "1805983.0", + "avg_total_tokens_all_models": "77510.21", + "total_total_tokens_all_models": "2170286.0", + "avg_turn_count": "27.39", + "total_turn_count": "767.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "reasoning_effort": "\"high\"" + } + } + }, + { + "evaluation_name": "std on healthbench/S-adaptive/+4ep/1582ffe1 for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "51", + "sample_ids": "1f548d5b-cd00-49a0-b327-283a2e00debd,0b8f1d60-2081-4562-98f7-b6a976fe1c6d,6f7a2ee9-e9c6-42d8-b79f-22dea966b8d2,19ec4833-86e9-4166-8b82-d1da09f31fd7,7ebc830a-8dbd-489b-9d61-4d8bacf0db8d,c971f9d1-5f6a-464e-b282-41c8f0d82f58,5a6e4a41-3ea6-4050-a971-93433fe34877,78510411-e03b-4f93-a369-c7e72d2e4908,bbd759b4-8e4d-4db7-bc36-f2e069396c1c,8f2a65de-dea7-48e8-8adb-6194eca26c08,6d5f483c-3e86-456d-bfd5-4e28de699ae5,f05491d8-d160-4b2b-bd5b-59d757585e39,85408180-238e-4c1b-92c1-55d8ce6c3717,ecd7453e-1dc1-4e52-bb95-67e8e5788573,0e073591-b3e7-4dc8-94d0-dc5aa93ca35c,7637b987-8fe2-49d0-8372-1ebf17284d70,5b294937-13e5-424e-8bb4-5d1904a2344a,c5bf7fc3-dfb0-4b47-9bbe-222e87b952e2,83cf8f2d-2857-4f01-a283-9595d8f4ae8e,7a4548e6-38b7-48dd-9088-d7d1219f7852,e0d1b955-1150-457e-96d7-91a9481999cd,24f9a6e7-b214-4011-94c4-6502f249a621,6a97773b-c33c-4f2e-9ac3-fdd53de0f175,f32ee0c1-ea9c-4cfb-8332-bd9579d23024,b6eb69cb-b911-4c81-a7a1-357d21d27109,c0dc053b-157b-4d13-9956-213b67ff6a36,3bb4a735-ace7-4342-b7a5-0bae083e6f82,66b448bf-2e49-488d-9d0b-f9912ab3f0f8,b12d2453-fd3b-462c-b9cc-554ff16bfd3f,0a548d04-0973-4343-a3bf-07f730bed2f6,e2cd9f23-e3dd-4b67-8984-a5412dd802e5,1fb6ab91-a999-437d-8c93-c5937c0a89a3,42af0c9f-a715-45c4-ab24-c2948f6f943d,a94aad40-387e-4a37-b92f-61badd57388b,596ba714-29f7-447e-aa73-88d46ecd6e88,1233737e-2674-41d3-8519-b817aeb0b32e,c80a2a84-281d-41cc-a10b-32ee48c584f6,bf7718f2-bceb-4702-8d74-a7fc1bd7fb78,b7a82324-73ee-4996-8f71-8f455f789e90,22fe3eec-03b0-4cdf-a06a-87e71d236082,6c568ea6-86bd-4286-94ba-e3ccc8b67b21,905949d2-7a0a-4461-8f4b-257de6be6eed,c653b6eb-f9b9-4626-b702-ff73d71e8253,cf09b209-5434-4317-aa2a-7b07537b475e,77c0be2e-364a-40de-ae21-a277154e3d60,af750930-5fa5-4efe-8f85-b499641377d4,af62e0b8-bbd0-4dcc-b55c-cb4c7bbcfb2c,f3729e05-61ef-4306-9491-6f5659d4eef9,126236de-9e7a-4814-b5d7-98cbe89b33b2,9bb31d40-c155-4213-868d-103fa0cc7c56,97bec88b-6ef7-4f24-84d6-a2bcc592548b" + } + }, + "evaluation_timestamp": "1776259833.0", + "metric_config": { + "evaluation_description": "std", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.04085821178066376, + "uncertainty": { + "standard_deviation": 0.04085821178066376, + "num_samples": 204 + }, + "details": { + "total_matched_trajectories": "28", + "stopping_reason_count_repetition_guard": "28", + "avg_total_tokens_target_model": "13010.82", + "total_total_tokens_target_model": "364303.0", + "avg_total_tokens_other_models": "64499.39", + "total_total_tokens_other_models": "1805983.0", + "avg_total_tokens_all_models": "77510.21", + "total_total_tokens_all_models": "2170286.0", + "avg_turn_count": "27.39", + "total_turn_count": "767.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "reasoning_effort": "\"high\"" + } + } + }, + { + "evaluation_name": "criteria_met_rate on healthbench/S-adaptive/+4ep/1582ffe1 for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "51", + "sample_ids": "1f548d5b-cd00-49a0-b327-283a2e00debd,0b8f1d60-2081-4562-98f7-b6a976fe1c6d,6f7a2ee9-e9c6-42d8-b79f-22dea966b8d2,19ec4833-86e9-4166-8b82-d1da09f31fd7,7ebc830a-8dbd-489b-9d61-4d8bacf0db8d,c971f9d1-5f6a-464e-b282-41c8f0d82f58,5a6e4a41-3ea6-4050-a971-93433fe34877,78510411-e03b-4f93-a369-c7e72d2e4908,bbd759b4-8e4d-4db7-bc36-f2e069396c1c,8f2a65de-dea7-48e8-8adb-6194eca26c08,6d5f483c-3e86-456d-bfd5-4e28de699ae5,f05491d8-d160-4b2b-bd5b-59d757585e39,85408180-238e-4c1b-92c1-55d8ce6c3717,ecd7453e-1dc1-4e52-bb95-67e8e5788573,0e073591-b3e7-4dc8-94d0-dc5aa93ca35c,7637b987-8fe2-49d0-8372-1ebf17284d70,5b294937-13e5-424e-8bb4-5d1904a2344a,c5bf7fc3-dfb0-4b47-9bbe-222e87b952e2,83cf8f2d-2857-4f01-a283-9595d8f4ae8e,7a4548e6-38b7-48dd-9088-d7d1219f7852,e0d1b955-1150-457e-96d7-91a9481999cd,24f9a6e7-b214-4011-94c4-6502f249a621,6a97773b-c33c-4f2e-9ac3-fdd53de0f175,f32ee0c1-ea9c-4cfb-8332-bd9579d23024,b6eb69cb-b911-4c81-a7a1-357d21d27109,c0dc053b-157b-4d13-9956-213b67ff6a36,3bb4a735-ace7-4342-b7a5-0bae083e6f82,66b448bf-2e49-488d-9d0b-f9912ab3f0f8,b12d2453-fd3b-462c-b9cc-554ff16bfd3f,0a548d04-0973-4343-a3bf-07f730bed2f6,e2cd9f23-e3dd-4b67-8984-a5412dd802e5,1fb6ab91-a999-437d-8c93-c5937c0a89a3,42af0c9f-a715-45c4-ab24-c2948f6f943d,a94aad40-387e-4a37-b92f-61badd57388b,596ba714-29f7-447e-aa73-88d46ecd6e88,1233737e-2674-41d3-8519-b817aeb0b32e,c80a2a84-281d-41cc-a10b-32ee48c584f6,bf7718f2-bceb-4702-8d74-a7fc1bd7fb78,b7a82324-73ee-4996-8f71-8f455f789e90,22fe3eec-03b0-4cdf-a06a-87e71d236082,6c568ea6-86bd-4286-94ba-e3ccc8b67b21,905949d2-7a0a-4461-8f4b-257de6be6eed,c653b6eb-f9b9-4626-b702-ff73d71e8253,cf09b209-5434-4317-aa2a-7b07537b475e,77c0be2e-364a-40de-ae21-a277154e3d60,af750930-5fa5-4efe-8f85-b499641377d4,af62e0b8-bbd0-4dcc-b55c-cb4c7bbcfb2c,f3729e05-61ef-4306-9491-6f5659d4eef9,126236de-9e7a-4814-b5d7-98cbe89b33b2,9bb31d40-c155-4213-868d-103fa0cc7c56,97bec88b-6ef7-4f24-84d6-a2bcc592548b" + } + }, + "evaluation_timestamp": "1776259833.0", + "metric_config": { + "evaluation_description": "criteria_met_rate", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.5013736263736264, + "uncertainty": { + "standard_deviation": 0.04085821178066376, + "num_samples": 204 + }, + "details": { + "total_matched_trajectories": "28", + "stopping_reason_count_repetition_guard": "28", + "avg_total_tokens_target_model": "13010.82", + "total_total_tokens_target_model": "364303.0", + "avg_total_tokens_other_models": "64499.39", + "total_total_tokens_other_models": "1805983.0", + "avg_total_tokens_all_models": "77510.21", + "total_total_tokens_all_models": "2170286.0", + "avg_turn_count": "27.39", + "total_turn_count": "767.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "reasoning_effort": "\"high\"" + } + } + }, + { + "evaluation_name": "mean on healthbench/S-adaptive/+4ep/1582ffe1 for scorer healthbench_score", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "51", + "sample_ids": "1f548d5b-cd00-49a0-b327-283a2e00debd,0b8f1d60-2081-4562-98f7-b6a976fe1c6d,6f7a2ee9-e9c6-42d8-b79f-22dea966b8d2,19ec4833-86e9-4166-8b82-d1da09f31fd7,7ebc830a-8dbd-489b-9d61-4d8bacf0db8d,c971f9d1-5f6a-464e-b282-41c8f0d82f58,5a6e4a41-3ea6-4050-a971-93433fe34877,78510411-e03b-4f93-a369-c7e72d2e4908,bbd759b4-8e4d-4db7-bc36-f2e069396c1c,8f2a65de-dea7-48e8-8adb-6194eca26c08,6d5f483c-3e86-456d-bfd5-4e28de699ae5,f05491d8-d160-4b2b-bd5b-59d757585e39,85408180-238e-4c1b-92c1-55d8ce6c3717,ecd7453e-1dc1-4e52-bb95-67e8e5788573,0e073591-b3e7-4dc8-94d0-dc5aa93ca35c,7637b987-8fe2-49d0-8372-1ebf17284d70,5b294937-13e5-424e-8bb4-5d1904a2344a,c5bf7fc3-dfb0-4b47-9bbe-222e87b952e2,83cf8f2d-2857-4f01-a283-9595d8f4ae8e,7a4548e6-38b7-48dd-9088-d7d1219f7852,e0d1b955-1150-457e-96d7-91a9481999cd,24f9a6e7-b214-4011-94c4-6502f249a621,6a97773b-c33c-4f2e-9ac3-fdd53de0f175,f32ee0c1-ea9c-4cfb-8332-bd9579d23024,b6eb69cb-b911-4c81-a7a1-357d21d27109,c0dc053b-157b-4d13-9956-213b67ff6a36,3bb4a735-ace7-4342-b7a5-0bae083e6f82,66b448bf-2e49-488d-9d0b-f9912ab3f0f8,b12d2453-fd3b-462c-b9cc-554ff16bfd3f,0a548d04-0973-4343-a3bf-07f730bed2f6,e2cd9f23-e3dd-4b67-8984-a5412dd802e5,1fb6ab91-a999-437d-8c93-c5937c0a89a3,42af0c9f-a715-45c4-ab24-c2948f6f943d,a94aad40-387e-4a37-b92f-61badd57388b,596ba714-29f7-447e-aa73-88d46ecd6e88,1233737e-2674-41d3-8519-b817aeb0b32e,c80a2a84-281d-41cc-a10b-32ee48c584f6,bf7718f2-bceb-4702-8d74-a7fc1bd7fb78,b7a82324-73ee-4996-8f71-8f455f789e90,22fe3eec-03b0-4cdf-a06a-87e71d236082,6c568ea6-86bd-4286-94ba-e3ccc8b67b21,905949d2-7a0a-4461-8f4b-257de6be6eed,c653b6eb-f9b9-4626-b702-ff73d71e8253,cf09b209-5434-4317-aa2a-7b07537b475e,77c0be2e-364a-40de-ae21-a277154e3d60,af750930-5fa5-4efe-8f85-b499641377d4,af62e0b8-bbd0-4dcc-b55c-cb4c7bbcfb2c,f3729e05-61ef-4306-9491-6f5659d4eef9,126236de-9e7a-4814-b5d7-98cbe89b33b2,9bb31d40-c155-4213-868d-103fa0cc7c56,97bec88b-6ef7-4f24-84d6-a2bcc592548b" + } + }, + "evaluation_timestamp": "1776259833.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.5276888876033371, + "uncertainty": { + "num_samples": 204 + }, + "details": { + "total_matched_trajectories": "28", + "stopping_reason_count_repetition_guard": "28", + "avg_total_tokens_target_model": "13010.82", + "total_total_tokens_target_model": "364303.0", + "avg_total_tokens_other_models": "64499.39", + "total_total_tokens_other_models": "1805983.0", + "avg_total_tokens_all_models": "77510.21", + "total_total_tokens_all_models": "2170286.0", + "avg_turn_count": "27.39", + "total_turn_count": "767.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/openai/gpt-5.2-2025-12-11/ba8ad251-c25c-42df-9496-5026e690eb4c_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "0c64aa886fa4f53064f842865a99dea3c989b8cdd0b571d8614feb5ff2478bdd", + "total_rows": 204 + } +} \ No newline at end of file diff --git a/data/healthbench/openai/gpt-5.2-2025-12-11/ba8ad251-c25c-42df-9496-5026e690eb4c_samples.jsonl b/data/healthbench/openai/gpt-5.2-2025-12-11/ba8ad251-c25c-42df-9496-5026e690eb4c_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..7d6732bcd739ec50271fc7dce4ec39cddc7df6e2 --- /dev/null +++ b/data/healthbench/openai/gpt-5.2-2025-12-11/ba8ad251-c25c-42df-9496-5026e690eb4c_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0c64aa886fa4f53064f842865a99dea3c989b8cdd0b571d8614feb5ff2478bdd +size 12594197 diff --git a/data/healthbench/openai/gpt-5.2-2025-12-11/d8b28eac-8491-4b0a-b23f-ed3b297c6ab8.json b/data/healthbench/openai/gpt-5.2-2025-12-11/d8b28eac-8491-4b0a-b23f-ed3b297c6ab8.json new file mode 100644 index 0000000000000000000000000000000000000000..dfb00e33fdc9faae2c06210555df314c76141517 --- /dev/null +++ b/data/healthbench/openai/gpt-5.2-2025-12-11/d8b28eac-8491-4b0a-b23f-ed3b297c6ab8.json @@ -0,0 +1,320 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/openai_gpt-5.2-2025-12-11/1786163790.099789", + "evaluation_timestamp": "1776725937.0", + "retrieved_timestamp": "1786163790.099789", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "bootstrap_score on healthbench/S-adaptive/c83d6e0c for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "26", + "sample_ids": "1dc0b953-ef66-4a78-9989-7cf1053b5b41,3cf35bfe-d881-4881-a268-b5fa0607bf89,63e4e0d4-0cb1-442e-8316-8c12a6e3ec14,99b22b39-30d5-40dc-82d9-c1cd3c793a44,8c342edc-619b-4079-a28c-b7129f3670a6,e8991792-c875-4f40-99a0-54c902d42a52,54349c88-e80a-4728-b181-305f97905d59,f93b0850-2c60-479d-831b-b8c8f9493d61,c63a4306-4003-45a6-8be1-d669457473a9,d3cbb579-70a5-46e8-95f6-72cc9afa4ba3,3852d050-c009-4e33-b953-fd734a5b5e73,5dc23430-dfbb-4794-9f77-0d4dd6093033,67e76447-f16a-4a65-aba7-fc7166fd9b63,a783e7a3-cecc-4d15-b18c-3f2ddfc2dc6c,62e9d35f-153d-4744-b97c-9abfb2a90b1e,04b6e0c8-2835-415e-865d-aca9f1cd3c57,7a283823-00ee-489d-8dbc-356e2857526f,e89e00c9-29e9-420c-813d-e06ddc39d585,51433193-b0c3-4725-a3a8-9dd15e969b0c,48c89cd1-0ecf-4738-8fcc-392269585bd3,92ea478e-6bf8-4ffb-b88a-f57ccf5f2346,36c6e780-d4a8-47b2-8a36-90ee4b97607f,5fb6812f-9816-4d5d-b7e0-4ec3e9df5dc1,c29c95d1-5b5e-4f53-9ec5-83e1f72b5aa5,b81c059a-1137-47cf-bab8-ab5eb8ff34d5,886864a6-a0e4-468f-aaf4-26cf8dc5e541" + } + }, + "evaluation_timestamp": "1776725937.0", + "metric_config": { + "evaluation_description": "bootstrap_score", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.31610998817076924, + "uncertainty": { + "standard_deviation": 0.03272473645227161, + "num_samples": 129 + }, + "details": { + "total_matched_trajectories": "129", + "stopping_reason_count_repetition_guard": "129", + "avg_total_tokens_target_model": "10996.45", + "total_total_tokens_target_model": "1418542.0", + "avg_total_tokens_other_models": "60331.32", + "total_total_tokens_other_models": "7782740.0", + "avg_total_tokens_all_models": "71327.77", + "total_total_tokens_all_models": "9201282.0", + "avg_turn_count": "27.35", + "total_turn_count": "3528.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "3600", + "attempt_timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + }, + { + "evaluation_name": "std on healthbench/S-adaptive/c83d6e0c for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "26", + "sample_ids": "1dc0b953-ef66-4a78-9989-7cf1053b5b41,3cf35bfe-d881-4881-a268-b5fa0607bf89,63e4e0d4-0cb1-442e-8316-8c12a6e3ec14,99b22b39-30d5-40dc-82d9-c1cd3c793a44,8c342edc-619b-4079-a28c-b7129f3670a6,e8991792-c875-4f40-99a0-54c902d42a52,54349c88-e80a-4728-b181-305f97905d59,f93b0850-2c60-479d-831b-b8c8f9493d61,c63a4306-4003-45a6-8be1-d669457473a9,d3cbb579-70a5-46e8-95f6-72cc9afa4ba3,3852d050-c009-4e33-b953-fd734a5b5e73,5dc23430-dfbb-4794-9f77-0d4dd6093033,67e76447-f16a-4a65-aba7-fc7166fd9b63,a783e7a3-cecc-4d15-b18c-3f2ddfc2dc6c,62e9d35f-153d-4744-b97c-9abfb2a90b1e,04b6e0c8-2835-415e-865d-aca9f1cd3c57,7a283823-00ee-489d-8dbc-356e2857526f,e89e00c9-29e9-420c-813d-e06ddc39d585,51433193-b0c3-4725-a3a8-9dd15e969b0c,48c89cd1-0ecf-4738-8fcc-392269585bd3,92ea478e-6bf8-4ffb-b88a-f57ccf5f2346,36c6e780-d4a8-47b2-8a36-90ee4b97607f,5fb6812f-9816-4d5d-b7e0-4ec3e9df5dc1,c29c95d1-5b5e-4f53-9ec5-83e1f72b5aa5,b81c059a-1137-47cf-bab8-ab5eb8ff34d5,886864a6-a0e4-468f-aaf4-26cf8dc5e541" + } + }, + "evaluation_timestamp": "1776725937.0", + "metric_config": { + "evaluation_description": "std", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.03272473645227161, + "uncertainty": { + "standard_deviation": 0.03272473645227161, + "num_samples": 129 + }, + "details": { + "total_matched_trajectories": "129", + "stopping_reason_count_repetition_guard": "129", + "avg_total_tokens_target_model": "10996.45", + "total_total_tokens_target_model": "1418542.0", + "avg_total_tokens_other_models": "60331.32", + "total_total_tokens_other_models": "7782740.0", + "avg_total_tokens_all_models": "71327.77", + "total_total_tokens_all_models": "9201282.0", + "avg_turn_count": "27.35", + "total_turn_count": "3528.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "3600", + "attempt_timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + }, + { + "evaluation_name": "criteria_met_rate on healthbench/S-adaptive/c83d6e0c for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "26", + "sample_ids": "1dc0b953-ef66-4a78-9989-7cf1053b5b41,3cf35bfe-d881-4881-a268-b5fa0607bf89,63e4e0d4-0cb1-442e-8316-8c12a6e3ec14,99b22b39-30d5-40dc-82d9-c1cd3c793a44,8c342edc-619b-4079-a28c-b7129f3670a6,e8991792-c875-4f40-99a0-54c902d42a52,54349c88-e80a-4728-b181-305f97905d59,f93b0850-2c60-479d-831b-b8c8f9493d61,c63a4306-4003-45a6-8be1-d669457473a9,d3cbb579-70a5-46e8-95f6-72cc9afa4ba3,3852d050-c009-4e33-b953-fd734a5b5e73,5dc23430-dfbb-4794-9f77-0d4dd6093033,67e76447-f16a-4a65-aba7-fc7166fd9b63,a783e7a3-cecc-4d15-b18c-3f2ddfc2dc6c,62e9d35f-153d-4744-b97c-9abfb2a90b1e,04b6e0c8-2835-415e-865d-aca9f1cd3c57,7a283823-00ee-489d-8dbc-356e2857526f,e89e00c9-29e9-420c-813d-e06ddc39d585,51433193-b0c3-4725-a3a8-9dd15e969b0c,48c89cd1-0ecf-4738-8fcc-392269585bd3,92ea478e-6bf8-4ffb-b88a-f57ccf5f2346,36c6e780-d4a8-47b2-8a36-90ee4b97607f,5fb6812f-9816-4d5d-b7e0-4ec3e9df5dc1,c29c95d1-5b5e-4f53-9ec5-83e1f72b5aa5,b81c059a-1137-47cf-bab8-ab5eb8ff34d5,886864a6-a0e4-468f-aaf4-26cf8dc5e541" + } + }, + "evaluation_timestamp": "1776725937.0", + "metric_config": { + "evaluation_description": "criteria_met_rate", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.4197115384615384, + "uncertainty": { + "standard_deviation": 0.03272473645227161, + "num_samples": 129 + }, + "details": { + "total_matched_trajectories": "129", + "stopping_reason_count_repetition_guard": "129", + "avg_total_tokens_target_model": "10996.45", + "total_total_tokens_target_model": "1418542.0", + "avg_total_tokens_other_models": "60331.32", + "total_total_tokens_other_models": "7782740.0", + "avg_total_tokens_all_models": "71327.77", + "total_total_tokens_all_models": "9201282.0", + "avg_turn_count": "27.35", + "total_turn_count": "3528.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "3600", + "attempt_timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + }, + { + "evaluation_name": "mean on healthbench/S-adaptive/c83d6e0c for scorer healthbench_score", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "26", + "sample_ids": "1dc0b953-ef66-4a78-9989-7cf1053b5b41,3cf35bfe-d881-4881-a268-b5fa0607bf89,63e4e0d4-0cb1-442e-8316-8c12a6e3ec14,99b22b39-30d5-40dc-82d9-c1cd3c793a44,8c342edc-619b-4079-a28c-b7129f3670a6,e8991792-c875-4f40-99a0-54c902d42a52,54349c88-e80a-4728-b181-305f97905d59,f93b0850-2c60-479d-831b-b8c8f9493d61,c63a4306-4003-45a6-8be1-d669457473a9,d3cbb579-70a5-46e8-95f6-72cc9afa4ba3,3852d050-c009-4e33-b953-fd734a5b5e73,5dc23430-dfbb-4794-9f77-0d4dd6093033,67e76447-f16a-4a65-aba7-fc7166fd9b63,a783e7a3-cecc-4d15-b18c-3f2ddfc2dc6c,62e9d35f-153d-4744-b97c-9abfb2a90b1e,04b6e0c8-2835-415e-865d-aca9f1cd3c57,7a283823-00ee-489d-8dbc-356e2857526f,e89e00c9-29e9-420c-813d-e06ddc39d585,51433193-b0c3-4725-a3a8-9dd15e969b0c,48c89cd1-0ecf-4738-8fcc-392269585bd3,92ea478e-6bf8-4ffb-b88a-f57ccf5f2346,36c6e780-d4a8-47b2-8a36-90ee4b97607f,5fb6812f-9816-4d5d-b7e0-4ec3e9df5dc1,c29c95d1-5b5e-4f53-9ec5-83e1f72b5aa5,b81c059a-1137-47cf-bab8-ab5eb8ff34d5,886864a6-a0e4-468f-aaf4-26cf8dc5e541" + } + }, + "evaluation_timestamp": "1776725937.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.31610998817076924, + "uncertainty": { + "num_samples": 129 + }, + "details": { + "total_matched_trajectories": "129", + "stopping_reason_count_repetition_guard": "129", + "avg_total_tokens_target_model": "10996.45", + "total_total_tokens_target_model": "1418542.0", + "avg_total_tokens_other_models": "60331.32", + "total_total_tokens_other_models": "7782740.0", + "avg_total_tokens_all_models": "71327.77", + "total_total_tokens_all_models": "9201282.0", + "avg_turn_count": "27.35", + "total_turn_count": "3528.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "3600", + "attempt_timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/openai/gpt-5.2-2025-12-11/d8b28eac-8491-4b0a-b23f-ed3b297c6ab8_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "4db2955cff31b7b3dd1bc7efcfb7879605ae81a01ad31366f560ede3bafec7f5", + "total_rows": 129 + } +} \ No newline at end of file diff --git a/data/healthbench/openai/gpt-5.2-2025-12-11/d8b28eac-8491-4b0a-b23f-ed3b297c6ab8_samples.jsonl b/data/healthbench/openai/gpt-5.2-2025-12-11/d8b28eac-8491-4b0a-b23f-ed3b297c6ab8_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..9ccd616172fa3495d8e7939ecfa07ebb02bb7eed --- /dev/null +++ b/data/healthbench/openai/gpt-5.2-2025-12-11/d8b28eac-8491-4b0a-b23f-ed3b297c6ab8_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4db2955cff31b7b3dd1bc7efcfb7879605ae81a01ad31366f560ede3bafec7f5 +size 9409781 diff --git a/data/healthbench/openai/gpt-5.2-2025-12-11/de990ae0-28f9-4827-b221-edf8e852211c.json b/data/healthbench/openai/gpt-5.2-2025-12-11/de990ae0-28f9-4827-b221-edf8e852211c.json new file mode 100644 index 0000000000000000000000000000000000000000..23a461bda51afd1a4e4de41499c8c41af77581cc --- /dev/null +++ b/data/healthbench/openai/gpt-5.2-2025-12-11/de990ae0-28f9-4827-b221-edf8e852211c.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/openai_gpt-5.2-2025-12-11/1786166412.080299", + "evaluation_timestamp": "1776696116.0", + "retrieved_timestamp": "1786166412.080299", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/openai/gpt-5.2-2025-12-11/de990ae0-28f9-4827-b221-edf8e852211c_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "f8e7ca442e2d15e0aae18b01c735a56e17a888afe2d00ec5b1ef88b0f0444e2c", + "total_rows": 190 + } +} \ No newline at end of file diff --git a/data/healthbench/openai/gpt-5.2-2025-12-11/de990ae0-28f9-4827-b221-edf8e852211c_samples.jsonl b/data/healthbench/openai/gpt-5.2-2025-12-11/de990ae0-28f9-4827-b221-edf8e852211c_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..88322e39cb1a3b5938158d9e6f45b4721090ae78 --- /dev/null +++ b/data/healthbench/openai/gpt-5.2-2025-12-11/de990ae0-28f9-4827-b221-edf8e852211c_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f8e7ca442e2d15e0aae18b01c735a56e17a888afe2d00ec5b1ef88b0f0444e2c +size 13598450 diff --git a/data/healthbench/openai/gpt-5.2-2025-12-11/e7eca9ef-5929-41a6-9216-1ec1bcbdab78.json b/data/healthbench/openai/gpt-5.2-2025-12-11/e7eca9ef-5929-41a6-9216-1ec1bcbdab78.json new file mode 100644 index 0000000000000000000000000000000000000000..7d54806772ba7e8ab6f966d92ee100d0d9460f61 --- /dev/null +++ b/data/healthbench/openai/gpt-5.2-2025-12-11/e7eca9ef-5929-41a6-9216-1ec1bcbdab78.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/openai_gpt-5.2-2025-12-11/1786168345.679857", + "evaluation_timestamp": "1776165241.0", + "retrieved_timestamp": "1786168345.679857", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/openai/gpt-5.2-2025-12-11/e7eca9ef-5929-41a6-9216-1ec1bcbdab78_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "53974f8b7f476644b783e5775027678c16ae1a7e5b8a0ea0f846fa54d80ec6c8", + "total_rows": 451 + } +} \ No newline at end of file diff --git a/data/healthbench/openai/gpt-5.2-2025-12-11/e7eca9ef-5929-41a6-9216-1ec1bcbdab78_samples.jsonl b/data/healthbench/openai/gpt-5.2-2025-12-11/e7eca9ef-5929-41a6-9216-1ec1bcbdab78_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..89899732a3f6fe84be30c282c6098e6244a780bf --- /dev/null +++ b/data/healthbench/openai/gpt-5.2-2025-12-11/e7eca9ef-5929-41a6-9216-1ec1bcbdab78_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:53974f8b7f476644b783e5775027678c16ae1a7e5b8a0ea0f846fa54d80ec6c8 +size 29060931 diff --git a/data/healthbench/openai/gpt-5.2-2025-12-11/ebb70e90-fae6-4e7c-a34c-f5ed13ff644d.json b/data/healthbench/openai/gpt-5.2-2025-12-11/ebb70e90-fae6-4e7c-a34c-f5ed13ff644d.json new file mode 100644 index 0000000000000000000000000000000000000000..4312802c334e4c3c55efaef206a7f1767f577387 --- /dev/null +++ b/data/healthbench/openai/gpt-5.2-2025-12-11/ebb70e90-fae6-4e7c-a34c-f5ed13ff644d.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/openai_gpt-5.2-2025-12-11/1786168295.418032", + "evaluation_timestamp": "1776682987.0", + "retrieved_timestamp": "1786168295.418032", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/openai/gpt-5.2-2025-12-11/ebb70e90-fae6-4e7c-a34c-f5ed13ff644d_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "8d72fe3bb62ab6891d69f45c864312f0d0372b33577dc70f86a21f2d3617fbf3", + "total_rows": 40 + } +} \ No newline at end of file diff --git a/data/healthbench/openai/gpt-5.2-2025-12-11/ebb70e90-fae6-4e7c-a34c-f5ed13ff644d_samples.jsonl b/data/healthbench/openai/gpt-5.2-2025-12-11/ebb70e90-fae6-4e7c-a34c-f5ed13ff644d_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..af5cb67cdc6b7bafcf17f4e1ad8d787c82b071a8 --- /dev/null +++ b/data/healthbench/openai/gpt-5.2-2025-12-11/ebb70e90-fae6-4e7c-a34c-f5ed13ff644d_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8d72fe3bb62ab6891d69f45c864312f0d0372b33577dc70f86a21f2d3617fbf3 +size 2680108 diff --git a/data/healthbench/openai/gpt-5.2-2025-12-11/ec4a379b-5b1a-4fff-a303-2b12c9b64305.json b/data/healthbench/openai/gpt-5.2-2025-12-11/ec4a379b-5b1a-4fff-a303-2b12c9b64305.json new file mode 100644 index 0000000000000000000000000000000000000000..a0f2dc046b6cf7b4b45a898d91601ab8015eb859 --- /dev/null +++ b/data/healthbench/openai/gpt-5.2-2025-12-11/ec4a379b-5b1a-4fff-a303-2b12c9b64305.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/openai_gpt-5.2-2025-12-11/1786168321.579167", + "evaluation_timestamp": "1776165242.0", + "retrieved_timestamp": "1786168321.579167", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/openai/gpt-5.2-2025-12-11/ec4a379b-5b1a-4fff-a303-2b12c9b64305_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "d05d15bf6912ed50f7091b2fc81b92aafcb656d22f971fd4b3709e16440e4d9b", + "total_rows": 300 + } +} \ No newline at end of file diff --git a/data/healthbench/openai/gpt-5.2-2025-12-11/ec4a379b-5b1a-4fff-a303-2b12c9b64305_samples.jsonl b/data/healthbench/openai/gpt-5.2-2025-12-11/ec4a379b-5b1a-4fff-a303-2b12c9b64305_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..22b3d0f8a0ffb44157b7c4497d9fcdcee6463016 --- /dev/null +++ b/data/healthbench/openai/gpt-5.2-2025-12-11/ec4a379b-5b1a-4fff-a303-2b12c9b64305_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d05d15bf6912ed50f7091b2fc81b92aafcb656d22f971fd4b3709e16440e4d9b +size 19205173 diff --git a/data/healthbench/openai/gpt-5.2-2025-12-11/ecb51382-86ee-4c6c-96c1-b647b2919644.json b/data/healthbench/openai/gpt-5.2-2025-12-11/ecb51382-86ee-4c6c-96c1-b647b2919644.json new file mode 100644 index 0000000000000000000000000000000000000000..34dc8220dbdee715194a01d1696ee6e994f7a50e --- /dev/null +++ b/data/healthbench/openai/gpt-5.2-2025-12-11/ecb51382-86ee-4c6c-96c1-b647b2919644.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/openai_gpt-5.2-2025-12-11/1786166364.166433", + "evaluation_timestamp": "1776682987.0", + "retrieved_timestamp": "1786166364.166433", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/openai/gpt-5.2-2025-12-11/ecb51382-86ee-4c6c-96c1-b647b2919644_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "dcac67919fd4a5b9c8f34c3a09d4298d1b336b6c0a5093e7411c6f9ee0cdb92d", + "total_rows": 28 + } +} \ No newline at end of file diff --git a/data/healthbench/openai/gpt-5.2-2025-12-11/ecb51382-86ee-4c6c-96c1-b647b2919644_samples.jsonl b/data/healthbench/openai/gpt-5.2-2025-12-11/ecb51382-86ee-4c6c-96c1-b647b2919644_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..2689ddc2b579ed5d73e0787b9564aa3342c66c29 --- /dev/null +++ b/data/healthbench/openai/gpt-5.2-2025-12-11/ecb51382-86ee-4c6c-96c1-b647b2919644_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dcac67919fd4a5b9c8f34c3a09d4298d1b336b6c0a5093e7411c6f9ee0cdb92d +size 1875756 diff --git a/data/healthbench/openai/gpt-5.2-2025-12-11/ed4fb961-785e-47c0-ad12-213c06dbe4cf.json b/data/healthbench/openai/gpt-5.2-2025-12-11/ed4fb961-785e-47c0-ad12-213c06dbe4cf.json new file mode 100644 index 0000000000000000000000000000000000000000..454429106aab21913174a865f47786799f80fd77 --- /dev/null +++ b/data/healthbench/openai/gpt-5.2-2025-12-11/ed4fb961-785e-47c0-ad12-213c06dbe4cf.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/openai_gpt-5.2-2025-12-11/1786168289.578732", + "evaluation_timestamp": "1775646127.0", + "retrieved_timestamp": "1786168289.578732", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/openai/gpt-5.2-2025-12-11/ed4fb961-785e-47c0-ad12-213c06dbe4cf_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "a7fb9d96b1f3e87f3551e3cb1afeeadf91873597bb37b682680fe6b926fd5c35", + "total_rows": 100 + } +} \ No newline at end of file diff --git a/data/healthbench/openai/gpt-5.2-2025-12-11/ed4fb961-785e-47c0-ad12-213c06dbe4cf_samples.jsonl b/data/healthbench/openai/gpt-5.2-2025-12-11/ed4fb961-785e-47c0-ad12-213c06dbe4cf_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..2327bd28786af398bd61b023a2498d945b293ff8 --- /dev/null +++ b/data/healthbench/openai/gpt-5.2-2025-12-11/ed4fb961-785e-47c0-ad12-213c06dbe4cf_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a7fb9d96b1f3e87f3551e3cb1afeeadf91873597bb37b682680fe6b926fd5c35 +size 6514357 diff --git a/data/healthbench/openai/gpt-5.2-2025-12-11/f95f585c-18e5-4723-8827-a5de54106bca.json b/data/healthbench/openai/gpt-5.2-2025-12-11/f95f585c-18e5-4723-8827-a5de54106bca.json new file mode 100644 index 0000000000000000000000000000000000000000..00f149ee7558ebb2e50f5a4ef449afb4ef7f9e24 --- /dev/null +++ b/data/healthbench/openai/gpt-5.2-2025-12-11/f95f585c-18e5-4723-8827-a5de54106bca.json @@ -0,0 +1,256 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/openai_gpt-5.2-2025-12-11/1786163795.048472", + "evaluation_timestamp": "1776301373.0", + "retrieved_timestamp": "1786163795.048472", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "bootstrap_score on healthbench/S-adaptive/+8ep/42982016 for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "1", + "sample_ids": "06792558-4afa-4aa8-b311-1fc8493ec61e" + } + }, + "evaluation_timestamp": "1776301373.0", + "metric_config": { + "evaluation_description": "bootstrap_score", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.5282258064516129, + "uncertainty": { + "standard_deviation": 0.0, + "num_samples": 8 + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "reasoning_effort": "\"high\"" + } + } + }, + { + "evaluation_name": "std on healthbench/S-adaptive/+8ep/42982016 for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "1", + "sample_ids": "06792558-4afa-4aa8-b311-1fc8493ec61e" + } + }, + "evaluation_timestamp": "1776301373.0", + "metric_config": { + "evaluation_description": "std", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.0, + "uncertainty": { + "standard_deviation": 0.0, + "num_samples": 8 + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "reasoning_effort": "\"high\"" + } + } + }, + { + "evaluation_name": "criteria_met_rate on healthbench/S-adaptive/+8ep/42982016 for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "1", + "sample_ids": "06792558-4afa-4aa8-b311-1fc8493ec61e" + } + }, + "evaluation_timestamp": "1776301373.0", + "metric_config": { + "evaluation_description": "criteria_met_rate", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.7109375, + "uncertainty": { + "standard_deviation": 0.0, + "num_samples": 8 + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "reasoning_effort": "\"high\"" + } + } + }, + { + "evaluation_name": "mean on healthbench/S-adaptive/+8ep/42982016 for scorer healthbench_score", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "1", + "sample_ids": "06792558-4afa-4aa8-b311-1fc8493ec61e" + } + }, + "evaluation_timestamp": "1776301373.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.5282258064516129, + "uncertainty": { + "num_samples": 8 + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/openai/gpt-5.2-2025-12-11/f95f585c-18e5-4723-8827-a5de54106bca_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "8e4de2dc5f3186a8c3301b462024944952827679e8bc5f932138b87e97729932", + "total_rows": 8 + } +} \ No newline at end of file diff --git a/data/healthbench/openai/gpt-5.2-2025-12-11/f95f585c-18e5-4723-8827-a5de54106bca_samples.jsonl b/data/healthbench/openai/gpt-5.2-2025-12-11/f95f585c-18e5-4723-8827-a5de54106bca_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..9be8923310be8ee4962787da95df0a3d2506d4f3 --- /dev/null +++ b/data/healthbench/openai/gpt-5.2-2025-12-11/f95f585c-18e5-4723-8827-a5de54106bca_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8e4de2dc5f3186a8c3301b462024944952827679e8bc5f932138b87e97729932 +size 681685 diff --git a/data/healthbench/openai/gpt-5.2-2025-12-11/f9853ab0-f743-4561-baf0-2e6a853639c9.json b/data/healthbench/openai/gpt-5.2-2025-12-11/f9853ab0-f743-4561-baf0-2e6a853639c9.json new file mode 100644 index 0000000000000000000000000000000000000000..eaf8988812981df689354a409cb8c0550caee9e4 --- /dev/null +++ b/data/healthbench/openai/gpt-5.2-2025-12-11/f9853ab0-f743-4561-baf0-2e6a853639c9.json @@ -0,0 +1,256 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/openai_gpt-5.2-2025-12-11/1786163793.02185", + "evaluation_timestamp": "1776290559.0", + "retrieved_timestamp": "1786163793.02185", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "bootstrap_score on healthbench/S-adaptive/+6ep/3818a4f2 for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "1", + "sample_ids": "3bcc3abf-5e76-4d3d-b0bb-c74cd0e5f4d8" + } + }, + "evaluation_timestamp": "1776290559.0", + "metric_config": { + "evaluation_description": "bootstrap_score", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.8, + "uncertainty": { + "standard_deviation": 1.1102230246251565e-16, + "num_samples": 6 + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "reasoning_effort": "\"high\"" + } + } + }, + { + "evaluation_name": "std on healthbench/S-adaptive/+6ep/3818a4f2 for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "1", + "sample_ids": "3bcc3abf-5e76-4d3d-b0bb-c74cd0e5f4d8" + } + }, + "evaluation_timestamp": "1776290559.0", + "metric_config": { + "evaluation_description": "std", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 1.1102230246251565e-16, + "uncertainty": { + "standard_deviation": 1.1102230246251565e-16, + "num_samples": 6 + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "reasoning_effort": "\"high\"" + } + } + }, + { + "evaluation_name": "criteria_met_rate on healthbench/S-adaptive/+6ep/3818a4f2 for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "1", + "sample_ids": "3bcc3abf-5e76-4d3d-b0bb-c74cd0e5f4d8" + } + }, + "evaluation_timestamp": "1776290559.0", + "metric_config": { + "evaluation_description": "criteria_met_rate", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.5208333333333334, + "uncertainty": { + "standard_deviation": 1.1102230246251565e-16, + "num_samples": 6 + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "reasoning_effort": "\"high\"" + } + } + }, + { + "evaluation_name": "mean on healthbench/S-adaptive/+6ep/3818a4f2 for scorer healthbench_score", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "1", + "sample_ids": "3bcc3abf-5e76-4d3d-b0bb-c74cd0e5f4d8" + } + }, + "evaluation_timestamp": "1776290559.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.8, + "uncertainty": { + "num_samples": 6 + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/openai/gpt-5.2-2025-12-11/f9853ab0-f743-4561-baf0-2e6a853639c9_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "a47b08ad06909ce35d97ce276e6b5f1978a73af78b120443f0e117d66e9b5ddc", + "total_rows": 6 + } +} \ No newline at end of file diff --git a/data/healthbench/openai/gpt-5.2-2025-12-11/f9853ab0-f743-4561-baf0-2e6a853639c9_samples.jsonl b/data/healthbench/openai/gpt-5.2-2025-12-11/f9853ab0-f743-4561-baf0-2e6a853639c9_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..969a79dec8fefed87892f134655dbb668aa35a17 --- /dev/null +++ b/data/healthbench/openai/gpt-5.2-2025-12-11/f9853ab0-f743-4561-baf0-2e6a853639c9_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a47b08ad06909ce35d97ce276e6b5f1978a73af78b120443f0e117d66e9b5ddc +size 248653 diff --git a/data/healthbench/openai/gpt-5.4-2026-03-05/007a969d-643b-439f-852b-36f8ee9f55ba.json b/data/healthbench/openai/gpt-5.4-2026-03-05/007a969d-643b-439f-852b-36f8ee9f55ba.json new file mode 100644 index 0000000000000000000000000000000000000000..101829d9f5c0c456d23c2e31a7ba638dd96d9037 --- /dev/null +++ b/data/healthbench/openai/gpt-5.4-2026-03-05/007a969d-643b-439f-852b-36f8ee9f55ba.json @@ -0,0 +1,320 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/openai_gpt-5.4-2026-03-05/1786163797.216629", + "evaluation_timestamp": "1776682987.0", + "retrieved_timestamp": "1786163797.216629", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "bootstrap_score on healthbench/S-adaptive/+1ep/f314a009 for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "15", + "sample_ids": "1049130c-e9c9-461d-b080-90027bc011c0,13bb02e4-844e-4972-9200-2befa5f911d8,2694a753-2f57-41eb-a7f9-8c4c7acdd2e8,b160df9f-be98-4a4d-9cc8-11a612ffaf87,5e0a84d7-b972-484c-a649-469b15f0a90f,c7606b1b-ca4d-4a71-8277-bcddff09a653,20cdebb7-9189-4f09-ad21-bfc8e8b75680,f24935ac-0530-4063-b2f1-ac45cd265037,1b8f338a-2e98-4a42-ba80-cba266f4b9fa,6f827e79-05c4-4383-8ed4-16d36260f29f,c835bebe-e998-4dcd-843a-2de03d4aabc6,70afd1ac-85ea-4455-a5a2-15b35378e10a,aa7a760a-5173-4a39-af83-48d6ae2230f0,f712eb09-82af-4fe8-800d-bc3af772a9ba,8ee6a990-8966-4b42-baa8-6c7e9f74ce74" + } + }, + "evaluation_timestamp": "1776682987.0", + "metric_config": { + "evaluation_description": "bootstrap_score", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.36588664714254887, + "uncertainty": { + "standard_deviation": 0.06691390058534903, + "num_samples": 15 + }, + "details": { + "total_matched_trajectories": "15", + "stopping_reason_count_repetition_guard": "15", + "avg_total_tokens_target_model": "10513.13", + "total_total_tokens_target_model": "157697.0", + "avg_total_tokens_other_models": "57858.87", + "total_total_tokens_other_models": "867883.0", + "avg_total_tokens_all_models": "68372.00", + "total_total_tokens_all_models": "1025580.0", + "avg_turn_count": "28.87", + "total_turn_count": "433.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "3600", + "attempt_timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + }, + { + "evaluation_name": "std on healthbench/S-adaptive/+1ep/f314a009 for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "15", + "sample_ids": "1049130c-e9c9-461d-b080-90027bc011c0,13bb02e4-844e-4972-9200-2befa5f911d8,2694a753-2f57-41eb-a7f9-8c4c7acdd2e8,b160df9f-be98-4a4d-9cc8-11a612ffaf87,5e0a84d7-b972-484c-a649-469b15f0a90f,c7606b1b-ca4d-4a71-8277-bcddff09a653,20cdebb7-9189-4f09-ad21-bfc8e8b75680,f24935ac-0530-4063-b2f1-ac45cd265037,1b8f338a-2e98-4a42-ba80-cba266f4b9fa,6f827e79-05c4-4383-8ed4-16d36260f29f,c835bebe-e998-4dcd-843a-2de03d4aabc6,70afd1ac-85ea-4455-a5a2-15b35378e10a,aa7a760a-5173-4a39-af83-48d6ae2230f0,f712eb09-82af-4fe8-800d-bc3af772a9ba,8ee6a990-8966-4b42-baa8-6c7e9f74ce74" + } + }, + "evaluation_timestamp": "1776682987.0", + "metric_config": { + "evaluation_description": "std", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.06691390058534903, + "uncertainty": { + "standard_deviation": 0.06691390058534903, + "num_samples": 15 + }, + "details": { + "total_matched_trajectories": "15", + "stopping_reason_count_repetition_guard": "15", + "avg_total_tokens_target_model": "10513.13", + "total_total_tokens_target_model": "157697.0", + "avg_total_tokens_other_models": "57858.87", + "total_total_tokens_other_models": "867883.0", + "avg_total_tokens_all_models": "68372.00", + "total_total_tokens_all_models": "1025580.0", + "avg_turn_count": "28.87", + "total_turn_count": "433.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "3600", + "attempt_timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + }, + { + "evaluation_name": "criteria_met_rate on healthbench/S-adaptive/+1ep/f314a009 for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "15", + "sample_ids": "1049130c-e9c9-461d-b080-90027bc011c0,13bb02e4-844e-4972-9200-2befa5f911d8,2694a753-2f57-41eb-a7f9-8c4c7acdd2e8,b160df9f-be98-4a4d-9cc8-11a612ffaf87,5e0a84d7-b972-484c-a649-469b15f0a90f,c7606b1b-ca4d-4a71-8277-bcddff09a653,20cdebb7-9189-4f09-ad21-bfc8e8b75680,f24935ac-0530-4063-b2f1-ac45cd265037,1b8f338a-2e98-4a42-ba80-cba266f4b9fa,6f827e79-05c4-4383-8ed4-16d36260f29f,c835bebe-e998-4dcd-843a-2de03d4aabc6,70afd1ac-85ea-4455-a5a2-15b35378e10a,aa7a760a-5173-4a39-af83-48d6ae2230f0,f712eb09-82af-4fe8-800d-bc3af772a9ba,8ee6a990-8966-4b42-baa8-6c7e9f74ce74" + } + }, + "evaluation_timestamp": "1776682987.0", + "metric_config": { + "evaluation_description": "criteria_met_rate", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.40625, + "uncertainty": { + "standard_deviation": 0.06691390058534903, + "num_samples": 15 + }, + "details": { + "total_matched_trajectories": "15", + "stopping_reason_count_repetition_guard": "15", + "avg_total_tokens_target_model": "10513.13", + "total_total_tokens_target_model": "157697.0", + "avg_total_tokens_other_models": "57858.87", + "total_total_tokens_other_models": "867883.0", + "avg_total_tokens_all_models": "68372.00", + "total_total_tokens_all_models": "1025580.0", + "avg_turn_count": "28.87", + "total_turn_count": "433.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "3600", + "attempt_timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + }, + { + "evaluation_name": "mean on healthbench/S-adaptive/+1ep/f314a009 for scorer healthbench_score", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "15", + "sample_ids": "1049130c-e9c9-461d-b080-90027bc011c0,13bb02e4-844e-4972-9200-2befa5f911d8,2694a753-2f57-41eb-a7f9-8c4c7acdd2e8,b160df9f-be98-4a4d-9cc8-11a612ffaf87,5e0a84d7-b972-484c-a649-469b15f0a90f,c7606b1b-ca4d-4a71-8277-bcddff09a653,20cdebb7-9189-4f09-ad21-bfc8e8b75680,f24935ac-0530-4063-b2f1-ac45cd265037,1b8f338a-2e98-4a42-ba80-cba266f4b9fa,6f827e79-05c4-4383-8ed4-16d36260f29f,c835bebe-e998-4dcd-843a-2de03d4aabc6,70afd1ac-85ea-4455-a5a2-15b35378e10a,aa7a760a-5173-4a39-af83-48d6ae2230f0,f712eb09-82af-4fe8-800d-bc3af772a9ba,8ee6a990-8966-4b42-baa8-6c7e9f74ce74" + } + }, + "evaluation_timestamp": "1776682987.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.36588664714254887, + "uncertainty": { + "num_samples": 15 + }, + "details": { + "total_matched_trajectories": "15", + "stopping_reason_count_repetition_guard": "15", + "avg_total_tokens_target_model": "10513.13", + "total_total_tokens_target_model": "157697.0", + "avg_total_tokens_other_models": "57858.87", + "total_total_tokens_other_models": "867883.0", + "avg_total_tokens_all_models": "68372.00", + "total_total_tokens_all_models": "1025580.0", + "avg_turn_count": "28.87", + "total_turn_count": "433.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "3600", + "attempt_timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/openai/gpt-5.4-2026-03-05/007a969d-643b-439f-852b-36f8ee9f55ba_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "b83ba1b9e469b67fcea3f1b8db63b1a70e5a00eed85045c7d6521262ade7e982", + "total_rows": 15 + } +} \ No newline at end of file diff --git a/data/healthbench/openai/gpt-5.4-2026-03-05/007a969d-643b-439f-852b-36f8ee9f55ba_samples.jsonl b/data/healthbench/openai/gpt-5.4-2026-03-05/007a969d-643b-439f-852b-36f8ee9f55ba_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..67304c7ed4dbc372932846fbf0ce4007b7292304 --- /dev/null +++ b/data/healthbench/openai/gpt-5.4-2026-03-05/007a969d-643b-439f-852b-36f8ee9f55ba_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b83ba1b9e469b67fcea3f1b8db63b1a70e5a00eed85045c7d6521262ade7e982 +size 1057280 diff --git a/data/healthbench/openai/gpt-5.4-2026-03-05/01983a11-a63b-4524-91e1-a9cba85f864d.json b/data/healthbench/openai/gpt-5.4-2026-03-05/01983a11-a63b-4524-91e1-a9cba85f864d.json new file mode 100644 index 0000000000000000000000000000000000000000..248160496074573eb0b13a5272cc7e3b0dc834cc --- /dev/null +++ b/data/healthbench/openai/gpt-5.4-2026-03-05/01983a11-a63b-4524-91e1-a9cba85f864d.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/openai_gpt-5.4-2026-03-05/1786166447.04052", + "evaluation_timestamp": "1776696116.0", + "retrieved_timestamp": "1786166447.04052", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/openai/gpt-5.4-2026-03-05/01983a11-a63b-4524-91e1-a9cba85f864d_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "81c98b1020b6db597a55a95b4e9ad8fbf380a10afc30494a240601cd0e090b4d", + "total_rows": 209 + } +} \ No newline at end of file diff --git a/data/healthbench/openai/gpt-5.4-2026-03-05/01983a11-a63b-4524-91e1-a9cba85f864d_samples.jsonl b/data/healthbench/openai/gpt-5.4-2026-03-05/01983a11-a63b-4524-91e1-a9cba85f864d_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..cd630fa2b4f61f5544bde7aee36f885e6702d4b7 --- /dev/null +++ b/data/healthbench/openai/gpt-5.4-2026-03-05/01983a11-a63b-4524-91e1-a9cba85f864d_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:81c98b1020b6db597a55a95b4e9ad8fbf380a10afc30494a240601cd0e090b4d +size 14647246 diff --git a/data/healthbench/openai/gpt-5.4-2026-03-05/1cd81328-5daa-4a52-8a03-c9f89fb328af.json b/data/healthbench/openai/gpt-5.4-2026-03-05/1cd81328-5daa-4a52-8a03-c9f89fb328af.json new file mode 100644 index 0000000000000000000000000000000000000000..07734e34e3208d5e5381863b095f4775cd048e75 --- /dev/null +++ b/data/healthbench/openai/gpt-5.4-2026-03-05/1cd81328-5daa-4a52-8a03-c9f89fb328af.json @@ -0,0 +1,320 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/openai_gpt-5.4-2026-03-05/1786163803.076373", + "evaluation_timestamp": "1776803719.0", + "retrieved_timestamp": "1786163803.076373", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "bootstrap_score on healthbench/S-adaptive/+1ep/9ef228bf for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "2", + "sample_ids": "85c11e2a-628b-464f-a8bb-e97ab04a73eb,c63a4306-4003-45a6-8be1-d669457473a9" + } + }, + "evaluation_timestamp": "1776803719.0", + "metric_config": { + "evaluation_description": "bootstrap_score", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.4147727272727273, + "uncertainty": { + "standard_deviation": 0.0763158467886872, + "num_samples": 2 + }, + "details": { + "total_matched_trajectories": "2", + "stopping_reason_count_repetition_guard": "2", + "avg_total_tokens_target_model": "36570.50", + "total_total_tokens_target_model": "73141.0", + "avg_total_tokens_other_models": "140115.50", + "total_total_tokens_other_models": "280231.0", + "avg_total_tokens_all_models": "176686.00", + "total_total_tokens_all_models": "353372.0", + "avg_turn_count": "34.00", + "total_turn_count": "68.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + }, + { + "evaluation_name": "std on healthbench/S-adaptive/+1ep/9ef228bf for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "2", + "sample_ids": "85c11e2a-628b-464f-a8bb-e97ab04a73eb,c63a4306-4003-45a6-8be1-d669457473a9" + } + }, + "evaluation_timestamp": "1776803719.0", + "metric_config": { + "evaluation_description": "std", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.0763158467886872, + "uncertainty": { + "standard_deviation": 0.0763158467886872, + "num_samples": 2 + }, + "details": { + "total_matched_trajectories": "2", + "stopping_reason_count_repetition_guard": "2", + "avg_total_tokens_target_model": "36570.50", + "total_total_tokens_target_model": "73141.0", + "avg_total_tokens_other_models": "140115.50", + "total_total_tokens_other_models": "280231.0", + "avg_total_tokens_all_models": "176686.00", + "total_total_tokens_all_models": "353372.0", + "avg_turn_count": "34.00", + "total_turn_count": "68.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + }, + { + "evaluation_name": "criteria_met_rate on healthbench/S-adaptive/+1ep/9ef228bf for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "2", + "sample_ids": "85c11e2a-628b-464f-a8bb-e97ab04a73eb,c63a4306-4003-45a6-8be1-d669457473a9" + } + }, + "evaluation_timestamp": "1776803719.0", + "metric_config": { + "evaluation_description": "criteria_met_rate", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.3870967741935484, + "uncertainty": { + "standard_deviation": 0.0763158467886872, + "num_samples": 2 + }, + "details": { + "total_matched_trajectories": "2", + "stopping_reason_count_repetition_guard": "2", + "avg_total_tokens_target_model": "36570.50", + "total_total_tokens_target_model": "73141.0", + "avg_total_tokens_other_models": "140115.50", + "total_total_tokens_other_models": "280231.0", + "avg_total_tokens_all_models": "176686.00", + "total_total_tokens_all_models": "353372.0", + "avg_turn_count": "34.00", + "total_turn_count": "68.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + }, + { + "evaluation_name": "mean on healthbench/S-adaptive/+1ep/9ef228bf for scorer healthbench_score", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "2", + "sample_ids": "85c11e2a-628b-464f-a8bb-e97ab04a73eb,c63a4306-4003-45a6-8be1-d669457473a9" + } + }, + "evaluation_timestamp": "1776803719.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.4147727272727273, + "uncertainty": { + "num_samples": 2 + }, + "details": { + "total_matched_trajectories": "2", + "stopping_reason_count_repetition_guard": "2", + "avg_total_tokens_target_model": "36570.50", + "total_total_tokens_target_model": "73141.0", + "avg_total_tokens_other_models": "140115.50", + "total_total_tokens_other_models": "280231.0", + "avg_total_tokens_all_models": "176686.00", + "total_total_tokens_all_models": "353372.0", + "avg_turn_count": "34.00", + "total_turn_count": "68.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/openai/gpt-5.4-2026-03-05/1cd81328-5daa-4a52-8a03-c9f89fb328af_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "58e89386e558bb12b3fdae158a7724cf7194626e906bc983f0fb9f86a9d66785", + "total_rows": 2 + } +} \ No newline at end of file diff --git a/data/healthbench/openai/gpt-5.4-2026-03-05/1cd81328-5daa-4a52-8a03-c9f89fb328af_samples.jsonl b/data/healthbench/openai/gpt-5.4-2026-03-05/1cd81328-5daa-4a52-8a03-c9f89fb328af_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..ae9596665ac3296b4a9eb2ebf17a7c133390766a --- /dev/null +++ b/data/healthbench/openai/gpt-5.4-2026-03-05/1cd81328-5daa-4a52-8a03-c9f89fb328af_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:58e89386e558bb12b3fdae158a7724cf7194626e906bc983f0fb9f86a9d66785 +size 195172 diff --git a/data/healthbench/openai/gpt-5.4-2026-03-05/2e2a8266-7847-4040-9981-d9e7213b47b5.json b/data/healthbench/openai/gpt-5.4-2026-03-05/2e2a8266-7847-4040-9981-d9e7213b47b5.json new file mode 100644 index 0000000000000000000000000000000000000000..c6273711e61761a2a6c20b614ee3555137652c67 --- /dev/null +++ b/data/healthbench/openai/gpt-5.4-2026-03-05/2e2a8266-7847-4040-9981-d9e7213b47b5.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/openai_gpt-5.4-2026-03-05/1786168370.23994", + "evaluation_timestamp": "1775604415.0", + "retrieved_timestamp": "1786168370.23994", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.202,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/openai/gpt-5.4-2026-03-05/2e2a8266-7847-4040-9981-d9e7213b47b5_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "8186d21df865cc6505a4dd51f0d9d79062b57bdd41fdf98f26fc28de14f2f862", + "total_rows": 100 + } +} \ No newline at end of file diff --git a/data/healthbench/openai/gpt-5.4-2026-03-05/2e2a8266-7847-4040-9981-d9e7213b47b5_samples.jsonl b/data/healthbench/openai/gpt-5.4-2026-03-05/2e2a8266-7847-4040-9981-d9e7213b47b5_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..22a08c375d7325243f13222dacacb5ca1a79dc2e --- /dev/null +++ b/data/healthbench/openai/gpt-5.4-2026-03-05/2e2a8266-7847-4040-9981-d9e7213b47b5_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8186d21df865cc6505a4dd51f0d9d79062b57bdd41fdf98f26fc28de14f2f862 +size 6249061 diff --git a/data/healthbench/openai/gpt-5.4-2026-03-05/4925dd4e-9ff2-4811-afb9-d2ad7d1161c6.json b/data/healthbench/openai/gpt-5.4-2026-03-05/4925dd4e-9ff2-4811-afb9-d2ad7d1161c6.json new file mode 100644 index 0000000000000000000000000000000000000000..a459f2a6e1c0c53acbb128d2a5de64228f502233 --- /dev/null +++ b/data/healthbench/openai/gpt-5.4-2026-03-05/4925dd4e-9ff2-4811-afb9-d2ad7d1161c6.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/openai_gpt-5.4-2026-03-05/1786168385.898087", + "evaluation_timestamp": "1776682987.0", + "retrieved_timestamp": "1786168385.898087", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/openai/gpt-5.4-2026-03-05/4925dd4e-9ff2-4811-afb9-d2ad7d1161c6_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "c5a942c116831a96bfad7e4b3f294115efdb0a84468d10e4bcfcc83825cd6c06", + "total_rows": 57 + } +} \ No newline at end of file diff --git a/data/healthbench/openai/gpt-5.4-2026-03-05/4925dd4e-9ff2-4811-afb9-d2ad7d1161c6_samples.jsonl b/data/healthbench/openai/gpt-5.4-2026-03-05/4925dd4e-9ff2-4811-afb9-d2ad7d1161c6_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..6e10c8ba6253c684404599be09c8837abca57dc6 --- /dev/null +++ b/data/healthbench/openai/gpt-5.4-2026-03-05/4925dd4e-9ff2-4811-afb9-d2ad7d1161c6_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c5a942c116831a96bfad7e4b3f294115efdb0a84468d10e4bcfcc83825cd6c06 +size 3721237 diff --git a/data/healthbench/openai/gpt-5.4-2026-03-05/5a6f7f95-a535-4c18-8e93-5d38024ea269.json b/data/healthbench/openai/gpt-5.4-2026-03-05/5a6f7f95-a535-4c18-8e93-5d38024ea269.json new file mode 100644 index 0000000000000000000000000000000000000000..66dec6e226104f9d016e2e6c6094fdcb4c8a799f --- /dev/null +++ b/data/healthbench/openai/gpt-5.4-2026-03-05/5a6f7f95-a535-4c18-8e93-5d38024ea269.json @@ -0,0 +1,312 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/openai_gpt-5.4-2026-03-05/1786182985.881935", + "evaluation_timestamp": "1774953547.0", + "retrieved_timestamp": "1786182985.881935", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.191.dev111+g7c0c4345,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "bootstrap_score on healthbench for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "5000", + "sample_ids": "1f548d5b-cd00-49a0-b327-283a2e00debd,0b8f1d60-2081-4562-98f7-b6a976fe1c6d,6f7a2ee9-e9c6-42d8-b79f-22dea966b8d2,19ec4833-86e9-4166-8b82-d1da09f31fd7,7ebc830a-8dbd-489b-9d61-4d8bacf0db8d" + } + }, + "evaluation_timestamp": "1774953547.0", + "metric_config": { + "evaluation_description": "bootstrap_score", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.4659809744555507, + "uncertainty": { + "standard_deviation": 0.09482835645300884, + "num_samples": 10 + }, + "details": { + "total_matched_trajectories": "2", + "stopping_reason_count_repetition_guard": "2", + "avg_total_tokens_target_model": "68089.00", + "total_total_tokens_target_model": "136178.0", + "avg_total_tokens_other_models": "126142.50", + "total_total_tokens_other_models": "252285.0", + "avg_total_tokens_all_models": "194231.50", + "total_total_tokens_all_models": "388463.0", + "avg_turn_count": "29.00", + "total_turn_count": "58.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\"}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "3", + "attempt_timeout": "1200", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {}, + "max_attempts": 3 + }, + "additional_details": { + "max_retries": "3" + } + } + }, + { + "evaluation_name": "std on healthbench for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "5000", + "sample_ids": "1f548d5b-cd00-49a0-b327-283a2e00debd,0b8f1d60-2081-4562-98f7-b6a976fe1c6d,6f7a2ee9-e9c6-42d8-b79f-22dea966b8d2,19ec4833-86e9-4166-8b82-d1da09f31fd7,7ebc830a-8dbd-489b-9d61-4d8bacf0db8d" + } + }, + "evaluation_timestamp": "1774953547.0", + "metric_config": { + "evaluation_description": "std", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.09482835645300884, + "uncertainty": { + "standard_deviation": 0.09482835645300884, + "num_samples": 10 + }, + "details": { + "total_matched_trajectories": "2", + "stopping_reason_count_repetition_guard": "2", + "avg_total_tokens_target_model": "68089.00", + "total_total_tokens_target_model": "136178.0", + "avg_total_tokens_other_models": "126142.50", + "total_total_tokens_other_models": "252285.0", + "avg_total_tokens_all_models": "194231.50", + "total_total_tokens_all_models": "388463.0", + "avg_turn_count": "29.00", + "total_turn_count": "58.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\"}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "3", + "attempt_timeout": "1200", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {}, + "max_attempts": 3 + }, + "additional_details": { + "max_retries": "3" + } + } + }, + { + "evaluation_name": "criteria_met_rate on healthbench for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "5000", + "sample_ids": "1f548d5b-cd00-49a0-b327-283a2e00debd,0b8f1d60-2081-4562-98f7-b6a976fe1c6d,6f7a2ee9-e9c6-42d8-b79f-22dea966b8d2,19ec4833-86e9-4166-8b82-d1da09f31fd7,7ebc830a-8dbd-489b-9d61-4d8bacf0db8d" + } + }, + "evaluation_timestamp": "1774953547.0", + "metric_config": { + "evaluation_description": "criteria_met_rate", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.4453125, + "uncertainty": { + "standard_deviation": 0.09482835645300884, + "num_samples": 10 + }, + "details": { + "total_matched_trajectories": "2", + "stopping_reason_count_repetition_guard": "2", + "avg_total_tokens_target_model": "68089.00", + "total_total_tokens_target_model": "136178.0", + "avg_total_tokens_other_models": "126142.50", + "total_total_tokens_other_models": "252285.0", + "avg_total_tokens_all_models": "194231.50", + "total_total_tokens_all_models": "388463.0", + "avg_turn_count": "29.00", + "total_turn_count": "58.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\"}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "3", + "attempt_timeout": "1200", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {}, + "max_attempts": 3 + }, + "additional_details": { + "max_retries": "3" + } + } + }, + { + "evaluation_name": "mean on healthbench for scorer healthbench_score", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "5000", + "sample_ids": "1f548d5b-cd00-49a0-b327-283a2e00debd,0b8f1d60-2081-4562-98f7-b6a976fe1c6d,6f7a2ee9-e9c6-42d8-b79f-22dea966b8d2,19ec4833-86e9-4166-8b82-d1da09f31fd7,7ebc830a-8dbd-489b-9d61-4d8bacf0db8d" + } + }, + "evaluation_timestamp": "1774953547.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.4659809744555507, + "uncertainty": { + "num_samples": 10 + }, + "details": { + "total_matched_trajectories": "2", + "stopping_reason_count_repetition_guard": "2", + "avg_total_tokens_target_model": "68089.00", + "total_total_tokens_target_model": "136178.0", + "avg_total_tokens_other_models": "126142.50", + "total_total_tokens_other_models": "252285.0", + "avg_total_tokens_all_models": "194231.50", + "total_total_tokens_all_models": "388463.0", + "avg_turn_count": "29.00", + "total_turn_count": "58.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\"}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "3", + "attempt_timeout": "1200", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {}, + "max_attempts": 3 + }, + "additional_details": { + "max_retries": "3" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/openai/gpt-5.4-2026-03-05/5a6f7f95-a535-4c18-8e93-5d38024ea269_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "511a8b3ceb2440a15d15a06173713a31cced604068f3708d23fda75b8477d99e", + "total_rows": 10 + } +} \ No newline at end of file diff --git a/data/healthbench/openai/gpt-5.4-2026-03-05/5a6f7f95-a535-4c18-8e93-5d38024ea269_samples.jsonl b/data/healthbench/openai/gpt-5.4-2026-03-05/5a6f7f95-a535-4c18-8e93-5d38024ea269_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..4613059b89f1ad9f2af7085fa46afe5db3772471 --- /dev/null +++ b/data/healthbench/openai/gpt-5.4-2026-03-05/5a6f7f95-a535-4c18-8e93-5d38024ea269_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:511a8b3ceb2440a15d15a06173713a31cced604068f3708d23fda75b8477d99e +size 7166993 diff --git a/data/healthbench/openai/gpt-5.4-2026-03-05/5d501243-8b76-48e2-a819-8cf2cb54ff15.json b/data/healthbench/openai/gpt-5.4-2026-03-05/5d501243-8b76-48e2-a819-8cf2cb54ff15.json new file mode 100644 index 0000000000000000000000000000000000000000..a69a67023d6e272b5fa9c417907e4f40cd9c77ab --- /dev/null +++ b/data/healthbench/openai/gpt-5.4-2026-03-05/5d501243-8b76-48e2-a819-8cf2cb54ff15.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/openai_gpt-5.4-2026-03-05/1786168412.317752", + "evaluation_timestamp": "1776165242.0", + "retrieved_timestamp": "1786168412.317752", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/openai/gpt-5.4-2026-03-05/5d501243-8b76-48e2-a819-8cf2cb54ff15_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "186a86c40b161b8f57df67efc2749c467c7df925620ef44d477eb2703419e427", + "total_rows": 300 + } +} \ No newline at end of file diff --git a/data/healthbench/openai/gpt-5.4-2026-03-05/5d501243-8b76-48e2-a819-8cf2cb54ff15_samples.jsonl b/data/healthbench/openai/gpt-5.4-2026-03-05/5d501243-8b76-48e2-a819-8cf2cb54ff15_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..aad17151f7b86e60aeb6c97fda3c5561e6d06370 --- /dev/null +++ b/data/healthbench/openai/gpt-5.4-2026-03-05/5d501243-8b76-48e2-a819-8cf2cb54ff15_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:186a86c40b161b8f57df67efc2749c467c7df925620ef44d477eb2703419e427 +size 18321825 diff --git a/data/healthbench/openai/gpt-5.4-2026-03-05/64e9b2b7-c070-46d4-a91d-b14cee456301.json b/data/healthbench/openai/gpt-5.4-2026-03-05/64e9b2b7-c070-46d4-a91d-b14cee456301.json new file mode 100644 index 0000000000000000000000000000000000000000..52200a6b9dbeec88f6876d21e2165b6977c7a6ba --- /dev/null +++ b/data/healthbench/openai/gpt-5.4-2026-03-05/64e9b2b7-c070-46d4-a91d-b14cee456301.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/openai_gpt-5.4-2026-03-05/1786166468.670361", + "evaluation_timestamp": "1776165241.0", + "retrieved_timestamp": "1786166468.670361", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/openai/gpt-5.4-2026-03-05/64e9b2b7-c070-46d4-a91d-b14cee456301_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "129aa8220a14424238a78dbf3f0d446a3fabab3deefb2fe15601e4fc11691d6c", + "total_rows": 381 + } +} \ No newline at end of file diff --git a/data/healthbench/openai/gpt-5.4-2026-03-05/64e9b2b7-c070-46d4-a91d-b14cee456301_samples.jsonl b/data/healthbench/openai/gpt-5.4-2026-03-05/64e9b2b7-c070-46d4-a91d-b14cee456301_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..5d89813bde223bd13893bf7f80997d8cf445d8e2 --- /dev/null +++ b/data/healthbench/openai/gpt-5.4-2026-03-05/64e9b2b7-c070-46d4-a91d-b14cee456301_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:129aa8220a14424238a78dbf3f0d446a3fabab3deefb2fe15601e4fc11691d6c +size 23940960 diff --git a/data/healthbench/openai/gpt-5.4-2026-03-05/65d619b9-dca4-4527-bce4-d1d6d5e08afd.json b/data/healthbench/openai/gpt-5.4-2026-03-05/65d619b9-dca4-4527-bce4-d1d6d5e08afd.json new file mode 100644 index 0000000000000000000000000000000000000000..3fe815b02c933ffa4a57cf401302bca03bd4e54d --- /dev/null +++ b/data/healthbench/openai/gpt-5.4-2026-03-05/65d619b9-dca4-4527-bce4-d1d6d5e08afd.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/openai_gpt-5.4-2026-03-05/1786168380.929513", + "evaluation_timestamp": "1776682987.0", + "retrieved_timestamp": "1786168380.929513", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/openai/gpt-5.4-2026-03-05/65d619b9-dca4-4527-bce4-d1d6d5e08afd_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "26389699d759a01493e3edcf9a6863eb1f93030ac9debd1ddbf3538f9222b10f", + "total_rows": 49 + } +} \ No newline at end of file diff --git a/data/healthbench/openai/gpt-5.4-2026-03-05/65d619b9-dca4-4527-bce4-d1d6d5e08afd_samples.jsonl b/data/healthbench/openai/gpt-5.4-2026-03-05/65d619b9-dca4-4527-bce4-d1d6d5e08afd_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..1ff537cd596f31ad1cbd3a77657adc33d4307f34 --- /dev/null +++ b/data/healthbench/openai/gpt-5.4-2026-03-05/65d619b9-dca4-4527-bce4-d1d6d5e08afd_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:26389699d759a01493e3edcf9a6863eb1f93030ac9debd1ddbf3538f9222b10f +size 3194291 diff --git a/data/healthbench/openai/gpt-5.4-2026-03-05/74cfb0e9-274f-4ecd-ab44-8f728055f25c.json b/data/healthbench/openai/gpt-5.4-2026-03-05/74cfb0e9-274f-4ecd-ab44-8f728055f25c.json new file mode 100644 index 0000000000000000000000000000000000000000..b48dacce1b70134fec15e4f8f8942b6f30c0cde6 --- /dev/null +++ b/data/healthbench/openai/gpt-5.4-2026-03-05/74cfb0e9-274f-4ecd-ab44-8f728055f25c.json @@ -0,0 +1,312 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/openai_gpt-5.4-2026-03-05/1786182982.041737", + "evaluation_timestamp": "1774959431.0", + "retrieved_timestamp": "1786182982.041737", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.191.dev111+g7c0c4345,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "bootstrap_score on healthbench for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "5000", + "sample_ids": "1f548d5b-cd00-49a0-b327-283a2e00debd,0b8f1d60-2081-4562-98f7-b6a976fe1c6d,6f7a2ee9-e9c6-42d8-b79f-22dea966b8d2,19ec4833-86e9-4166-8b82-d1da09f31fd7,7ebc830a-8dbd-489b-9d61-4d8bacf0db8d" + } + }, + "evaluation_timestamp": "1774959431.0", + "metric_config": { + "evaluation_description": "bootstrap_score", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.7455609592332756, + "uncertainty": { + "standard_deviation": 0.09611048714401113, + "num_samples": 10 + }, + "details": { + "total_matched_trajectories": "2", + "stopping_reason_count_repetition_guard": "2", + "avg_total_tokens_target_model": "275145.00", + "total_total_tokens_target_model": "550290.0", + "avg_total_tokens_other_models": "251592.50", + "total_total_tokens_other_models": "503185.0", + "avg_total_tokens_all_models": "526737.50", + "total_total_tokens_all_models": "1053475.0", + "avg_turn_count": "77.00", + "total_turn_count": "154.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"healthbench_score_value\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\"}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"healthbench_score_value\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "3", + "attempt_timeout": "1200", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {}, + "max_attempts": 3 + }, + "additional_details": { + "max_retries": "3" + } + } + }, + { + "evaluation_name": "std on healthbench for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "5000", + "sample_ids": "1f548d5b-cd00-49a0-b327-283a2e00debd,0b8f1d60-2081-4562-98f7-b6a976fe1c6d,6f7a2ee9-e9c6-42d8-b79f-22dea966b8d2,19ec4833-86e9-4166-8b82-d1da09f31fd7,7ebc830a-8dbd-489b-9d61-4d8bacf0db8d" + } + }, + "evaluation_timestamp": "1774959431.0", + "metric_config": { + "evaluation_description": "std", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.09611048714401113, + "uncertainty": { + "standard_deviation": 0.09611048714401113, + "num_samples": 10 + }, + "details": { + "total_matched_trajectories": "2", + "stopping_reason_count_repetition_guard": "2", + "avg_total_tokens_target_model": "275145.00", + "total_total_tokens_target_model": "550290.0", + "avg_total_tokens_other_models": "251592.50", + "total_total_tokens_other_models": "503185.0", + "avg_total_tokens_all_models": "526737.50", + "total_total_tokens_all_models": "1053475.0", + "avg_turn_count": "77.00", + "total_turn_count": "154.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"healthbench_score_value\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\"}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"healthbench_score_value\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "3", + "attempt_timeout": "1200", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {}, + "max_attempts": 3 + }, + "additional_details": { + "max_retries": "3" + } + } + }, + { + "evaluation_name": "criteria_met_rate on healthbench for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "5000", + "sample_ids": "1f548d5b-cd00-49a0-b327-283a2e00debd,0b8f1d60-2081-4562-98f7-b6a976fe1c6d,6f7a2ee9-e9c6-42d8-b79f-22dea966b8d2,19ec4833-86e9-4166-8b82-d1da09f31fd7,7ebc830a-8dbd-489b-9d61-4d8bacf0db8d" + } + }, + "evaluation_timestamp": "1774959431.0", + "metric_config": { + "evaluation_description": "criteria_met_rate", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.609375, + "uncertainty": { + "standard_deviation": 0.09611048714401113, + "num_samples": 10 + }, + "details": { + "total_matched_trajectories": "2", + "stopping_reason_count_repetition_guard": "2", + "avg_total_tokens_target_model": "275145.00", + "total_total_tokens_target_model": "550290.0", + "avg_total_tokens_other_models": "251592.50", + "total_total_tokens_other_models": "503185.0", + "avg_total_tokens_all_models": "526737.50", + "total_total_tokens_all_models": "1053475.0", + "avg_turn_count": "77.00", + "total_turn_count": "154.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"healthbench_score_value\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\"}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"healthbench_score_value\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "3", + "attempt_timeout": "1200", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {}, + "max_attempts": 3 + }, + "additional_details": { + "max_retries": "3" + } + } + }, + { + "evaluation_name": "mean on healthbench for scorer healthbench_score", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "5000", + "sample_ids": "1f548d5b-cd00-49a0-b327-283a2e00debd,0b8f1d60-2081-4562-98f7-b6a976fe1c6d,6f7a2ee9-e9c6-42d8-b79f-22dea966b8d2,19ec4833-86e9-4166-8b82-d1da09f31fd7,7ebc830a-8dbd-489b-9d61-4d8bacf0db8d" + } + }, + "evaluation_timestamp": "1774959431.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.7455609592332756, + "uncertainty": { + "num_samples": 10 + }, + "details": { + "total_matched_trajectories": "2", + "stopping_reason_count_repetition_guard": "2", + "avg_total_tokens_target_model": "275145.00", + "total_total_tokens_target_model": "550290.0", + "avg_total_tokens_other_models": "251592.50", + "total_total_tokens_other_models": "503185.0", + "avg_total_tokens_all_models": "526737.50", + "total_total_tokens_all_models": "1053475.0", + "avg_turn_count": "77.00", + "total_turn_count": "154.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"healthbench_score_value\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\"}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"healthbench_score_value\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "3", + "attempt_timeout": "1200", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {}, + "max_attempts": 3 + }, + "additional_details": { + "max_retries": "3" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/openai/gpt-5.4-2026-03-05/74cfb0e9-274f-4ecd-ab44-8f728055f25c_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "6f2da67d22d5d2d0adea880e04d9d1f059857f5035e9938727dae798d6ba0739", + "total_rows": 10 + } +} \ No newline at end of file diff --git a/data/healthbench/openai/gpt-5.4-2026-03-05/74cfb0e9-274f-4ecd-ab44-8f728055f25c_samples.jsonl b/data/healthbench/openai/gpt-5.4-2026-03-05/74cfb0e9-274f-4ecd-ab44-8f728055f25c_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..0e7b9f68b5360cb98d80a91c868beffda41b0bbf --- /dev/null +++ b/data/healthbench/openai/gpt-5.4-2026-03-05/74cfb0e9-274f-4ecd-ab44-8f728055f25c_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6f2da67d22d5d2d0adea880e04d9d1f059857f5035e9938727dae798d6ba0739 +size 1852283 diff --git a/data/healthbench/openai/gpt-5.4-2026-03-05/74d97efa-8b96-4da6-a86e-78cc45349320.json b/data/healthbench/openai/gpt-5.4-2026-03-05/74d97efa-8b96-4da6-a86e-78cc45349320.json new file mode 100644 index 0000000000000000000000000000000000000000..f4f58a49b046bf05c5f1df4fc0a39ba7207627be --- /dev/null +++ b/data/healthbench/openai/gpt-5.4-2026-03-05/74d97efa-8b96-4da6-a86e-78cc45349320.json @@ -0,0 +1,320 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/openai_gpt-5.4-2026-03-05/1786163798.959596", + "evaluation_timestamp": "1776779722.0", + "retrieved_timestamp": "1786163798.959596", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "bootstrap_score on healthbench/S-adaptive/+1ep/37d3534f for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "2", + "sample_ids": "54349c88-e80a-4728-b181-305f97905d59,5dc23430-dfbb-4794-9f77-0d4dd6093033" + } + }, + "evaluation_timestamp": "1776779722.0", + "metric_config": { + "evaluation_description": "bootstrap_score", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.1205547652916074, + "uncertainty": { + "standard_deviation": 0.028972584158551893, + "num_samples": 2 + }, + "details": { + "total_matched_trajectories": "2", + "stopping_reason_count_repetition_guard": "2", + "avg_total_tokens_target_model": "14611.50", + "total_total_tokens_target_model": "29223.0", + "avg_total_tokens_other_models": "47543.50", + "total_total_tokens_other_models": "95087.0", + "avg_total_tokens_all_models": "62155.00", + "total_total_tokens_all_models": "124310.0", + "avg_turn_count": "21.00", + "total_turn_count": "42.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + }, + { + "evaluation_name": "std on healthbench/S-adaptive/+1ep/37d3534f for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "2", + "sample_ids": "54349c88-e80a-4728-b181-305f97905d59,5dc23430-dfbb-4794-9f77-0d4dd6093033" + } + }, + "evaluation_timestamp": "1776779722.0", + "metric_config": { + "evaluation_description": "std", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.028972584158551893, + "uncertainty": { + "standard_deviation": 0.028972584158551893, + "num_samples": 2 + }, + "details": { + "total_matched_trajectories": "2", + "stopping_reason_count_repetition_guard": "2", + "avg_total_tokens_target_model": "14611.50", + "total_total_tokens_target_model": "29223.0", + "avg_total_tokens_other_models": "47543.50", + "total_total_tokens_other_models": "95087.0", + "avg_total_tokens_all_models": "62155.00", + "total_total_tokens_all_models": "124310.0", + "avg_turn_count": "21.00", + "total_turn_count": "42.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + }, + { + "evaluation_name": "criteria_met_rate on healthbench/S-adaptive/+1ep/37d3534f for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "2", + "sample_ids": "54349c88-e80a-4728-b181-305f97905d59,5dc23430-dfbb-4794-9f77-0d4dd6093033" + } + }, + "evaluation_timestamp": "1776779722.0", + "metric_config": { + "evaluation_description": "criteria_met_rate", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.3333333333333333, + "uncertainty": { + "standard_deviation": 0.028972584158551893, + "num_samples": 2 + }, + "details": { + "total_matched_trajectories": "2", + "stopping_reason_count_repetition_guard": "2", + "avg_total_tokens_target_model": "14611.50", + "total_total_tokens_target_model": "29223.0", + "avg_total_tokens_other_models": "47543.50", + "total_total_tokens_other_models": "95087.0", + "avg_total_tokens_all_models": "62155.00", + "total_total_tokens_all_models": "124310.0", + "avg_turn_count": "21.00", + "total_turn_count": "42.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + }, + { + "evaluation_name": "mean on healthbench/S-adaptive/+1ep/37d3534f for scorer healthbench_score", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "2", + "sample_ids": "54349c88-e80a-4728-b181-305f97905d59,5dc23430-dfbb-4794-9f77-0d4dd6093033" + } + }, + "evaluation_timestamp": "1776779722.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.1205547652916074, + "uncertainty": { + "num_samples": 2 + }, + "details": { + "total_matched_trajectories": "2", + "stopping_reason_count_repetition_guard": "2", + "avg_total_tokens_target_model": "14611.50", + "total_total_tokens_target_model": "29223.0", + "avg_total_tokens_other_models": "47543.50", + "total_total_tokens_other_models": "95087.0", + "avg_total_tokens_all_models": "62155.00", + "total_total_tokens_all_models": "124310.0", + "avg_turn_count": "21.00", + "total_turn_count": "42.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/openai/gpt-5.4-2026-03-05/74d97efa-8b96-4da6-a86e-78cc45349320_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "1421503e7b341da69d11fe028026fa182ae8703dc6e8b551d97d5d656b45ba7e", + "total_rows": 2 + } +} \ No newline at end of file diff --git a/data/healthbench/openai/gpt-5.4-2026-03-05/74d97efa-8b96-4da6-a86e-78cc45349320_samples.jsonl b/data/healthbench/openai/gpt-5.4-2026-03-05/74d97efa-8b96-4da6-a86e-78cc45349320_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..edbf8fc90f564f7b4d5f8a08977e98944f38678e --- /dev/null +++ b/data/healthbench/openai/gpt-5.4-2026-03-05/74d97efa-8b96-4da6-a86e-78cc45349320_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1421503e7b341da69d11fe028026fa182ae8703dc6e8b551d97d5d656b45ba7e +size 124384 diff --git a/data/healthbench/openai/gpt-5.4-2026-03-05/9074f6e8-71ca-4b54-9379-b9cff702cdf9.json b/data/healthbench/openai/gpt-5.4-2026-03-05/9074f6e8-71ca-4b54-9379-b9cff702cdf9.json new file mode 100644 index 0000000000000000000000000000000000000000..1b8014a3ab2be824e8589c9c3dabaab8c3e5e017 --- /dev/null +++ b/data/healthbench/openai/gpt-5.4-2026-03-05/9074f6e8-71ca-4b54-9379-b9cff702cdf9.json @@ -0,0 +1,304 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/openai_gpt-5.4-2026-03-05/1786163801.263358", + "evaluation_timestamp": "1776335705.0", + "retrieved_timestamp": "1786163801.263358", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "bootstrap_score on healthbench/S-adaptive/+1ep/6774f3ae for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "25", + "sample_ids": "762d9eca-4835-4447-a649-f91740c74bd0,b267896f-cfc0-4bbe-9fda-fdfa287d1f43,620b9bf1-f0f3-4d68-aa74-1560fb406a68,9385d26b-bf44-4ff7-9410-e181413540dd,cf8040e2-39e3-4200-ab7f-b9b532310ac3,4fc7b71f-9591-4ae1-bfca-2929c7225182,da05b57a-d776-45b6-96af-1750e0254eec,ef446d15-d019-4c5a-a052-ae8cc3ea3b81,0fe4ea94-633d-4f73-985b-cbcd9f5a0270,cbb0ca6e-aa5e-4683-b107-d7dfa4cb60ce,ae6ec2ce-5c23-4e65-802f-cef9f1624d71,cd3dbc9f-a39c-4d13-af60-cba7dbc3baad,02423ab9-f3ed-4096-a7df-bfa3ebb40c7c,084b0c27-79c5-42f6-9cbd-690d0e79167a,42465540-8af1-423b-9b35-a31de1e1d49e,74a50705-7db2-44be-84dc-92b2d960d344,3b79284e-1143-46b4-b6a0-78e5448ff773,afbfc79d-6455-40b8-b475-1bcde4ef0cc9,6d705f80-8259-4ca1-8645-163f42c85a3c,129057dd-4bf4-472d-a8cf-a063dbed60ad,2833344d-bff9-4bbe-8fab-6b13d64c6b2a,48669cdc-1a64-4c99-a6e8-9a7181341fa2,be70739f-234a-4a0f-9b14-d15b8800577b,4047884e-4629-457c-9d77-130e937fae60,eb97bae4-430e-45cd-a065-2df3ab5c600e" + } + }, + "evaluation_timestamp": "1776335705.0", + "metric_config": { + "evaluation_description": "bootstrap_score", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.34952086347173456, + "uncertainty": { + "standard_deviation": 0.05424134088247199, + "num_samples": 25 + }, + "details": { + "total_matched_trajectories": "6", + "stopping_reason_count_repetition_guard": "6", + "avg_total_tokens_target_model": "7971.33", + "total_total_tokens_target_model": "47828.0", + "avg_total_tokens_other_models": "56320.50", + "total_total_tokens_other_models": "337923.0", + "avg_total_tokens_all_models": "64291.83", + "total_total_tokens_all_models": "385751.0", + "avg_turn_count": "29.50", + "total_turn_count": "177.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "reasoning_effort": "\"high\"" + } + } + }, + { + "evaluation_name": "std on healthbench/S-adaptive/+1ep/6774f3ae for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "25", + "sample_ids": "762d9eca-4835-4447-a649-f91740c74bd0,b267896f-cfc0-4bbe-9fda-fdfa287d1f43,620b9bf1-f0f3-4d68-aa74-1560fb406a68,9385d26b-bf44-4ff7-9410-e181413540dd,cf8040e2-39e3-4200-ab7f-b9b532310ac3,4fc7b71f-9591-4ae1-bfca-2929c7225182,da05b57a-d776-45b6-96af-1750e0254eec,ef446d15-d019-4c5a-a052-ae8cc3ea3b81,0fe4ea94-633d-4f73-985b-cbcd9f5a0270,cbb0ca6e-aa5e-4683-b107-d7dfa4cb60ce,ae6ec2ce-5c23-4e65-802f-cef9f1624d71,cd3dbc9f-a39c-4d13-af60-cba7dbc3baad,02423ab9-f3ed-4096-a7df-bfa3ebb40c7c,084b0c27-79c5-42f6-9cbd-690d0e79167a,42465540-8af1-423b-9b35-a31de1e1d49e,74a50705-7db2-44be-84dc-92b2d960d344,3b79284e-1143-46b4-b6a0-78e5448ff773,afbfc79d-6455-40b8-b475-1bcde4ef0cc9,6d705f80-8259-4ca1-8645-163f42c85a3c,129057dd-4bf4-472d-a8cf-a063dbed60ad,2833344d-bff9-4bbe-8fab-6b13d64c6b2a,48669cdc-1a64-4c99-a6e8-9a7181341fa2,be70739f-234a-4a0f-9b14-d15b8800577b,4047884e-4629-457c-9d77-130e937fae60,eb97bae4-430e-45cd-a065-2df3ab5c600e" + } + }, + "evaluation_timestamp": "1776335705.0", + "metric_config": { + "evaluation_description": "std", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.05424134088247199, + "uncertainty": { + "standard_deviation": 0.05424134088247199, + "num_samples": 25 + }, + "details": { + "total_matched_trajectories": "6", + "stopping_reason_count_repetition_guard": "6", + "avg_total_tokens_target_model": "7971.33", + "total_total_tokens_target_model": "47828.0", + "avg_total_tokens_other_models": "56320.50", + "total_total_tokens_other_models": "337923.0", + "avg_total_tokens_all_models": "64291.83", + "total_total_tokens_all_models": "385751.0", + "avg_turn_count": "29.50", + "total_turn_count": "177.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "reasoning_effort": "\"high\"" + } + } + }, + { + "evaluation_name": "criteria_met_rate on healthbench/S-adaptive/+1ep/6774f3ae for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "25", + "sample_ids": "762d9eca-4835-4447-a649-f91740c74bd0,b267896f-cfc0-4bbe-9fda-fdfa287d1f43,620b9bf1-f0f3-4d68-aa74-1560fb406a68,9385d26b-bf44-4ff7-9410-e181413540dd,cf8040e2-39e3-4200-ab7f-b9b532310ac3,4fc7b71f-9591-4ae1-bfca-2929c7225182,da05b57a-d776-45b6-96af-1750e0254eec,ef446d15-d019-4c5a-a052-ae8cc3ea3b81,0fe4ea94-633d-4f73-985b-cbcd9f5a0270,cbb0ca6e-aa5e-4683-b107-d7dfa4cb60ce,ae6ec2ce-5c23-4e65-802f-cef9f1624d71,cd3dbc9f-a39c-4d13-af60-cba7dbc3baad,02423ab9-f3ed-4096-a7df-bfa3ebb40c7c,084b0c27-79c5-42f6-9cbd-690d0e79167a,42465540-8af1-423b-9b35-a31de1e1d49e,74a50705-7db2-44be-84dc-92b2d960d344,3b79284e-1143-46b4-b6a0-78e5448ff773,afbfc79d-6455-40b8-b475-1bcde4ef0cc9,6d705f80-8259-4ca1-8645-163f42c85a3c,129057dd-4bf4-472d-a8cf-a063dbed60ad,2833344d-bff9-4bbe-8fab-6b13d64c6b2a,48669cdc-1a64-4c99-a6e8-9a7181341fa2,be70739f-234a-4a0f-9b14-d15b8800577b,4047884e-4629-457c-9d77-130e937fae60,eb97bae4-430e-45cd-a065-2df3ab5c600e" + } + }, + "evaluation_timestamp": "1776335705.0", + "metric_config": { + "evaluation_description": "criteria_met_rate", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.39296187683284456, + "uncertainty": { + "standard_deviation": 0.05424134088247199, + "num_samples": 25 + }, + "details": { + "total_matched_trajectories": "6", + "stopping_reason_count_repetition_guard": "6", + "avg_total_tokens_target_model": "7971.33", + "total_total_tokens_target_model": "47828.0", + "avg_total_tokens_other_models": "56320.50", + "total_total_tokens_other_models": "337923.0", + "avg_total_tokens_all_models": "64291.83", + "total_total_tokens_all_models": "385751.0", + "avg_turn_count": "29.50", + "total_turn_count": "177.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "reasoning_effort": "\"high\"" + } + } + }, + { + "evaluation_name": "mean on healthbench/S-adaptive/+1ep/6774f3ae for scorer healthbench_score", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "25", + "sample_ids": "762d9eca-4835-4447-a649-f91740c74bd0,b267896f-cfc0-4bbe-9fda-fdfa287d1f43,620b9bf1-f0f3-4d68-aa74-1560fb406a68,9385d26b-bf44-4ff7-9410-e181413540dd,cf8040e2-39e3-4200-ab7f-b9b532310ac3,4fc7b71f-9591-4ae1-bfca-2929c7225182,da05b57a-d776-45b6-96af-1750e0254eec,ef446d15-d019-4c5a-a052-ae8cc3ea3b81,0fe4ea94-633d-4f73-985b-cbcd9f5a0270,cbb0ca6e-aa5e-4683-b107-d7dfa4cb60ce,ae6ec2ce-5c23-4e65-802f-cef9f1624d71,cd3dbc9f-a39c-4d13-af60-cba7dbc3baad,02423ab9-f3ed-4096-a7df-bfa3ebb40c7c,084b0c27-79c5-42f6-9cbd-690d0e79167a,42465540-8af1-423b-9b35-a31de1e1d49e,74a50705-7db2-44be-84dc-92b2d960d344,3b79284e-1143-46b4-b6a0-78e5448ff773,afbfc79d-6455-40b8-b475-1bcde4ef0cc9,6d705f80-8259-4ca1-8645-163f42c85a3c,129057dd-4bf4-472d-a8cf-a063dbed60ad,2833344d-bff9-4bbe-8fab-6b13d64c6b2a,48669cdc-1a64-4c99-a6e8-9a7181341fa2,be70739f-234a-4a0f-9b14-d15b8800577b,4047884e-4629-457c-9d77-130e937fae60,eb97bae4-430e-45cd-a065-2df3ab5c600e" + } + }, + "evaluation_timestamp": "1776335705.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.34952086347173456, + "uncertainty": { + "num_samples": 25 + }, + "details": { + "total_matched_trajectories": "6", + "stopping_reason_count_repetition_guard": "6", + "avg_total_tokens_target_model": "7971.33", + "total_total_tokens_target_model": "47828.0", + "avg_total_tokens_other_models": "56320.50", + "total_total_tokens_other_models": "337923.0", + "avg_total_tokens_all_models": "64291.83", + "total_total_tokens_all_models": "385751.0", + "avg_turn_count": "29.50", + "total_turn_count": "177.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/openai/gpt-5.4-2026-03-05/9074f6e8-71ca-4b54-9379-b9cff702cdf9_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "840c8c3492aed68a64984b972e629581769f3b8db83cd756b736e4d702dee04d", + "total_rows": 25 + } +} \ No newline at end of file diff --git a/data/healthbench/openai/gpt-5.4-2026-03-05/9074f6e8-71ca-4b54-9379-b9cff702cdf9_samples.jsonl b/data/healthbench/openai/gpt-5.4-2026-03-05/9074f6e8-71ca-4b54-9379-b9cff702cdf9_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..e101875f1673b4c3be8f2d266989acf2c3387ec6 --- /dev/null +++ b/data/healthbench/openai/gpt-5.4-2026-03-05/9074f6e8-71ca-4b54-9379-b9cff702cdf9_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:840c8c3492aed68a64984b972e629581769f3b8db83cd756b736e4d702dee04d +size 1739653 diff --git a/data/healthbench/openai/gpt-5.4-2026-03-05/94c9779f-10b7-4bd1-a7ac-32bc0328ba0e.json b/data/healthbench/openai/gpt-5.4-2026-03-05/94c9779f-10b7-4bd1-a7ac-32bc0328ba0e.json new file mode 100644 index 0000000000000000000000000000000000000000..68c73cece37f75e2583be540d32eb4bea4bac689 --- /dev/null +++ b/data/healthbench/openai/gpt-5.4-2026-03-05/94c9779f-10b7-4bd1-a7ac-32bc0328ba0e.json @@ -0,0 +1,304 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/openai_gpt-5.4-2026-03-05/1786163854.734384", + "evaluation_timestamp": "1776261057.0", + "retrieved_timestamp": "1786163854.734384", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "bootstrap_score on healthbench/S-adaptive/+4ep/34f5e26e for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "50", + "sample_ids": "c971f9d1-5f6a-464e-b282-41c8f0d82f58,5a6e4a41-3ea6-4050-a971-93433fe34877,78510411-e03b-4f93-a369-c7e72d2e4908,bbd759b4-8e4d-4db7-bc36-f2e069396c1c,8f2a65de-dea7-48e8-8adb-6194eca26c08,6d5f483c-3e86-456d-bfd5-4e28de699ae5,f05491d8-d160-4b2b-bd5b-59d757585e39,85408180-238e-4c1b-92c1-55d8ce6c3717,ecd7453e-1dc1-4e52-bb95-67e8e5788573,0e073591-b3e7-4dc8-94d0-dc5aa93ca35c,7637b987-8fe2-49d0-8372-1ebf17284d70,5b294937-13e5-424e-8bb4-5d1904a2344a,c5bf7fc3-dfb0-4b47-9bbe-222e87b952e2,83cf8f2d-2857-4f01-a283-9595d8f4ae8e,7a4548e6-38b7-48dd-9088-d7d1219f7852,e0d1b955-1150-457e-96d7-91a9481999cd,24f9a6e7-b214-4011-94c4-6502f249a621,6a97773b-c33c-4f2e-9ac3-fdd53de0f175,f32ee0c1-ea9c-4cfb-8332-bd9579d23024,b6eb69cb-b911-4c81-a7a1-357d21d27109,c0dc053b-157b-4d13-9956-213b67ff6a36,3bb4a735-ace7-4342-b7a5-0bae083e6f82,66b448bf-2e49-488d-9d0b-f9912ab3f0f8,b12d2453-fd3b-462c-b9cc-554ff16bfd3f,0a548d04-0973-4343-a3bf-07f730bed2f6,e2cd9f23-e3dd-4b67-8984-a5412dd802e5,1fb6ab91-a999-437d-8c93-c5937c0a89a3,42af0c9f-a715-45c4-ab24-c2948f6f943d,a94aad40-387e-4a37-b92f-61badd57388b,596ba714-29f7-447e-aa73-88d46ecd6e88,1233737e-2674-41d3-8519-b817aeb0b32e,c80a2a84-281d-41cc-a10b-32ee48c584f6,bf7718f2-bceb-4702-8d74-a7fc1bd7fb78,b7a82324-73ee-4996-8f71-8f455f789e90,22fe3eec-03b0-4cdf-a06a-87e71d236082,6c568ea6-86bd-4286-94ba-e3ccc8b67b21,905949d2-7a0a-4461-8f4b-257de6be6eed,c653b6eb-f9b9-4626-b702-ff73d71e8253,cf09b209-5434-4317-aa2a-7b07537b475e,77c0be2e-364a-40de-ae21-a277154e3d60,af750930-5fa5-4efe-8f85-b499641377d4,af62e0b8-bbd0-4dcc-b55c-cb4c7bbcfb2c,f3729e05-61ef-4306-9491-6f5659d4eef9,126236de-9e7a-4814-b5d7-98cbe89b33b2,9bb31d40-c155-4213-868d-103fa0cc7c56,97bec88b-6ef7-4f24-84d6-a2bcc592548b,779b5d14-1a6f-4d0a-8763-e0a25961a811,62b4d946-e447-4fb4-96ed-b6d36c8e2c62,d475913b-0b18-47e9-90b2-28449e1d6541,3a0eec54-fd86-47c1-883b-d8e828d2a327" + } + }, + "evaluation_timestamp": "1776261057.0", + "metric_config": { + "evaluation_description": "bootstrap_score", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.4784794918327914, + "uncertainty": { + "standard_deviation": 0.04373893161222908, + "num_samples": 200 + }, + "details": { + "total_matched_trajectories": "28", + "stopping_reason_count_repetition_guard": "28", + "avg_total_tokens_target_model": "12524.11", + "total_total_tokens_target_model": "350675.0", + "avg_total_tokens_other_models": "60450.39", + "total_total_tokens_other_models": "1692611.0", + "avg_total_tokens_all_models": "72974.50", + "total_total_tokens_all_models": "2043286.0", + "avg_turn_count": "27.36", + "total_turn_count": "766.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "reasoning_effort": "\"high\"" + } + } + }, + { + "evaluation_name": "std on healthbench/S-adaptive/+4ep/34f5e26e for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "50", + "sample_ids": "c971f9d1-5f6a-464e-b282-41c8f0d82f58,5a6e4a41-3ea6-4050-a971-93433fe34877,78510411-e03b-4f93-a369-c7e72d2e4908,bbd759b4-8e4d-4db7-bc36-f2e069396c1c,8f2a65de-dea7-48e8-8adb-6194eca26c08,6d5f483c-3e86-456d-bfd5-4e28de699ae5,f05491d8-d160-4b2b-bd5b-59d757585e39,85408180-238e-4c1b-92c1-55d8ce6c3717,ecd7453e-1dc1-4e52-bb95-67e8e5788573,0e073591-b3e7-4dc8-94d0-dc5aa93ca35c,7637b987-8fe2-49d0-8372-1ebf17284d70,5b294937-13e5-424e-8bb4-5d1904a2344a,c5bf7fc3-dfb0-4b47-9bbe-222e87b952e2,83cf8f2d-2857-4f01-a283-9595d8f4ae8e,7a4548e6-38b7-48dd-9088-d7d1219f7852,e0d1b955-1150-457e-96d7-91a9481999cd,24f9a6e7-b214-4011-94c4-6502f249a621,6a97773b-c33c-4f2e-9ac3-fdd53de0f175,f32ee0c1-ea9c-4cfb-8332-bd9579d23024,b6eb69cb-b911-4c81-a7a1-357d21d27109,c0dc053b-157b-4d13-9956-213b67ff6a36,3bb4a735-ace7-4342-b7a5-0bae083e6f82,66b448bf-2e49-488d-9d0b-f9912ab3f0f8,b12d2453-fd3b-462c-b9cc-554ff16bfd3f,0a548d04-0973-4343-a3bf-07f730bed2f6,e2cd9f23-e3dd-4b67-8984-a5412dd802e5,1fb6ab91-a999-437d-8c93-c5937c0a89a3,42af0c9f-a715-45c4-ab24-c2948f6f943d,a94aad40-387e-4a37-b92f-61badd57388b,596ba714-29f7-447e-aa73-88d46ecd6e88,1233737e-2674-41d3-8519-b817aeb0b32e,c80a2a84-281d-41cc-a10b-32ee48c584f6,bf7718f2-bceb-4702-8d74-a7fc1bd7fb78,b7a82324-73ee-4996-8f71-8f455f789e90,22fe3eec-03b0-4cdf-a06a-87e71d236082,6c568ea6-86bd-4286-94ba-e3ccc8b67b21,905949d2-7a0a-4461-8f4b-257de6be6eed,c653b6eb-f9b9-4626-b702-ff73d71e8253,cf09b209-5434-4317-aa2a-7b07537b475e,77c0be2e-364a-40de-ae21-a277154e3d60,af750930-5fa5-4efe-8f85-b499641377d4,af62e0b8-bbd0-4dcc-b55c-cb4c7bbcfb2c,f3729e05-61ef-4306-9491-6f5659d4eef9,126236de-9e7a-4814-b5d7-98cbe89b33b2,9bb31d40-c155-4213-868d-103fa0cc7c56,97bec88b-6ef7-4f24-84d6-a2bcc592548b,779b5d14-1a6f-4d0a-8763-e0a25961a811,62b4d946-e447-4fb4-96ed-b6d36c8e2c62,d475913b-0b18-47e9-90b2-28449e1d6541,3a0eec54-fd86-47c1-883b-d8e828d2a327" + } + }, + "evaluation_timestamp": "1776261057.0", + "metric_config": { + "evaluation_description": "std", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.04373893161222908, + "uncertainty": { + "standard_deviation": 0.04373893161222908, + "num_samples": 200 + }, + "details": { + "total_matched_trajectories": "28", + "stopping_reason_count_repetition_guard": "28", + "avg_total_tokens_target_model": "12524.11", + "total_total_tokens_target_model": "350675.0", + "avg_total_tokens_other_models": "60450.39", + "total_total_tokens_other_models": "1692611.0", + "avg_total_tokens_all_models": "72974.50", + "total_total_tokens_all_models": "2043286.0", + "avg_turn_count": "27.36", + "total_turn_count": "766.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "reasoning_effort": "\"high\"" + } + } + }, + { + "evaluation_name": "criteria_met_rate on healthbench/S-adaptive/+4ep/34f5e26e for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "50", + "sample_ids": "c971f9d1-5f6a-464e-b282-41c8f0d82f58,5a6e4a41-3ea6-4050-a971-93433fe34877,78510411-e03b-4f93-a369-c7e72d2e4908,bbd759b4-8e4d-4db7-bc36-f2e069396c1c,8f2a65de-dea7-48e8-8adb-6194eca26c08,6d5f483c-3e86-456d-bfd5-4e28de699ae5,f05491d8-d160-4b2b-bd5b-59d757585e39,85408180-238e-4c1b-92c1-55d8ce6c3717,ecd7453e-1dc1-4e52-bb95-67e8e5788573,0e073591-b3e7-4dc8-94d0-dc5aa93ca35c,7637b987-8fe2-49d0-8372-1ebf17284d70,5b294937-13e5-424e-8bb4-5d1904a2344a,c5bf7fc3-dfb0-4b47-9bbe-222e87b952e2,83cf8f2d-2857-4f01-a283-9595d8f4ae8e,7a4548e6-38b7-48dd-9088-d7d1219f7852,e0d1b955-1150-457e-96d7-91a9481999cd,24f9a6e7-b214-4011-94c4-6502f249a621,6a97773b-c33c-4f2e-9ac3-fdd53de0f175,f32ee0c1-ea9c-4cfb-8332-bd9579d23024,b6eb69cb-b911-4c81-a7a1-357d21d27109,c0dc053b-157b-4d13-9956-213b67ff6a36,3bb4a735-ace7-4342-b7a5-0bae083e6f82,66b448bf-2e49-488d-9d0b-f9912ab3f0f8,b12d2453-fd3b-462c-b9cc-554ff16bfd3f,0a548d04-0973-4343-a3bf-07f730bed2f6,e2cd9f23-e3dd-4b67-8984-a5412dd802e5,1fb6ab91-a999-437d-8c93-c5937c0a89a3,42af0c9f-a715-45c4-ab24-c2948f6f943d,a94aad40-387e-4a37-b92f-61badd57388b,596ba714-29f7-447e-aa73-88d46ecd6e88,1233737e-2674-41d3-8519-b817aeb0b32e,c80a2a84-281d-41cc-a10b-32ee48c584f6,bf7718f2-bceb-4702-8d74-a7fc1bd7fb78,b7a82324-73ee-4996-8f71-8f455f789e90,22fe3eec-03b0-4cdf-a06a-87e71d236082,6c568ea6-86bd-4286-94ba-e3ccc8b67b21,905949d2-7a0a-4461-8f4b-257de6be6eed,c653b6eb-f9b9-4626-b702-ff73d71e8253,cf09b209-5434-4317-aa2a-7b07537b475e,77c0be2e-364a-40de-ae21-a277154e3d60,af750930-5fa5-4efe-8f85-b499641377d4,af62e0b8-bbd0-4dcc-b55c-cb4c7bbcfb2c,f3729e05-61ef-4306-9491-6f5659d4eef9,126236de-9e7a-4814-b5d7-98cbe89b33b2,9bb31d40-c155-4213-868d-103fa0cc7c56,97bec88b-6ef7-4f24-84d6-a2bcc592548b,779b5d14-1a6f-4d0a-8763-e0a25961a811,62b4d946-e447-4fb4-96ed-b6d36c8e2c62,d475913b-0b18-47e9-90b2-28449e1d6541,3a0eec54-fd86-47c1-883b-d8e828d2a327" + } + }, + "evaluation_timestamp": "1776261057.0", + "metric_config": { + "evaluation_description": "criteria_met_rate", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.44953271028037384, + "uncertainty": { + "standard_deviation": 0.04373893161222908, + "num_samples": 200 + }, + "details": { + "total_matched_trajectories": "28", + "stopping_reason_count_repetition_guard": "28", + "avg_total_tokens_target_model": "12524.11", + "total_total_tokens_target_model": "350675.0", + "avg_total_tokens_other_models": "60450.39", + "total_total_tokens_other_models": "1692611.0", + "avg_total_tokens_all_models": "72974.50", + "total_total_tokens_all_models": "2043286.0", + "avg_turn_count": "27.36", + "total_turn_count": "766.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "reasoning_effort": "\"high\"" + } + } + }, + { + "evaluation_name": "mean on healthbench/S-adaptive/+4ep/34f5e26e for scorer healthbench_score", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "50", + "sample_ids": "c971f9d1-5f6a-464e-b282-41c8f0d82f58,5a6e4a41-3ea6-4050-a971-93433fe34877,78510411-e03b-4f93-a369-c7e72d2e4908,bbd759b4-8e4d-4db7-bc36-f2e069396c1c,8f2a65de-dea7-48e8-8adb-6194eca26c08,6d5f483c-3e86-456d-bfd5-4e28de699ae5,f05491d8-d160-4b2b-bd5b-59d757585e39,85408180-238e-4c1b-92c1-55d8ce6c3717,ecd7453e-1dc1-4e52-bb95-67e8e5788573,0e073591-b3e7-4dc8-94d0-dc5aa93ca35c,7637b987-8fe2-49d0-8372-1ebf17284d70,5b294937-13e5-424e-8bb4-5d1904a2344a,c5bf7fc3-dfb0-4b47-9bbe-222e87b952e2,83cf8f2d-2857-4f01-a283-9595d8f4ae8e,7a4548e6-38b7-48dd-9088-d7d1219f7852,e0d1b955-1150-457e-96d7-91a9481999cd,24f9a6e7-b214-4011-94c4-6502f249a621,6a97773b-c33c-4f2e-9ac3-fdd53de0f175,f32ee0c1-ea9c-4cfb-8332-bd9579d23024,b6eb69cb-b911-4c81-a7a1-357d21d27109,c0dc053b-157b-4d13-9956-213b67ff6a36,3bb4a735-ace7-4342-b7a5-0bae083e6f82,66b448bf-2e49-488d-9d0b-f9912ab3f0f8,b12d2453-fd3b-462c-b9cc-554ff16bfd3f,0a548d04-0973-4343-a3bf-07f730bed2f6,e2cd9f23-e3dd-4b67-8984-a5412dd802e5,1fb6ab91-a999-437d-8c93-c5937c0a89a3,42af0c9f-a715-45c4-ab24-c2948f6f943d,a94aad40-387e-4a37-b92f-61badd57388b,596ba714-29f7-447e-aa73-88d46ecd6e88,1233737e-2674-41d3-8519-b817aeb0b32e,c80a2a84-281d-41cc-a10b-32ee48c584f6,bf7718f2-bceb-4702-8d74-a7fc1bd7fb78,b7a82324-73ee-4996-8f71-8f455f789e90,22fe3eec-03b0-4cdf-a06a-87e71d236082,6c568ea6-86bd-4286-94ba-e3ccc8b67b21,905949d2-7a0a-4461-8f4b-257de6be6eed,c653b6eb-f9b9-4626-b702-ff73d71e8253,cf09b209-5434-4317-aa2a-7b07537b475e,77c0be2e-364a-40de-ae21-a277154e3d60,af750930-5fa5-4efe-8f85-b499641377d4,af62e0b8-bbd0-4dcc-b55c-cb4c7bbcfb2c,f3729e05-61ef-4306-9491-6f5659d4eef9,126236de-9e7a-4814-b5d7-98cbe89b33b2,9bb31d40-c155-4213-868d-103fa0cc7c56,97bec88b-6ef7-4f24-84d6-a2bcc592548b,779b5d14-1a6f-4d0a-8763-e0a25961a811,62b4d946-e447-4fb4-96ed-b6d36c8e2c62,d475913b-0b18-47e9-90b2-28449e1d6541,3a0eec54-fd86-47c1-883b-d8e828d2a327" + } + }, + "evaluation_timestamp": "1776261057.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.4784794918327914, + "uncertainty": { + "num_samples": 200 + }, + "details": { + "total_matched_trajectories": "28", + "stopping_reason_count_repetition_guard": "28", + "avg_total_tokens_target_model": "12524.11", + "total_total_tokens_target_model": "350675.0", + "avg_total_tokens_other_models": "60450.39", + "total_total_tokens_other_models": "1692611.0", + "avg_total_tokens_all_models": "72974.50", + "total_total_tokens_all_models": "2043286.0", + "avg_turn_count": "27.36", + "total_turn_count": "766.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/openai/gpt-5.4-2026-03-05/94c9779f-10b7-4bd1-a7ac-32bc0328ba0e_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "0afce663bcb28f5f89f2c49259b15c84b5658d219c5bdee2d87ce5032028d98f", + "total_rows": 200 + } +} \ No newline at end of file diff --git a/data/healthbench/openai/gpt-5.4-2026-03-05/94c9779f-10b7-4bd1-a7ac-32bc0328ba0e_samples.jsonl b/data/healthbench/openai/gpt-5.4-2026-03-05/94c9779f-10b7-4bd1-a7ac-32bc0328ba0e_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..213fde5ae98041e23504e92dc3f98d1e9a725ed1 --- /dev/null +++ b/data/healthbench/openai/gpt-5.4-2026-03-05/94c9779f-10b7-4bd1-a7ac-32bc0328ba0e_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0afce663bcb28f5f89f2c49259b15c84b5658d219c5bdee2d87ce5032028d98f +size 11892450 diff --git a/data/healthbench/openai/gpt-5.4-2026-03-05/9af7b8b0-3245-4d14-b3e9-b49ee1559e0a.json b/data/healthbench/openai/gpt-5.4-2026-03-05/9af7b8b0-3245-4d14-b3e9-b49ee1559e0a.json new file mode 100644 index 0000000000000000000000000000000000000000..c91bfea6245b70be692be376a9a50b9b819de645 --- /dev/null +++ b/data/healthbench/openai/gpt-5.4-2026-03-05/9af7b8b0-3245-4d14-b3e9-b49ee1559e0a.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/openai_gpt-5.4-2026-03-05/1786166431.9528", + "evaluation_timestamp": "1775665083.0", + "retrieved_timestamp": "1786166431.9528", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/openai/gpt-5.4-2026-03-05/9af7b8b0-3245-4d14-b3e9-b49ee1559e0a_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "f952f8dc8f31dc4c4515d59e5d1da2fd5bb938355573c826fdae27a2c54c5e08", + "total_rows": 237 + } +} \ No newline at end of file diff --git a/data/healthbench/openai/gpt-5.4-2026-03-05/9af7b8b0-3245-4d14-b3e9-b49ee1559e0a_samples.jsonl b/data/healthbench/openai/gpt-5.4-2026-03-05/9af7b8b0-3245-4d14-b3e9-b49ee1559e0a_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..3be7acec1aecd20bd7f14d5eb21860e6e95eeb28 --- /dev/null +++ b/data/healthbench/openai/gpt-5.4-2026-03-05/9af7b8b0-3245-4d14-b3e9-b49ee1559e0a_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f952f8dc8f31dc4c4515d59e5d1da2fd5bb938355573c826fdae27a2c54c5e08 +size 14686493 diff --git a/data/healthbench/openai/gpt-5.4-2026-03-05/9b2e8bff-4fc8-4618-9e47-f11d633220fc.json b/data/healthbench/openai/gpt-5.4-2026-03-05/9b2e8bff-4fc8-4618-9e47-f11d633220fc.json new file mode 100644 index 0000000000000000000000000000000000000000..d2788676ffc52084026adff186beae63c764cc6f --- /dev/null +++ b/data/healthbench/openai/gpt-5.4-2026-03-05/9b2e8bff-4fc8-4618-9e47-f11d633220fc.json @@ -0,0 +1,320 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/openai_gpt-5.4-2026-03-05/1786163857.664767", + "evaluation_timestamp": "1776725924.0", + "retrieved_timestamp": "1786163857.664767", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "bootstrap_score on healthbench/S-adaptive/8fda534a for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "4", + "sample_ids": "51433193-b0c3-4725-a3a8-9dd15e969b0c,48c89cd1-0ecf-4738-8fcc-392269585bd3,36c6e780-d4a8-47b2-8a36-90ee4b97607f,b81c059a-1137-47cf-bab8-ab5eb8ff34d5" + } + }, + "evaluation_timestamp": "1776725924.0", + "metric_config": { + "evaluation_description": "bootstrap_score", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.24310030679529415, + "uncertainty": { + "standard_deviation": 0.0922357467516836, + "num_samples": 20 + }, + "details": { + "total_matched_trajectories": "20", + "stopping_reason_count_repetition_guard": "20", + "avg_total_tokens_target_model": "13984.10", + "total_total_tokens_target_model": "279682.0", + "avg_total_tokens_other_models": "74869.40", + "total_total_tokens_other_models": "1497388.0", + "avg_total_tokens_all_models": "88853.50", + "total_total_tokens_all_models": "1777070.0", + "avg_turn_count": "33.15", + "total_turn_count": "663.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "3600", + "attempt_timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + }, + { + "evaluation_name": "std on healthbench/S-adaptive/8fda534a for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "4", + "sample_ids": "51433193-b0c3-4725-a3a8-9dd15e969b0c,48c89cd1-0ecf-4738-8fcc-392269585bd3,36c6e780-d4a8-47b2-8a36-90ee4b97607f,b81c059a-1137-47cf-bab8-ab5eb8ff34d5" + } + }, + "evaluation_timestamp": "1776725924.0", + "metric_config": { + "evaluation_description": "std", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.0922357467516836, + "uncertainty": { + "standard_deviation": 0.0922357467516836, + "num_samples": 20 + }, + "details": { + "total_matched_trajectories": "20", + "stopping_reason_count_repetition_guard": "20", + "avg_total_tokens_target_model": "13984.10", + "total_total_tokens_target_model": "279682.0", + "avg_total_tokens_other_models": "74869.40", + "total_total_tokens_other_models": "1497388.0", + "avg_total_tokens_all_models": "88853.50", + "total_total_tokens_all_models": "1777070.0", + "avg_turn_count": "33.15", + "total_turn_count": "663.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "3600", + "attempt_timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + }, + { + "evaluation_name": "criteria_met_rate on healthbench/S-adaptive/8fda534a for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "4", + "sample_ids": "51433193-b0c3-4725-a3a8-9dd15e969b0c,48c89cd1-0ecf-4738-8fcc-392269585bd3,36c6e780-d4a8-47b2-8a36-90ee4b97607f,b81c059a-1137-47cf-bab8-ab5eb8ff34d5" + } + }, + "evaluation_timestamp": "1776725924.0", + "metric_config": { + "evaluation_description": "criteria_met_rate", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.3566666666666667, + "uncertainty": { + "standard_deviation": 0.0922357467516836, + "num_samples": 20 + }, + "details": { + "total_matched_trajectories": "20", + "stopping_reason_count_repetition_guard": "20", + "avg_total_tokens_target_model": "13984.10", + "total_total_tokens_target_model": "279682.0", + "avg_total_tokens_other_models": "74869.40", + "total_total_tokens_other_models": "1497388.0", + "avg_total_tokens_all_models": "88853.50", + "total_total_tokens_all_models": "1777070.0", + "avg_turn_count": "33.15", + "total_turn_count": "663.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "3600", + "attempt_timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + }, + { + "evaluation_name": "mean on healthbench/S-adaptive/8fda534a for scorer healthbench_score", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "4", + "sample_ids": "51433193-b0c3-4725-a3a8-9dd15e969b0c,48c89cd1-0ecf-4738-8fcc-392269585bd3,36c6e780-d4a8-47b2-8a36-90ee4b97607f,b81c059a-1137-47cf-bab8-ab5eb8ff34d5" + } + }, + "evaluation_timestamp": "1776725924.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.24310030679529415, + "uncertainty": { + "num_samples": 20 + }, + "details": { + "total_matched_trajectories": "20", + "stopping_reason_count_repetition_guard": "20", + "avg_total_tokens_target_model": "13984.10", + "total_total_tokens_target_model": "279682.0", + "avg_total_tokens_other_models": "74869.40", + "total_total_tokens_other_models": "1497388.0", + "avg_total_tokens_all_models": "88853.50", + "total_total_tokens_all_models": "1777070.0", + "avg_turn_count": "33.15", + "total_turn_count": "663.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "3600", + "attempt_timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/openai/gpt-5.4-2026-03-05/9b2e8bff-4fc8-4618-9e47-f11d633220fc_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "07ee2c59efa09f97864d4b341af4817f02b7690bad58b6c418f8d47430188ea6", + "total_rows": 20 + } +} \ No newline at end of file diff --git a/data/healthbench/openai/gpt-5.4-2026-03-05/9b2e8bff-4fc8-4618-9e47-f11d633220fc_samples.jsonl b/data/healthbench/openai/gpt-5.4-2026-03-05/9b2e8bff-4fc8-4618-9e47-f11d633220fc_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..901db754e599ae8ff9d8a6119265dd49a58181b0 --- /dev/null +++ b/data/healthbench/openai/gpt-5.4-2026-03-05/9b2e8bff-4fc8-4618-9e47-f11d633220fc_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:07ee2c59efa09f97864d4b341af4817f02b7690bad58b6c418f8d47430188ea6 +size 1924029 diff --git a/data/healthbench/openai/gpt-5.4-2026-03-05/a2c7c36b-e0f2-4bef-9247-051b9652b608.json b/data/healthbench/openai/gpt-5.4-2026-03-05/a2c7c36b-e0f2-4bef-9247-051b9652b608.json new file mode 100644 index 0000000000000000000000000000000000000000..d7c84250f4e109b119a1906251ac18e5980a56cc --- /dev/null +++ b/data/healthbench/openai/gpt-5.4-2026-03-05/a2c7c36b-e0f2-4bef-9247-051b9652b608.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/openai_gpt-5.4-2026-03-05/1786166417.700785", + "evaluation_timestamp": "1776682987.0", + "retrieved_timestamp": "1786166417.700785", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/openai/gpt-5.4-2026-03-05/a2c7c36b-e0f2-4bef-9247-051b9652b608_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "e81bd80f565acda0858de4ab7445388f253c89f0e3eb68803bef45ce0256189e", + "total_rows": 30 + } +} \ No newline at end of file diff --git a/data/healthbench/openai/gpt-5.4-2026-03-05/a2c7c36b-e0f2-4bef-9247-051b9652b608_samples.jsonl b/data/healthbench/openai/gpt-5.4-2026-03-05/a2c7c36b-e0f2-4bef-9247-051b9652b608_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..a0e4d0e32218ce2ad91b915cd3eb4b7a63558900 --- /dev/null +++ b/data/healthbench/openai/gpt-5.4-2026-03-05/a2c7c36b-e0f2-4bef-9247-051b9652b608_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e81bd80f565acda0858de4ab7445388f253c89f0e3eb68803bef45ce0256189e +size 1936672 diff --git a/data/healthbench/openai/gpt-5.4-2026-03-05/a439cdb7-89b7-485c-816a-67dc53bb306b.json b/data/healthbench/openai/gpt-5.4-2026-03-05/a439cdb7-89b7-485c-816a-67dc53bb306b.json new file mode 100644 index 0000000000000000000000000000000000000000..510361ae54ef6eb402dc815453febd69c10bed4f --- /dev/null +++ b/data/healthbench/openai/gpt-5.4-2026-03-05/a439cdb7-89b7-485c-816a-67dc53bb306b.json @@ -0,0 +1,320 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/openai_gpt-5.4-2026-03-05/1786163804.857751", + "evaluation_timestamp": "1776785226.0", + "retrieved_timestamp": "1786163804.857751", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "bootstrap_score on healthbench/S-adaptive/+2ep/4aa2120a for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "1", + "sample_ids": "b7ddde65-702e-4255-b1d7-09d01248e76d" + } + }, + "evaluation_timestamp": "1776785226.0", + "metric_config": { + "evaluation_description": "bootstrap_score", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.5813953488372092, + "uncertainty": { + "standard_deviation": 0.0, + "num_samples": 2 + }, + "details": { + "total_matched_trajectories": "2", + "stopping_reason_count_repetition_guard": "2", + "avg_total_tokens_target_model": "17167.00", + "total_total_tokens_target_model": "34334.0", + "avg_total_tokens_other_models": "156337.50", + "total_total_tokens_other_models": "312675.0", + "avg_total_tokens_all_models": "173504.50", + "total_total_tokens_all_models": "347009.0", + "avg_turn_count": "43.00", + "total_turn_count": "86.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + }, + { + "evaluation_name": "std on healthbench/S-adaptive/+2ep/4aa2120a for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "1", + "sample_ids": "b7ddde65-702e-4255-b1d7-09d01248e76d" + } + }, + "evaluation_timestamp": "1776785226.0", + "metric_config": { + "evaluation_description": "std", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.0, + "uncertainty": { + "standard_deviation": 0.0, + "num_samples": 2 + }, + "details": { + "total_matched_trajectories": "2", + "stopping_reason_count_repetition_guard": "2", + "avg_total_tokens_target_model": "17167.00", + "total_total_tokens_target_model": "34334.0", + "avg_total_tokens_other_models": "156337.50", + "total_total_tokens_other_models": "312675.0", + "avg_total_tokens_all_models": "173504.50", + "total_total_tokens_all_models": "347009.0", + "avg_turn_count": "43.00", + "total_turn_count": "86.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + }, + { + "evaluation_name": "criteria_met_rate on healthbench/S-adaptive/+2ep/4aa2120a for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "1", + "sample_ids": "b7ddde65-702e-4255-b1d7-09d01248e76d" + } + }, + "evaluation_timestamp": "1776785226.0", + "metric_config": { + "evaluation_description": "criteria_met_rate", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.475, + "uncertainty": { + "standard_deviation": 0.0, + "num_samples": 2 + }, + "details": { + "total_matched_trajectories": "2", + "stopping_reason_count_repetition_guard": "2", + "avg_total_tokens_target_model": "17167.00", + "total_total_tokens_target_model": "34334.0", + "avg_total_tokens_other_models": "156337.50", + "total_total_tokens_other_models": "312675.0", + "avg_total_tokens_all_models": "173504.50", + "total_total_tokens_all_models": "347009.0", + "avg_turn_count": "43.00", + "total_turn_count": "86.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + }, + { + "evaluation_name": "mean on healthbench/S-adaptive/+2ep/4aa2120a for scorer healthbench_score", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "1", + "sample_ids": "b7ddde65-702e-4255-b1d7-09d01248e76d" + } + }, + "evaluation_timestamp": "1776785226.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.5813953488372092, + "uncertainty": { + "num_samples": 2 + }, + "details": { + "total_matched_trajectories": "2", + "stopping_reason_count_repetition_guard": "2", + "avg_total_tokens_target_model": "17167.00", + "total_total_tokens_target_model": "34334.0", + "avg_total_tokens_other_models": "156337.50", + "total_total_tokens_other_models": "312675.0", + "avg_total_tokens_all_models": "173504.50", + "total_total_tokens_all_models": "347009.0", + "avg_turn_count": "43.00", + "total_turn_count": "86.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/openai/gpt-5.4-2026-03-05/a439cdb7-89b7-485c-816a-67dc53bb306b_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "cd7f5b94f164dd59da0db8da1126db170dee5dc391b066233a2144494cc020c4", + "total_rows": 2 + } +} \ No newline at end of file diff --git a/data/healthbench/openai/gpt-5.4-2026-03-05/a439cdb7-89b7-485c-816a-67dc53bb306b_samples.jsonl b/data/healthbench/openai/gpt-5.4-2026-03-05/a439cdb7-89b7-485c-816a-67dc53bb306b_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..25fb2dac13fdba449135252ad0296c18bac24b18 --- /dev/null +++ b/data/healthbench/openai/gpt-5.4-2026-03-05/a439cdb7-89b7-485c-816a-67dc53bb306b_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cd7f5b94f164dd59da0db8da1126db170dee5dc391b066233a2144494cc020c4 +size 214805 diff --git a/data/healthbench/openai/gpt-5.4-2026-03-05/c02710ce-4b95-4a4a-b780-46f8769fb8dc.json b/data/healthbench/openai/gpt-5.4-2026-03-05/c02710ce-4b95-4a4a-b780-46f8769fb8dc.json new file mode 100644 index 0000000000000000000000000000000000000000..a90f210f55540a512313291ff82a8ad6c55e180f --- /dev/null +++ b/data/healthbench/openai/gpt-5.4-2026-03-05/c02710ce-4b95-4a4a-b780-46f8769fb8dc.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/openai_gpt-5.4-2026-03-05/1786168394.223893", + "evaluation_timestamp": "1776259832.0", + "retrieved_timestamp": "1786168394.223893", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/openai/gpt-5.4-2026-03-05/c02710ce-4b95-4a4a-b780-46f8769fb8dc_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "3ede32da1c10592a105e930bae740285eb96557c54d1df73ec584152ea3b8070", + "total_rows": 120 + } +} \ No newline at end of file diff --git a/data/healthbench/openai/gpt-5.4-2026-03-05/c02710ce-4b95-4a4a-b780-46f8769fb8dc_samples.jsonl b/data/healthbench/openai/gpt-5.4-2026-03-05/c02710ce-4b95-4a4a-b780-46f8769fb8dc_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..3574e74f82b1e43ab08ba7be67b3b205285acb38 --- /dev/null +++ b/data/healthbench/openai/gpt-5.4-2026-03-05/c02710ce-4b95-4a4a-b780-46f8769fb8dc_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ede32da1c10592a105e930bae740285eb96557c54d1df73ec584152ea3b8070 +size 7394542 diff --git a/data/healthbench/openai/gpt-5.4-2026-03-05/c22aa536-3246-47f5-8baa-8c87850b3982.json b/data/healthbench/openai/gpt-5.4-2026-03-05/c22aa536-3246-47f5-8baa-8c87850b3982.json new file mode 100644 index 0000000000000000000000000000000000000000..4ac8698700c442bd7bb0d95728e501ca2a6bbff5 --- /dev/null +++ b/data/healthbench/openai/gpt-5.4-2026-03-05/c22aa536-3246-47f5-8baa-8c87850b3982.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/openai_gpt-5.4-2026-03-05/1786168452.978392", + "evaluation_timestamp": "1776259832.0", + "retrieved_timestamp": "1786168452.978392", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/openai/gpt-5.4-2026-03-05/c22aa536-3246-47f5-8baa-8c87850b3982_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "af7c537d155ebd6dd71b8aa9006330f1031fe32f8c5710c38570ddc0aed078c8", + "total_rows": 220 + } +} \ No newline at end of file diff --git a/data/healthbench/openai/gpt-5.4-2026-03-05/c22aa536-3246-47f5-8baa-8c87850b3982_samples.jsonl b/data/healthbench/openai/gpt-5.4-2026-03-05/c22aa536-3246-47f5-8baa-8c87850b3982_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..7df3bc80ab09f9fbfb5ec7778812fec2a70c9488 --- /dev/null +++ b/data/healthbench/openai/gpt-5.4-2026-03-05/c22aa536-3246-47f5-8baa-8c87850b3982_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:af7c537d155ebd6dd71b8aa9006330f1031fe32f8c5710c38570ddc0aed078c8 +size 13913564 diff --git a/data/healthbench/openai/gpt-5.4-2026-03-05/c283d8e7-bc54-42b1-b007-3af2ef7bd40e.json b/data/healthbench/openai/gpt-5.4-2026-03-05/c283d8e7-bc54-42b1-b007-3af2ef7bd40e.json new file mode 100644 index 0000000000000000000000000000000000000000..9193018da11d83171e2c6e7b1b5687b139c8d1bb --- /dev/null +++ b/data/healthbench/openai/gpt-5.4-2026-03-05/c283d8e7-bc54-42b1-b007-3af2ef7bd40e.json @@ -0,0 +1,304 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/openai_gpt-5.4-2026-03-05/1786163807.030287", + "evaluation_timestamp": "1775604415.0", + "retrieved_timestamp": "1786163807.030287", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.202,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "bootstrap_score on healthbench/S-adaptive/+2ep for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "5", + "sample_ids": "1f548d5b-cd00-49a0-b327-283a2e00debd,0b8f1d60-2081-4562-98f7-b6a976fe1c6d,6f7a2ee9-e9c6-42d8-b79f-22dea966b8d2,19ec4833-86e9-4166-8b82-d1da09f31fd7,7ebc830a-8dbd-489b-9d61-4d8bacf0db8d" + } + }, + "evaluation_timestamp": "1775604415.0", + "metric_config": { + "evaluation_description": "bootstrap_score", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.6679258615416808, + "uncertainty": { + "standard_deviation": 0.06613298282330451, + "num_samples": 10 + }, + "details": { + "total_matched_trajectories": "2", + "stopping_reason_count_repetition_guard": "2", + "avg_total_tokens_target_model": "45439.00", + "total_total_tokens_target_model": "90878.0", + "avg_total_tokens_other_models": "104583.00", + "total_total_tokens_other_models": "209166.0", + "avg_total_tokens_all_models": "150022.00", + "total_total_tokens_all_models": "300044.0", + "avg_turn_count": "23.00", + "total_turn_count": "46.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "reasoning_effort": "\"high\"" + } + } + }, + { + "evaluation_name": "std on healthbench/S-adaptive/+2ep for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "5", + "sample_ids": "1f548d5b-cd00-49a0-b327-283a2e00debd,0b8f1d60-2081-4562-98f7-b6a976fe1c6d,6f7a2ee9-e9c6-42d8-b79f-22dea966b8d2,19ec4833-86e9-4166-8b82-d1da09f31fd7,7ebc830a-8dbd-489b-9d61-4d8bacf0db8d" + } + }, + "evaluation_timestamp": "1775604415.0", + "metric_config": { + "evaluation_description": "std", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.06613298282330451, + "uncertainty": { + "standard_deviation": 0.06613298282330451, + "num_samples": 10 + }, + "details": { + "total_matched_trajectories": "2", + "stopping_reason_count_repetition_guard": "2", + "avg_total_tokens_target_model": "45439.00", + "total_total_tokens_target_model": "90878.0", + "avg_total_tokens_other_models": "104583.00", + "total_total_tokens_other_models": "209166.0", + "avg_total_tokens_all_models": "150022.00", + "total_total_tokens_all_models": "300044.0", + "avg_turn_count": "23.00", + "total_turn_count": "46.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "reasoning_effort": "\"high\"" + } + } + }, + { + "evaluation_name": "criteria_met_rate on healthbench/S-adaptive/+2ep for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "5", + "sample_ids": "1f548d5b-cd00-49a0-b327-283a2e00debd,0b8f1d60-2081-4562-98f7-b6a976fe1c6d,6f7a2ee9-e9c6-42d8-b79f-22dea966b8d2,19ec4833-86e9-4166-8b82-d1da09f31fd7,7ebc830a-8dbd-489b-9d61-4d8bacf0db8d" + } + }, + "evaluation_timestamp": "1775604415.0", + "metric_config": { + "evaluation_description": "criteria_met_rate", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.625, + "uncertainty": { + "standard_deviation": 0.06613298282330451, + "num_samples": 10 + }, + "details": { + "total_matched_trajectories": "2", + "stopping_reason_count_repetition_guard": "2", + "avg_total_tokens_target_model": "45439.00", + "total_total_tokens_target_model": "90878.0", + "avg_total_tokens_other_models": "104583.00", + "total_total_tokens_other_models": "209166.0", + "avg_total_tokens_all_models": "150022.00", + "total_total_tokens_all_models": "300044.0", + "avg_turn_count": "23.00", + "total_turn_count": "46.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "reasoning_effort": "\"high\"" + } + } + }, + { + "evaluation_name": "mean on healthbench/S-adaptive/+2ep for scorer healthbench_score", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "5", + "sample_ids": "1f548d5b-cd00-49a0-b327-283a2e00debd,0b8f1d60-2081-4562-98f7-b6a976fe1c6d,6f7a2ee9-e9c6-42d8-b79f-22dea966b8d2,19ec4833-86e9-4166-8b82-d1da09f31fd7,7ebc830a-8dbd-489b-9d61-4d8bacf0db8d" + } + }, + "evaluation_timestamp": "1775604415.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.6679258615416808, + "uncertainty": { + "num_samples": 10 + }, + "details": { + "total_matched_trajectories": "2", + "stopping_reason_count_repetition_guard": "2", + "avg_total_tokens_target_model": "45439.00", + "total_total_tokens_target_model": "90878.0", + "avg_total_tokens_other_models": "104583.00", + "total_total_tokens_other_models": "209166.0", + "avg_total_tokens_all_models": "150022.00", + "total_total_tokens_all_models": "300044.0", + "avg_turn_count": "23.00", + "total_turn_count": "46.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/openai/gpt-5.4-2026-03-05/c283d8e7-bc54-42b1-b007-3af2ef7bd40e_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "61b52a62e0be2b531817a3314b92e0a8d96d5526ed6a869c5dd9e9d526b19534", + "total_rows": 10 + } +} \ No newline at end of file diff --git a/data/healthbench/openai/gpt-5.4-2026-03-05/c283d8e7-bc54-42b1-b007-3af2ef7bd40e_samples.jsonl b/data/healthbench/openai/gpt-5.4-2026-03-05/c283d8e7-bc54-42b1-b007-3af2ef7bd40e_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..bf7ff2f2b4b64e2680c25b33703df5c640a85775 --- /dev/null +++ b/data/healthbench/openai/gpt-5.4-2026-03-05/c283d8e7-bc54-42b1-b007-3af2ef7bd40e_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:61b52a62e0be2b531817a3314b92e0a8d96d5526ed6a869c5dd9e9d526b19534 +size 748332 diff --git a/data/healthbench/openai/gpt-5.4-2026-03-05/c633b600-cbe4-4663-a5d0-10b29bc7511c.json b/data/healthbench/openai/gpt-5.4-2026-03-05/c633b600-cbe4-4663-a5d0-10b29bc7511c.json new file mode 100644 index 0000000000000000000000000000000000000000..85cd32e05b72af841ee11a996ddf4b8a49c03b66 --- /dev/null +++ b/data/healthbench/openai/gpt-5.4-2026-03-05/c633b600-cbe4-4663-a5d0-10b29bc7511c.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/openai_gpt-5.4-2026-03-05/1786168438.965917", + "evaluation_timestamp": "1776165241.0", + "retrieved_timestamp": "1786168438.965917", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/openai/gpt-5.4-2026-03-05/c633b600-cbe4-4663-a5d0-10b29bc7511c_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "31a4e761ed83fcaf2975ebe1d56b6a6360b5bd6d9c8aa51c39017aae2051a808", + "total_rows": 450 + } +} \ No newline at end of file diff --git a/data/healthbench/openai/gpt-5.4-2026-03-05/c633b600-cbe4-4663-a5d0-10b29bc7511c_samples.jsonl b/data/healthbench/openai/gpt-5.4-2026-03-05/c633b600-cbe4-4663-a5d0-10b29bc7511c_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..b2eeb857ae552431d73cdf42911d27cea8cea92b --- /dev/null +++ b/data/healthbench/openai/gpt-5.4-2026-03-05/c633b600-cbe4-4663-a5d0-10b29bc7511c_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:31a4e761ed83fcaf2975ebe1d56b6a6360b5bd6d9c8aa51c39017aae2051a808 +size 27617078 diff --git a/data/healthbench/openai/gpt-5.4-2026-03-05/e5667bf4-69d1-4f30-af32-410d6cb4922a.json b/data/healthbench/openai/gpt-5.4-2026-03-05/e5667bf4-69d1-4f30-af32-410d6cb4922a.json new file mode 100644 index 0000000000000000000000000000000000000000..13e45ffc37c2baec34b31d7d886fbdda53c42491 --- /dev/null +++ b/data/healthbench/openai/gpt-5.4-2026-03-05/e5667bf4-69d1-4f30-af32-410d6cb4922a.json @@ -0,0 +1,320 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/openai_gpt-5.4-2026-03-05/1786163816.533052", + "evaluation_timestamp": "1776720756.0", + "retrieved_timestamp": "1786163816.533052", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "bootstrap_score on healthbench/S-adaptive/+4ep/da714960 for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "10", + "sample_ids": "54349c88-e80a-4728-b181-305f97905d59,3852d050-c009-4e33-b953-fd734a5b5e73,5dc23430-dfbb-4794-9f77-0d4dd6093033,04b6e0c8-2835-415e-865d-aca9f1cd3c57,7a283823-00ee-489d-8dbc-356e2857526f,e89e00c9-29e9-420c-813d-e06ddc39d585,92ea478e-6bf8-4ffb-b88a-f57ccf5f2346,5fb6812f-9816-4d5d-b7e0-4ec3e9df5dc1,c29c95d1-5b5e-4f53-9ec5-83e1f72b5aa5,886864a6-a0e4-468f-aaf4-26cf8dc5e541" + } + }, + "evaluation_timestamp": "1776720756.0", + "metric_config": { + "evaluation_description": "bootstrap_score", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.25838226631031425, + "uncertainty": { + "standard_deviation": 0.07534070907239619, + "num_samples": 40 + }, + "details": { + "total_matched_trajectories": "40", + "stopping_reason_count_repetition_guard": "40", + "avg_total_tokens_target_model": "12545.05", + "total_total_tokens_target_model": "501802.0", + "avg_total_tokens_other_models": "48480.40", + "total_total_tokens_other_models": "1939216.0", + "avg_total_tokens_all_models": "61025.45", + "total_total_tokens_all_models": "2441018.0", + "avg_turn_count": "22.43", + "total_turn_count": "897.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "3600", + "attempt_timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + }, + { + "evaluation_name": "std on healthbench/S-adaptive/+4ep/da714960 for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "10", + "sample_ids": "54349c88-e80a-4728-b181-305f97905d59,3852d050-c009-4e33-b953-fd734a5b5e73,5dc23430-dfbb-4794-9f77-0d4dd6093033,04b6e0c8-2835-415e-865d-aca9f1cd3c57,7a283823-00ee-489d-8dbc-356e2857526f,e89e00c9-29e9-420c-813d-e06ddc39d585,92ea478e-6bf8-4ffb-b88a-f57ccf5f2346,5fb6812f-9816-4d5d-b7e0-4ec3e9df5dc1,c29c95d1-5b5e-4f53-9ec5-83e1f72b5aa5,886864a6-a0e4-468f-aaf4-26cf8dc5e541" + } + }, + "evaluation_timestamp": "1776720756.0", + "metric_config": { + "evaluation_description": "std", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.07534070907239619, + "uncertainty": { + "standard_deviation": 0.07534070907239619, + "num_samples": 40 + }, + "details": { + "total_matched_trajectories": "40", + "stopping_reason_count_repetition_guard": "40", + "avg_total_tokens_target_model": "12545.05", + "total_total_tokens_target_model": "501802.0", + "avg_total_tokens_other_models": "48480.40", + "total_total_tokens_other_models": "1939216.0", + "avg_total_tokens_all_models": "61025.45", + "total_total_tokens_all_models": "2441018.0", + "avg_turn_count": "22.43", + "total_turn_count": "897.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "3600", + "attempt_timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + }, + { + "evaluation_name": "criteria_met_rate on healthbench/S-adaptive/+4ep/da714960 for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "10", + "sample_ids": "54349c88-e80a-4728-b181-305f97905d59,3852d050-c009-4e33-b953-fd734a5b5e73,5dc23430-dfbb-4794-9f77-0d4dd6093033,04b6e0c8-2835-415e-865d-aca9f1cd3c57,7a283823-00ee-489d-8dbc-356e2857526f,e89e00c9-29e9-420c-813d-e06ddc39d585,92ea478e-6bf8-4ffb-b88a-f57ccf5f2346,5fb6812f-9816-4d5d-b7e0-4ec3e9df5dc1,c29c95d1-5b5e-4f53-9ec5-83e1f72b5aa5,886864a6-a0e4-468f-aaf4-26cf8dc5e541" + } + }, + "evaluation_timestamp": "1776720756.0", + "metric_config": { + "evaluation_description": "criteria_met_rate", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.3118556701030928, + "uncertainty": { + "standard_deviation": 0.07534070907239619, + "num_samples": 40 + }, + "details": { + "total_matched_trajectories": "40", + "stopping_reason_count_repetition_guard": "40", + "avg_total_tokens_target_model": "12545.05", + "total_total_tokens_target_model": "501802.0", + "avg_total_tokens_other_models": "48480.40", + "total_total_tokens_other_models": "1939216.0", + "avg_total_tokens_all_models": "61025.45", + "total_total_tokens_all_models": "2441018.0", + "avg_turn_count": "22.43", + "total_turn_count": "897.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "3600", + "attempt_timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + }, + { + "evaluation_name": "mean on healthbench/S-adaptive/+4ep/da714960 for scorer healthbench_score", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "10", + "sample_ids": "54349c88-e80a-4728-b181-305f97905d59,3852d050-c009-4e33-b953-fd734a5b5e73,5dc23430-dfbb-4794-9f77-0d4dd6093033,04b6e0c8-2835-415e-865d-aca9f1cd3c57,7a283823-00ee-489d-8dbc-356e2857526f,e89e00c9-29e9-420c-813d-e06ddc39d585,92ea478e-6bf8-4ffb-b88a-f57ccf5f2346,5fb6812f-9816-4d5d-b7e0-4ec3e9df5dc1,c29c95d1-5b5e-4f53-9ec5-83e1f72b5aa5,886864a6-a0e4-468f-aaf4-26cf8dc5e541" + } + }, + "evaluation_timestamp": "1776720756.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.25838226631031425, + "uncertainty": { + "num_samples": 40 + }, + "details": { + "total_matched_trajectories": "40", + "stopping_reason_count_repetition_guard": "40", + "avg_total_tokens_target_model": "12545.05", + "total_total_tokens_target_model": "501802.0", + "avg_total_tokens_other_models": "48480.40", + "total_total_tokens_other_models": "1939216.0", + "avg_total_tokens_all_models": "61025.45", + "total_total_tokens_all_models": "2441018.0", + "avg_turn_count": "22.43", + "total_turn_count": "897.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "3600", + "attempt_timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/openai/gpt-5.4-2026-03-05/e5667bf4-69d1-4f30-af32-410d6cb4922a_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "14508652fe20dc372323ff17e1058632c3b48cca60ac1b18026fc7711984cbfe", + "total_rows": 40 + } +} \ No newline at end of file diff --git a/data/healthbench/openai/gpt-5.4-2026-03-05/e5667bf4-69d1-4f30-af32-410d6cb4922a_samples.jsonl b/data/healthbench/openai/gpt-5.4-2026-03-05/e5667bf4-69d1-4f30-af32-410d6cb4922a_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..0ca000122377596eff6dc4fdbbcba2e19bea7a0e --- /dev/null +++ b/data/healthbench/openai/gpt-5.4-2026-03-05/e5667bf4-69d1-4f30-af32-410d6cb4922a_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:14508652fe20dc372323ff17e1058632c3b48cca60ac1b18026fc7711984cbfe +size 2594682 diff --git a/data/healthbench/openai/gpt-5.4-2026-03-05/eb35e621-2690-4a01-b45f-47cf72e367e4.json b/data/healthbench/openai/gpt-5.4-2026-03-05/eb35e621-2690-4a01-b45f-47cf72e367e4.json new file mode 100644 index 0000000000000000000000000000000000000000..baa366782139d8b2f46aeffc858ec9cdacf4f472 --- /dev/null +++ b/data/healthbench/openai/gpt-5.4-2026-03-05/eb35e621-2690-4a01-b45f-47cf72e367e4.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/openai_gpt-5.4-2026-03-05/1786168376.079192", + "evaluation_timestamp": "1775646127.0", + "retrieved_timestamp": "1786168376.079192", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/openai/gpt-5.4-2026-03-05/eb35e621-2690-4a01-b45f-47cf72e367e4_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "e2b2dca81a22f2b0f37bbd7e95abe6424d5672e404192668914bc2f43728ac10", + "total_rows": 100 + } +} \ No newline at end of file diff --git a/data/healthbench/openai/gpt-5.4-2026-03-05/eb35e621-2690-4a01-b45f-47cf72e367e4_samples.jsonl b/data/healthbench/openai/gpt-5.4-2026-03-05/eb35e621-2690-4a01-b45f-47cf72e367e4_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..6e181b965de05c95d0cdf7e88de8f2aac64a8edc --- /dev/null +++ b/data/healthbench/openai/gpt-5.4-2026-03-05/eb35e621-2690-4a01-b45f-47cf72e367e4_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e2b2dca81a22f2b0f37bbd7e95abe6424d5672e404192668914bc2f43728ac10 +size 6198940 diff --git a/data/healthbench/openai/gpt-5.4-2026-03-05/ef8ab69f-1c1c-409f-84fe-fab6c9380774.json b/data/healthbench/openai/gpt-5.4-2026-03-05/ef8ab69f-1c1c-409f-84fe-fab6c9380774.json new file mode 100644 index 0000000000000000000000000000000000000000..a6aaf86d5f7731228aba5659206f6b2c79749377 --- /dev/null +++ b/data/healthbench/openai/gpt-5.4-2026-03-05/ef8ab69f-1c1c-409f-84fe-fab6c9380774.json @@ -0,0 +1,320 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/openai_gpt-5.4-2026-03-05/1786163812.805094", + "evaluation_timestamp": "1776696116.0", + "retrieved_timestamp": "1786163812.805094", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "bootstrap_score on healthbench/S-adaptive/+3ep/114707e5 for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "50", + "sample_ids": "267bda16-fd4b-4a9e-b3ac-5602c2532eed,0e819a9c-851d-4a7a-9263-62cfc8ce1b48,328c55cf-0b57-4992-99b1-f6a21af62ade,7b415a2f-22b0-4735-a38e-567f1891acc5,c49ac752-3693-4be1-ab4d-91fbdd02f827,a079d928-7b78-4a85-a72b-4274d7fc2914,a20cbbf3-330b-46d9-a35b-92e89a022282,be6150a5-8060-4434-9724-afede307d495,3a896889-9029-4ef8-8684-ca378da4702b,7074bb7a-a595-434d-883e-95ef6f5f6667,9f7b555a-fc30-40c3-88d3-f2ece06454a2,a6204079-9e72-42ad-8f39-a65552fa2e68,c0d80b93-b2d5-4129-b998-b1312f98e2a8,3dc41419-2565-428e-b9b9-02e3fd93c470,a91a22f8-d2e6-4593-bcae-6d479ed38031,d0d3426f-8b7a-4501-adb4-86d8fa9099c8,a7dac6a9-eab5-4c01-882c-dbf6c774d7c3,c1240427-ae38-4335-9e80-492c13ef4893,b7ddde65-702e-4255-b1d7-09d01248e76d,00d34549-134c-48e2-b4bb-b54946da70cc,85c11e2a-628b-464f-a8bb-e97ab04a73eb,69a544f8-957d-4dda-a777-608ec7b95617,90789a49-ab1f-4229-bd89-e1c56ea8be3a,cd53e21a-e5ed-4d9b-a258-42a23e9715bb,8d409c7b-29d2-4df2-aa69-ab56c9339bc5,6566f489-f254-4bca-a86b-90f5ac7471d7,d9a44a36-6a3c-4d5d-9c66-b2c0afb479d1,9cf87956-37e6-4bed-9cea-a64b13ae0de5,1dc0b953-ef66-4a78-9989-7cf1053b5b41,a1a08e1d-1353-4d6f-94ce-f90872074c7c,51d9c72b-ba60-47c5-9cc9-3ed6212a3982,3cf35bfe-d881-4881-a268-b5fa0607bf89,bc8d3e4a-8ba0-4cab-be1c-b20ae3af91e4,99306c29-bcef-49d5-bcd1-78d4e541e197,d0fcfaf0-7c6d-4ead-915b-6d9d9130e747,59c68d10-55c1-4629-b05a-bedde7f343b5,d86ee090-3a2e-44d0-abed-6f0f40a38c08,25cf07b8-606f-4448-99ef-ea87ca3097ba,1ca222ed-cc79-4e78-9ca3-3547e7b37e3a,63e4e0d4-0cb1-442e-8316-8c12a6e3ec14,0eb6fd6d-eaaa-46db-ab16-02e610f238a9,99b22b39-30d5-40dc-82d9-c1cd3c793a44,8c342edc-619b-4079-a28c-b7129f3670a6,e8991792-c875-4f40-99a0-54c902d42a52,f93b0850-2c60-479d-831b-b8c8f9493d61,c63a4306-4003-45a6-8be1-d669457473a9,d3cbb579-70a5-46e8-95f6-72cc9afa4ba3,67e76447-f16a-4a65-aba7-fc7166fd9b63,a783e7a3-cecc-4d15-b18c-3f2ddfc2dc6c,62e9d35f-153d-4744-b97c-9abfb2a90b1e" + } + }, + "evaluation_timestamp": "1776696116.0", + "metric_config": { + "evaluation_description": "bootstrap_score", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.2964017949707683, + "uncertainty": { + "standard_deviation": 0.03706635839637665, + "num_samples": 148 + }, + "details": { + "total_matched_trajectories": "148", + "stopping_reason_count_repetition_guard": "148", + "avg_total_tokens_target_model": "10812.39", + "total_total_tokens_target_model": "1600234.0", + "avg_total_tokens_other_models": "56916.10", + "total_total_tokens_other_models": "8423583.0", + "avg_total_tokens_all_models": "67728.49", + "total_total_tokens_all_models": "10023817.0", + "avg_turn_count": "25.86", + "total_turn_count": "3827.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "3600", + "attempt_timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + }, + { + "evaluation_name": "std on healthbench/S-adaptive/+3ep/114707e5 for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "50", + "sample_ids": "267bda16-fd4b-4a9e-b3ac-5602c2532eed,0e819a9c-851d-4a7a-9263-62cfc8ce1b48,328c55cf-0b57-4992-99b1-f6a21af62ade,7b415a2f-22b0-4735-a38e-567f1891acc5,c49ac752-3693-4be1-ab4d-91fbdd02f827,a079d928-7b78-4a85-a72b-4274d7fc2914,a20cbbf3-330b-46d9-a35b-92e89a022282,be6150a5-8060-4434-9724-afede307d495,3a896889-9029-4ef8-8684-ca378da4702b,7074bb7a-a595-434d-883e-95ef6f5f6667,9f7b555a-fc30-40c3-88d3-f2ece06454a2,a6204079-9e72-42ad-8f39-a65552fa2e68,c0d80b93-b2d5-4129-b998-b1312f98e2a8,3dc41419-2565-428e-b9b9-02e3fd93c470,a91a22f8-d2e6-4593-bcae-6d479ed38031,d0d3426f-8b7a-4501-adb4-86d8fa9099c8,a7dac6a9-eab5-4c01-882c-dbf6c774d7c3,c1240427-ae38-4335-9e80-492c13ef4893,b7ddde65-702e-4255-b1d7-09d01248e76d,00d34549-134c-48e2-b4bb-b54946da70cc,85c11e2a-628b-464f-a8bb-e97ab04a73eb,69a544f8-957d-4dda-a777-608ec7b95617,90789a49-ab1f-4229-bd89-e1c56ea8be3a,cd53e21a-e5ed-4d9b-a258-42a23e9715bb,8d409c7b-29d2-4df2-aa69-ab56c9339bc5,6566f489-f254-4bca-a86b-90f5ac7471d7,d9a44a36-6a3c-4d5d-9c66-b2c0afb479d1,9cf87956-37e6-4bed-9cea-a64b13ae0de5,1dc0b953-ef66-4a78-9989-7cf1053b5b41,a1a08e1d-1353-4d6f-94ce-f90872074c7c,51d9c72b-ba60-47c5-9cc9-3ed6212a3982,3cf35bfe-d881-4881-a268-b5fa0607bf89,bc8d3e4a-8ba0-4cab-be1c-b20ae3af91e4,99306c29-bcef-49d5-bcd1-78d4e541e197,d0fcfaf0-7c6d-4ead-915b-6d9d9130e747,59c68d10-55c1-4629-b05a-bedde7f343b5,d86ee090-3a2e-44d0-abed-6f0f40a38c08,25cf07b8-606f-4448-99ef-ea87ca3097ba,1ca222ed-cc79-4e78-9ca3-3547e7b37e3a,63e4e0d4-0cb1-442e-8316-8c12a6e3ec14,0eb6fd6d-eaaa-46db-ab16-02e610f238a9,99b22b39-30d5-40dc-82d9-c1cd3c793a44,8c342edc-619b-4079-a28c-b7129f3670a6,e8991792-c875-4f40-99a0-54c902d42a52,f93b0850-2c60-479d-831b-b8c8f9493d61,c63a4306-4003-45a6-8be1-d669457473a9,d3cbb579-70a5-46e8-95f6-72cc9afa4ba3,67e76447-f16a-4a65-aba7-fc7166fd9b63,a783e7a3-cecc-4d15-b18c-3f2ddfc2dc6c,62e9d35f-153d-4744-b97c-9abfb2a90b1e" + } + }, + "evaluation_timestamp": "1776696116.0", + "metric_config": { + "evaluation_description": "std", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.03706635839637665, + "uncertainty": { + "standard_deviation": 0.03706635839637665, + "num_samples": 148 + }, + "details": { + "total_matched_trajectories": "148", + "stopping_reason_count_repetition_guard": "148", + "avg_total_tokens_target_model": "10812.39", + "total_total_tokens_target_model": "1600234.0", + "avg_total_tokens_other_models": "56916.10", + "total_total_tokens_other_models": "8423583.0", + "avg_total_tokens_all_models": "67728.49", + "total_total_tokens_all_models": "10023817.0", + "avg_turn_count": "25.86", + "total_turn_count": "3827.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "3600", + "attempt_timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + }, + { + "evaluation_name": "criteria_met_rate on healthbench/S-adaptive/+3ep/114707e5 for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "50", + "sample_ids": "267bda16-fd4b-4a9e-b3ac-5602c2532eed,0e819a9c-851d-4a7a-9263-62cfc8ce1b48,328c55cf-0b57-4992-99b1-f6a21af62ade,7b415a2f-22b0-4735-a38e-567f1891acc5,c49ac752-3693-4be1-ab4d-91fbdd02f827,a079d928-7b78-4a85-a72b-4274d7fc2914,a20cbbf3-330b-46d9-a35b-92e89a022282,be6150a5-8060-4434-9724-afede307d495,3a896889-9029-4ef8-8684-ca378da4702b,7074bb7a-a595-434d-883e-95ef6f5f6667,9f7b555a-fc30-40c3-88d3-f2ece06454a2,a6204079-9e72-42ad-8f39-a65552fa2e68,c0d80b93-b2d5-4129-b998-b1312f98e2a8,3dc41419-2565-428e-b9b9-02e3fd93c470,a91a22f8-d2e6-4593-bcae-6d479ed38031,d0d3426f-8b7a-4501-adb4-86d8fa9099c8,a7dac6a9-eab5-4c01-882c-dbf6c774d7c3,c1240427-ae38-4335-9e80-492c13ef4893,b7ddde65-702e-4255-b1d7-09d01248e76d,00d34549-134c-48e2-b4bb-b54946da70cc,85c11e2a-628b-464f-a8bb-e97ab04a73eb,69a544f8-957d-4dda-a777-608ec7b95617,90789a49-ab1f-4229-bd89-e1c56ea8be3a,cd53e21a-e5ed-4d9b-a258-42a23e9715bb,8d409c7b-29d2-4df2-aa69-ab56c9339bc5,6566f489-f254-4bca-a86b-90f5ac7471d7,d9a44a36-6a3c-4d5d-9c66-b2c0afb479d1,9cf87956-37e6-4bed-9cea-a64b13ae0de5,1dc0b953-ef66-4a78-9989-7cf1053b5b41,a1a08e1d-1353-4d6f-94ce-f90872074c7c,51d9c72b-ba60-47c5-9cc9-3ed6212a3982,3cf35bfe-d881-4881-a268-b5fa0607bf89,bc8d3e4a-8ba0-4cab-be1c-b20ae3af91e4,99306c29-bcef-49d5-bcd1-78d4e541e197,d0fcfaf0-7c6d-4ead-915b-6d9d9130e747,59c68d10-55c1-4629-b05a-bedde7f343b5,d86ee090-3a2e-44d0-abed-6f0f40a38c08,25cf07b8-606f-4448-99ef-ea87ca3097ba,1ca222ed-cc79-4e78-9ca3-3547e7b37e3a,63e4e0d4-0cb1-442e-8316-8c12a6e3ec14,0eb6fd6d-eaaa-46db-ab16-02e610f238a9,99b22b39-30d5-40dc-82d9-c1cd3c793a44,8c342edc-619b-4079-a28c-b7129f3670a6,e8991792-c875-4f40-99a0-54c902d42a52,f93b0850-2c60-479d-831b-b8c8f9493d61,c63a4306-4003-45a6-8be1-d669457473a9,d3cbb579-70a5-46e8-95f6-72cc9afa4ba3,67e76447-f16a-4a65-aba7-fc7166fd9b63,a783e7a3-cecc-4d15-b18c-3f2ddfc2dc6c,62e9d35f-153d-4744-b97c-9abfb2a90b1e" + } + }, + "evaluation_timestamp": "1776696116.0", + "metric_config": { + "evaluation_description": "criteria_met_rate", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.3595113438045376, + "uncertainty": { + "standard_deviation": 0.03706635839637665, + "num_samples": 148 + }, + "details": { + "total_matched_trajectories": "148", + "stopping_reason_count_repetition_guard": "148", + "avg_total_tokens_target_model": "10812.39", + "total_total_tokens_target_model": "1600234.0", + "avg_total_tokens_other_models": "56916.10", + "total_total_tokens_other_models": "8423583.0", + "avg_total_tokens_all_models": "67728.49", + "total_total_tokens_all_models": "10023817.0", + "avg_turn_count": "25.86", + "total_turn_count": "3827.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "3600", + "attempt_timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + }, + { + "evaluation_name": "mean on healthbench/S-adaptive/+3ep/114707e5 for scorer healthbench_score", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "50", + "sample_ids": "267bda16-fd4b-4a9e-b3ac-5602c2532eed,0e819a9c-851d-4a7a-9263-62cfc8ce1b48,328c55cf-0b57-4992-99b1-f6a21af62ade,7b415a2f-22b0-4735-a38e-567f1891acc5,c49ac752-3693-4be1-ab4d-91fbdd02f827,a079d928-7b78-4a85-a72b-4274d7fc2914,a20cbbf3-330b-46d9-a35b-92e89a022282,be6150a5-8060-4434-9724-afede307d495,3a896889-9029-4ef8-8684-ca378da4702b,7074bb7a-a595-434d-883e-95ef6f5f6667,9f7b555a-fc30-40c3-88d3-f2ece06454a2,a6204079-9e72-42ad-8f39-a65552fa2e68,c0d80b93-b2d5-4129-b998-b1312f98e2a8,3dc41419-2565-428e-b9b9-02e3fd93c470,a91a22f8-d2e6-4593-bcae-6d479ed38031,d0d3426f-8b7a-4501-adb4-86d8fa9099c8,a7dac6a9-eab5-4c01-882c-dbf6c774d7c3,c1240427-ae38-4335-9e80-492c13ef4893,b7ddde65-702e-4255-b1d7-09d01248e76d,00d34549-134c-48e2-b4bb-b54946da70cc,85c11e2a-628b-464f-a8bb-e97ab04a73eb,69a544f8-957d-4dda-a777-608ec7b95617,90789a49-ab1f-4229-bd89-e1c56ea8be3a,cd53e21a-e5ed-4d9b-a258-42a23e9715bb,8d409c7b-29d2-4df2-aa69-ab56c9339bc5,6566f489-f254-4bca-a86b-90f5ac7471d7,d9a44a36-6a3c-4d5d-9c66-b2c0afb479d1,9cf87956-37e6-4bed-9cea-a64b13ae0de5,1dc0b953-ef66-4a78-9989-7cf1053b5b41,a1a08e1d-1353-4d6f-94ce-f90872074c7c,51d9c72b-ba60-47c5-9cc9-3ed6212a3982,3cf35bfe-d881-4881-a268-b5fa0607bf89,bc8d3e4a-8ba0-4cab-be1c-b20ae3af91e4,99306c29-bcef-49d5-bcd1-78d4e541e197,d0fcfaf0-7c6d-4ead-915b-6d9d9130e747,59c68d10-55c1-4629-b05a-bedde7f343b5,d86ee090-3a2e-44d0-abed-6f0f40a38c08,25cf07b8-606f-4448-99ef-ea87ca3097ba,1ca222ed-cc79-4e78-9ca3-3547e7b37e3a,63e4e0d4-0cb1-442e-8316-8c12a6e3ec14,0eb6fd6d-eaaa-46db-ab16-02e610f238a9,99b22b39-30d5-40dc-82d9-c1cd3c793a44,8c342edc-619b-4079-a28c-b7129f3670a6,e8991792-c875-4f40-99a0-54c902d42a52,f93b0850-2c60-479d-831b-b8c8f9493d61,c63a4306-4003-45a6-8be1-d669457473a9,d3cbb579-70a5-46e8-95f6-72cc9afa4ba3,67e76447-f16a-4a65-aba7-fc7166fd9b63,a783e7a3-cecc-4d15-b18c-3f2ddfc2dc6c,62e9d35f-153d-4744-b97c-9abfb2a90b1e" + } + }, + "evaluation_timestamp": "1776696116.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.2964017949707683, + "uncertainty": { + "num_samples": 148 + }, + "details": { + "total_matched_trajectories": "148", + "stopping_reason_count_repetition_guard": "148", + "avg_total_tokens_target_model": "10812.39", + "total_total_tokens_target_model": "1600234.0", + "avg_total_tokens_other_models": "56916.10", + "total_total_tokens_other_models": "8423583.0", + "avg_total_tokens_all_models": "67728.49", + "total_total_tokens_all_models": "10023817.0", + "avg_turn_count": "25.86", + "total_turn_count": "3827.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "3600", + "attempt_timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/openai/gpt-5.4-2026-03-05/ef8ab69f-1c1c-409f-84fe-fab6c9380774_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "17fea7fc391373319923461a8139514b7fb587742fddf4bc5acb41dd5a7e3dba", + "total_rows": 148 + } +} \ No newline at end of file diff --git a/data/healthbench/openai/gpt-5.4-2026-03-05/ef8ab69f-1c1c-409f-84fe-fab6c9380774_samples.jsonl b/data/healthbench/openai/gpt-5.4-2026-03-05/ef8ab69f-1c1c-409f-84fe-fab6c9380774_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..bab3467c92878eb03e71b71f14064b74d97cf12e --- /dev/null +++ b/data/healthbench/openai/gpt-5.4-2026-03-05/ef8ab69f-1c1c-409f-84fe-fab6c9380774_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:17fea7fc391373319923461a8139514b7fb587742fddf4bc5acb41dd5a7e3dba +size 9682897 diff --git a/data/healthbench/openai/gpt-5.4-2026-03-05/fd44cf57-b1cd-4f7c-82ef-f5ea753b78c0.json b/data/healthbench/openai/gpt-5.4-2026-03-05/fd44cf57-b1cd-4f7c-82ef-f5ea753b78c0.json new file mode 100644 index 0000000000000000000000000000000000000000..fc934809a262cdc6bd3bab44ec6f7d0515c89c76 --- /dev/null +++ b/data/healthbench/openai/gpt-5.4-2026-03-05/fd44cf57-b1cd-4f7c-82ef-f5ea753b78c0.json @@ -0,0 +1,304 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "healthbench/openai_gpt-5.4-2026-03-05/1786163860.737885", + "evaluation_timestamp": "1775558527.0", + "retrieved_timestamp": "1786163860.737885", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.202,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "bootstrap_score on healthbench/S-adaptive/+8ep for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "5", + "sample_ids": "1f548d5b-cd00-49a0-b327-283a2e00debd,0b8f1d60-2081-4562-98f7-b6a976fe1c6d,6f7a2ee9-e9c6-42d8-b79f-22dea966b8d2,19ec4833-86e9-4166-8b82-d1da09f31fd7,7ebc830a-8dbd-489b-9d61-4d8bacf0db8d" + } + }, + "evaluation_timestamp": "1775558527.0", + "metric_config": { + "evaluation_description": "bootstrap_score", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.6299411918903445, + "uncertainty": { + "standard_deviation": 0.06504824053547809, + "num_samples": 40 + }, + "details": { + "total_matched_trajectories": "1", + "stopping_reason_count_repetition_guard": "1", + "avg_total_tokens_target_model": "46955.00", + "total_total_tokens_target_model": "46955.0", + "avg_total_tokens_other_models": "98284.00", + "total_total_tokens_other_models": "98284.0", + "avg_total_tokens_all_models": "145239.00", + "total_total_tokens_all_models": "145239.0", + "avg_turn_count": "23.00", + "total_turn_count": "23.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "reasoning_effort": "\"high\"" + } + } + }, + { + "evaluation_name": "std on healthbench/S-adaptive/+8ep for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "5", + "sample_ids": "1f548d5b-cd00-49a0-b327-283a2e00debd,0b8f1d60-2081-4562-98f7-b6a976fe1c6d,6f7a2ee9-e9c6-42d8-b79f-22dea966b8d2,19ec4833-86e9-4166-8b82-d1da09f31fd7,7ebc830a-8dbd-489b-9d61-4d8bacf0db8d" + } + }, + "evaluation_timestamp": "1775558527.0", + "metric_config": { + "evaluation_description": "std", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.06504824053547809, + "uncertainty": { + "standard_deviation": 0.06504824053547809, + "num_samples": 40 + }, + "details": { + "total_matched_trajectories": "1", + "stopping_reason_count_repetition_guard": "1", + "avg_total_tokens_target_model": "46955.00", + "total_total_tokens_target_model": "46955.0", + "avg_total_tokens_other_models": "98284.00", + "total_total_tokens_other_models": "98284.0", + "avg_total_tokens_all_models": "145239.00", + "total_total_tokens_all_models": "145239.0", + "avg_turn_count": "23.00", + "total_turn_count": "23.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "reasoning_effort": "\"high\"" + } + } + }, + { + "evaluation_name": "criteria_met_rate on healthbench/S-adaptive/+8ep for scorer _scorer", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "5", + "sample_ids": "1f548d5b-cd00-49a0-b327-283a2e00debd,0b8f1d60-2081-4562-98f7-b6a976fe1c6d,6f7a2ee9-e9c6-42d8-b79f-22dea966b8d2,19ec4833-86e9-4166-8b82-d1da09f31fd7,7ebc830a-8dbd-489b-9d61-4d8bacf0db8d" + } + }, + "evaluation_timestamp": "1775558527.0", + "metric_config": { + "evaluation_description": "criteria_met_rate", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.59375, + "uncertainty": { + "standard_deviation": 0.06504824053547809, + "num_samples": 40 + }, + "details": { + "total_matched_trajectories": "1", + "stopping_reason_count_repetition_guard": "1", + "avg_total_tokens_target_model": "46955.00", + "total_total_tokens_target_model": "46955.0", + "avg_total_tokens_other_models": "98284.00", + "total_total_tokens_other_models": "98284.0", + "avg_total_tokens_all_models": "145239.00", + "total_total_tokens_all_models": "145239.0", + "avg_turn_count": "23.00", + "total_turn_count": "23.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "reasoning_effort": "\"high\"" + } + } + }, + { + "evaluation_name": "mean on healthbench/S-adaptive/+8ep for scorer healthbench_score", + "source_data": { + "dataset_name": "healthbench", + "source_type": "other", + "additional_details": { + "shuffled": "False", + "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl", + "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06", + "samples_number": "5", + "sample_ids": "1f548d5b-cd00-49a0-b327-283a2e00debd,0b8f1d60-2081-4562-98f7-b6a976fe1c6d,6f7a2ee9-e9c6-42d8-b79f-22dea966b8d2,19ec4833-86e9-4166-8b82-d1da09f31fd7,7ebc830a-8dbd-489b-9d61-4d8bacf0db8d" + } + }, + "evaluation_timestamp": "1775558527.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.6299411918903445, + "uncertainty": { + "num_samples": 40 + }, + "details": { + "total_matched_trajectories": "1", + "stopping_reason_count_repetition_guard": "1", + "avg_total_tokens_target_model": "46955.00", + "total_total_tokens_target_model": "46955.0", + "avg_total_tokens_other_models": "98284.00", + "total_total_tokens_other_models": "98284.0", + "avg_total_tokens_all_models": "145239.00", + "total_total_tokens_all_models": "145239.0", + "avg_turn_count": "23.00", + "total_turn_count": "23.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/healthbench/openai/gpt-5.4-2026-03-05/fd44cf57-b1cd-4f7c-82ef-f5ea753b78c0_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "3d21c1d87ab20d397bb573d4a151f6aff0801614ca052b802478c77f16c9c6ce", + "total_rows": 40 + } +} \ No newline at end of file diff --git a/data/healthbench/openai/gpt-5.4-2026-03-05/fd44cf57-b1cd-4f7c-82ef-f5ea753b78c0_samples.jsonl b/data/healthbench/openai/gpt-5.4-2026-03-05/fd44cf57-b1cd-4f7c-82ef-f5ea753b78c0_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..aef275e16a5d952b425a634a6d61e31d41145b42 --- /dev/null +++ b/data/healthbench/openai/gpt-5.4-2026-03-05/fd44cf57-b1cd-4f7c-82ef-f5ea753b78c0_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3d21c1d87ab20d397bb573d4a151f6aff0801614ca052b802478c77f16c9c6ce +size 2984518 diff --git a/data/hle/anthropic/claude-opus-4-20250514/02951aba-98d9-435e-8400-8b469a0dd815.json b/data/hle/anthropic/claude-opus-4-20250514/02951aba-98d9-435e-8400-8b469a0dd815.json new file mode 100644 index 0000000000000000000000000000000000000000..d90b1151b4f42b19df79b24549ce7f4b9c417af7 --- /dev/null +++ b/data/hle/anthropic/claude-opus-4-20250514/02951aba-98d9-435e-8400-8b469a0dd815.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/anthropic_claude-opus-4-20250514/1786167072.128841", + "evaluation_timestamp": "1775078458.0", + "retrieved_timestamp": "1786167072.128841", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.202,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-20250514", + "id": "anthropic/claude-opus-4-20250514", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/anthropic/claude-opus-4-20250514/02951aba-98d9-435e-8400-8b469a0dd815_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "ca8b9e44e5bf2c9829ffaca382fed0c1da5a2822315c2a81c8209d6e421cabd1", + "total_rows": 100 + } +} \ No newline at end of file diff --git a/data/hle/anthropic/claude-opus-4-20250514/02951aba-98d9-435e-8400-8b469a0dd815_samples.jsonl b/data/hle/anthropic/claude-opus-4-20250514/02951aba-98d9-435e-8400-8b469a0dd815_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..453320ad1f87417eedd30849b08069c4b9e9d2ee --- /dev/null +++ b/data/hle/anthropic/claude-opus-4-20250514/02951aba-98d9-435e-8400-8b469a0dd815_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ca8b9e44e5bf2c9829ffaca382fed0c1da5a2822315c2a81c8209d6e421cabd1 +size 20933818 diff --git a/data/hle/anthropic/claude-opus-4-20250514/2ecd74e8-9979-49a1-a073-3ca298ebd69c.json b/data/hle/anthropic/claude-opus-4-20250514/2ecd74e8-9979-49a1-a073-3ca298ebd69c.json new file mode 100644 index 0000000000000000000000000000000000000000..57dd462ece59f96453ecd387f0fb40aad78ee1c8 --- /dev/null +++ b/data/hle/anthropic/claude-opus-4-20250514/2ecd74e8-9979-49a1-a073-3ca298ebd69c.json @@ -0,0 +1,211 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/anthropic_claude-opus-4-20250514/1786183218.562696", + "evaluation_timestamp": "1775080047.0", + "retrieved_timestamp": "1786183218.562696", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.202,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-20250514", + "id": "anthropic/claude-opus-4-20250514", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "mean on hle for scorer _llm_grader", + "source_data": { + "dataset_name": "hle", + "source_type": "hf_dataset", + "hf_repo": "cais/hle", + "samples_number": 1909, + "sample_ids": [ + "hle_d6b73839", + "hle_07662c9c", + "hle_764ce81c", + "hle_4e3440f7", + "hle_baccd6da", + "hle_f46c2857", + "hle_86a18059", + "hle_8770ab31", + "hle_d1bd2767", + "hle_6499b500", + "hle_e1296346", + "hle_ca6c436b", + "hle_a559277e", + "hle_8a419425", + "hle_85e95ef4", + "hle_28643173", + "hle_bc5fbce2", + "hle_25e0c71d", + "hle_76cdecee", + "hle_dfdb78f5", + "hle_142ffcff", + "hle_ffb12d83", + "hle_37c97863", + "hle_209b6b64", + "hle_0702accc", + "hle_16388cc8", + "hle_4418bea4", + "hle_ebd62885", + "hle_4485802f", + "hle_9019c023", + "hle_9282e7ae", + "hle_44007b97", + "hle_7808792d", + "hle_e3b8f5c2", + "hle_3cf4483d", + "hle_f85a0fa1", + "hle_c3ca0bda", + "hle_376b8247", + "hle_31497469", + "hle_c43a27a3", + "hle_ea41e785", + "hle_d0c8b07e", + "hle_0291c889", + "hle_b1a7811d", + "hle_dab7fbb1", + "hle_cf3d65bd", + "hle_5e235b48", + "hle_0ff76ff3", + "hle_70e5aebc", + "hle_aa2fed90", + "hle_9df5d371", + "hle_5759e7d7", + "hle_ca59b32c", + "hle_33bc94e2", + "hle_e3817b7e", + "hle_d7b7c3dc", + "hle_1dc141bb", + "hle_a764505b", + "hle_b6cec502", + "hle_e62ba103", + "hle_9255683e", + "hle_34c69126", + "hle_5f04b314", + "hle_fcd56401", + "hle_b5c681d9", + "hle_d5a55900", + "hle_3e11b38a", + "hle_0a1eec3f", + "hle_5d49423c", + "hle_1d580261", + "hle_82f5c665", + "hle_cc7f25c0", + "hle_226e134b", + "hle_c1158fb8", + "hle_d0ea1019", + "hle_fcd42d0c", + "hle_b33012cf", + "hle_3697d2bc", + "hle_5e349785", + "hle_8b0384c4", + "hle_ee938b26", + "hle_d190d8c2", + "hle_b26bf8aa", + "hle_c2ff3a43", + "hle_141cd3ef", + "hle_5671e386", + "hle_a634e89e", + "hle_670287ad", + "hle_298b5616", + "hle_e14d8c23", + "hle_510b6040", + "hle_c6adc1b4", + "hle_0ec247c5", + "hle_aa343d51", + "hle_d52d3dae", + "hle_6d963e58", + "hle_35a40c52", + "hle_45d1c140", + "hle_c0ed25f7", + "hle_cd4aa5b3" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1775080047.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.3543333333333333, + "uncertainty": { + "standard_error": { + "value": 0.039510096168123475 + }, + "num_samples": 962 + }, + "details": { + "total_matched_trajectories": "468", + "stopping_reason_count_repetition_guard": "279", + "stopping_reason_count_completed_on_successful_submit": "166", + "stopping_reason_count_token_limit": "19", + "stopping_reason_count_completed_with_submit": "4", + "avg_total_tokens_target_model": "457868.43", + "total_total_tokens_target_model": "214282424.0", + "avg_total_tokens_other_models": "36041.89", + "total_total_tokens_other_models": "16867604.0", + "avg_total_tokens_all_models": "493910.32", + "total_total_tokens_all_models": "231150028.0", + "avg_turn_count": "78.35", + "total_turn_count": "36668.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\"}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "3", + "attempt_timeout": "1200", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 5000000 + }, + "sandbox": {}, + "max_attempts": 3 + }, + "additional_details": { + "max_retries": "3" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/anthropic/claude-opus-4-20250514/2ecd74e8-9979-49a1-a073-3ca298ebd69c_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "b8d694aa08908edf6ef3616e226af70ab4fbd4fc6927b4368fa5c7faba951911", + "total_rows": 962 + } +} \ No newline at end of file diff --git a/data/hle/anthropic/claude-opus-4-20250514/2ecd74e8-9979-49a1-a073-3ca298ebd69c_samples.jsonl b/data/hle/anthropic/claude-opus-4-20250514/2ecd74e8-9979-49a1-a073-3ca298ebd69c_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..1d74b9e6de0d57022d6f50fa994dcce1a8579ebd --- /dev/null +++ b/data/hle/anthropic/claude-opus-4-20250514/2ecd74e8-9979-49a1-a073-3ca298ebd69c_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b8d694aa08908edf6ef3616e226af70ab4fbd4fc6927b4368fa5c7faba951911 +size 168400023 diff --git a/data/hle/anthropic/claude-opus-4-20250514/89a7bd40-98e7-441f-9dd0-dd873e72f421.json b/data/hle/anthropic/claude-opus-4-20250514/89a7bd40-98e7-441f-9dd0-dd873e72f421.json new file mode 100644 index 0000000000000000000000000000000000000000..39aa619a04066799a95062f1816490f79680cb8e --- /dev/null +++ b/data/hle/anthropic/claude-opus-4-20250514/89a7bd40-98e7-441f-9dd0-dd873e72f421.json @@ -0,0 +1,209 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/anthropic_claude-opus-4-20250514/1786183524.356333", + "evaluation_timestamp": "1774990283.0", + "retrieved_timestamp": "1786183524.356333", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.191.dev111+g7c0c4345,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-20250514", + "id": "anthropic/claude-opus-4-20250514", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "mean on hle for scorer _llm_grader", + "source_data": { + "dataset_name": "hle", + "source_type": "hf_dataset", + "hf_repo": "cais/hle", + "samples_number": 1909, + "sample_ids": [ + "hle_d6b73839", + "hle_07662c9c", + "hle_764ce81c", + "hle_4e3440f7", + "hle_baccd6da", + "hle_f46c2857", + "hle_86a18059", + "hle_8770ab31", + "hle_d1bd2767", + "hle_6499b500", + "hle_e1296346", + "hle_ca6c436b", + "hle_a559277e", + "hle_8a419425", + "hle_85e95ef4", + "hle_28643173", + "hle_bc5fbce2", + "hle_25e0c71d", + "hle_76cdecee", + "hle_dfdb78f5", + "hle_142ffcff", + "hle_ffb12d83", + "hle_37c97863", + "hle_209b6b64", + "hle_0702accc", + "hle_16388cc8", + "hle_4418bea4", + "hle_ebd62885", + "hle_4485802f", + "hle_9019c023", + "hle_9282e7ae", + "hle_44007b97", + "hle_7808792d", + "hle_e3b8f5c2", + "hle_3cf4483d", + "hle_f85a0fa1", + "hle_c3ca0bda", + "hle_376b8247", + "hle_31497469", + "hle_c43a27a3", + "hle_ea41e785", + "hle_d0c8b07e", + "hle_0291c889", + "hle_b1a7811d", + "hle_dab7fbb1", + "hle_cf3d65bd", + "hle_5e235b48", + "hle_0ff76ff3", + "hle_70e5aebc", + "hle_aa2fed90", + "hle_9df5d371", + "hle_5759e7d7", + "hle_ca59b32c", + "hle_33bc94e2", + "hle_e3817b7e", + "hle_d7b7c3dc", + "hle_1dc141bb", + "hle_a764505b", + "hle_b6cec502", + "hle_e62ba103", + "hle_9255683e", + "hle_34c69126", + "hle_5f04b314", + "hle_fcd56401", + "hle_b5c681d9", + "hle_d5a55900", + "hle_3e11b38a", + "hle_0a1eec3f", + "hle_5d49423c", + "hle_1d580261", + "hle_82f5c665", + "hle_cc7f25c0", + "hle_226e134b", + "hle_c1158fb8", + "hle_d0ea1019", + "hle_fcd42d0c", + "hle_b33012cf", + "hle_3697d2bc", + "hle_5e349785", + "hle_8b0384c4", + "hle_ee938b26", + "hle_d190d8c2", + "hle_b26bf8aa", + "hle_c2ff3a43", + "hle_141cd3ef", + "hle_5671e386", + "hle_a634e89e", + "hle_670287ad", + "hle_298b5616", + "hle_e14d8c23", + "hle_510b6040", + "hle_c6adc1b4", + "hle_0ec247c5", + "hle_aa343d51", + "hle_d52d3dae", + "hle_6d963e58", + "hle_35a40c52", + "hle_45d1c140", + "hle_c0ed25f7", + "hle_cd4aa5b3" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1774990283.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.077, + "uncertainty": { + "standard_error": { + "value": 0.01797051457080164 + }, + "num_samples": 999 + }, + "details": { + "total_matched_trajectories": "500", + "stopping_reason_count_repetition_guard": "491", + "stopping_reason_count_token_limit": "9", + "avg_total_tokens_target_model": "246982.47", + "total_total_tokens_target_model": "123491233.0", + "avg_total_tokens_other_models": "18687.59", + "total_total_tokens_other_models": "9343796.0", + "avg_total_tokens_all_models": "265670.06", + "total_total_tokens_all_models": "132835029.0", + "avg_turn_count": "41.21", + "total_turn_count": "20605.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\"}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "3", + "attempt_timeout": "1200", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 5000000 + }, + "sandbox": {}, + "max_attempts": 3 + }, + "additional_details": { + "max_retries": "3" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/anthropic/claude-opus-4-20250514/89a7bd40-98e7-441f-9dd0-dd873e72f421_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "196396ff0783971697c160502200ab5f499cffc2062fa70a617fb442284d0a9e", + "total_rows": 999 + } +} \ No newline at end of file diff --git a/data/hle/anthropic/claude-opus-4-20250514/89a7bd40-98e7-441f-9dd0-dd873e72f421_samples.jsonl b/data/hle/anthropic/claude-opus-4-20250514/89a7bd40-98e7-441f-9dd0-dd873e72f421_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..6b67f2517b69c198f2eba7b9bb14c41a7d4600bb --- /dev/null +++ b/data/hle/anthropic/claude-opus-4-20250514/89a7bd40-98e7-441f-9dd0-dd873e72f421_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:196396ff0783971697c160502200ab5f499cffc2062fa70a617fb442284d0a9e +size 91663732 diff --git a/data/hle/anthropic/claude-opus-4-5-20251101/1df6e8fa-dedb-44c1-ae28-c012002be427.json b/data/hle/anthropic/claude-opus-4-5-20251101/1df6e8fa-dedb-44c1-ae28-c012002be427.json new file mode 100644 index 0000000000000000000000000000000000000000..48942b75fd045cae9a02f382a6a95820f30256ae --- /dev/null +++ b/data/hle/anthropic/claude-opus-4-5-20251101/1df6e8fa-dedb-44c1-ae28-c012002be427.json @@ -0,0 +1,118 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/anthropic_claude-opus-4-5-20251101/1786164257.371688", + "evaluation_timestamp": "1776642929.0", + "retrieved_timestamp": "1786164257.371688", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-5-20251101", + "id": "anthropic/claude-opus-4-5-20251101", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "mean on hle/S-adaptive/+2ep/b2283bee for scorer _llm_grader", + "source_data": { + "dataset_name": "hle", + "source_type": "hf_dataset", + "hf_repo": "cais/hle", + "samples_number": 8, + "sample_ids": [ + "hle_86a18059", + "hle_8770ab31", + "hle_209b6b64", + "hle_4418bea4", + "hle_33bc94e2", + "hle_d7b7c3dc", + "hle_1dc141bb", + "hle_a764505b" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1776642929.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.0625, + "uncertainty": { + "standard_error": { + "value": 0.0625 + }, + "num_samples": 16 + }, + "details": { + "total_matched_trajectories": "16", + "stopping_reason_count_repetition_guard": "14", + "stopping_reason_count_token_limit": "2", + "avg_total_tokens_target_model": "1699971.50", + "total_total_tokens_target_model": "27199544.0", + "avg_total_tokens_other_models": "29643.44", + "total_total_tokens_other_models": "474295.0", + "avg_total_tokens_all_models": "1729614.94", + "total_total_tokens_all_models": "27673839.0", + "avg_turn_count": "84.12", + "total_turn_count": "1346.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "3600", + "attempt_timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 5000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "3600", + "max_connections": "10" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/anthropic/claude-opus-4-5-20251101/1df6e8fa-dedb-44c1-ae28-c012002be427_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "23f8a040af18966e117f58e8deaf405da73c3a8ec0932b92d0956248f129601f", + "total_rows": 16 + } +} \ No newline at end of file diff --git a/data/hle/anthropic/claude-opus-4-5-20251101/1df6e8fa-dedb-44c1-ae28-c012002be427_samples.jsonl b/data/hle/anthropic/claude-opus-4-5-20251101/1df6e8fa-dedb-44c1-ae28-c012002be427_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..d422058ec8bb2b338875a145541724250e9e3786 --- /dev/null +++ b/data/hle/anthropic/claude-opus-4-5-20251101/1df6e8fa-dedb-44c1-ae28-c012002be427_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:23f8a040af18966e117f58e8deaf405da73c3a8ec0932b92d0956248f129601f +size 2878975 diff --git a/data/hle/anthropic/claude-opus-4-5-20251101/26d05853-6958-48bd-91cd-8b8160b38834.json b/data/hle/anthropic/claude-opus-4-5-20251101/26d05853-6958-48bd-91cd-8b8160b38834.json new file mode 100644 index 0000000000000000000000000000000000000000..5c425125df73efec7333a1330657cd3bcf40c63b --- /dev/null +++ b/data/hle/anthropic/claude-opus-4-5-20251101/26d05853-6958-48bd-91cd-8b8160b38834.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/anthropic_claude-opus-4-5-20251101/1786166912.041754", + "evaluation_timestamp": "1776519735.0", + "retrieved_timestamp": "1786166912.041754", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-5-20251101", + "id": "anthropic/claude-opus-4-5-20251101", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/anthropic/claude-opus-4-5-20251101/26d05853-6958-48bd-91cd-8b8160b38834_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "32a35056f8982d77bc9d1a65aeaf60b9f8ddf110408bccc876fc6f673ca9e3cb", + "total_rows": 70 + } +} \ No newline at end of file diff --git a/data/hle/anthropic/claude-opus-4-5-20251101/26d05853-6958-48bd-91cd-8b8160b38834_samples.jsonl b/data/hle/anthropic/claude-opus-4-5-20251101/26d05853-6958-48bd-91cd-8b8160b38834_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..9aa94b73d3efdcca39a7818a5775995954c02f46 --- /dev/null +++ b/data/hle/anthropic/claude-opus-4-5-20251101/26d05853-6958-48bd-91cd-8b8160b38834_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:32a35056f8982d77bc9d1a65aeaf60b9f8ddf110408bccc876fc6f673ca9e3cb +size 3879697 diff --git a/data/hle/anthropic/claude-opus-4-5-20251101/26d2a3da-ebfd-479b-b4a9-14012ef237ff.json b/data/hle/anthropic/claude-opus-4-5-20251101/26d2a3da-ebfd-479b-b4a9-14012ef237ff.json new file mode 100644 index 0000000000000000000000000000000000000000..dd1b92f1fce56e37fa7bfca4a748a6b134479f3e --- /dev/null +++ b/data/hle/anthropic/claude-opus-4-5-20251101/26d2a3da-ebfd-479b-b4a9-14012ef237ff.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/anthropic_claude-opus-4-5-20251101/1786166920.33676", + "evaluation_timestamp": "1776834390.0", + "retrieved_timestamp": "1786166920.33676", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-5-20251101", + "id": "anthropic/claude-opus-4-5-20251101", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/anthropic/claude-opus-4-5-20251101/26d2a3da-ebfd-479b-b4a9-14012ef237ff_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "5e9951d2acbd786be3aaca66e42ef5e445ca9268bd1ae8a3b61b23578f6ea108", + "total_rows": 115 + } +} \ No newline at end of file diff --git a/data/hle/anthropic/claude-opus-4-5-20251101/26d2a3da-ebfd-479b-b4a9-14012ef237ff_samples.jsonl b/data/hle/anthropic/claude-opus-4-5-20251101/26d2a3da-ebfd-479b-b4a9-14012ef237ff_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..47ae03262ef4a5ae925f25a75cb7bfa5210a97e2 --- /dev/null +++ b/data/hle/anthropic/claude-opus-4-5-20251101/26d2a3da-ebfd-479b-b4a9-14012ef237ff_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5e9951d2acbd786be3aaca66e42ef5e445ca9268bd1ae8a3b61b23578f6ea108 +size 7994780 diff --git a/data/hle/anthropic/claude-opus-4-5-20251101/3a838c40-64cb-40f6-9f24-d88127942ac8.json b/data/hle/anthropic/claude-opus-4-5-20251101/3a838c40-64cb-40f6-9f24-d88127942ac8.json new file mode 100644 index 0000000000000000000000000000000000000000..77f53cc6e5d4697fba7ff3aab8cf689d8380bad1 --- /dev/null +++ b/data/hle/anthropic/claude-opus-4-5-20251101/3a838c40-64cb-40f6-9f24-d88127942ac8.json @@ -0,0 +1,112 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/anthropic_claude-opus-4-5-20251101/1786164250.427356", + "evaluation_timestamp": "1776803719.0", + "retrieved_timestamp": "1786164250.427356", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-5-20251101", + "id": "anthropic/claude-opus-4-5-20251101", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "mean on hle/S-adaptive/+1ep/af55fb4e for scorer _llm_grader", + "source_data": { + "dataset_name": "hle", + "source_type": "hf_dataset", + "hf_repo": "cais/hle", + "samples_number": 4, + "sample_ids": [ + "hle_f85a0fa1", + "hle_c3ca0bda", + "hle_d190d8c2", + "hle_45d1c140" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1776803719.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.0, + "uncertainty": { + "num_samples": 4 + }, + "details": { + "total_matched_trajectories": "4", + "stopping_reason_count_token_limit": "3", + "stopping_reason_count_repetition_guard": "1", + "avg_total_tokens_target_model": "4068923.50", + "total_total_tokens_target_model": "16275694.0", + "avg_total_tokens_other_models": "65775.50", + "total_total_tokens_other_models": "263102.0", + "avg_total_tokens_all_models": "4134699.00", + "total_total_tokens_all_models": "16538796.0", + "avg_turn_count": "174.50", + "total_turn_count": "698.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 5000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/anthropic/claude-opus-4-5-20251101/3a838c40-64cb-40f6-9f24-d88127942ac8_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "edfa5d7eb5c65750ff5b599f9b8cd3d581bad8e2127819e288a386e806631b53", + "total_rows": 4 + } +} \ No newline at end of file diff --git a/data/hle/anthropic/claude-opus-4-5-20251101/3a838c40-64cb-40f6-9f24-d88127942ac8_samples.jsonl b/data/hle/anthropic/claude-opus-4-5-20251101/3a838c40-64cb-40f6-9f24-d88127942ac8_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..ba132c7d6e9f36565b99a58b221e4b57012cffaf --- /dev/null +++ b/data/hle/anthropic/claude-opus-4-5-20251101/3a838c40-64cb-40f6-9f24-d88127942ac8_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:edfa5d7eb5c65750ff5b599f9b8cd3d581bad8e2127819e288a386e806631b53 +size 1856970 diff --git a/data/hle/anthropic/claude-opus-4-5-20251101/3cd03b72-2204-464c-b091-65a9d1f1cb9b.json b/data/hle/anthropic/claude-opus-4-5-20251101/3cd03b72-2204-464c-b091-65a9d1f1cb9b.json new file mode 100644 index 0000000000000000000000000000000000000000..14c993ddbef500c44fdb795a56b47eafe332c84a --- /dev/null +++ b/data/hle/anthropic/claude-opus-4-5-20251101/3cd03b72-2204-464c-b091-65a9d1f1cb9b.json @@ -0,0 +1,108 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/anthropic_claude-opus-4-5-20251101/1786164259.825259", + "evaluation_timestamp": "1776519735.0", + "retrieved_timestamp": "1786164259.825259", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-5-20251101", + "id": "anthropic/claude-opus-4-5-20251101", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "mean on hle/S-adaptive/+3ep/591b3559 for scorer _llm_grader", + "source_data": { + "dataset_name": "hle", + "source_type": "hf_dataset", + "hf_repo": "cais/hle", + "samples_number": 4, + "sample_ids": [ + "hle_ca6c436b", + "hle_8a419425", + "hle_85e95ef4", + "hle_28643173" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1776519735.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.0, + "uncertainty": { + "num_samples": 12 + }, + "details": { + "total_matched_trajectories": "12", + "stopping_reason_count_repetition_guard": "12", + "avg_total_tokens_target_model": "208370.92", + "total_total_tokens_target_model": "2500451.0", + "avg_total_tokens_other_models": "7613.58", + "total_total_tokens_other_models": "91363.0", + "avg_total_tokens_all_models": "215984.50", + "total_total_tokens_all_models": "2591814.0", + "avg_turn_count": "11.00", + "total_turn_count": "132.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "max_connections": "10", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 5000000 + }, + "sandbox": {} + }, + "additional_details": { + "max_connections": "10" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/anthropic/claude-opus-4-5-20251101/3cd03b72-2204-464c-b091-65a9d1f1cb9b_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "57273a66cbe75fff5ac4c7a6e35bf16eaba0be9b4f523074b345cc25a3de8ad8", + "total_rows": 12 + } +} \ No newline at end of file diff --git a/data/hle/anthropic/claude-opus-4-5-20251101/3cd03b72-2204-464c-b091-65a9d1f1cb9b_samples.jsonl b/data/hle/anthropic/claude-opus-4-5-20251101/3cd03b72-2204-464c-b091-65a9d1f1cb9b_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..fe316d6cd624d9e678549795460e2c2cc69cfd03 --- /dev/null +++ b/data/hle/anthropic/claude-opus-4-5-20251101/3cd03b72-2204-464c-b091-65a9d1f1cb9b_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:57273a66cbe75fff5ac4c7a6e35bf16eaba0be9b4f523074b345cc25a3de8ad8 +size 319279 diff --git a/data/hle/anthropic/claude-opus-4-5-20251101/3d753e1b-3677-4546-a812-4e553a1d0d90.json b/data/hle/anthropic/claude-opus-4-5-20251101/3d753e1b-3677-4546-a812-4e553a1d0d90.json new file mode 100644 index 0000000000000000000000000000000000000000..a36ef84f1df2cd728812aa0062ccd9994db88ce7 --- /dev/null +++ b/data/hle/anthropic/claude-opus-4-5-20251101/3d753e1b-3677-4546-a812-4e553a1d0d90.json @@ -0,0 +1,118 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/anthropic_claude-opus-4-5-20251101/1786164263.380245", + "evaluation_timestamp": "1776430851.0", + "retrieved_timestamp": "1786164263.380245", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-5-20251101", + "id": "anthropic/claude-opus-4-5-20251101", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "mean on hle/S-adaptive/+4ep/3e749b09 for scorer _llm_grader", + "source_data": { + "dataset_name": "hle", + "source_type": "hf_dataset", + "hf_repo": "cais/hle", + "samples_number": 10, + "sample_ids": [ + "hle_d6b73839", + "hle_07662c9c", + "hle_764ce81c", + "hle_4e3440f7", + "hle_baccd6da", + "hle_f46c2857", + "hle_d1bd2767", + "hle_6499b500", + "hle_e1296346", + "hle_a559277e" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1776430851.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.25, + "uncertainty": { + "standard_error": { + "value": 0.13437096247164249 + }, + "num_samples": 40 + }, + "details": { + "total_matched_trajectories": "40", + "stopping_reason_count_repetition_guard": "39", + "stopping_reason_count_token_limit": "1", + "avg_total_tokens_target_model": "310395.55", + "total_total_tokens_target_model": "12415822.0", + "avg_total_tokens_other_models": "9381.35", + "total_total_tokens_other_models": "375254.0", + "avg_total_tokens_all_models": "319776.90", + "total_total_tokens_all_models": "12791076.0", + "avg_turn_count": "17.82", + "total_turn_count": "713.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "max_connections": "10", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 5000000 + }, + "sandbox": {} + }, + "additional_details": { + "max_connections": "10" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/anthropic/claude-opus-4-5-20251101/3d753e1b-3677-4546-a812-4e553a1d0d90_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "55d80d9abdaffccd6516b98ea6c36ba721a5fb38624d9a8e5b145d7a05b77d96", + "total_rows": 40 + } +} \ No newline at end of file diff --git a/data/hle/anthropic/claude-opus-4-5-20251101/3d753e1b-3677-4546-a812-4e553a1d0d90_samples.jsonl b/data/hle/anthropic/claude-opus-4-5-20251101/3d753e1b-3677-4546-a812-4e553a1d0d90_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..16febfe6d35cd67dd17e47abc7c5a3910c87af1b --- /dev/null +++ b/data/hle/anthropic/claude-opus-4-5-20251101/3d753e1b-3677-4546-a812-4e553a1d0d90_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:55d80d9abdaffccd6516b98ea6c36ba721a5fb38624d9a8e5b145d7a05b77d96 +size 1682150 diff --git a/data/hle/anthropic/claude-opus-4-5-20251101/48c19385-3a72-4a99-b36d-2e1a7a64ab22.json b/data/hle/anthropic/claude-opus-4-5-20251101/48c19385-3a72-4a99-b36d-2e1a7a64ab22.json new file mode 100644 index 0000000000000000000000000000000000000000..1c277c73c6e469f14e7ff07f3fd56ce131416ba1 --- /dev/null +++ b/data/hle/anthropic/claude-opus-4-5-20251101/48c19385-3a72-4a99-b36d-2e1a7a64ab22.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/anthropic_claude-opus-4-5-20251101/1786168891.160574", + "evaluation_timestamp": "1776677380.0", + "retrieved_timestamp": "1786168891.160574", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-5-20251101", + "id": "anthropic/claude-opus-4-5-20251101", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/anthropic/claude-opus-4-5-20251101/48c19385-3a72-4a99-b36d-2e1a7a64ab22_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "22231d54fcda637062afd4324dc7ed4317c80e0e7ca5506397d3507edf23b398", + "total_rows": 5 + } +} \ No newline at end of file diff --git a/data/hle/anthropic/claude-opus-4-5-20251101/48c19385-3a72-4a99-b36d-2e1a7a64ab22_samples.jsonl b/data/hle/anthropic/claude-opus-4-5-20251101/48c19385-3a72-4a99-b36d-2e1a7a64ab22_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..97279e434d3b939d8d30dad238351467d20bea2b --- /dev/null +++ b/data/hle/anthropic/claude-opus-4-5-20251101/48c19385-3a72-4a99-b36d-2e1a7a64ab22_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:22231d54fcda637062afd4324dc7ed4317c80e0e7ca5506397d3507edf23b398 +size 130960 diff --git a/data/hle/anthropic/claude-opus-4-5-20251101/4d820ade-3fbf-4612-9f4d-570e48adfa6f.json b/data/hle/anthropic/claude-opus-4-5-20251101/4d820ade-3fbf-4612-9f4d-570e48adfa6f.json new file mode 100644 index 0000000000000000000000000000000000000000..956ef88b536ee8fa2d3df745f318e323c371efce --- /dev/null +++ b/data/hle/anthropic/claude-opus-4-5-20251101/4d820ade-3fbf-4612-9f4d-570e48adfa6f.json @@ -0,0 +1,140 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/anthropic_claude-opus-4-5-20251101/1786164245.836301", + "evaluation_timestamp": "1776628389.0", + "retrieved_timestamp": "1786164245.836301", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-5-20251101", + "id": "anthropic/claude-opus-4-5-20251101", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "mean on hle/S-adaptive/+1ep/a176383e for scorer _llm_grader", + "source_data": { + "dataset_name": "hle", + "source_type": "hf_dataset", + "hf_repo": "cais/hle", + "samples_number": 30, + "sample_ids": [ + "hle_bc5fbce2", + "hle_25e0c71d", + "hle_76cdecee", + "hle_dfdb78f5", + "hle_142ffcff", + "hle_ffb12d83", + "hle_37c97863", + "hle_0702accc", + "hle_ebd62885", + "hle_4485802f", + "hle_9019c023", + "hle_9282e7ae", + "hle_44007b97", + "hle_7808792d", + "hle_e3b8f5c2", + "hle_3cf4483d", + "hle_376b8247", + "hle_31497469", + "hle_c43a27a3", + "hle_ea41e785", + "hle_d0c8b07e", + "hle_0291c889", + "hle_b1a7811d", + "hle_cf3d65bd", + "hle_0ff76ff3", + "hle_70e5aebc", + "hle_aa2fed90", + "hle_9df5d371", + "hle_5759e7d7", + "hle_ca59b32c" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1776628389.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.43333333333333335, + "uncertainty": { + "standard_error": { + "value": 0.0920186554465537 + }, + "num_samples": 30 + }, + "details": { + "total_matched_trajectories": "30", + "stopping_reason_count_repetition_guard": "29", + "stopping_reason_count_token_limit": "1", + "avg_total_tokens_target_model": "413450.27", + "total_total_tokens_target_model": "12403508.0", + "avg_total_tokens_other_models": "11265.60", + "total_total_tokens_other_models": "337968.0", + "avg_total_tokens_all_models": "424715.87", + "total_total_tokens_all_models": "12741476.0", + "avg_turn_count": "19.17", + "total_turn_count": "575.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "3600", + "attempt_timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 5000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "3600", + "max_connections": "10" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/anthropic/claude-opus-4-5-20251101/4d820ade-3fbf-4612-9f4d-570e48adfa6f_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "f7e84f7dbd28b161f1cc1d6d5be323fa53e1c59333600f1f96404a2f7d391f65", + "total_rows": 30 + } +} \ No newline at end of file diff --git a/data/hle/anthropic/claude-opus-4-5-20251101/4d820ade-3fbf-4612-9f4d-570e48adfa6f_samples.jsonl b/data/hle/anthropic/claude-opus-4-5-20251101/4d820ade-3fbf-4612-9f4d-570e48adfa6f_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..b05031dae00549378f02adc5b78104189fb5e194 --- /dev/null +++ b/data/hle/anthropic/claude-opus-4-5-20251101/4d820ade-3fbf-4612-9f4d-570e48adfa6f_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f7e84f7dbd28b161f1cc1d6d5be323fa53e1c59333600f1f96404a2f7d391f65 +size 1761791 diff --git a/data/hle/anthropic/claude-opus-4-5-20251101/6b0c6707-d569-4275-9769-44c3db843b92.json b/data/hle/anthropic/claude-opus-4-5-20251101/6b0c6707-d569-4275-9769-44c3db843b92.json new file mode 100644 index 0000000000000000000000000000000000000000..56739390588fd4f6e72bbdddfff4f3bab3dbde1c --- /dev/null +++ b/data/hle/anthropic/claude-opus-4-5-20251101/6b0c6707-d569-4275-9769-44c3db843b92.json @@ -0,0 +1,110 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/anthropic_claude-opus-4-5-20251101/1786164267.903438", + "evaluation_timestamp": "1776885821.0", + "retrieved_timestamp": "1786164267.903438", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-5-20251101", + "id": "anthropic/claude-opus-4-5-20251101", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "mean on hle/S-adaptive/b929445b for scorer _llm_grader", + "source_data": { + "dataset_name": "hle", + "source_type": "hf_dataset", + "hf_repo": "cais/hle", + "samples_number": 2, + "sample_ids": [ + "hle_5e235b48", + "hle_8b0384c4" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1776885821.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.0, + "uncertainty": { + "num_samples": 10 + }, + "details": { + "total_matched_trajectories": "10", + "stopping_reason_count_token_limit": "6", + "stopping_reason_count_repetition_guard": "4", + "avg_total_tokens_target_model": "4375170.30", + "total_total_tokens_target_model": "43751703.0", + "avg_total_tokens_other_models": "126073.50", + "total_total_tokens_other_models": "1260735.0", + "avg_total_tokens_all_models": "4501243.80", + "total_total_tokens_all_models": "45012438.0", + "avg_turn_count": "333.20", + "total_turn_count": "3332.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 5000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/anthropic/claude-opus-4-5-20251101/6b0c6707-d569-4275-9769-44c3db843b92_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "46f46989ddba157170354b5769695e3068f80e9abb848df0292c3873a90fd275", + "total_rows": 10 + } +} \ No newline at end of file diff --git a/data/hle/anthropic/claude-opus-4-5-20251101/6b0c6707-d569-4275-9769-44c3db843b92_samples.jsonl b/data/hle/anthropic/claude-opus-4-5-20251101/6b0c6707-d569-4275-9769-44c3db843b92_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..df303dc8d9fd00a07d47f91ed5faa89ede6618f3 --- /dev/null +++ b/data/hle/anthropic/claude-opus-4-5-20251101/6b0c6707-d569-4275-9769-44c3db843b92_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:46f46989ddba157170354b5769695e3068f80e9abb848df0292c3873a90fd275 +size 7609497 diff --git a/data/hle/anthropic/claude-opus-4-5-20251101/6caa1301-632a-4208-a084-12466c7887de.json b/data/hle/anthropic/claude-opus-4-5-20251101/6caa1301-632a-4208-a084-12466c7887de.json new file mode 100644 index 0000000000000000000000000000000000000000..d7e9e606e22bede314f171846008a7fef62c2200 --- /dev/null +++ b/data/hle/anthropic/claude-opus-4-5-20251101/6caa1301-632a-4208-a084-12466c7887de.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/anthropic_claude-opus-4-5-20251101/1786168888.514257", + "evaluation_timestamp": "1776628389.0", + "retrieved_timestamp": "1786168888.514257", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-5-20251101", + "id": "anthropic/claude-opus-4-5-20251101", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/anthropic/claude-opus-4-5-20251101/6caa1301-632a-4208-a084-12466c7887de_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "b3f7dea09b2def9c6e2166546209a48d138e4bb7abae399aa7518cb460e62d36", + "total_rows": 100 + } +} \ No newline at end of file diff --git a/data/hle/anthropic/claude-opus-4-5-20251101/6caa1301-632a-4208-a084-12466c7887de_samples.jsonl b/data/hle/anthropic/claude-opus-4-5-20251101/6caa1301-632a-4208-a084-12466c7887de_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..e03e444afe50e6a9c9ad19292cf15566abb38b8b --- /dev/null +++ b/data/hle/anthropic/claude-opus-4-5-20251101/6caa1301-632a-4208-a084-12466c7887de_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b3f7dea09b2def9c6e2166546209a48d138e4bb7abae399aa7518cb460e62d36 +size 6606948 diff --git a/data/hle/anthropic/claude-opus-4-5-20251101/791dea96-952b-43c0-9f84-d067f0eca4f4.json b/data/hle/anthropic/claude-opus-4-5-20251101/791dea96-952b-43c0-9f84-d067f0eca4f4.json new file mode 100644 index 0000000000000000000000000000000000000000..55962b7c2722d1e70a71623cee05c46293f8fa29 --- /dev/null +++ b/data/hle/anthropic/claude-opus-4-5-20251101/791dea96-952b-43c0-9f84-d067f0eca4f4.json @@ -0,0 +1,108 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/anthropic_claude-opus-4-5-20251101/1786164252.337791", + "evaluation_timestamp": "1776686674.0", + "retrieved_timestamp": "1786164252.337791", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-5-20251101", + "id": "anthropic/claude-opus-4-5-20251101", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "mean on hle/S-adaptive/+2ep/9cd54746 for scorer _llm_grader", + "source_data": { + "dataset_name": "hle", + "source_type": "hf_dataset", + "hf_repo": "cais/hle", + "samples_number": 1, + "sample_ids": [ + "hle_45d1c140" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1776686674.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.0, + "uncertainty": { + "num_samples": 2 + }, + "details": { + "total_matched_trajectories": "2", + "stopping_reason_count_repetition_guard": "1", + "stopping_reason_count_token_limit": "1", + "avg_total_tokens_target_model": "2519338.00", + "total_total_tokens_target_model": "5038676.0", + "avg_total_tokens_other_models": "25407.00", + "total_total_tokens_other_models": "50814.0", + "avg_total_tokens_all_models": "2544745.00", + "total_total_tokens_all_models": "5089490.0", + "avg_turn_count": "70.00", + "total_turn_count": "140.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "3600", + "attempt_timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 5000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "3600", + "max_connections": "10" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/anthropic/claude-opus-4-5-20251101/791dea96-952b-43c0-9f84-d067f0eca4f4_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "d289a8956b17d7f970c6cbc4d83a000978b880d7c6c4db6002c79190cb5fdffc", + "total_rows": 2 + } +} \ No newline at end of file diff --git a/data/hle/anthropic/claude-opus-4-5-20251101/791dea96-952b-43c0-9f84-d067f0eca4f4_samples.jsonl b/data/hle/anthropic/claude-opus-4-5-20251101/791dea96-952b-43c0-9f84-d067f0eca4f4_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..9cca45139df1010c98e0f9b3e8e5e0721874319a --- /dev/null +++ b/data/hle/anthropic/claude-opus-4-5-20251101/791dea96-952b-43c0-9f84-d067f0eca4f4_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d289a8956b17d7f970c6cbc4d83a000978b880d7c6c4db6002c79190cb5fdffc +size 578158 diff --git a/data/hle/anthropic/claude-opus-4-5-20251101/a5c685a3-e09b-4daf-9a5b-b749cb3a2b79.json b/data/hle/anthropic/claude-opus-4-5-20251101/a5c685a3-e09b-4daf-9a5b-b749cb3a2b79.json new file mode 100644 index 0000000000000000000000000000000000000000..fbc64435b1a42089513ffc0cc64582a3a8a86156 --- /dev/null +++ b/data/hle/anthropic/claude-opus-4-5-20251101/a5c685a3-e09b-4daf-9a5b-b749cb3a2b79.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/anthropic_claude-opus-4-5-20251101/1786168834.690721", + "evaluation_timestamp": "1776359245.0", + "retrieved_timestamp": "1786168834.690721", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-5-20251101", + "id": "anthropic/claude-opus-4-5-20251101", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/anthropic/claude-opus-4-5-20251101/a5c685a3-e09b-4daf-9a5b-b749cb3a2b79_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "c6425e50a6837a5319b6bfd12d63ca9d33f5d06d4ddf4ad59d8a3e550df79990", + "total_rows": 10 + } +} \ No newline at end of file diff --git a/data/hle/anthropic/claude-opus-4-5-20251101/a5c685a3-e09b-4daf-9a5b-b749cb3a2b79_samples.jsonl b/data/hle/anthropic/claude-opus-4-5-20251101/a5c685a3-e09b-4daf-9a5b-b749cb3a2b79_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..1933b79afdc328975d67d6057a04412c46749e7f --- /dev/null +++ b/data/hle/anthropic/claude-opus-4-5-20251101/a5c685a3-e09b-4daf-9a5b-b749cb3a2b79_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c6425e50a6837a5319b6bfd12d63ca9d33f5d06d4ddf4ad59d8a3e550df79990 +size 184729 diff --git a/data/hle/anthropic/claude-opus-4-5-20251101/b25cdb7b-41c2-4d6a-8484-bc860a8bfb21.json b/data/hle/anthropic/claude-opus-4-5-20251101/b25cdb7b-41c2-4d6a-8484-bc860a8bfb21.json new file mode 100644 index 0000000000000000000000000000000000000000..c736f0252e37c0aad4920f4ad7a1f016630b1752 --- /dev/null +++ b/data/hle/anthropic/claude-opus-4-5-20251101/b25cdb7b-41c2-4d6a-8484-bc860a8bfb21.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/anthropic_claude-opus-4-5-20251101/1786166899.405862", + "evaluation_timestamp": "1776430851.0", + "retrieved_timestamp": "1786166899.405862", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-5-20251101", + "id": "anthropic/claude-opus-4-5-20251101", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/anthropic/claude-opus-4-5-20251101/b25cdb7b-41c2-4d6a-8484-bc860a8bfb21_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "644de2da8819e296a7397acc95c11d6f54e2a3261133b800c1a8614cfe19ccc3", + "total_rows": 20 + } +} \ No newline at end of file diff --git a/data/hle/anthropic/claude-opus-4-5-20251101/b25cdb7b-41c2-4d6a-8484-bc860a8bfb21_samples.jsonl b/data/hle/anthropic/claude-opus-4-5-20251101/b25cdb7b-41c2-4d6a-8484-bc860a8bfb21_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..35ad34344deb3687118142bf8c5231b0cb1b5685 --- /dev/null +++ b/data/hle/anthropic/claude-opus-4-5-20251101/b25cdb7b-41c2-4d6a-8484-bc860a8bfb21_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:644de2da8819e296a7397acc95c11d6f54e2a3261133b800c1a8614cfe19ccc3 +size 1116059 diff --git a/data/hle/anthropic/claude-opus-4-5-20251101/b8f1952f-fb80-497b-a97c-f613415a631a.json b/data/hle/anthropic/claude-opus-4-5-20251101/b8f1952f-fb80-497b-a97c-f613415a631a.json new file mode 100644 index 0000000000000000000000000000000000000000..15424a0fc58d9dc36ddcd87a7070b4e2ef7711cf --- /dev/null +++ b/data/hle/anthropic/claude-opus-4-5-20251101/b8f1952f-fb80-497b-a97c-f613415a631a.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/anthropic_claude-opus-4-5-20251101/1786168848.023521", + "evaluation_timestamp": "1776677380.0", + "retrieved_timestamp": "1786168848.023521", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-5-20251101", + "id": "anthropic/claude-opus-4-5-20251101", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/anthropic/claude-opus-4-5-20251101/b8f1952f-fb80-497b-a97c-f613415a631a_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "b5819ea4b86702a6048e8f6112fe56571533d178076c1ea379eccbd86fc95fa1", + "total_rows": 29 + } +} \ No newline at end of file diff --git a/data/hle/anthropic/claude-opus-4-5-20251101/b8f1952f-fb80-497b-a97c-f613415a631a_samples.jsonl b/data/hle/anthropic/claude-opus-4-5-20251101/b8f1952f-fb80-497b-a97c-f613415a631a_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..5048dec5773fd84402449d789b306131bd2b8fe4 --- /dev/null +++ b/data/hle/anthropic/claude-opus-4-5-20251101/b8f1952f-fb80-497b-a97c-f613415a631a_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b5819ea4b86702a6048e8f6112fe56571533d178076c1ea379eccbd86fc95fa1 +size 780683 diff --git a/data/hle/anthropic/claude-opus-4-5-20251101/c63d68b9-9989-4af3-b748-bb5be1280c8f.json b/data/hle/anthropic/claude-opus-4-5-20251101/c63d68b9-9989-4af3-b748-bb5be1280c8f.json new file mode 100644 index 0000000000000000000000000000000000000000..da565284174763a52f48e88a63c4eb110a17eb75 --- /dev/null +++ b/data/hle/anthropic/claude-opus-4-5-20251101/c63d68b9-9989-4af3-b748-bb5be1280c8f.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/anthropic_claude-opus-4-5-20251101/1786166896.935212", + "evaluation_timestamp": "1776359245.0", + "retrieved_timestamp": "1786166896.935212", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-5-20251101", + "id": "anthropic/claude-opus-4-5-20251101", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/anthropic/claude-opus-4-5-20251101/c63d68b9-9989-4af3-b748-bb5be1280c8f_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "98ecd75fe2c83278da3cb1affe4c4c1d34ef956ada2f35301a7de4084a52fc79", + "total_rows": 10 + } +} \ No newline at end of file diff --git a/data/hle/anthropic/claude-opus-4-5-20251101/c63d68b9-9989-4af3-b748-bb5be1280c8f_samples.jsonl b/data/hle/anthropic/claude-opus-4-5-20251101/c63d68b9-9989-4af3-b748-bb5be1280c8f_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..91fcdb4b333ee65212dfded20a87ea0e70fb0442 --- /dev/null +++ b/data/hle/anthropic/claude-opus-4-5-20251101/c63d68b9-9989-4af3-b748-bb5be1280c8f_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:98ecd75fe2c83278da3cb1affe4c4c1d34ef956ada2f35301a7de4084a52fc79 +size 322770 diff --git a/data/hle/anthropic/claude-opus-4-5-20251101/d25b5e3c-7703-4491-a5b5-5701899618c3.json b/data/hle/anthropic/claude-opus-4-5-20251101/d25b5e3c-7703-4491-a5b5-5701899618c3.json new file mode 100644 index 0000000000000000000000000000000000000000..5deff924e6c0b53093592b4a0353cbbc6fde1f35 --- /dev/null +++ b/data/hle/anthropic/claude-opus-4-5-20251101/d25b5e3c-7703-4491-a5b5-5701899618c3.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/anthropic_claude-opus-4-5-20251101/1786168877.133616", + "evaluation_timestamp": "1776430851.0", + "retrieved_timestamp": "1786168877.133616", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-5-20251101", + "id": "anthropic/claude-opus-4-5-20251101", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/anthropic/claude-opus-4-5-20251101/d25b5e3c-7703-4491-a5b5-5701899618c3_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "c5203c2494b55f73d020d5a4c3a2bea8b72b975e454f7a673f2af901551b5685", + "total_rows": 90 + } +} \ No newline at end of file diff --git a/data/hle/anthropic/claude-opus-4-5-20251101/d25b5e3c-7703-4491-a5b5-5701899618c3_samples.jsonl b/data/hle/anthropic/claude-opus-4-5-20251101/d25b5e3c-7703-4491-a5b5-5701899618c3_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..b8b94aa4b557f5e590e2789bc645a307772b1984 --- /dev/null +++ b/data/hle/anthropic/claude-opus-4-5-20251101/d25b5e3c-7703-4491-a5b5-5701899618c3_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c5203c2494b55f73d020d5a4c3a2bea8b72b975e454f7a673f2af901551b5685 +size 4295946 diff --git a/data/hle/anthropic/claude-opus-4-5-20251101/d2cfb255-dee4-4691-9c7f-cea6b6fb1706.json b/data/hle/anthropic/claude-opus-4-5-20251101/d2cfb255-dee4-4691-9c7f-cea6b6fb1706.json new file mode 100644 index 0000000000000000000000000000000000000000..1cd67840147d0cabe0e46e70c9ba76a02070dd16 --- /dev/null +++ b/data/hle/anthropic/claude-opus-4-5-20251101/d2cfb255-dee4-4691-9c7f-cea6b6fb1706.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/anthropic_claude-opus-4-5-20251101/1786168840.292156", + "evaluation_timestamp": "1776599540.0", + "retrieved_timestamp": "1786168840.292156", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-5-20251101", + "id": "anthropic/claude-opus-4-5-20251101", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/anthropic/claude-opus-4-5-20251101/d2cfb255-dee4-4691-9c7f-cea6b6fb1706_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "614a2874f3b5e1195107f7c8211870120f333d6ca563b40f97585094b6c2929c", + "total_rows": 30 + } +} \ No newline at end of file diff --git a/data/hle/anthropic/claude-opus-4-5-20251101/d2cfb255-dee4-4691-9c7f-cea6b6fb1706_samples.jsonl b/data/hle/anthropic/claude-opus-4-5-20251101/d2cfb255-dee4-4691-9c7f-cea6b6fb1706_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..759faab7a703797d524b85e72317ced6d39cdbd0 --- /dev/null +++ b/data/hle/anthropic/claude-opus-4-5-20251101/d2cfb255-dee4-4691-9c7f-cea6b6fb1706_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:614a2874f3b5e1195107f7c8211870120f333d6ca563b40f97585094b6c2929c +size 751847 diff --git a/data/hle/anthropic/claude-opus-4-5-20251101/e14adcd9-44f4-41b5-929a-655649f1985f.json b/data/hle/anthropic/claude-opus-4-5-20251101/e14adcd9-44f4-41b5-929a-655649f1985f.json new file mode 100644 index 0000000000000000000000000000000000000000..029aea6637877795571a708644dc62d436d7009f --- /dev/null +++ b/data/hle/anthropic/claude-opus-4-5-20251101/e14adcd9-44f4-41b5-929a-655649f1985f.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/anthropic_claude-opus-4-5-20251101/1786166903.363663", + "evaluation_timestamp": "1776686674.0", + "retrieved_timestamp": "1786166903.363663", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-5-20251101", + "id": "anthropic/claude-opus-4-5-20251101", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/anthropic/claude-opus-4-5-20251101/e14adcd9-44f4-41b5-929a-655649f1985f_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "0dcf1e2d5bd1710a4220bdf98cfcee1c3332a65386734eec1070e1b9ab5ca4ae", + "total_rows": 40 + } +} \ No newline at end of file diff --git a/data/hle/anthropic/claude-opus-4-5-20251101/e14adcd9-44f4-41b5-929a-655649f1985f_samples.jsonl b/data/hle/anthropic/claude-opus-4-5-20251101/e14adcd9-44f4-41b5-929a-655649f1985f_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..096f940e84be7754cb7175cb88346d5d9420b885 --- /dev/null +++ b/data/hle/anthropic/claude-opus-4-5-20251101/e14adcd9-44f4-41b5-929a-655649f1985f_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0dcf1e2d5bd1710a4220bdf98cfcee1c3332a65386734eec1070e1b9ab5ca4ae +size 4332038 diff --git a/data/hle/anthropic/claude-opus-4-5-20251101/e236fef2-7572-4931-9e3d-de58ce41e196.json b/data/hle/anthropic/claude-opus-4-5-20251101/e236fef2-7572-4931-9e3d-de58ce41e196.json new file mode 100644 index 0000000000000000000000000000000000000000..45d60687d3e9e2748aff07dd3b0e4c3fc2129003 --- /dev/null +++ b/data/hle/anthropic/claude-opus-4-5-20251101/e236fef2-7572-4931-9e3d-de58ce41e196.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/anthropic_claude-opus-4-5-20251101/1786168863.097679", + "evaluation_timestamp": "1776519735.0", + "retrieved_timestamp": "1786168863.097679", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-5-20251101", + "id": "anthropic/claude-opus-4-5-20251101", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/anthropic/claude-opus-4-5-20251101/e236fef2-7572-4931-9e3d-de58ce41e196_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "237f3f293bf360b0616b16efea9e966febeff99cc9ed93b5266f933041470425", + "total_rows": 120 + } +} \ No newline at end of file diff --git a/data/hle/anthropic/claude-opus-4-5-20251101/e236fef2-7572-4931-9e3d-de58ce41e196_samples.jsonl b/data/hle/anthropic/claude-opus-4-5-20251101/e236fef2-7572-4931-9e3d-de58ce41e196_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..cfdcd8b467fde979bf9b6628877f62e81a430a16 --- /dev/null +++ b/data/hle/anthropic/claude-opus-4-5-20251101/e236fef2-7572-4931-9e3d-de58ce41e196_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:237f3f293bf360b0616b16efea9e966febeff99cc9ed93b5266f933041470425 +size 9477751 diff --git a/data/hle/anthropic/claude-opus-4-5-20251101/f39ac3e9-ad48-49cb-a56f-35b832d2d8b8.json b/data/hle/anthropic/claude-opus-4-5-20251101/f39ac3e9-ad48-49cb-a56f-35b832d2d8b8.json new file mode 100644 index 0000000000000000000000000000000000000000..0c01ab62a163daacb44901e376cd352b541b01fb --- /dev/null +++ b/data/hle/anthropic/claude-opus-4-5-20251101/f39ac3e9-ad48-49cb-a56f-35b832d2d8b8.json @@ -0,0 +1,108 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/anthropic_claude-opus-4-5-20251101/1786164254.113053", + "evaluation_timestamp": "1776823496.0", + "retrieved_timestamp": "1786164254.113053", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-5-20251101", + "id": "anthropic/claude-opus-4-5-20251101", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "mean on hle/S-adaptive/+2ep/05654ec7 for scorer _llm_grader", + "source_data": { + "dataset_name": "hle", + "source_type": "hf_dataset", + "hf_repo": "cais/hle", + "samples_number": 1, + "sample_ids": [ + "hle_670287ad" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1776823496.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.0, + "uncertainty": { + "num_samples": 2 + }, + "details": { + "total_matched_trajectories": "2", + "stopping_reason_count_repetition_guard": "2", + "avg_total_tokens_target_model": "97603.00", + "total_total_tokens_target_model": "195206.0", + "avg_total_tokens_other_models": "7522.50", + "total_total_tokens_other_models": "15045.0", + "avg_total_tokens_all_models": "105125.50", + "total_total_tokens_all_models": "210251.0", + "avg_turn_count": "20.00", + "total_turn_count": "40.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 5000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/anthropic/claude-opus-4-5-20251101/f39ac3e9-ad48-49cb-a56f-35b832d2d8b8_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "43b779de4274f9cdb48cedbef4379cf54f407ed3bee923c535a550359d70b572", + "total_rows": 2 + } +} \ No newline at end of file diff --git a/data/hle/anthropic/claude-opus-4-5-20251101/f39ac3e9-ad48-49cb-a56f-35b832d2d8b8_samples.jsonl b/data/hle/anthropic/claude-opus-4-5-20251101/f39ac3e9-ad48-49cb-a56f-35b832d2d8b8_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..c2f5e419df1599e6c57c8139d70b756091f29178 --- /dev/null +++ b/data/hle/anthropic/claude-opus-4-5-20251101/f39ac3e9-ad48-49cb-a56f-35b832d2d8b8_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:43b779de4274f9cdb48cedbef4379cf54f407ed3bee923c535a550359d70b572 +size 74258 diff --git a/data/hle/anthropic/claude-opus-4-6/1533fb55-ef2e-4568-a202-4683b1465cae.json b/data/hle/anthropic/claude-opus-4-6/1533fb55-ef2e-4568-a202-4683b1465cae.json new file mode 100644 index 0000000000000000000000000000000000000000..ef959d733b37354573e5ac0bd478ef47eb2e5035 --- /dev/null +++ b/data/hle/anthropic/claude-opus-4-6/1533fb55-ef2e-4568-a202-4683b1465cae.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/anthropic_claude-opus-4-6/1786168931.439943", + "evaluation_timestamp": "1776628389.0", + "retrieved_timestamp": "1786168931.439943", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-6", + "id": "anthropic/claude-opus-4-6", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/anthropic/claude-opus-4-6/1533fb55-ef2e-4568-a202-4683b1465cae_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "5bdae7d7986f031c7d1abdeb7b72423442160fcbd77958d869a35ff938b9602e", + "total_rows": 60 + } +} \ No newline at end of file diff --git a/data/hle/anthropic/claude-opus-4-6/1533fb55-ef2e-4568-a202-4683b1465cae_samples.jsonl b/data/hle/anthropic/claude-opus-4-6/1533fb55-ef2e-4568-a202-4683b1465cae_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..7a59ab1115f6e8233bb2e35497f86b3db485fc35 --- /dev/null +++ b/data/hle/anthropic/claude-opus-4-6/1533fb55-ef2e-4568-a202-4683b1465cae_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5bdae7d7986f031c7d1abdeb7b72423442160fcbd77958d869a35ff938b9602e +size 15800658 diff --git a/data/hle/anthropic/claude-opus-4-6/17670952-0f62-4a80-aaf7-eb7d37b7c227.json b/data/hle/anthropic/claude-opus-4-6/17670952-0f62-4a80-aaf7-eb7d37b7c227.json new file mode 100644 index 0000000000000000000000000000000000000000..128e4470708098391da066b6181429df54b3f2f4 --- /dev/null +++ b/data/hle/anthropic/claude-opus-4-6/17670952-0f62-4a80-aaf7-eb7d37b7c227.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/anthropic_claude-opus-4-6/1786168953.843776", + "evaluation_timestamp": "1776519735.0", + "retrieved_timestamp": "1786168953.843776", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-6", + "id": "anthropic/claude-opus-4-6", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/anthropic/claude-opus-4-6/17670952-0f62-4a80-aaf7-eb7d37b7c227_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "98e8feac6223e67410786f9c1d6af552924f333fe98ea614305f8a0e42dcc391", + "total_rows": 120 + } +} \ No newline at end of file diff --git a/data/hle/anthropic/claude-opus-4-6/17670952-0f62-4a80-aaf7-eb7d37b7c227_samples.jsonl b/data/hle/anthropic/claude-opus-4-6/17670952-0f62-4a80-aaf7-eb7d37b7c227_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..8cc66f9ef0fc11dd52a2598d052c38035049aac2 --- /dev/null +++ b/data/hle/anthropic/claude-opus-4-6/17670952-0f62-4a80-aaf7-eb7d37b7c227_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:98e8feac6223e67410786f9c1d6af552924f333fe98ea614305f8a0e42dcc391 +size 36720209 diff --git a/data/hle/anthropic/claude-opus-4-6/2dd2e399-9335-44d6-834d-5717d8581629.json b/data/hle/anthropic/claude-opus-4-6/2dd2e399-9335-44d6-834d-5717d8581629.json new file mode 100644 index 0000000000000000000000000000000000000000..30a072a7f03ce996232150f8f00782d58a7db3fe --- /dev/null +++ b/data/hle/anthropic/claude-opus-4-6/2dd2e399-9335-44d6-834d-5717d8581629.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/anthropic_claude-opus-4-6/1786168900.410823", + "evaluation_timestamp": "1776599540.0", + "retrieved_timestamp": "1786168900.410823", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-6", + "id": "anthropic/claude-opus-4-6", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/anthropic/claude-opus-4-6/2dd2e399-9335-44d6-834d-5717d8581629_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "80fe9b8af7295722738038ef7dfc12490745f1a82c415f949a41c8545bc76a02", + "total_rows": 40 + } +} \ No newline at end of file diff --git a/data/hle/anthropic/claude-opus-4-6/2dd2e399-9335-44d6-834d-5717d8581629_samples.jsonl b/data/hle/anthropic/claude-opus-4-6/2dd2e399-9335-44d6-834d-5717d8581629_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..9720b578ebe60a6dc30d09845819db35b0380609 --- /dev/null +++ b/data/hle/anthropic/claude-opus-4-6/2dd2e399-9335-44d6-834d-5717d8581629_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:80fe9b8af7295722738038ef7dfc12490745f1a82c415f949a41c8545bc76a02 +size 6234402 diff --git a/data/hle/anthropic/claude-opus-4-6/2ff53d4b-f449-4458-9ee3-0da16c73d717.json b/data/hle/anthropic/claude-opus-4-6/2ff53d4b-f449-4458-9ee3-0da16c73d717.json new file mode 100644 index 0000000000000000000000000000000000000000..8676d13404e5737daaa27478bb8eee714b2396db --- /dev/null +++ b/data/hle/anthropic/claude-opus-4-6/2ff53d4b-f449-4458-9ee3-0da16c73d717.json @@ -0,0 +1,117 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/anthropic_claude-opus-4-6/1786164309.63985", + "evaluation_timestamp": "1776430851.0", + "retrieved_timestamp": "1786164309.63985", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-6", + "id": "anthropic/claude-opus-4-6", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "mean on hle/S-adaptive/+4ep/08e2c4f0 for scorer _llm_grader", + "source_data": { + "dataset_name": "hle", + "source_type": "hf_dataset", + "hf_repo": "cais/hle", + "samples_number": 10, + "sample_ids": [ + "hle_d6b73839", + "hle_07662c9c", + "hle_764ce81c", + "hle_4e3440f7", + "hle_baccd6da", + "hle_f46c2857", + "hle_86a18059", + "hle_8770ab31", + "hle_e1296346", + "hle_28643173" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1776430851.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.4, + "uncertainty": { + "standard_error": { + "value": 0.1632993161855452 + }, + "num_samples": 40 + }, + "details": { + "total_matched_trajectories": "40", + "stopping_reason_count_repetition_guard": "40", + "avg_total_tokens_target_model": "339461.62", + "total_total_tokens_target_model": "13578465.0", + "avg_total_tokens_other_models": "5246.38", + "total_total_tokens_other_models": "209855.0", + "avg_total_tokens_all_models": "344708.00", + "total_total_tokens_all_models": "13788320.0", + "avg_turn_count": "7.05", + "total_turn_count": "282.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "max_connections": "10", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 5000000 + }, + "sandbox": {} + }, + "additional_details": { + "max_connections": "10" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/anthropic/claude-opus-4-6/2ff53d4b-f449-4458-9ee3-0da16c73d717_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "106f3bfa01442e73bc9dfb0b26c78250965b2463da4033c50e476d9ee215e52c", + "total_rows": 40 + } +} \ No newline at end of file diff --git a/data/hle/anthropic/claude-opus-4-6/2ff53d4b-f449-4458-9ee3-0da16c73d717_samples.jsonl b/data/hle/anthropic/claude-opus-4-6/2ff53d4b-f449-4458-9ee3-0da16c73d717_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..1f67346bdf963991de5645e041d8cdc934b58f52 --- /dev/null +++ b/data/hle/anthropic/claude-opus-4-6/2ff53d4b-f449-4458-9ee3-0da16c73d717_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:106f3bfa01442e73bc9dfb0b26c78250965b2463da4033c50e476d9ee215e52c +size 8278541 diff --git a/data/hle/anthropic/claude-opus-4-6/3370bd7f-193c-4aa7-a957-70b9ddb2c963.json b/data/hle/anthropic/claude-opus-4-6/3370bd7f-193c-4aa7-a957-70b9ddb2c963.json new file mode 100644 index 0000000000000000000000000000000000000000..7b3351ebdd1153b6b624317a3407e2a49e77cd6d --- /dev/null +++ b/data/hle/anthropic/claude-opus-4-6/3370bd7f-193c-4aa7-a957-70b9ddb2c963.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/anthropic_claude-opus-4-6/1786168987.253714", + "evaluation_timestamp": "1776430851.0", + "retrieved_timestamp": "1786168987.253714", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-6", + "id": "anthropic/claude-opus-4-6", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/anthropic/claude-opus-4-6/3370bd7f-193c-4aa7-a957-70b9ddb2c963_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "261e0b0ddb1a0f4ed1728056ad9aba099ac35d69419cf5678ff64563cb64b8c6", + "total_rows": 100 + } +} \ No newline at end of file diff --git a/data/hle/anthropic/claude-opus-4-6/3370bd7f-193c-4aa7-a957-70b9ddb2c963_samples.jsonl b/data/hle/anthropic/claude-opus-4-6/3370bd7f-193c-4aa7-a957-70b9ddb2c963_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..e15b7729598642074a52fe214f883d26f7f3a57e --- /dev/null +++ b/data/hle/anthropic/claude-opus-4-6/3370bd7f-193c-4aa7-a957-70b9ddb2c963_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:261e0b0ddb1a0f4ed1728056ad9aba099ac35d69419cf5678ff64563cb64b8c6 +size 22673691 diff --git a/data/hle/anthropic/claude-opus-4-6/42f0c40c-bd62-4cdb-b10f-3fdee7f76c84.json b/data/hle/anthropic/claude-opus-4-6/42f0c40c-bd62-4cdb-b10f-3fdee7f76c84.json new file mode 100644 index 0000000000000000000000000000000000000000..22cd2a2f293ab44aa9a04a732290beb84edd0a45 --- /dev/null +++ b/data/hle/anthropic/claude-opus-4-6/42f0c40c-bd62-4cdb-b10f-3fdee7f76c84.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/anthropic_claude-opus-4-6/1786166967.928078", + "evaluation_timestamp": "1776955657.0", + "retrieved_timestamp": "1786166967.928078", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-6", + "id": "anthropic/claude-opus-4-6", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/anthropic/claude-opus-4-6/42f0c40c-bd62-4cdb-b10f-3fdee7f76c84_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "b0d70df43495db012654ba1e2a38beacad845cbfefb452bf6f75706bfef996df", + "total_rows": 69 + } +} \ No newline at end of file diff --git a/data/hle/anthropic/claude-opus-4-6/42f0c40c-bd62-4cdb-b10f-3fdee7f76c84_samples.jsonl b/data/hle/anthropic/claude-opus-4-6/42f0c40c-bd62-4cdb-b10f-3fdee7f76c84_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..12f3d487b84da1e314630945bae45f87ae8b4657 --- /dev/null +++ b/data/hle/anthropic/claude-opus-4-6/42f0c40c-bd62-4cdb-b10f-3fdee7f76c84_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b0d70df43495db012654ba1e2a38beacad845cbfefb452bf6f75706bfef996df +size 45734953 diff --git a/data/hle/anthropic/claude-opus-4-6/48cc7d13-bb69-4d6e-a9f6-eb3d9c2aaeb2.json b/data/hle/anthropic/claude-opus-4-6/48cc7d13-bb69-4d6e-a9f6-eb3d9c2aaeb2.json new file mode 100644 index 0000000000000000000000000000000000000000..95d5000d4d1acbd876011414c65463e6409396eb --- /dev/null +++ b/data/hle/anthropic/claude-opus-4-6/48cc7d13-bb69-4d6e-a9f6-eb3d9c2aaeb2.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/anthropic_claude-opus-4-6/1786168902.481662", + "evaluation_timestamp": "1776677380.0", + "retrieved_timestamp": "1786168902.481662", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-6", + "id": "anthropic/claude-opus-4-6", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/anthropic/claude-opus-4-6/48cc7d13-bb69-4d6e-a9f6-eb3d9c2aaeb2_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "48ec561eeb43c1885daa4677211edc0a4fc9aec3673768e59fd9d4c1ffbf6f24", + "total_rows": 1 + } +} \ No newline at end of file diff --git a/data/hle/anthropic/claude-opus-4-6/48cc7d13-bb69-4d6e-a9f6-eb3d9c2aaeb2_samples.jsonl b/data/hle/anthropic/claude-opus-4-6/48cc7d13-bb69-4d6e-a9f6-eb3d9c2aaeb2_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..0b882f8e5c62063fd6151f2fd947a0b66b9a8541 --- /dev/null +++ b/data/hle/anthropic/claude-opus-4-6/48cc7d13-bb69-4d6e-a9f6-eb3d9c2aaeb2_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:48ec561eeb43c1885daa4677211edc0a4fc9aec3673768e59fd9d4c1ffbf6f24 +size 167165 diff --git a/data/hle/anthropic/claude-opus-4-6/4fc939dd-4014-42ec-b8af-68f001b4dc32.json b/data/hle/anthropic/claude-opus-4-6/4fc939dd-4014-42ec-b8af-68f001b4dc32.json new file mode 100644 index 0000000000000000000000000000000000000000..5aee8ed0e1062cd6c7c6fc2bd5e08ecf09c4a244 --- /dev/null +++ b/data/hle/anthropic/claude-opus-4-6/4fc939dd-4014-42ec-b8af-68f001b4dc32.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/anthropic_claude-opus-4-6/1786167012.317766", + "evaluation_timestamp": "1778489058.0", + "retrieved_timestamp": "1786167012.317766", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-6", + "id": "anthropic/claude-opus-4-6", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/anthropic/claude-opus-4-6/4fc939dd-4014-42ec-b8af-68f001b4dc32_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "43c36625570d470869d6f551ccdacb487d4d49c707868f563364660825063cc4", + "total_rows": 4 + } +} \ No newline at end of file diff --git a/data/hle/anthropic/claude-opus-4-6/4fc939dd-4014-42ec-b8af-68f001b4dc32_samples.jsonl b/data/hle/anthropic/claude-opus-4-6/4fc939dd-4014-42ec-b8af-68f001b4dc32_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..a49a5139f283fd0cb2638b51ec9ade8661d3b8a9 --- /dev/null +++ b/data/hle/anthropic/claude-opus-4-6/4fc939dd-4014-42ec-b8af-68f001b4dc32_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:43c36625570d470869d6f551ccdacb487d4d49c707868f563364660825063cc4 +size 22376435 diff --git a/data/hle/anthropic/claude-opus-4-6/6e289d11-1e8a-4d84-8cd0-96f0d5a99d80.json b/data/hle/anthropic/claude-opus-4-6/6e289d11-1e8a-4d84-8cd0-96f0d5a99d80.json new file mode 100644 index 0000000000000000000000000000000000000000..4336164f97480c753627617c0acc892e46beb6ec --- /dev/null +++ b/data/hle/anthropic/claude-opus-4-6/6e289d11-1e8a-4d84-8cd0-96f0d5a99d80.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/anthropic_claude-opus-4-6/1786166930.433662", + "evaluation_timestamp": "1776519735.0", + "retrieved_timestamp": "1786166930.433662", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-6", + "id": "anthropic/claude-opus-4-6", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/anthropic/claude-opus-4-6/6e289d11-1e8a-4d84-8cd0-96f0d5a99d80_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "9f1e679627a120331a6c35799e52840bf7e2a07e55e99f53698c8ed1eac658e6", + "total_rows": 40 + } +} \ No newline at end of file diff --git a/data/hle/anthropic/claude-opus-4-6/6e289d11-1e8a-4d84-8cd0-96f0d5a99d80_samples.jsonl b/data/hle/anthropic/claude-opus-4-6/6e289d11-1e8a-4d84-8cd0-96f0d5a99d80_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..e7bc50ecf223b9bf0ccb1c83162538e913bffb51 --- /dev/null +++ b/data/hle/anthropic/claude-opus-4-6/6e289d11-1e8a-4d84-8cd0-96f0d5a99d80_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9f1e679627a120331a6c35799e52840bf7e2a07e55e99f53698c8ed1eac658e6 +size 14024386 diff --git a/data/hle/anthropic/claude-opus-4-6/6f06f5c8-bf0c-4145-b205-334a9fa47045.json b/data/hle/anthropic/claude-opus-4-6/6f06f5c8-bf0c-4145-b205-334a9fa47045.json new file mode 100644 index 0000000000000000000000000000000000000000..e2e788ff4d80f32cd7e0533ab6d47d879f649fde --- /dev/null +++ b/data/hle/anthropic/claude-opus-4-6/6f06f5c8-bf0c-4145-b205-334a9fa47045.json @@ -0,0 +1,114 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/anthropic_claude-opus-4-6/1786164291.118421", + "evaluation_timestamp": "1776955657.0", + "retrieved_timestamp": "1786164291.118421", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-6", + "id": "anthropic/claude-opus-4-6", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "mean on hle/S-adaptive/+2ep/99f25d69 for scorer _llm_grader", + "source_data": { + "dataset_name": "hle", + "source_type": "hf_dataset", + "hf_repo": "cais/hle", + "samples_number": 2, + "sample_ids": [ + "hle_5e235b48", + "hle_45d1c140" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1776955657.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.25, + "uncertainty": { + "standard_error": { + "value": 0.25 + }, + "num_samples": 4 + }, + "details": { + "total_matched_trajectories": "4", + "stopping_reason_count_repetition_guard": "2", + "stopping_reason_count_token_limit": "1", + "stopping_reason_count_completed_with_submit": "1", + "avg_total_tokens_target_model": "3207849.00", + "total_total_tokens_target_model": "12831396.0", + "avg_total_tokens_other_models": "13198.75", + "total_total_tokens_other_models": "52795.0", + "avg_total_tokens_all_models": "3221047.75", + "total_total_tokens_all_models": "12884191.0", + "avg_turn_count": "59.50", + "total_turn_count": "238.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 5000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/anthropic/claude-opus-4-6/6f06f5c8-bf0c-4145-b205-334a9fa47045_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "68825164a997e2caccff7f50059108a623a69f1278668c815939e0709a0fe451", + "total_rows": 4 + } +} \ No newline at end of file diff --git a/data/hle/anthropic/claude-opus-4-6/6f06f5c8-bf0c-4145-b205-334a9fa47045_samples.jsonl b/data/hle/anthropic/claude-opus-4-6/6f06f5c8-bf0c-4145-b205-334a9fa47045_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..a60ceed5ad8af448ce0359c8c2e12d9d6a873751 --- /dev/null +++ b/data/hle/anthropic/claude-opus-4-6/6f06f5c8-bf0c-4145-b205-334a9fa47045_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68825164a997e2caccff7f50059108a623a69f1278668c815939e0709a0fe451 +size 10185225 diff --git a/data/hle/anthropic/claude-opus-4-6/6ffc0f32-946f-4db1-a6f9-ee630e30c9bf.json b/data/hle/anthropic/claude-opus-4-6/6ffc0f32-946f-4db1-a6f9-ee630e30c9bf.json new file mode 100644 index 0000000000000000000000000000000000000000..2a816e53a2b3b98f9213b3522773a5821aa9b2bd --- /dev/null +++ b/data/hle/anthropic/claude-opus-4-6/6ffc0f32-946f-4db1-a6f9-ee630e30c9bf.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/anthropic_claude-opus-4-6/1786166989.94826", + "evaluation_timestamp": "1776686673.0", + "retrieved_timestamp": "1786166989.94826", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-6", + "id": "anthropic/claude-opus-4-6", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/anthropic/claude-opus-4-6/6ffc0f32-946f-4db1-a6f9-ee630e30c9bf_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "48bd9c98e6d3f9fe567ef9b037bd4e5356105ea978a8a1889b788b601d255fef", + "total_rows": 80 + } +} \ No newline at end of file diff --git a/data/hle/anthropic/claude-opus-4-6/6ffc0f32-946f-4db1-a6f9-ee630e30c9bf_samples.jsonl b/data/hle/anthropic/claude-opus-4-6/6ffc0f32-946f-4db1-a6f9-ee630e30c9bf_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..904843bdd86bbf75e92ac8a5f7688f572ef26e34 --- /dev/null +++ b/data/hle/anthropic/claude-opus-4-6/6ffc0f32-946f-4db1-a6f9-ee630e30c9bf_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:48bd9c98e6d3f9fe567ef9b037bd4e5356105ea978a8a1889b788b601d255fef +size 31221511 diff --git a/data/hle/anthropic/claude-opus-4-6/76ef17ce-c739-4d32-bbac-07c30875c626.json b/data/hle/anthropic/claude-opus-4-6/76ef17ce-c739-4d32-bbac-07c30875c626.json new file mode 100644 index 0000000000000000000000000000000000000000..4c5e6389792881370b6baf69b1dc152cc6fc24ff --- /dev/null +++ b/data/hle/anthropic/claude-opus-4-6/76ef17ce-c739-4d32-bbac-07c30875c626.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/anthropic_claude-opus-4-6/1786168909.515972", + "evaluation_timestamp": "1776599540.0", + "retrieved_timestamp": "1786168909.515972", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-6", + "id": "anthropic/claude-opus-4-6", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/anthropic/claude-opus-4-6/76ef17ce-c739-4d32-bbac-07c30875c626_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "261de9fb5f5e1fc1093ed5e60b8acf4550ba0234fcad3c71edc2ff500b0bbaa8", + "total_rows": 30 + } +} \ No newline at end of file diff --git a/data/hle/anthropic/claude-opus-4-6/76ef17ce-c739-4d32-bbac-07c30875c626_samples.jsonl b/data/hle/anthropic/claude-opus-4-6/76ef17ce-c739-4d32-bbac-07c30875c626_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..7aa84608ef28f36d5b808c5015b3d669f80ab92e --- /dev/null +++ b/data/hle/anthropic/claude-opus-4-6/76ef17ce-c739-4d32-bbac-07c30875c626_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:261de9fb5f5e1fc1093ed5e60b8acf4550ba0234fcad3c71edc2ff500b0bbaa8 +size 5194156 diff --git a/data/hle/anthropic/claude-opus-4-6/8147b6e7-0e23-4368-99ff-8eeaef8dd8c2.json b/data/hle/anthropic/claude-opus-4-6/8147b6e7-0e23-4368-99ff-8eeaef8dd8c2.json new file mode 100644 index 0000000000000000000000000000000000000000..2a20a8f036feea120ab9a6d317ae9dbbb42be11e --- /dev/null +++ b/data/hle/anthropic/claude-opus-4-6/8147b6e7-0e23-4368-99ff-8eeaef8dd8c2.json @@ -0,0 +1,119 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/anthropic_claude-opus-4-6/1786164297.157886", + "evaluation_timestamp": "1776519735.0", + "retrieved_timestamp": "1786164297.157886", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-6", + "id": "anthropic/claude-opus-4-6", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "mean on hle/S-adaptive/+3ep/753f04cf for scorer _llm_grader", + "source_data": { + "dataset_name": "hle", + "source_type": "hf_dataset", + "hf_repo": "cais/hle", + "samples_number": 12, + "sample_ids": [ + "hle_d1bd2767", + "hle_ca6c436b", + "hle_a559277e", + "hle_8a419425", + "hle_85e95ef4", + "hle_25e0c71d", + "hle_76cdecee", + "hle_142ffcff", + "hle_ffb12d83", + "hle_37c97863", + "hle_0702accc", + "hle_16388cc8" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1776519735.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.38888888888888884, + "uncertainty": { + "standard_error": { + "value": 0.14114331931800123 + }, + "num_samples": 36 + }, + "details": { + "total_matched_trajectories": "36", + "stopping_reason_count_repetition_guard": "36", + "avg_total_tokens_target_model": "456865.97", + "total_total_tokens_target_model": "16447175.0", + "avg_total_tokens_other_models": "6961.69", + "total_total_tokens_other_models": "250621.0", + "avg_total_tokens_all_models": "463827.67", + "total_total_tokens_all_models": "16697796.0", + "avg_turn_count": "10.67", + "total_turn_count": "384.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "max_connections": "10", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 5000000 + }, + "sandbox": {} + }, + "additional_details": { + "max_connections": "10" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/anthropic/claude-opus-4-6/8147b6e7-0e23-4368-99ff-8eeaef8dd8c2_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "5eb0a325e150e4c472bafb4f51ba989b1cc8f761bff4b100224be1c8d2c2d055", + "total_rows": 36 + } +} \ No newline at end of file diff --git a/data/hle/anthropic/claude-opus-4-6/8147b6e7-0e23-4368-99ff-8eeaef8dd8c2_samples.jsonl b/data/hle/anthropic/claude-opus-4-6/8147b6e7-0e23-4368-99ff-8eeaef8dd8c2_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..e59063013a01641ffcc830539b4ea108e66e1c0e --- /dev/null +++ b/data/hle/anthropic/claude-opus-4-6/8147b6e7-0e23-4368-99ff-8eeaef8dd8c2_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5eb0a325e150e4c472bafb4f51ba989b1cc8f761bff4b100224be1c8d2c2d055 +size 7151576 diff --git a/data/hle/anthropic/claude-opus-4-6/874eaf63-a176-4f44-b7dd-3a41d12f44a8.json b/data/hle/anthropic/claude-opus-4-6/874eaf63-a176-4f44-b7dd-3a41d12f44a8.json new file mode 100644 index 0000000000000000000000000000000000000000..85f5b55e3c8c21b3b74b9bb40b29fdcb38970bd8 --- /dev/null +++ b/data/hle/anthropic/claude-opus-4-6/874eaf63-a176-4f44-b7dd-3a41d12f44a8.json @@ -0,0 +1,121 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/anthropic_claude-opus-4-6/1786162852.860742", + "evaluation_timestamp": "1777089067.0", + "retrieved_timestamp": "1786162852.860742", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-6", + "id": "anthropic/claude-opus-4-6", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "mean on hle/S-adaptive+C/+4ep/1a90d6fc for scorer _llm_grader", + "source_data": { + "dataset_name": "hle", + "source_type": "hf_dataset", + "hf_repo": "cais/hle", + "samples_number": 9, + "sample_ids": [ + "hle_5759e7d7", + "hle_9255683e", + "hle_b5c681d9", + "hle_b33012cf", + "hle_141cd3ef", + "hle_6d963e58", + "hle_45d1c140", + "hle_c0ed25f7", + "hle_cd4aa5b3" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1777089067.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.19444444444444445, + "uncertainty": { + "standard_error": { + "value": 0.13028932666176193 + }, + "num_samples": 35 + }, + "details": { + "total_matched_trajectories": "35", + "stopping_reason_count_repetition_guard": "15", + "stopping_reason_count_token_limit": "13", + "stopping_reason_count_completed_on_successful_submit": "7", + "avg_total_tokens_target_model": "2581525.66", + "total_total_tokens_target_model": "90353398.0", + "avg_total_tokens_other_models": "10105.86", + "total_total_tokens_other_models": "353705.0", + "avg_total_tokens_all_models": "2591631.51", + "total_total_tokens_all_models": "90707103.0", + "avg_turn_count": "40.00", + "total_turn_count": "1400.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 5000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/anthropic/claude-opus-4-6/874eaf63-a176-4f44-b7dd-3a41d12f44a8_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "dd07f3fb564b45d9beb81adca9c4d30b202318ba8d2ed6e1e37bd468ee3cb9cb", + "total_rows": 35 + } +} \ No newline at end of file diff --git a/data/hle/anthropic/claude-opus-4-6/874eaf63-a176-4f44-b7dd-3a41d12f44a8_samples.jsonl b/data/hle/anthropic/claude-opus-4-6/874eaf63-a176-4f44-b7dd-3a41d12f44a8_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..8b1e57890e5b1181bdcbedf86434729fa748af90 --- /dev/null +++ b/data/hle/anthropic/claude-opus-4-6/874eaf63-a176-4f44-b7dd-3a41d12f44a8_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dd07f3fb564b45d9beb81adca9c4d30b202318ba8d2ed6e1e37bd468ee3cb9cb +size 77759717 diff --git a/data/hle/anthropic/claude-opus-4-6/9f1abcf9-dafc-4ffd-a5b7-82c61cbb4abb.json b/data/hle/anthropic/claude-opus-4-6/9f1abcf9-dafc-4ffd-a5b7-82c61cbb4abb.json new file mode 100644 index 0000000000000000000000000000000000000000..157b7ab9835cfd163c7e6e0ef1c8f266180049d2 --- /dev/null +++ b/data/hle/anthropic/claude-opus-4-6/9f1abcf9-dafc-4ffd-a5b7-82c61cbb4abb.json @@ -0,0 +1,113 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/anthropic_claude-opus-4-6/1786164315.87958", + "evaluation_timestamp": "1776955657.0", + "retrieved_timestamp": "1786164315.87958", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-6", + "id": "anthropic/claude-opus-4-6", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "mean on hle/S-adaptive/6877b0e3 for scorer _llm_grader", + "source_data": { + "dataset_name": "hle", + "source_type": "hf_dataset", + "hf_repo": "cais/hle", + "samples_number": 2, + "sample_ids": [ + "hle_ca59b32c", + "hle_0ec247c5" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1776955657.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.4, + "uncertainty": { + "standard_error": { + "value": 0.4 + }, + "num_samples": 10 + }, + "details": { + "total_matched_trajectories": "10", + "stopping_reason_count_repetition_guard": "8", + "stopping_reason_count_token_limit": "2", + "avg_total_tokens_target_model": "1754654.60", + "total_total_tokens_target_model": "17546546.0", + "avg_total_tokens_other_models": "4668.00", + "total_total_tokens_other_models": "46680.0", + "avg_total_tokens_all_models": "1759322.60", + "total_total_tokens_all_models": "17593226.0", + "avg_turn_count": "19.00", + "total_turn_count": "190.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 5000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/anthropic/claude-opus-4-6/9f1abcf9-dafc-4ffd-a5b7-82c61cbb4abb_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "debc70327c30f6d7ec31ff593348a4c33bcd5b06a75edc06b1d88dbcac936073", + "total_rows": 10 + } +} \ No newline at end of file diff --git a/data/hle/anthropic/claude-opus-4-6/9f1abcf9-dafc-4ffd-a5b7-82c61cbb4abb_samples.jsonl b/data/hle/anthropic/claude-opus-4-6/9f1abcf9-dafc-4ffd-a5b7-82c61cbb4abb_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..f596c21632646cd1fb4d55436529fc912242426a --- /dev/null +++ b/data/hle/anthropic/claude-opus-4-6/9f1abcf9-dafc-4ffd-a5b7-82c61cbb4abb_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:debc70327c30f6d7ec31ff593348a4c33bcd5b06a75edc06b1d88dbcac936073 +size 18693253 diff --git a/data/hle/anthropic/claude-opus-4-6/a30e2144-f7e5-4a91-8d53-d6626be34a70.json b/data/hle/anthropic/claude-opus-4-6/a30e2144-f7e5-4a91-8d53-d6626be34a70.json new file mode 100644 index 0000000000000000000000000000000000000000..d2a395b5f9eda4cad930d453b375178ad3584423 --- /dev/null +++ b/data/hle/anthropic/claude-opus-4-6/a30e2144-f7e5-4a91-8d53-d6626be34a70.json @@ -0,0 +1,115 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/anthropic_claude-opus-4-6/1786164274.250151", + "evaluation_timestamp": "1776803719.0", + "retrieved_timestamp": "1786164274.250151", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-6", + "id": "anthropic/claude-opus-4-6", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "mean on hle/S-adaptive/+1ep/f307bc5e for scorer _llm_grader", + "source_data": { + "dataset_name": "hle", + "source_type": "hf_dataset", + "hf_repo": "cais/hle", + "samples_number": 4, + "sample_ids": [ + "hle_bc5fbce2", + "hle_dfdb78f5", + "hle_209b6b64", + "hle_8b0384c4" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1776803719.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.25, + "uncertainty": { + "standard_error": { + "value": 0.25 + }, + "num_samples": 4 + }, + "details": { + "total_matched_trajectories": "4", + "stopping_reason_count_token_limit": "3", + "stopping_reason_count_repetition_guard": "1", + "avg_total_tokens_target_model": "4974974.25", + "total_total_tokens_target_model": "19899897.0", + "avg_total_tokens_other_models": "14524.50", + "total_total_tokens_other_models": "58098.0", + "avg_total_tokens_all_models": "4989498.75", + "total_total_tokens_all_models": "19957995.0", + "avg_turn_count": "74.50", + "total_turn_count": "298.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 5000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/anthropic/claude-opus-4-6/a30e2144-f7e5-4a91-8d53-d6626be34a70_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "da0fb3de20c1e14a377e6456aed9ea041824b405103a2e1098c245ab20e066c5", + "total_rows": 4 + } +} \ No newline at end of file diff --git a/data/hle/anthropic/claude-opus-4-6/a30e2144-f7e5-4a91-8d53-d6626be34a70_samples.jsonl b/data/hle/anthropic/claude-opus-4-6/a30e2144-f7e5-4a91-8d53-d6626be34a70_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..de88f3b487702ab07f1e4bb2b2b52eafeb0bfc29 --- /dev/null +++ b/data/hle/anthropic/claude-opus-4-6/a30e2144-f7e5-4a91-8d53-d6626be34a70_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:da0fb3de20c1e14a377e6456aed9ea041824b405103a2e1098c245ab20e066c5 +size 15380408 diff --git a/data/hle/anthropic/claude-opus-4-6/a812fd93-8465-4ba2-9d67-fbd181f7da81.json b/data/hle/anthropic/claude-opus-4-6/a812fd93-8465-4ba2-9d67-fbd181f7da81.json new file mode 100644 index 0000000000000000000000000000000000000000..f527e38151aafa2f5574825d2d4315e8ce4defa6 --- /dev/null +++ b/data/hle/anthropic/claude-opus-4-6/a812fd93-8465-4ba2-9d67-fbd181f7da81.json @@ -0,0 +1,151 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/anthropic_claude-opus-4-6/1786164280.93765", + "evaluation_timestamp": "1776628389.0", + "retrieved_timestamp": "1786164280.93765", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-6", + "id": "anthropic/claude-opus-4-6", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "mean on hle/S-adaptive/+1ep/8e2df467 for scorer _llm_grader", + "source_data": { + "dataset_name": "hle", + "source_type": "hf_dataset", + "hf_repo": "cais/hle", + "samples_number": 41, + "sample_ids": [ + "hle_6499b500", + "hle_4418bea4", + "hle_ebd62885", + "hle_4485802f", + "hle_9019c023", + "hle_9282e7ae", + "hle_44007b97", + "hle_7808792d", + "hle_e3b8f5c2", + "hle_3cf4483d", + "hle_f85a0fa1", + "hle_c3ca0bda", + "hle_376b8247", + "hle_31497469", + "hle_c43a27a3", + "hle_ea41e785", + "hle_d0c8b07e", + "hle_0291c889", + "hle_b1a7811d", + "hle_dab7fbb1", + "hle_cf3d65bd", + "hle_0ff76ff3", + "hle_70e5aebc", + "hle_aa2fed90", + "hle_9df5d371", + "hle_5759e7d7", + "hle_33bc94e2", + "hle_e3817b7e", + "hle_d7b7c3dc", + "hle_1dc141bb", + "hle_a764505b", + "hle_e62ba103", + "hle_34c69126", + "hle_5f04b314", + "hle_d5a55900", + "hle_3e11b38a", + "hle_5d49423c", + "hle_82f5c665", + "hle_cc7f25c0", + "hle_226e134b", + "hle_c1158fb8" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1776628389.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.43902439024390244, + "uncertainty": { + "standard_error": { + "value": 0.07846686801046543 + }, + "num_samples": 41 + }, + "details": { + "total_matched_trajectories": "41", + "stopping_reason_count_repetition_guard": "40", + "stopping_reason_count_token_limit": "1", + "avg_total_tokens_target_model": "490788.37", + "total_total_tokens_target_model": "20122323.0", + "avg_total_tokens_other_models": "6254.93", + "total_total_tokens_other_models": "256452.0", + "avg_total_tokens_all_models": "497043.29", + "total_total_tokens_all_models": "20378775.0", + "avg_turn_count": "11.80", + "total_turn_count": "484.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "3600", + "attempt_timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 5000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "3600", + "max_connections": "10" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/anthropic/claude-opus-4-6/a812fd93-8465-4ba2-9d67-fbd181f7da81_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "1feca8871c1b02b5947024246ce08002ce5b33870183d86da74a417f3263c254", + "total_rows": 41 + } +} \ No newline at end of file diff --git a/data/hle/anthropic/claude-opus-4-6/a812fd93-8465-4ba2-9d67-fbd181f7da81_samples.jsonl b/data/hle/anthropic/claude-opus-4-6/a812fd93-8465-4ba2-9d67-fbd181f7da81_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..2cda4bf9711f6de15cf27bc27748545203c1ec96 --- /dev/null +++ b/data/hle/anthropic/claude-opus-4-6/a812fd93-8465-4ba2-9d67-fbd181f7da81_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1feca8871c1b02b5947024246ce08002ce5b33870183d86da74a417f3263c254 +size 11700110 diff --git a/data/hle/anthropic/claude-opus-4-6/aaa95591-0edd-4415-9683-8286ec538e4e.json b/data/hle/anthropic/claude-opus-4-6/aaa95591-0edd-4415-9683-8286ec538e4e.json new file mode 100644 index 0000000000000000000000000000000000000000..d827c7aedd7df092547eebcb95ccce665e7876e9 --- /dev/null +++ b/data/hle/anthropic/claude-opus-4-6/aaa95591-0edd-4415-9683-8286ec538e4e.json @@ -0,0 +1,118 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/anthropic_claude-opus-4-6/1786162838.262278", + "evaluation_timestamp": "1776430851.0", + "retrieved_timestamp": "1786162838.262278", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-6", + "id": "anthropic/claude-opus-4-6", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "mean on hle/S-adaptive+C/+4ep/c9b99b37 for scorer _llm_grader", + "source_data": { + "dataset_name": "hle", + "source_type": "hf_dataset", + "hf_repo": "cais/hle", + "samples_number": 10, + "sample_ids": [ + "hle_d6b73839", + "hle_764ce81c", + "hle_4e3440f7", + "hle_baccd6da", + "hle_f46c2857", + "hle_6499b500", + "hle_a559277e", + "hle_8a419425", + "hle_28643173", + "hle_25e0c71d" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1776430851.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.8, + "uncertainty": { + "standard_error": { + "value": 0.13333333333333333 + }, + "num_samples": 40 + }, + "details": { + "total_matched_trajectories": "40", + "stopping_reason_count_completed_on_successful_submit": "32", + "stopping_reason_count_repetition_guard": "8", + "avg_total_tokens_target_model": "231455.55", + "total_total_tokens_target_model": "9258222.0", + "avg_total_tokens_other_models": "3319.15", + "total_total_tokens_other_models": "132766.0", + "avg_total_tokens_all_models": "234774.70", + "total_total_tokens_all_models": "9390988.0", + "avg_turn_count": "3.85", + "total_turn_count": "154.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "max_connections": "10", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 5000000 + }, + "sandbox": {} + }, + "additional_details": { + "max_connections": "10" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/anthropic/claude-opus-4-6/aaa95591-0edd-4415-9683-8286ec538e4e_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "972f362fa4e0bf83d2af4a8910c064bf971860fe0157432357cfc61e2674b6fa", + "total_rows": 40 + } +} \ No newline at end of file diff --git a/data/hle/anthropic/claude-opus-4-6/aaa95591-0edd-4415-9683-8286ec538e4e_samples.jsonl b/data/hle/anthropic/claude-opus-4-6/aaa95591-0edd-4415-9683-8286ec538e4e_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..2db7dacb423da1d015c5ffd1dbd38e6c5b9e0d93 --- /dev/null +++ b/data/hle/anthropic/claude-opus-4-6/aaa95591-0edd-4415-9683-8286ec538e4e_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:972f362fa4e0bf83d2af4a8910c064bf971860fe0157432357cfc61e2674b6fa +size 8168012 diff --git a/data/hle/anthropic/claude-opus-4-6/b2ca31d2-d749-4969-acbb-f9e931d5d3a7.json b/data/hle/anthropic/claude-opus-4-6/b2ca31d2-d749-4969-acbb-f9e931d5d3a7.json new file mode 100644 index 0000000000000000000000000000000000000000..0f59d8b1215b6319f7db7d26f579545d791f734f --- /dev/null +++ b/data/hle/anthropic/claude-opus-4-6/b2ca31d2-d749-4969-acbb-f9e931d5d3a7.json @@ -0,0 +1,144 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/anthropic_claude-opus-4-6/1786162832.974201", + "evaluation_timestamp": "1777036568.0", + "retrieved_timestamp": "1786162832.974201", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-6", + "id": "anthropic/claude-opus-4-6", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "mean on hle/S-adaptive+C/+3ep/5b3bbe5a for scorer _llm_grader", + "source_data": { + "dataset_name": "hle", + "source_type": "hf_dataset", + "hf_repo": "cais/hle", + "samples_number": 30, + "sample_ids": [ + "hle_c3ca0bda", + "hle_0291c889", + "hle_b6cec502", + "hle_3e11b38a", + "hle_0a1eec3f", + "hle_5d49423c", + "hle_1d580261", + "hle_82f5c665", + "hle_cc7f25c0", + "hle_226e134b", + "hle_c1158fb8", + "hle_d0ea1019", + "hle_fcd42d0c", + "hle_3697d2bc", + "hle_5e349785", + "hle_8b0384c4", + "hle_ee938b26", + "hle_d190d8c2", + "hle_b26bf8aa", + "hle_c2ff3a43", + "hle_5671e386", + "hle_a634e89e", + "hle_670287ad", + "hle_298b5616", + "hle_e14d8c23", + "hle_510b6040", + "hle_c6adc1b4", + "hle_aa343d51", + "hle_d52d3dae", + "hle_35a40c52" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1777036568.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.4666666666666667, + "uncertainty": { + "standard_error": { + "value": 0.0869520794023251 + }, + "num_samples": 89 + }, + "details": { + "total_matched_trajectories": "89", + "stopping_reason_count_repetition_guard": "37", + "stopping_reason_count_completed_on_successful_submit": "42", + "stopping_reason_count_completed_with_submit": "1", + "stopping_reason_count_completed_no_submit": "3", + "stopping_reason_count_token_limit": "6", + "avg_total_tokens_target_model": "681381.99", + "total_total_tokens_target_model": "60642997.0", + "avg_total_tokens_other_models": "6723.76", + "total_total_tokens_other_models": "598415.0", + "avg_total_tokens_all_models": "688105.75", + "total_total_tokens_all_models": "61241412.0", + "avg_turn_count": "17.48", + "total_turn_count": "1556.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 5000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/anthropic/claude-opus-4-6/b2ca31d2-d749-4969-acbb-f9e931d5d3a7_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "2ccf4bc97f52059bb21f3ebfdb848300aa14682b7b2098678ee37e4a1dbc1fa1", + "total_rows": 89 + } +} \ No newline at end of file diff --git a/data/hle/anthropic/claude-opus-4-6/b2ca31d2-d749-4969-acbb-f9e931d5d3a7_samples.jsonl b/data/hle/anthropic/claude-opus-4-6/b2ca31d2-d749-4969-acbb-f9e931d5d3a7_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..ba3ff27f78bf8e1f734878e0113c9d96a6913aab --- /dev/null +++ b/data/hle/anthropic/claude-opus-4-6/b2ca31d2-d749-4969-acbb-f9e931d5d3a7_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2ccf4bc97f52059bb21f3ebfdb848300aa14682b7b2098678ee37e4a1dbc1fa1 +size 41338551 diff --git a/data/hle/anthropic/claude-opus-4-6/d663dacb-13d8-4433-98cb-674f807a3682.json b/data/hle/anthropic/claude-opus-4-6/d663dacb-13d8-4433-98cb-674f807a3682.json new file mode 100644 index 0000000000000000000000000000000000000000..b31cb405468bcecea676e6dee7b874b472b616be --- /dev/null +++ b/data/hle/anthropic/claude-opus-4-6/d663dacb-13d8-4433-98cb-674f807a3682.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/anthropic_claude-opus-4-6/1786166924.16392", + "evaluation_timestamp": "1776359245.0", + "retrieved_timestamp": "1786166924.16392", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-6", + "id": "anthropic/claude-opus-4-6", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/anthropic/claude-opus-4-6/d663dacb-13d8-4433-98cb-674f807a3682_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "7e3acb1b09a1cd30605428661e64b213c8f5c022e57e9f43944dfc1145f1b8b0", + "total_rows": 10 + } +} \ No newline at end of file diff --git a/data/hle/anthropic/claude-opus-4-6/d663dacb-13d8-4433-98cb-674f807a3682_samples.jsonl b/data/hle/anthropic/claude-opus-4-6/d663dacb-13d8-4433-98cb-674f807a3682_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..5f70253424a5f8fd2a7de064f15bb393ac2a8369 --- /dev/null +++ b/data/hle/anthropic/claude-opus-4-6/d663dacb-13d8-4433-98cb-674f807a3682_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7e3acb1b09a1cd30605428661e64b213c8f5c022e57e9f43944dfc1145f1b8b0 +size 2099158 diff --git a/data/hle/anthropic/claude-opus-4-6/e0a9f9a9-59c3-4125-978a-26e908fb8357.json b/data/hle/anthropic/claude-opus-4-6/e0a9f9a9-59c3-4125-978a-26e908fb8357.json new file mode 100644 index 0000000000000000000000000000000000000000..6e910632c580d1ee3820e9127da2d1f14dcc83ca --- /dev/null +++ b/data/hle/anthropic/claude-opus-4-6/e0a9f9a9-59c3-4125-978a-26e908fb8357.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/anthropic_claude-opus-4-6/1786168894.06337", + "evaluation_timestamp": "1776359245.0", + "retrieved_timestamp": "1786168894.06337", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-6", + "id": "anthropic/claude-opus-4-6", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/anthropic/claude-opus-4-6/e0a9f9a9-59c3-4125-978a-26e908fb8357_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "1eb8ccba36c2488696abbe229617cfaf4546920a5c9914820708870585199626", + "total_rows": 10 + } +} \ No newline at end of file diff --git a/data/hle/anthropic/claude-opus-4-6/e0a9f9a9-59c3-4125-978a-26e908fb8357_samples.jsonl b/data/hle/anthropic/claude-opus-4-6/e0a9f9a9-59c3-4125-978a-26e908fb8357_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..229a485796fd54536d0506842167c3662ee5e791 --- /dev/null +++ b/data/hle/anthropic/claude-opus-4-6/e0a9f9a9-59c3-4125-978a-26e908fb8357_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1eb8ccba36c2488696abbe229617cfaf4546920a5c9914820708870585199626 +size 1001216 diff --git a/data/hle/anthropic/claude-opus-4-6/f72d7cdd-28fc-4208-a175-c02ff04a526b.json b/data/hle/anthropic/claude-opus-4-6/f72d7cdd-28fc-4208-a175-c02ff04a526b.json new file mode 100644 index 0000000000000000000000000000000000000000..84b17c8e66af8b13a434ea72e7b537a11f25cff0 --- /dev/null +++ b/data/hle/anthropic/claude-opus-4-6/f72d7cdd-28fc-4208-a175-c02ff04a526b.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/anthropic_claude-opus-4-6/1786167001.542226", + "evaluation_timestamp": "1776686674.0", + "retrieved_timestamp": "1786167001.542226", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-6", + "id": "anthropic/claude-opus-4-6", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/anthropic/claude-opus-4-6/f72d7cdd-28fc-4208-a175-c02ff04a526b_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "7adc3db6e80c7354e06b60545b771cce7babbfccb990341c567bb5f680417e51", + "total_rows": 30 + } +} \ No newline at end of file diff --git a/data/hle/anthropic/claude-opus-4-6/f72d7cdd-28fc-4208-a175-c02ff04a526b_samples.jsonl b/data/hle/anthropic/claude-opus-4-6/f72d7cdd-28fc-4208-a175-c02ff04a526b_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..e53ba7847e30aebcf09a11c78fc38307911e29f1 --- /dev/null +++ b/data/hle/anthropic/claude-opus-4-6/f72d7cdd-28fc-4208-a175-c02ff04a526b_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7adc3db6e80c7354e06b60545b771cce7babbfccb990341c567bb5f680417e51 +size 28457723 diff --git a/data/hle/anthropic/claude-opus-4-6/fb0f992a-6f5a-434a-acc1-190164a86003.json b/data/hle/anthropic/claude-opus-4-6/fb0f992a-6f5a-434a-acc1-190164a86003.json new file mode 100644 index 0000000000000000000000000000000000000000..74d512be18578ad38bca72dc71fafaade41db67d --- /dev/null +++ b/data/hle/anthropic/claude-opus-4-6/fb0f992a-6f5a-434a-acc1-190164a86003.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/anthropic_claude-opus-4-6/1786168992.55218", + "evaluation_timestamp": "1776677380.0", + "retrieved_timestamp": "1786168992.55218", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-6", + "id": "anthropic/claude-opus-4-6", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/anthropic/claude-opus-4-6/fb0f992a-6f5a-434a-acc1-190164a86003_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "59b777e2d1d5c5d6e575c3348456c76dcdc47b11fc5a0c6a334d562d3c79749c", + "total_rows": 4 + } +} \ No newline at end of file diff --git a/data/hle/anthropic/claude-opus-4-6/fb0f992a-6f5a-434a-acc1-190164a86003_samples.jsonl b/data/hle/anthropic/claude-opus-4-6/fb0f992a-6f5a-434a-acc1-190164a86003_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..9fc61cd4c99f4adf8f937bcc9a770d1bffc75e29 --- /dev/null +++ b/data/hle/anthropic/claude-opus-4-6/fb0f992a-6f5a-434a-acc1-190164a86003_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:59b777e2d1d5c5d6e575c3348456c76dcdc47b11fc5a0c6a334d562d3c79749c +size 1030190 diff --git a/data/hle/anthropic/claude-opus-4-6/fd92a145-6df3-42c1-92fb-bbe0db55d08f.json b/data/hle/anthropic/claude-opus-4-6/fd92a145-6df3-42c1-92fb-bbe0db55d08f.json new file mode 100644 index 0000000000000000000000000000000000000000..bcb9b879ddcb7cd2ed86d117d23923dcb2e59ec1 --- /dev/null +++ b/data/hle/anthropic/claude-opus-4-6/fd92a145-6df3-42c1-92fb-bbe0db55d08f.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/anthropic_claude-opus-4-6/1786166942.295465", + "evaluation_timestamp": "1776430851.0", + "retrieved_timestamp": "1786166942.295465", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-6", + "id": "anthropic/claude-opus-4-6", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/anthropic/claude-opus-4-6/fd92a145-6df3-42c1-92fb-bbe0db55d08f_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "7e7c627594e66d10e1b4f1c9f9094b9639b5da94f054662df3067afe07b076aa", + "total_rows": 50 + } +} \ No newline at end of file diff --git a/data/hle/anthropic/claude-opus-4-6/fd92a145-6df3-42c1-92fb-bbe0db55d08f_samples.jsonl b/data/hle/anthropic/claude-opus-4-6/fd92a145-6df3-42c1-92fb-bbe0db55d08f_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..876c8f91014fce2f098f40bc46f1b97d3711df96 --- /dev/null +++ b/data/hle/anthropic/claude-opus-4-6/fd92a145-6df3-42c1-92fb-bbe0db55d08f_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7e7c627594e66d10e1b4f1c9f9094b9639b5da94f054662df3067afe07b076aa +size 29297049 diff --git a/data/hle/openai/gpt-5-2025-08-07/23a40fda-c48d-433a-a195-e4d4ea5a87db.json b/data/hle/openai/gpt-5-2025-08-07/23a40fda-c48d-433a-a195-e4d4ea5a87db.json new file mode 100644 index 0000000000000000000000000000000000000000..28b0ab89a7beb45ac9b8066a51eb7508dca3578f --- /dev/null +++ b/data/hle/openai/gpt-5-2025-08-07/23a40fda-c48d-433a-a195-e4d4ea5a87db.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5-2025-08-07/1786166841.218639", + "evaluation_timestamp": "1777455567.0", + "retrieved_timestamp": "1786166841.218639", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5-2025-08-07", + "id": "openai/gpt-5-2025-08-07", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5-2025-08-07/23a40fda-c48d-433a-a195-e4d4ea5a87db_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "37073b84f4c111d9735478168ab3ea3583e0120c2225694b9f02736b7bc34d68", + "total_rows": 1 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5-2025-08-07/23a40fda-c48d-433a-a195-e4d4ea5a87db_samples.jsonl b/data/hle/openai/gpt-5-2025-08-07/23a40fda-c48d-433a-a195-e4d4ea5a87db_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..670bca6e0cce8b9ca392c3a71dfdb0fe7377c829 --- /dev/null +++ b/data/hle/openai/gpt-5-2025-08-07/23a40fda-c48d-433a-a195-e4d4ea5a87db_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:37073b84f4c111d9735478168ab3ea3583e0120c2225694b9f02736b7bc34d68 +size 122540 diff --git a/data/hle/openai/gpt-5-2025-08-07/24cb61b6-21d3-423c-9250-8517aa3e0eae.json b/data/hle/openai/gpt-5-2025-08-07/24cb61b6-21d3-423c-9250-8517aa3e0eae.json new file mode 100644 index 0000000000000000000000000000000000000000..a56e889a38e3bad9751cd95b1305d4e715bc4295 --- /dev/null +++ b/data/hle/openai/gpt-5-2025-08-07/24cb61b6-21d3-423c-9250-8517aa3e0eae.json @@ -0,0 +1,114 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5-2025-08-07/1786164238.287393", + "evaluation_timestamp": "1783093685.0", + "retrieved_timestamp": "1786164238.287393", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5-2025-08-07", + "id": "openai/gpt-5-2025-08-07", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "mean on hle/S-adaptive/+3ep/029ecef8 for scorer _llm_grader", + "source_data": { + "dataset_name": "hle", + "source_type": "hf_dataset", + "hf_repo": "cais/hle", + "samples_number": 3, + "sample_ids": [ + "hle_ca6c436b", + "hle_bc5fbce2", + "hle_34c69126" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1783093685.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.0, + "uncertainty": { + "num_samples": 7 + }, + "details": { + "total_matched_trajectories": "7", + "stopping_reason_count_repetition_guard": "7", + "avg_total_tokens_target_model": "250921.43", + "total_total_tokens_target_model": "1756450.0", + "avg_total_tokens_other_models": "8239.14", + "total_total_tokens_other_models": "57674.0", + "avg_total_tokens_all_models": "259160.57", + "total_total_tokens_all_models": "1814124.0", + "avg_turn_count": "5.57", + "total_turn_count": "39.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "1", + "timeout": "1200", + "attempt_timeout": "3600", + "max_connections": "16", + "max_tokens": "64000", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 5000000 + }, + "sandbox": {}, + "max_attempts": 1 + }, + "additional_details": { + "max_retries": "1", + "timeout": "1200", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5-2025-08-07/24cb61b6-21d3-423c-9250-8517aa3e0eae_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "f3abb2cc3bc49d9bd3987501ca4bdec75d5bcb84705cc5a58ab8ddcaa8ebb834", + "total_rows": 7 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5-2025-08-07/24cb61b6-21d3-423c-9250-8517aa3e0eae_samples.jsonl b/data/hle/openai/gpt-5-2025-08-07/24cb61b6-21d3-423c-9250-8517aa3e0eae_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..88622e32cc79d5d90520f88eea4166d0d612d0db --- /dev/null +++ b/data/hle/openai/gpt-5-2025-08-07/24cb61b6-21d3-423c-9250-8517aa3e0eae_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f3abb2cc3bc49d9bd3987501ca4bdec75d5bcb84705cc5a58ab8ddcaa8ebb834 +size 183716 diff --git a/data/hle/openai/gpt-5-2025-08-07/3406e9b7-01fe-4ca6-88c9-12c22ca6770c.json b/data/hle/openai/gpt-5-2025-08-07/3406e9b7-01fe-4ca6-88c9-12c22ca6770c.json new file mode 100644 index 0000000000000000000000000000000000000000..7c80a8499057c818178d85876b9c379ac00f1d72 --- /dev/null +++ b/data/hle/openai/gpt-5-2025-08-07/3406e9b7-01fe-4ca6-88c9-12c22ca6770c.json @@ -0,0 +1,116 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5-2025-08-07/1786164236.479996", + "evaluation_timestamp": "1782419110.0", + "retrieved_timestamp": "1786164236.479996", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5-2025-08-07", + "id": "openai/gpt-5-2025-08-07", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "mean on hle/S-adaptive/+2ep/df585680 for scorer _llm_grader", + "source_data": { + "dataset_name": "hle", + "source_type": "hf_dataset", + "hf_repo": "cais/hle", + "samples_number": 2, + "sample_ids": [ + "hle_28643173", + "hle_ea41e785" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1782419110.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.25, + "uncertainty": { + "standard_error": { + "value": 0.25 + }, + "num_samples": 3 + }, + "details": { + "total_matched_trajectories": "3", + "stopping_reason_count_repetition_guard": "3", + "avg_total_tokens_target_model": "790135.00", + "total_total_tokens_target_model": "2370405.0", + "avg_total_tokens_other_models": "4259.00", + "total_total_tokens_other_models": "12777.0", + "avg_total_tokens_all_models": "794394.00", + "total_total_tokens_all_models": "2383182.0", + "avg_turn_count": "36.00", + "total_turn_count": "108.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "1200", + "max_connections": "16", + "max_tokens": "20000", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 5000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5-2025-08-07/3406e9b7-01fe-4ca6-88c9-12c22ca6770c_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "35d8ef0d252e657c0020cfe89964b5250d4c0236b966743fc0707a91d605ccce", + "total_rows": 3 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5-2025-08-07/3406e9b7-01fe-4ca6-88c9-12c22ca6770c_samples.jsonl b/data/hle/openai/gpt-5-2025-08-07/3406e9b7-01fe-4ca6-88c9-12c22ca6770c_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..2612288d6a842f9fcd3240f4638672a4beeae27b --- /dev/null +++ b/data/hle/openai/gpt-5-2025-08-07/3406e9b7-01fe-4ca6-88c9-12c22ca6770c_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:35d8ef0d252e657c0020cfe89964b5250d4c0236b966743fc0707a91d605ccce +size 181211 diff --git a/data/hle/openai/gpt-5-2025-08-07/348bd7e6-81ae-450c-95ed-c73bff6e01dd.json b/data/hle/openai/gpt-5-2025-08-07/348bd7e6-81ae-450c-95ed-c73bff6e01dd.json new file mode 100644 index 0000000000000000000000000000000000000000..d680f77ed69c2830bacc4dfe88859c16136a4861 --- /dev/null +++ b/data/hle/openai/gpt-5-2025-08-07/348bd7e6-81ae-450c-95ed-c73bff6e01dd.json @@ -0,0 +1,118 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5-2025-08-07/1786164242.174258", + "evaluation_timestamp": "1783101762.0", + "retrieved_timestamp": "1786164242.174258", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5-2025-08-07", + "id": "openai/gpt-5-2025-08-07", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "mean on hle/S-adaptive/33a69589 for scorer _llm_grader", + "source_data": { + "dataset_name": "hle", + "source_type": "hf_dataset", + "hf_repo": "cais/hle", + "samples_number": 4, + "sample_ids": [ + "hle_764ce81c", + "hle_e3b8f5c2", + "hle_fcd56401", + "hle_5d49423c" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1783101762.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.6375, + "uncertainty": { + "standard_error": { + "value": 0.21925536861538114 + }, + "num_samples": 17 + }, + "details": { + "total_matched_trajectories": "17", + "stopping_reason_count_repetition_guard": "17", + "avg_total_tokens_target_model": "282031.82", + "total_total_tokens_target_model": "4794541.0", + "avg_total_tokens_other_models": "8429.18", + "total_total_tokens_other_models": "143296.0", + "avg_total_tokens_all_models": "290461.00", + "total_total_tokens_all_models": "4937837.0", + "avg_turn_count": "5.59", + "total_turn_count": "95.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "1", + "timeout": "1200", + "attempt_timeout": "3600", + "max_connections": "16", + "max_tokens": "64000", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 5000000 + }, + "sandbox": {}, + "max_attempts": 1 + }, + "additional_details": { + "max_retries": "1", + "timeout": "1200", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5-2025-08-07/348bd7e6-81ae-450c-95ed-c73bff6e01dd_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "03cae54a4497fc2162639443a21a4eb38b13751ab30f27d0d4947c963ce2640e", + "total_rows": 17 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5-2025-08-07/348bd7e6-81ae-450c-95ed-c73bff6e01dd_samples.jsonl b/data/hle/openai/gpt-5-2025-08-07/348bd7e6-81ae-450c-95ed-c73bff6e01dd_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..0462344dcb89fe4ef9e05029356024da9cfa615a --- /dev/null +++ b/data/hle/openai/gpt-5-2025-08-07/348bd7e6-81ae-450c-95ed-c73bff6e01dd_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:03cae54a4497fc2162639443a21a4eb38b13751ab30f27d0d4947c963ce2640e +size 409754 diff --git a/data/hle/openai/gpt-5-2025-08-07/3df22ad1-c4df-45f4-9d19-5902ed842f2d.json b/data/hle/openai/gpt-5-2025-08-07/3df22ad1-c4df-45f4-9d19-5902ed842f2d.json new file mode 100644 index 0000000000000000000000000000000000000000..8a65c60b99133fcd6ffebbf832ee0f864089aaf9 --- /dev/null +++ b/data/hle/openai/gpt-5-2025-08-07/3df22ad1-c4df-45f4-9d19-5902ed842f2d.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5-2025-08-07/1786166888.904171", + "evaluation_timestamp": "1779017113.0", + "retrieved_timestamp": "1786166888.904171", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5-2025-08-07", + "id": "openai/gpt-5-2025-08-07", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5-2025-08-07/3df22ad1-c4df-45f4-9d19-5902ed842f2d_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "4dbd63e15202b2759c7eb41807a4f8c28326a0029ce5d2369f4c450f9fa15340", + "total_rows": 2 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5-2025-08-07/3df22ad1-c4df-45f4-9d19-5902ed842f2d_samples.jsonl b/data/hle/openai/gpt-5-2025-08-07/3df22ad1-c4df-45f4-9d19-5902ed842f2d_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..89aa9e986863d9a9db693b66aeb84d7448a1d605 --- /dev/null +++ b/data/hle/openai/gpt-5-2025-08-07/3df22ad1-c4df-45f4-9d19-5902ed842f2d_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4dbd63e15202b2759c7eb41807a4f8c28326a0029ce5d2369f4c450f9fa15340 +size 105867 diff --git a/data/hle/openai/gpt-5-2025-08-07/43e4c28f-b646-4311-bb52-5520d0f1e0c3.json b/data/hle/openai/gpt-5-2025-08-07/43e4c28f-b646-4311-bb52-5520d0f1e0c3.json new file mode 100644 index 0000000000000000000000000000000000000000..0ecb869a0d07649e3f38c52f951b650b88ef5859 --- /dev/null +++ b/data/hle/openai/gpt-5-2025-08-07/43e4c28f-b646-4311-bb52-5520d0f1e0c3.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5-2025-08-07/1786166836.49723", + "evaluation_timestamp": "1776430851.0", + "retrieved_timestamp": "1786166836.49723", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5-2025-08-07", + "id": "openai/gpt-5-2025-08-07", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5-2025-08-07/43e4c28f-b646-4311-bb52-5520d0f1e0c3_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "ea651922b7ea2f70ccd75cc13ba9ecf90e06696d14a27eded7e2628903e61987", + "total_rows": 20 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5-2025-08-07/43e4c28f-b646-4311-bb52-5520d0f1e0c3_samples.jsonl b/data/hle/openai/gpt-5-2025-08-07/43e4c28f-b646-4311-bb52-5520d0f1e0c3_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..e4a8b3b9cccb1c08ea7c2880eb511f3c701f5d20 --- /dev/null +++ b/data/hle/openai/gpt-5-2025-08-07/43e4c28f-b646-4311-bb52-5520d0f1e0c3_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ea651922b7ea2f70ccd75cc13ba9ecf90e06696d14a27eded7e2628903e61987 +size 1051695 diff --git a/data/hle/openai/gpt-5-2025-08-07/58a68801-a713-4c69-89f1-825a5f764b2b.json b/data/hle/openai/gpt-5-2025-08-07/58a68801-a713-4c69-89f1-825a5f764b2b.json new file mode 100644 index 0000000000000000000000000000000000000000..4b3b8d564f30434c6dc42f5718a2009f0dcbbb1f --- /dev/null +++ b/data/hle/openai/gpt-5-2025-08-07/58a68801-a713-4c69-89f1-825a5f764b2b.json @@ -0,0 +1,112 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5-2025-08-07/1786164226.277286", + "evaluation_timestamp": "1783092534.0", + "retrieved_timestamp": "1786164226.277286", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5-2025-08-07", + "id": "openai/gpt-5-2025-08-07", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "mean on hle/S-adaptive/+1ep/57058abc for scorer _llm_grader", + "source_data": { + "dataset_name": "hle", + "source_type": "hf_dataset", + "hf_repo": "cais/hle", + "samples_number": 1, + "sample_ids": [ + "hle_c2ff3a43" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1783092534.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.0, + "uncertainty": { + "num_samples": 1 + }, + "details": { + "total_matched_trajectories": "1", + "stopping_reason_count_repetition_guard": "1", + "avg_total_tokens_target_model": "158638.00", + "total_total_tokens_target_model": "158638.0", + "avg_total_tokens_other_models": "5977.00", + "total_total_tokens_other_models": "5977.0", + "avg_total_tokens_all_models": "164615.00", + "total_total_tokens_all_models": "164615.0", + "avg_turn_count": "5.00", + "total_turn_count": "5.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "1", + "timeout": "1200", + "attempt_timeout": "3600", + "max_connections": "16", + "max_tokens": "64000", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 5000000 + }, + "sandbox": {}, + "max_attempts": 1 + }, + "additional_details": { + "max_retries": "1", + "timeout": "1200", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5-2025-08-07/58a68801-a713-4c69-89f1-825a5f764b2b_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "f05f10bb52c9b8c686a483f4cd138ea3f43dc5baa10cf4882f0bd01435cde57c", + "total_rows": 1 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5-2025-08-07/58a68801-a713-4c69-89f1-825a5f764b2b_samples.jsonl b/data/hle/openai/gpt-5-2025-08-07/58a68801-a713-4c69-89f1-825a5f764b2b_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..6740e9504f0e3b9ec358e234bdf1378c20447fea --- /dev/null +++ b/data/hle/openai/gpt-5-2025-08-07/58a68801-a713-4c69-89f1-825a5f764b2b_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f05f10bb52c9b8c686a483f4cd138ea3f43dc5baa10cf4882f0bd01435cde57c +size 21780 diff --git a/data/hle/openai/gpt-5-2025-08-07/58da7a7c-a520-4dcc-8543-1f8d586c1e2b.json b/data/hle/openai/gpt-5-2025-08-07/58da7a7c-a520-4dcc-8543-1f8d586c1e2b.json new file mode 100644 index 0000000000000000000000000000000000000000..b2eb5d7846b336487678d7ebf6c926adc931510e --- /dev/null +++ b/data/hle/openai/gpt-5-2025-08-07/58da7a7c-a520-4dcc-8543-1f8d586c1e2b.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5-2025-08-07/1786166843.962181", + "evaluation_timestamp": "1776686674.0", + "retrieved_timestamp": "1786166843.962181", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5-2025-08-07", + "id": "openai/gpt-5-2025-08-07", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5-2025-08-07/58da7a7c-a520-4dcc-8543-1f8d586c1e2b_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "0315cd6a0a6a201811803627fd5a5030219c18e04ed8d8936c88a0c503b10548", + "total_rows": 32 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5-2025-08-07/58da7a7c-a520-4dcc-8543-1f8d586c1e2b_samples.jsonl b/data/hle/openai/gpt-5-2025-08-07/58da7a7c-a520-4dcc-8543-1f8d586c1e2b_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..1708b8e59f61ce7720eb4d2a1432c16efdeeb0bb --- /dev/null +++ b/data/hle/openai/gpt-5-2025-08-07/58da7a7c-a520-4dcc-8543-1f8d586c1e2b_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0315cd6a0a6a201811803627fd5a5030219c18e04ed8d8936c88a0c503b10548 +size 1200156 diff --git a/data/hle/openai/gpt-5-2025-08-07/5d9a8dd5-cf71-4bfc-b3e4-2008a351a5ec.json b/data/hle/openai/gpt-5-2025-08-07/5d9a8dd5-cf71-4bfc-b3e4-2008a351a5ec.json new file mode 100644 index 0000000000000000000000000000000000000000..f6184dcf08e9cc1d3f7974ac41d2a47f31395c93 --- /dev/null +++ b/data/hle/openai/gpt-5-2025-08-07/5d9a8dd5-cf71-4bfc-b3e4-2008a351a5ec.json @@ -0,0 +1,112 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5-2025-08-07/1786164227.788162", + "evaluation_timestamp": "1783327709.0", + "retrieved_timestamp": "1786164227.788162", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5-2025-08-07", + "id": "openai/gpt-5-2025-08-07", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "mean on hle/S-adaptive/+1ep/745ed6e5 for scorer _llm_grader", + "source_data": { + "dataset_name": "hle", + "source_type": "hf_dataset", + "hf_repo": "cais/hle", + "samples_number": 1, + "sample_ids": [ + "hle_764ce81c" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1783327709.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 1.0, + "uncertainty": { + "num_samples": 1 + }, + "details": { + "total_matched_trajectories": "1", + "stopping_reason_count_repetition_guard": "1", + "avg_total_tokens_target_model": "236503.00", + "total_total_tokens_target_model": "236503.0", + "avg_total_tokens_other_models": "7110.00", + "total_total_tokens_other_models": "7110.0", + "avg_total_tokens_all_models": "243613.00", + "total_total_tokens_all_models": "243613.0", + "avg_turn_count": "5.00", + "total_turn_count": "5.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "1", + "timeout": "1200", + "attempt_timeout": "3600", + "max_connections": "16", + "max_tokens": "64000", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 5000000 + }, + "sandbox": {}, + "max_attempts": 1 + }, + "additional_details": { + "max_retries": "1", + "timeout": "1200", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5-2025-08-07/5d9a8dd5-cf71-4bfc-b3e4-2008a351a5ec_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "8d282e7d41fea4f2497709edb151f2b7f83e416b53d06e1b419dea8abcd7e012", + "total_rows": 1 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5-2025-08-07/5d9a8dd5-cf71-4bfc-b3e4-2008a351a5ec_samples.jsonl b/data/hle/openai/gpt-5-2025-08-07/5d9a8dd5-cf71-4bfc-b3e4-2008a351a5ec_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..2db1cd83e8b177d1156f0b1e6eba6ceb331af958 --- /dev/null +++ b/data/hle/openai/gpt-5-2025-08-07/5d9a8dd5-cf71-4bfc-b3e4-2008a351a5ec_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8d282e7d41fea4f2497709edb151f2b7f83e416b53d06e1b419dea8abcd7e012 +size 23914 diff --git a/data/hle/openai/gpt-5-2025-08-07/6ef27d4a-efa8-43ef-9780-cb9d955168a1.json b/data/hle/openai/gpt-5-2025-08-07/6ef27d4a-efa8-43ef-9780-cb9d955168a1.json new file mode 100644 index 0000000000000000000000000000000000000000..1a242b823b2fed652d3c095884577e96bfa8cab3 --- /dev/null +++ b/data/hle/openai/gpt-5-2025-08-07/6ef27d4a-efa8-43ef-9780-cb9d955168a1.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5-2025-08-07/1786166851.51517", + "evaluation_timestamp": "1776842823.0", + "retrieved_timestamp": "1786166851.51517", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5-2025-08-07", + "id": "openai/gpt-5-2025-08-07", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5-2025-08-07/6ef27d4a-efa8-43ef-9780-cb9d955168a1_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "cf7598364f470ea41a7621e10f516bd59a8b78ebc88bc176d42ae26b1969689b", + "total_rows": 7 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5-2025-08-07/6ef27d4a-efa8-43ef-9780-cb9d955168a1_samples.jsonl b/data/hle/openai/gpt-5-2025-08-07/6ef27d4a-efa8-43ef-9780-cb9d955168a1_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..91bd679c5213abd608d43eb6a8cfb808117355ff --- /dev/null +++ b/data/hle/openai/gpt-5-2025-08-07/6ef27d4a-efa8-43ef-9780-cb9d955168a1_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cf7598364f470ea41a7621e10f516bd59a8b78ebc88bc176d42ae26b1969689b +size 310937 diff --git a/data/hle/openai/gpt-5-2025-08-07/6f1642bf-6170-4dfb-9f1b-a937e2a8d22f.json b/data/hle/openai/gpt-5-2025-08-07/6f1642bf-6170-4dfb-9f1b-a937e2a8d22f.json new file mode 100644 index 0000000000000000000000000000000000000000..931ab485c195f711765179e5cdc02a79ae9980f6 --- /dev/null +++ b/data/hle/openai/gpt-5-2025-08-07/6f1642bf-6170-4dfb-9f1b-a937e2a8d22f.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5-2025-08-07/1786166894.77431", + "evaluation_timestamp": "1777172789.0", + "retrieved_timestamp": "1786166894.77431", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5-2025-08-07", + "id": "openai/gpt-5-2025-08-07", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5-2025-08-07/6f1642bf-6170-4dfb-9f1b-a937e2a8d22f_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "4fde68bb049183e4b85ba3f2dc09f1c0562a4cdabd674ce99252a2c19e67473d", + "total_rows": 14 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5-2025-08-07/6f1642bf-6170-4dfb-9f1b-a937e2a8d22f_samples.jsonl b/data/hle/openai/gpt-5-2025-08-07/6f1642bf-6170-4dfb-9f1b-a937e2a8d22f_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..7a148be16c62b4d04cf50802432ec17e5d6264f7 --- /dev/null +++ b/data/hle/openai/gpt-5-2025-08-07/6f1642bf-6170-4dfb-9f1b-a937e2a8d22f_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4fde68bb049183e4b85ba3f2dc09f1c0562a4cdabd674ce99252a2c19e67473d +size 468060 diff --git a/data/hle/openai/gpt-5-2025-08-07/7986254a-2228-4e29-852c-ad6c8d2ad13d.json b/data/hle/openai/gpt-5-2025-08-07/7986254a-2228-4e29-852c-ad6c8d2ad13d.json new file mode 100644 index 0000000000000000000000000000000000000000..031e99bfa2ce740e6ce2eac9171dfd0cd36b94e5 --- /dev/null +++ b/data/hle/openai/gpt-5-2025-08-07/7986254a-2228-4e29-852c-ad6c8d2ad13d.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5-2025-08-07/1786166831.013844", + "evaluation_timestamp": "1776359245.0", + "retrieved_timestamp": "1786166831.013844", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5-2025-08-07", + "id": "openai/gpt-5-2025-08-07", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5-2025-08-07/7986254a-2228-4e29-852c-ad6c8d2ad13d_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "02faa8176766d1ef5478172ab7d6a5eabe7826e7ab51b00a1bdd37e94e4ce895", + "total_rows": 10 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5-2025-08-07/7986254a-2228-4e29-852c-ad6c8d2ad13d_samples.jsonl b/data/hle/openai/gpt-5-2025-08-07/7986254a-2228-4e29-852c-ad6c8d2ad13d_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..85fd08bdc50f33ba94993161e72db5ebca6b0794 --- /dev/null +++ b/data/hle/openai/gpt-5-2025-08-07/7986254a-2228-4e29-852c-ad6c8d2ad13d_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:02faa8176766d1ef5478172ab7d6a5eabe7826e7ab51b00a1bdd37e94e4ce895 +size 273447 diff --git a/data/hle/openai/gpt-5-2025-08-07/96f27c80-6d98-40f6-b750-c0a46d6dde3e.json b/data/hle/openai/gpt-5-2025-08-07/96f27c80-6d98-40f6-b750-c0a46d6dde3e.json new file mode 100644 index 0000000000000000000000000000000000000000..917bd9278aeaa39cbda511aa11a20a22b4d98518 --- /dev/null +++ b/data/hle/openai/gpt-5-2025-08-07/96f27c80-6d98-40f6-b750-c0a46d6dde3e.json @@ -0,0 +1,117 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5-2025-08-07/1786164232.245249", + "evaluation_timestamp": "1782322677.0", + "retrieved_timestamp": "1786164232.245249", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5-2025-08-07", + "id": "openai/gpt-5-2025-08-07", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "mean on hle/S-adaptive/+1ep/4a951ab4 for scorer _llm_grader", + "source_data": { + "dataset_name": "hle", + "source_type": "hf_dataset", + "hf_repo": "cais/hle", + "samples_number": 3, + "sample_ids": [ + "hle_76cdecee", + "hle_1dc141bb", + "hle_c2ff3a43" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1782322677.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.6666666666666666, + "uncertainty": { + "standard_error": { + "value": 0.33333333333333337 + }, + "num_samples": 1 + }, + "details": { + "total_matched_trajectories": "1", + "stopping_reason_count_repetition_guard": "1", + "avg_total_tokens_target_model": "3392189.00", + "total_total_tokens_target_model": "3392189.0", + "avg_total_tokens_other_models": "4461.00", + "total_total_tokens_other_models": "4461.0", + "avg_total_tokens_all_models": "3396650.00", + "total_total_tokens_all_models": "3396650.0", + "avg_turn_count": "146.00", + "total_turn_count": "146.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "1200", + "max_connections": "16", + "max_tokens": "20000", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 5000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5-2025-08-07/96f27c80-6d98-40f6-b750-c0a46d6dde3e_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "c7c80d060fe4c9793a06299737e0c17d0e325a07cb9b217f6b34e0ff7a391433", + "total_rows": 1 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5-2025-08-07/96f27c80-6d98-40f6-b750-c0a46d6dde3e_samples.jsonl b/data/hle/openai/gpt-5-2025-08-07/96f27c80-6d98-40f6-b750-c0a46d6dde3e_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..180e12301f62e5c1079c10dfe5a7b5dd635b9c57 --- /dev/null +++ b/data/hle/openai/gpt-5-2025-08-07/96f27c80-6d98-40f6-b750-c0a46d6dde3e_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c7c80d060fe4c9793a06299737e0c17d0e325a07cb9b217f6b34e0ff7a391433 +size 208241 diff --git a/data/hle/openai/gpt-5-2025-08-07/970150c9-2fd0-4b7b-a2bd-ef5caf72e7df.json b/data/hle/openai/gpt-5-2025-08-07/970150c9-2fd0-4b7b-a2bd-ef5caf72e7df.json new file mode 100644 index 0000000000000000000000000000000000000000..1313b36fd0f49df79c980ab3abfb77d087099165 --- /dev/null +++ b/data/hle/openai/gpt-5-2025-08-07/970150c9-2fd0-4b7b-a2bd-ef5caf72e7df.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5-2025-08-07/1786166834.335691", + "evaluation_timestamp": "1776993879.0", + "retrieved_timestamp": "1786166834.335691", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5-2025-08-07", + "id": "openai/gpt-5-2025-08-07", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5-2025-08-07/970150c9-2fd0-4b7b-a2bd-ef5caf72e7df_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "588514a248b52f812eddec3376b4876ccff367e0504402da71ac533337064907", + "total_rows": 1 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5-2025-08-07/970150c9-2fd0-4b7b-a2bd-ef5caf72e7df_samples.jsonl b/data/hle/openai/gpt-5-2025-08-07/970150c9-2fd0-4b7b-a2bd-ef5caf72e7df_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..430b79d6dac6539cabfbca5d19e5d6f457e91002 --- /dev/null +++ b/data/hle/openai/gpt-5-2025-08-07/970150c9-2fd0-4b7b-a2bd-ef5caf72e7df_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:588514a248b52f812eddec3376b4876ccff367e0504402da71ac533337064907 +size 15776 diff --git a/data/hle/openai/gpt-5-2025-08-07/9b28d79e-e0c7-4822-82cc-ab7256bfae26.json b/data/hle/openai/gpt-5-2025-08-07/9b28d79e-e0c7-4822-82cc-ab7256bfae26.json new file mode 100644 index 0000000000000000000000000000000000000000..2c4e77d70d238d76302856a1cbe97fa01a706b88 --- /dev/null +++ b/data/hle/openai/gpt-5-2025-08-07/9b28d79e-e0c7-4822-82cc-ab7256bfae26.json @@ -0,0 +1,117 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5-2025-08-07/1786164240.223412", + "evaluation_timestamp": "1783097052.0", + "retrieved_timestamp": "1786164240.223412", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5-2025-08-07", + "id": "openai/gpt-5-2025-08-07", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "mean on hle/S-adaptive/+4ep/614fc868 for scorer _llm_grader", + "source_data": { + "dataset_name": "hle", + "source_type": "hf_dataset", + "hf_repo": "cais/hle", + "samples_number": 3, + "sample_ids": [ + "hle_dfdb78f5", + "hle_209b6b64", + "hle_141cd3ef" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1783097052.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.3333333333333333, + "uncertainty": { + "standard_error": { + "value": 0.33333333333333337 + }, + "num_samples": 12 + }, + "details": { + "total_matched_trajectories": "12", + "stopping_reason_count_repetition_guard": "12", + "avg_total_tokens_target_model": "259720.25", + "total_total_tokens_target_model": "3116643.0", + "avg_total_tokens_other_models": "3922.83", + "total_total_tokens_other_models": "47074.0", + "avg_total_tokens_all_models": "263643.08", + "total_total_tokens_all_models": "3163717.0", + "avg_turn_count": "6.17", + "total_turn_count": "74.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "1", + "timeout": "1200", + "attempt_timeout": "3600", + "max_connections": "16", + "max_tokens": "64000", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 5000000 + }, + "sandbox": {}, + "max_attempts": 1 + }, + "additional_details": { + "max_retries": "1", + "timeout": "1200", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5-2025-08-07/9b28d79e-e0c7-4822-82cc-ab7256bfae26_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "1029445db796117c50e94a866cc460c561d0557e6f92cf4d44c8649b15f7d79e", + "total_rows": 12 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5-2025-08-07/9b28d79e-e0c7-4822-82cc-ab7256bfae26_samples.jsonl b/data/hle/openai/gpt-5-2025-08-07/9b28d79e-e0c7-4822-82cc-ab7256bfae26_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..612b966d535021559b29461c2a2ca789a139884d --- /dev/null +++ b/data/hle/openai/gpt-5-2025-08-07/9b28d79e-e0c7-4822-82cc-ab7256bfae26_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1029445db796117c50e94a866cc460c561d0557e6f92cf4d44c8649b15f7d79e +size 247896 diff --git a/data/hle/openai/gpt-5-2025-08-07/a6f29f9c-750a-4216-965c-a1bc3ee266f3.json b/data/hle/openai/gpt-5-2025-08-07/a6f29f9c-750a-4216-965c-a1bc3ee266f3.json new file mode 100644 index 0000000000000000000000000000000000000000..90e2f6d85a36e4e79490763944a37e3b780705ec --- /dev/null +++ b/data/hle/openai/gpt-5-2025-08-07/a6f29f9c-750a-4216-965c-a1bc3ee266f3.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5-2025-08-07/1786166832.488799", + "evaluation_timestamp": "1776677380.0", + "retrieved_timestamp": "1786166832.488799", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5-2025-08-07", + "id": "openai/gpt-5-2025-08-07", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5-2025-08-07/a6f29f9c-750a-4216-965c-a1bc3ee266f3_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "65c81e12168c626848b3ab549eab96208aaa19b6a00b482d3ea89420c7b07524", + "total_rows": 1 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5-2025-08-07/a6f29f9c-750a-4216-965c-a1bc3ee266f3_samples.jsonl b/data/hle/openai/gpt-5-2025-08-07/a6f29f9c-750a-4216-965c-a1bc3ee266f3_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..d1325518dc0a53c4c3304eabe9a86c1720070306 --- /dev/null +++ b/data/hle/openai/gpt-5-2025-08-07/a6f29f9c-750a-4216-965c-a1bc3ee266f3_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:65c81e12168c626848b3ab549eab96208aaa19b6a00b482d3ea89420c7b07524 +size 64533 diff --git a/data/hle/openai/gpt-5-2025-08-07/a99e5193-714c-4680-a85b-18609056967e.json b/data/hle/openai/gpt-5-2025-08-07/a99e5193-714c-4680-a85b-18609056967e.json new file mode 100644 index 0000000000000000000000000000000000000000..f11c6a14e1125180ce67becd42c0691aa316f128 --- /dev/null +++ b/data/hle/openai/gpt-5-2025-08-07/a99e5193-714c-4680-a85b-18609056967e.json @@ -0,0 +1,208 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5-2025-08-07/1786183381.543466", + "evaluation_timestamp": "1775229625.0", + "retrieved_timestamp": "1786183381.543466", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.202,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5-2025-08-07", + "id": "openai/gpt-5-2025-08-07", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "mean on hle for scorer _llm_grader", + "source_data": { + "dataset_name": "hle", + "source_type": "hf_dataset", + "hf_repo": "cais/hle", + "samples_number": 1909, + "sample_ids": [ + "hle_d6b73839", + "hle_07662c9c", + "hle_764ce81c", + "hle_4e3440f7", + "hle_baccd6da", + "hle_f46c2857", + "hle_86a18059", + "hle_8770ab31", + "hle_d1bd2767", + "hle_6499b500", + "hle_e1296346", + "hle_ca6c436b", + "hle_a559277e", + "hle_8a419425", + "hle_85e95ef4", + "hle_28643173", + "hle_bc5fbce2", + "hle_25e0c71d", + "hle_76cdecee", + "hle_dfdb78f5", + "hle_142ffcff", + "hle_ffb12d83", + "hle_37c97863", + "hle_209b6b64", + "hle_0702accc", + "hle_16388cc8", + "hle_4418bea4", + "hle_ebd62885", + "hle_4485802f", + "hle_9019c023", + "hle_9282e7ae", + "hle_44007b97", + "hle_7808792d", + "hle_e3b8f5c2", + "hle_3cf4483d", + "hle_f85a0fa1", + "hle_c3ca0bda", + "hle_376b8247", + "hle_31497469", + "hle_c43a27a3", + "hle_ea41e785", + "hle_d0c8b07e", + "hle_0291c889", + "hle_b1a7811d", + "hle_dab7fbb1", + "hle_cf3d65bd", + "hle_5e235b48", + "hle_0ff76ff3", + "hle_70e5aebc", + "hle_aa2fed90", + "hle_9df5d371", + "hle_5759e7d7", + "hle_ca59b32c", + "hle_33bc94e2", + "hle_e3817b7e", + "hle_d7b7c3dc", + "hle_1dc141bb", + "hle_a764505b", + "hle_b6cec502", + "hle_e62ba103", + "hle_9255683e", + "hle_34c69126", + "hle_5f04b314", + "hle_fcd56401", + "hle_b5c681d9", + "hle_d5a55900", + "hle_3e11b38a", + "hle_0a1eec3f", + "hle_5d49423c", + "hle_1d580261", + "hle_82f5c665", + "hle_cc7f25c0", + "hle_226e134b", + "hle_c1158fb8", + "hle_d0ea1019", + "hle_fcd42d0c", + "hle_b33012cf", + "hle_3697d2bc", + "hle_5e349785", + "hle_8b0384c4", + "hle_ee938b26", + "hle_d190d8c2", + "hle_b26bf8aa", + "hle_c2ff3a43", + "hle_141cd3ef", + "hle_5671e386", + "hle_a634e89e", + "hle_670287ad", + "hle_298b5616", + "hle_e14d8c23", + "hle_510b6040", + "hle_c6adc1b4", + "hle_0ec247c5", + "hle_aa343d51", + "hle_d52d3dae", + "hle_6d963e58", + "hle_35a40c52", + "hle_45d1c140", + "hle_c0ed25f7", + "hle_cd4aa5b3" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1775229625.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.2646288029100529, + "uncertainty": { + "standard_error": { + "value": 0.039022604722590774 + }, + "num_samples": 837 + }, + "details": { + "total_matched_trajectories": "452", + "stopping_reason_count_repetition_guard": "452", + "avg_total_tokens_target_model": "41920.85", + "total_total_tokens_target_model": "18948224.0", + "avg_total_tokens_other_models": "4630.29", + "total_total_tokens_other_models": "2092891.0", + "avg_total_tokens_all_models": "46551.14", + "total_total_tokens_all_models": "21041115.0", + "avg_turn_count": "5.72", + "total_turn_count": "2585.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\"}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "3", + "attempt_timeout": "1200", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 5000000 + }, + "sandbox": {}, + "max_attempts": 3 + }, + "additional_details": { + "max_retries": "3" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5-2025-08-07/a99e5193-714c-4680-a85b-18609056967e_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "5b592e041d065749632065293f92a5fbedd5bcecf0b4f7b20b1f2af3667b8d92", + "total_rows": 837 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5-2025-08-07/a99e5193-714c-4680-a85b-18609056967e_samples.jsonl b/data/hle/openai/gpt-5-2025-08-07/a99e5193-714c-4680-a85b-18609056967e_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..cafb37e544a2e68e21e05f5ebbf07b84a55f45dd --- /dev/null +++ b/data/hle/openai/gpt-5-2025-08-07/a99e5193-714c-4680-a85b-18609056967e_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5b592e041d065749632065293f92a5fbedd5bcecf0b4f7b20b1f2af3667b8d92 +size 19281413 diff --git a/data/hle/openai/gpt-5-2025-08-07/b408af5a-e5c5-421f-9f55-3da497351bd3.json b/data/hle/openai/gpt-5-2025-08-07/b408af5a-e5c5-421f-9f55-3da497351bd3.json new file mode 100644 index 0000000000000000000000000000000000000000..0cc87f8a72be3e6ea5fd26b087a640a541da8a12 --- /dev/null +++ b/data/hle/openai/gpt-5-2025-08-07/b408af5a-e5c5-421f-9f55-3da497351bd3.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5-2025-08-07/1786166891.528768", + "evaluation_timestamp": "1777431373.0", + "retrieved_timestamp": "1786166891.528768", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5-2025-08-07", + "id": "openai/gpt-5-2025-08-07", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5-2025-08-07/b408af5a-e5c5-421f-9f55-3da497351bd3_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "8d416e2391a7b539be4bbfb66aeda5b25ed7bdb9740752d5494511072ddd5820", + "total_rows": 7 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5-2025-08-07/b408af5a-e5c5-421f-9f55-3da497351bd3_samples.jsonl b/data/hle/openai/gpt-5-2025-08-07/b408af5a-e5c5-421f-9f55-3da497351bd3_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..3a86785135b2244d0f3a0ff8084cdd5dc7194b9b --- /dev/null +++ b/data/hle/openai/gpt-5-2025-08-07/b408af5a-e5c5-421f-9f55-3da497351bd3_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8d416e2391a7b539be4bbfb66aeda5b25ed7bdb9740752d5494511072ddd5820 +size 469883 diff --git a/data/hle/openai/gpt-5-2025-08-07/b596bbfd-33f6-4f7e-be57-15e11a7441fb.json b/data/hle/openai/gpt-5-2025-08-07/b596bbfd-33f6-4f7e-be57-15e11a7441fb.json new file mode 100644 index 0000000000000000000000000000000000000000..456ed3bbe116212174dbb47382e14c0c7cf844fe --- /dev/null +++ b/data/hle/openai/gpt-5-2025-08-07/b596bbfd-33f6-4f7e-be57-15e11a7441fb.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5-2025-08-07/1786166839.468688", + "evaluation_timestamp": "1776519735.0", + "retrieved_timestamp": "1786166839.468688", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5-2025-08-07", + "id": "openai/gpt-5-2025-08-07", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5-2025-08-07/b596bbfd-33f6-4f7e-be57-15e11a7441fb_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "53814f565de41dd855e84690d7f3505b53fb6bb95a6fe5e74f811b4421c10145", + "total_rows": 20 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5-2025-08-07/b596bbfd-33f6-4f7e-be57-15e11a7441fb_samples.jsonl b/data/hle/openai/gpt-5-2025-08-07/b596bbfd-33f6-4f7e-be57-15e11a7441fb_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..584cc05eb1e30d2c28271421820d5b34834c9029 --- /dev/null +++ b/data/hle/openai/gpt-5-2025-08-07/b596bbfd-33f6-4f7e-be57-15e11a7441fb_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:53814f565de41dd855e84690d7f3505b53fb6bb95a6fe5e74f811b4421c10145 +size 1496889 diff --git a/data/hle/openai/gpt-5-2025-08-07/b8d32e0b-771f-414c-a996-338b177b9a15.json b/data/hle/openai/gpt-5-2025-08-07/b8d32e0b-771f-414c-a996-338b177b9a15.json new file mode 100644 index 0000000000000000000000000000000000000000..4926d42f734b8fddfb8778306b571e6208d6dc30 --- /dev/null +++ b/data/hle/openai/gpt-5-2025-08-07/b8d32e0b-771f-414c-a996-338b177b9a15.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5-2025-08-07/1786166849.488058", + "evaluation_timestamp": "1776519735.0", + "retrieved_timestamp": "1786166849.488058", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5-2025-08-07", + "id": "openai/gpt-5-2025-08-07", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5-2025-08-07/b8d32e0b-771f-414c-a996-338b177b9a15_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "53cc59297a7b9fee8f95060ffe07c2ec734f64d4fd0e55b0c6f4a96027f513e9", + "total_rows": 20 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5-2025-08-07/b8d32e0b-771f-414c-a996-338b177b9a15_samples.jsonl b/data/hle/openai/gpt-5-2025-08-07/b8d32e0b-771f-414c-a996-338b177b9a15_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..9a15d346fd6475c7a8c23693897879e169cf282c --- /dev/null +++ b/data/hle/openai/gpt-5-2025-08-07/b8d32e0b-771f-414c-a996-338b177b9a15_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:53cc59297a7b9fee8f95060ffe07c2ec734f64d4fd0e55b0c6f4a96027f513e9 +size 1296297 diff --git a/data/hle/openai/gpt-5-2025-08-07/d81a8f4d-26ac-4db8-8df1-f6c0bb4c0df4.json b/data/hle/openai/gpt-5-2025-08-07/d81a8f4d-26ac-4db8-8df1-f6c0bb4c0df4.json new file mode 100644 index 0000000000000000000000000000000000000000..1e7fef40ef44261cf8838fe5e842b2630c497f33 --- /dev/null +++ b/data/hle/openai/gpt-5-2025-08-07/d81a8f4d-26ac-4db8-8df1-f6c0bb4c0df4.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5-2025-08-07/1786166845.96413", + "evaluation_timestamp": "1776962159.0", + "retrieved_timestamp": "1786166845.96413", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5-2025-08-07", + "id": "openai/gpt-5-2025-08-07", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5-2025-08-07/d81a8f4d-26ac-4db8-8df1-f6c0bb4c0df4_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "601a5451f941ac24451b1e92edf7b2e24181663e2de81375b46fc26d164b0ba9", + "total_rows": 13 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5-2025-08-07/d81a8f4d-26ac-4db8-8df1-f6c0bb4c0df4_samples.jsonl b/data/hle/openai/gpt-5-2025-08-07/d81a8f4d-26ac-4db8-8df1-f6c0bb4c0df4_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..506d5905759f5f5e68549b10be25d9e684be3e5d --- /dev/null +++ b/data/hle/openai/gpt-5-2025-08-07/d81a8f4d-26ac-4db8-8df1-f6c0bb4c0df4_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:601a5451f941ac24451b1e92edf7b2e24181663e2de81375b46fc26d164b0ba9 +size 553288 diff --git a/data/hle/openai/gpt-5-2025-08-07/e1cd9f5f-b3f9-44a9-af8a-95dabf555170.json b/data/hle/openai/gpt-5-2025-08-07/e1cd9f5f-b3f9-44a9-af8a-95dabf555170.json new file mode 100644 index 0000000000000000000000000000000000000000..b86561f0c581d9e4de9c76209643c84d77c1a91d --- /dev/null +++ b/data/hle/openai/gpt-5-2025-08-07/e1cd9f5f-b3f9-44a9-af8a-95dabf555170.json @@ -0,0 +1,113 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5-2025-08-07/1786164229.511041", + "evaluation_timestamp": "1783023406.0", + "retrieved_timestamp": "1786164229.511041", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5-2025-08-07", + "id": "openai/gpt-5-2025-08-07", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "mean on hle/S-adaptive/+1ep/73ba8e75 for scorer _llm_grader", + "source_data": { + "dataset_name": "hle", + "source_type": "hf_dataset", + "hf_repo": "cais/hle", + "samples_number": 2, + "sample_ids": [ + "hle_76cdecee", + "hle_c2ff3a43" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1783023406.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 1.0, + "uncertainty": { + "num_samples": 1 + }, + "details": { + "total_matched_trajectories": "1", + "stopping_reason_count_repetition_guard": "1", + "avg_total_tokens_target_model": "59511.00", + "total_total_tokens_target_model": "59511.0", + "avg_total_tokens_other_models": "6007.00", + "total_total_tokens_other_models": "6007.0", + "avg_total_tokens_all_models": "65518.00", + "total_total_tokens_all_models": "65518.0", + "avg_turn_count": "10.00", + "total_turn_count": "10.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "1200", + "max_connections": "16", + "max_tokens": "20000", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 5000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5-2025-08-07/e1cd9f5f-b3f9-44a9-af8a-95dabf555170_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "d9a310ea3f15e69c585e7ec269ce86aa7069e42f9573527af21ba530dcfa570e", + "total_rows": 1 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5-2025-08-07/e1cd9f5f-b3f9-44a9-af8a-95dabf555170_samples.jsonl b/data/hle/openai/gpt-5-2025-08-07/e1cd9f5f-b3f9-44a9-af8a-95dabf555170_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..94802d136f944414513842c635db5d7c3c5feee3 --- /dev/null +++ b/data/hle/openai/gpt-5-2025-08-07/e1cd9f5f-b3f9-44a9-af8a-95dabf555170_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d9a310ea3f15e69c585e7ec269ce86aa7069e42f9573527af21ba530dcfa570e +size 24213 diff --git a/data/hle/openai/gpt-5-2025-08-07/ed106b8f-b3eb-44bb-a6e6-717ffdf8cd5d.json b/data/hle/openai/gpt-5-2025-08-07/ed106b8f-b3eb-44bb-a6e6-717ffdf8cd5d.json new file mode 100644 index 0000000000000000000000000000000000000000..ca0ef55e49154bb08fdc4397d50fe8813fa81be3 --- /dev/null +++ b/data/hle/openai/gpt-5-2025-08-07/ed106b8f-b3eb-44bb-a6e6-717ffdf8cd5d.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5-2025-08-07/1786168829.864453", + "evaluation_timestamp": "1782728944.0", + "retrieved_timestamp": "1786168829.864453", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5-2025-08-07", + "id": "openai/gpt-5-2025-08-07", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5-2025-08-07/ed106b8f-b3eb-44bb-a6e6-717ffdf8cd5d_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "a7d26ba59f41c2ba7f47f794d0b1c78cce3e827cbc3e209f72c79a96b67dac9e", + "total_rows": 1 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5-2025-08-07/ed106b8f-b3eb-44bb-a6e6-717ffdf8cd5d_samples.jsonl b/data/hle/openai/gpt-5-2025-08-07/ed106b8f-b3eb-44bb-a6e6-717ffdf8cd5d_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..13e395d56342034bd037151d011418adcee7e4bc --- /dev/null +++ b/data/hle/openai/gpt-5-2025-08-07/ed106b8f-b3eb-44bb-a6e6-717ffdf8cd5d_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a7d26ba59f41c2ba7f47f794d0b1c78cce3e827cbc3e209f72c79a96b67dac9e +size 81305 diff --git a/data/hle/openai/gpt-5-2025-08-07/fc92fa58-9237-440e-a0b0-51b90c61b935.json b/data/hle/openai/gpt-5-2025-08-07/fc92fa58-9237-440e-a0b0-51b90c61b935.json new file mode 100644 index 0000000000000000000000000000000000000000..5fa538c94cf7ca1e1e666ef0213c1b4aa55a02a5 --- /dev/null +++ b/data/hle/openai/gpt-5-2025-08-07/fc92fa58-9237-440e-a0b0-51b90c61b935.json @@ -0,0 +1,112 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5-2025-08-07/1786164233.895257", + "evaluation_timestamp": "1783334521.0", + "retrieved_timestamp": "1786164233.895257", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5-2025-08-07", + "id": "openai/gpt-5-2025-08-07", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "mean on hle/S-adaptive/+2ep/70f636a7 for scorer _llm_grader", + "source_data": { + "dataset_name": "hle", + "source_type": "hf_dataset", + "hf_repo": "cais/hle", + "samples_number": 1, + "sample_ids": [ + "hle_5d49423c" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1783334521.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 1.0, + "uncertainty": { + "num_samples": 2 + }, + "details": { + "total_matched_trajectories": "2", + "stopping_reason_count_repetition_guard": "2", + "avg_total_tokens_target_model": "361475.00", + "total_total_tokens_target_model": "722950.0", + "avg_total_tokens_other_models": "8090.50", + "total_total_tokens_other_models": "16181.0", + "avg_total_tokens_all_models": "369565.50", + "total_total_tokens_all_models": "739131.0", + "avg_turn_count": "6.50", + "total_turn_count": "13.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "1", + "timeout": "1200", + "attempt_timeout": "3600", + "max_connections": "16", + "max_tokens": "64000", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 5000000 + }, + "sandbox": {}, + "max_attempts": 1 + }, + "additional_details": { + "max_retries": "1", + "timeout": "1200", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5-2025-08-07/fc92fa58-9237-440e-a0b0-51b90c61b935_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "d9f8978b9ff2acac443dd8b154f64fefdf8610b14314904419a10493b2b71ffc", + "total_rows": 2 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5-2025-08-07/fc92fa58-9237-440e-a0b0-51b90c61b935_samples.jsonl b/data/hle/openai/gpt-5-2025-08-07/fc92fa58-9237-440e-a0b0-51b90c61b935_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..947cad076f9856cdc6b757c176b29ff398c38375 --- /dev/null +++ b/data/hle/openai/gpt-5-2025-08-07/fc92fa58-9237-440e-a0b0-51b90c61b935_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d9f8978b9ff2acac443dd8b154f64fefdf8610b14314904419a10493b2b71ffc +size 49303 diff --git a/data/hle/openai/gpt-5.2-2025-12-11/03399bdd-c71f-447d-8694-85bcd61cfdce.json b/data/hle/openai/gpt-5.2-2025-12-11/03399bdd-c71f-447d-8694-85bcd61cfdce.json new file mode 100644 index 0000000000000000000000000000000000000000..25aa7dde4a9acdc9ca94141a14d13a92a8dac26b --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/03399bdd-c71f-447d-8694-85bcd61cfdce.json @@ -0,0 +1,111 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5.2-2025-12-11/1786164033.275709", + "evaluation_timestamp": "1778511463.0", + "retrieved_timestamp": "1786164033.275709", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "mean on hle/S-adaptive/+1ep/a5627fac for scorer _llm_grader", + "source_data": { + "dataset_name": "hle", + "source_type": "hf_dataset", + "hf_repo": "cais/hle", + "samples_number": 1, + "sample_ids": [ + "hle_8a419425" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1778511463.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.0, + "uncertainty": { + "num_samples": 1 + }, + "details": { + "total_matched_trajectories": "1", + "stopping_reason_count_repetition_guard": "1", + "avg_total_tokens_target_model": "48563.00", + "total_total_tokens_target_model": "48563.0", + "avg_total_tokens_other_models": "6787.00", + "total_total_tokens_other_models": "6787.0", + "avg_total_tokens_all_models": "55350.00", + "total_total_tokens_all_models": "55350.0", + "avg_turn_count": "5.00", + "total_turn_count": "5.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 5000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5.2-2025-12-11/03399bdd-c71f-447d-8694-85bcd61cfdce_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "5034a3cc24e9dcd7dd8bb982259548503edca420ecfd5cead22113e4bc909050", + "total_rows": 1 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5.2-2025-12-11/03399bdd-c71f-447d-8694-85bcd61cfdce_samples.jsonl b/data/hle/openai/gpt-5.2-2025-12-11/03399bdd-c71f-447d-8694-85bcd61cfdce_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..e9310aa25ef72c411b40974913f5b330b9adc55a --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/03399bdd-c71f-447d-8694-85bcd61cfdce_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5034a3cc24e9dcd7dd8bb982259548503edca420ecfd5cead22113e4bc909050 +size 23003 diff --git a/data/hle/openai/gpt-5.2-2025-12-11/048ccf2e-7d28-41d1-8af7-bda456b201f5.json b/data/hle/openai/gpt-5.2-2025-12-11/048ccf2e-7d28-41d1-8af7-bda456b201f5.json new file mode 100644 index 0000000000000000000000000000000000000000..133c97f15fc18514d0b3e223aee1d2a65aac5d94 --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/048ccf2e-7d28-41d1-8af7-bda456b201f5.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5.2-2025-12-11/1786168715.201185", + "evaluation_timestamp": "1776519735.0", + "retrieved_timestamp": "1786168715.201185", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5.2-2025-12-11/048ccf2e-7d28-41d1-8af7-bda456b201f5_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "aee6867fa72562062c95f2beaf971a415bcd32b5f3034b266b60b652c72e5c7b", + "total_rows": 10 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5.2-2025-12-11/048ccf2e-7d28-41d1-8af7-bda456b201f5_samples.jsonl b/data/hle/openai/gpt-5.2-2025-12-11/048ccf2e-7d28-41d1-8af7-bda456b201f5_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..e53b5977290d47edad3e403df422a260e8ecc05f --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/048ccf2e-7d28-41d1-8af7-bda456b201f5_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:aee6867fa72562062c95f2beaf971a415bcd32b5f3034b266b60b652c72e5c7b +size 182368 diff --git a/data/hle/openai/gpt-5.2-2025-12-11/0b4ee198-4b7b-4b3f-b1df-e609fbf2abda.json b/data/hle/openai/gpt-5.2-2025-12-11/0b4ee198-4b7b-4b3f-b1df-e609fbf2abda.json new file mode 100644 index 0000000000000000000000000000000000000000..fb3d51e909db603f35f928b85869549e2c778d9c --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/0b4ee198-4b7b-4b3f-b1df-e609fbf2abda.json @@ -0,0 +1,112 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5.2-2025-12-11/1786164121.60635", + "evaluation_timestamp": "1783344151.0", + "retrieved_timestamp": "1786164121.60635", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "mean on hle/S-adaptive/+4ep/da3eda81 for scorer _llm_grader", + "source_data": { + "dataset_name": "hle", + "source_type": "hf_dataset", + "hf_repo": "cais/hle", + "samples_number": 1, + "sample_ids": [ + "hle_e3b8f5c2" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1783344151.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.0, + "uncertainty": { + "num_samples": 3 + }, + "details": { + "total_matched_trajectories": "3", + "stopping_reason_count_repetition_guard": "3", + "avg_total_tokens_target_model": "513990.00", + "total_total_tokens_target_model": "1541970.0", + "avg_total_tokens_other_models": "6062.00", + "total_total_tokens_other_models": "18186.0", + "avg_total_tokens_all_models": "520052.00", + "total_total_tokens_all_models": "1560156.0", + "avg_turn_count": "9.33", + "total_turn_count": "28.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "1", + "timeout": "1200", + "attempt_timeout": "3600", + "max_connections": "16", + "max_tokens": "64000", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 5000000 + }, + "sandbox": {}, + "max_attempts": 1 + }, + "additional_details": { + "max_retries": "1", + "timeout": "1200", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5.2-2025-12-11/0b4ee198-4b7b-4b3f-b1df-e609fbf2abda_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "91f3026fda5dcb0d86dfaa8460559bd8f269ce879106c2290b2d6e454a077f70", + "total_rows": 3 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5.2-2025-12-11/0b4ee198-4b7b-4b3f-b1df-e609fbf2abda_samples.jsonl b/data/hle/openai/gpt-5.2-2025-12-11/0b4ee198-4b7b-4b3f-b1df-e609fbf2abda_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..88cd3a79ef5aa22cf677df766abdd90dbca0ad79 --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/0b4ee198-4b7b-4b3f-b1df-e609fbf2abda_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:91f3026fda5dcb0d86dfaa8460559bd8f269ce879106c2290b2d6e454a077f70 +size 80681 diff --git a/data/hle/openai/gpt-5.2-2025-12-11/0fdb3aba-be61-47df-b5a1-1440a4165e82.json b/data/hle/openai/gpt-5.2-2025-12-11/0fdb3aba-be61-47df-b5a1-1440a4165e82.json new file mode 100644 index 0000000000000000000000000000000000000000..f712b965fab800bf06137476589b14441d2d835d --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/0fdb3aba-be61-47df-b5a1-1440a4165e82.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5.2-2025-12-11/1786166765.212794", + "evaluation_timestamp": "1776677380.0", + "retrieved_timestamp": "1786166765.212794", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5.2-2025-12-11/0fdb3aba-be61-47df-b5a1-1440a4165e82_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "6f6d862eca49b8c5366796126cd288218c7769f3a2cb96d92b2c17b2ee8d39cc", + "total_rows": 1 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5.2-2025-12-11/0fdb3aba-be61-47df-b5a1-1440a4165e82_samples.jsonl b/data/hle/openai/gpt-5.2-2025-12-11/0fdb3aba-be61-47df-b5a1-1440a4165e82_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..97da9ebd5d020bd7804cf2e89e3eba7329dc9950 --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/0fdb3aba-be61-47df-b5a1-1440a4165e82_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6f6d862eca49b8c5366796126cd288218c7769f3a2cb96d92b2c17b2ee8d39cc +size 44406 diff --git a/data/hle/openai/gpt-5.2-2025-12-11/13bf46d9-1382-435a-ac7d-0341a81c0072.json b/data/hle/openai/gpt-5.2-2025-12-11/13bf46d9-1382-435a-ac7d-0341a81c0072.json new file mode 100644 index 0000000000000000000000000000000000000000..902a5c6da0c0c9cae1c6334650dca85932219c45 --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/13bf46d9-1382-435a-ac7d-0341a81c0072.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5.2-2025-12-11/1786166763.681421", + "evaluation_timestamp": "1776359245.0", + "retrieved_timestamp": "1786166763.681421", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5.2-2025-12-11/13bf46d9-1382-435a-ac7d-0341a81c0072_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "af68281e8cd9b87889a0aa735e4d4fdab426cb5c4c755b261506eaea989005f2", + "total_rows": 10 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5.2-2025-12-11/13bf46d9-1382-435a-ac7d-0341a81c0072_samples.jsonl b/data/hle/openai/gpt-5.2-2025-12-11/13bf46d9-1382-435a-ac7d-0341a81c0072_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..2b91fe7449c4c235e474ac2bc4095c4b138787e7 --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/13bf46d9-1382-435a-ac7d-0341a81c0072_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:af68281e8cd9b87889a0aa735e4d4fdab426cb5c4c755b261506eaea989005f2 +size 292978 diff --git a/data/hle/openai/gpt-5.2-2025-12-11/167a461a-785b-4848-a615-48123bddf940.json b/data/hle/openai/gpt-5.2-2025-12-11/167a461a-785b-4848-a615-48123bddf940.json new file mode 100644 index 0000000000000000000000000000000000000000..7f92da524c84bbffc07afa8a99fd897261bc9ee7 --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/167a461a-785b-4848-a615-48123bddf940.json @@ -0,0 +1,125 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5.2-2025-12-11/1786164136.393154", + "evaluation_timestamp": "1783114899.0", + "retrieved_timestamp": "1786164136.393154", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "mean on hle/S-adaptive/a80e351c for scorer _llm_grader", + "source_data": { + "dataset_name": "hle", + "source_type": "hf_dataset", + "hf_repo": "cais/hle", + "samples_number": 11, + "sample_ids": [ + "hle_764ce81c", + "hle_bc5fbce2", + "hle_dfdb78f5", + "hle_209b6b64", + "hle_7808792d", + "hle_e3b8f5c2", + "hle_9df5d371", + "hle_9255683e", + "hle_fcd56401", + "hle_5d49423c", + "hle_45d1c140" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1783114899.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.5348484848484848, + "uncertainty": { + "standard_error": { + "value": 0.14035407148025156 + }, + "num_samples": 46 + }, + "details": { + "total_matched_trajectories": "46", + "stopping_reason_count_repetition_guard": "46", + "avg_total_tokens_target_model": "337715.59", + "total_total_tokens_target_model": "15534917.0", + "avg_total_tokens_other_models": "6176.02", + "total_total_tokens_other_models": "284097.0", + "avg_total_tokens_all_models": "343891.61", + "total_total_tokens_all_models": "15819014.0", + "avg_turn_count": "6.80", + "total_turn_count": "313.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "1", + "timeout": "1200", + "attempt_timeout": "3600", + "max_connections": "16", + "max_tokens": "64000", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 5000000 + }, + "sandbox": {}, + "max_attempts": 1 + }, + "additional_details": { + "max_retries": "1", + "timeout": "1200", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5.2-2025-12-11/167a461a-785b-4848-a615-48123bddf940_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "516a374af8d9a8ae0ef4e1b2097e6b00a7cc7449a8543a8616c5a5ef97b8d9ae", + "total_rows": 46 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5.2-2025-12-11/167a461a-785b-4848-a615-48123bddf940_samples.jsonl b/data/hle/openai/gpt-5.2-2025-12-11/167a461a-785b-4848-a615-48123bddf940_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..29a9b5c5dd0752428acffd6d37783376d0a957dc --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/167a461a-785b-4848-a615-48123bddf940_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:516a374af8d9a8ae0ef4e1b2097e6b00a7cc7449a8543a8616c5a5ef97b8d9ae +size 1094627 diff --git a/data/hle/openai/gpt-5.2-2025-12-11/1c0e902a-9277-4052-8cac-9689e2056146.json b/data/hle/openai/gpt-5.2-2025-12-11/1c0e902a-9277-4052-8cac-9689e2056146.json new file mode 100644 index 0000000000000000000000000000000000000000..5261cf22af00d1778f045a70d8df3d522ad0e3a0 --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/1c0e902a-9277-4052-8cac-9689e2056146.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5.2-2025-12-11/1786166770.485662", + "evaluation_timestamp": "1776848562.0", + "retrieved_timestamp": "1786166770.485662", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5.2-2025-12-11/1c0e902a-9277-4052-8cac-9689e2056146_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "ad6d5f95866f28b87d176650c0faa53d50fda4b8c8e861f6c5b405ba9ace5fd0", + "total_rows": 6 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5.2-2025-12-11/1c0e902a-9277-4052-8cac-9689e2056146_samples.jsonl b/data/hle/openai/gpt-5.2-2025-12-11/1c0e902a-9277-4052-8cac-9689e2056146_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..a7f9a11f82e2368c51f79f93bc724f0f23fc3984 --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/1c0e902a-9277-4052-8cac-9689e2056146_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ad6d5f95866f28b87d176650c0faa53d50fda4b8c8e861f6c5b405ba9ace5fd0 +size 180967 diff --git a/data/hle/openai/gpt-5.2-2025-12-11/1ce91e56-1290-4ccd-bb2c-56a946844615.json b/data/hle/openai/gpt-5.2-2025-12-11/1ce91e56-1290-4ccd-bb2c-56a946844615.json new file mode 100644 index 0000000000000000000000000000000000000000..82385db785452635520f81385813821f5bb22047 --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/1ce91e56-1290-4ccd-bb2c-56a946844615.json @@ -0,0 +1,114 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5.2-2025-12-11/1786164090.803535", + "evaluation_timestamp": "1781619827.0", + "retrieved_timestamp": "1786164090.803535", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "mean on hle/S-adaptive/+2ep/07b6875f for scorer _llm_grader", + "source_data": { + "dataset_name": "hle", + "source_type": "hf_dataset", + "hf_repo": "cais/hle", + "samples_number": 3, + "sample_ids": [ + "hle_70e5aebc", + "hle_b6cec502", + "hle_1d580261" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1781619827.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.0, + "uncertainty": { + "num_samples": 3 + }, + "details": { + "total_matched_trajectories": "3", + "stopping_reason_count_repetition_guard": "3", + "avg_total_tokens_target_model": "24334.33", + "total_total_tokens_target_model": "73003.0", + "avg_total_tokens_other_models": "4312.33", + "total_total_tokens_other_models": "12937.0", + "avg_total_tokens_all_models": "28646.67", + "total_total_tokens_all_models": "85940.0", + "avg_turn_count": "5.00", + "total_turn_count": "15.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "16", + "max_tokens": "20000", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 5000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5.2-2025-12-11/1ce91e56-1290-4ccd-bb2c-56a946844615_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "9b0d97dc0de4fe7171316ca20e21991eecc023cfe5ad2bdc62419b64437a00cb", + "total_rows": 3 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5.2-2025-12-11/1ce91e56-1290-4ccd-bb2c-56a946844615_samples.jsonl b/data/hle/openai/gpt-5.2-2025-12-11/1ce91e56-1290-4ccd-bb2c-56a946844615_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..f2b1bb99eb23f6d791830da42b811b724157682b --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/1ce91e56-1290-4ccd-bb2c-56a946844615_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9b0d97dc0de4fe7171316ca20e21991eecc023cfe5ad2bdc62419b64437a00cb +size 57975 diff --git a/data/hle/openai/gpt-5.2-2025-12-11/1e281136-d9ea-4a1d-a222-4fe71c46ddd5.json b/data/hle/openai/gpt-5.2-2025-12-11/1e281136-d9ea-4a1d-a222-4fe71c46ddd5.json new file mode 100644 index 0000000000000000000000000000000000000000..398836fc89f98b1cd767a404ae4f000af8851e24 --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/1e281136-d9ea-4a1d-a222-4fe71c46ddd5.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5.2-2025-12-11/1786166779.140989", + "evaluation_timestamp": "1777355509.0", + "retrieved_timestamp": "1786166779.140989", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5.2-2025-12-11/1e281136-d9ea-4a1d-a222-4fe71c46ddd5_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "f66baf0e94af62164ea2b0e9c52e8cfbd9afabb8467be1d6c960a7c8d05c001c", + "total_rows": 3 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5.2-2025-12-11/1e281136-d9ea-4a1d-a222-4fe71c46ddd5_samples.jsonl b/data/hle/openai/gpt-5.2-2025-12-11/1e281136-d9ea-4a1d-a222-4fe71c46ddd5_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..1ae6a8265315317f1eec629d2e778a0591c2843c --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/1e281136-d9ea-4a1d-a222-4fe71c46ddd5_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f66baf0e94af62164ea2b0e9c52e8cfbd9afabb8467be1d6c960a7c8d05c001c +size 131996 diff --git a/data/hle/openai/gpt-5.2-2025-12-11/202a1e8f-3030-4c58-af71-cf187fd07748.json b/data/hle/openai/gpt-5.2-2025-12-11/202a1e8f-3030-4c58-af71-cf187fd07748.json new file mode 100644 index 0000000000000000000000000000000000000000..e4636ae401f33abb8f93ba4c3063633b6c9a84e0 --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/202a1e8f-3030-4c58-af71-cf187fd07748.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5.2-2025-12-11/1786168719.342133", + "evaluation_timestamp": "1778563675.0", + "retrieved_timestamp": "1786168719.342133", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5.2-2025-12-11/202a1e8f-3030-4c58-af71-cf187fd07748_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "bcb97019281c58acf3f67b38887fb0bf0c05bb80b55675ffff22c2479e34e316", + "total_rows": 3 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5.2-2025-12-11/202a1e8f-3030-4c58-af71-cf187fd07748_samples.jsonl b/data/hle/openai/gpt-5.2-2025-12-11/202a1e8f-3030-4c58-af71-cf187fd07748_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..09d428587d8c5b0a2b1282f285b1892093600412 --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/202a1e8f-3030-4c58-af71-cf187fd07748_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bcb97019281c58acf3f67b38887fb0bf0c05bb80b55675ffff22c2479e34e316 +size 54627 diff --git a/data/hle/openai/gpt-5.2-2025-12-11/257d9f58-e21e-4b10-9a76-03bd0f0157a3.json b/data/hle/openai/gpt-5.2-2025-12-11/257d9f58-e21e-4b10-9a76-03bd0f0157a3.json new file mode 100644 index 0000000000000000000000000000000000000000..18a8224b996d47a367e819c17f1adf54322e9763 --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/257d9f58-e21e-4b10-9a76-03bd0f0157a3.json @@ -0,0 +1,110 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5.2-2025-12-11/1786164130.429003", + "evaluation_timestamp": "1776966483.0", + "retrieved_timestamp": "1786164130.429003", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "mean on hle/S-adaptive/+4ep/3510a9a9 for scorer _llm_grader", + "source_data": { + "dataset_name": "hle", + "source_type": "hf_dataset", + "hf_repo": "cais/hle", + "samples_number": 3, + "sample_ids": [ + "hle_ca6c436b", + "hle_25e0c71d", + "hle_376b8247" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1776966483.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.0, + "uncertainty": { + "num_samples": 6 + }, + "details": { + "total_matched_trajectories": "6", + "stopping_reason_count_repetition_guard": "6", + "avg_total_tokens_target_model": "38365.83", + "total_total_tokens_target_model": "230195.0", + "avg_total_tokens_other_models": "4493.33", + "total_total_tokens_other_models": "26960.0", + "avg_total_tokens_all_models": "42859.17", + "total_total_tokens_all_models": "257155.0", + "avg_turn_count": "5.00", + "total_turn_count": "30.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 5000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5.2-2025-12-11/257d9f58-e21e-4b10-9a76-03bd0f0157a3_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "991b6a78e9b193b03c9e21ca69aee80d9820a43cc1293b2f0c03e3982929f7b5", + "total_rows": 6 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5.2-2025-12-11/257d9f58-e21e-4b10-9a76-03bd0f0157a3_samples.jsonl b/data/hle/openai/gpt-5.2-2025-12-11/257d9f58-e21e-4b10-9a76-03bd0f0157a3_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..3a0423935c22a439e2062eeea96be9631c7bfaf7 --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/257d9f58-e21e-4b10-9a76-03bd0f0157a3_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:991b6a78e9b193b03c9e21ca69aee80d9820a43cc1293b2f0c03e3982929f7b5 +size 123124 diff --git a/data/hle/openai/gpt-5.2-2025-12-11/283fc7f6-63a4-4018-8984-e07568cfe63b.json b/data/hle/openai/gpt-5.2-2025-12-11/283fc7f6-63a4-4018-8984-e07568cfe63b.json new file mode 100644 index 0000000000000000000000000000000000000000..a4bc090f7ecf7640b814391cce05ec39e2b597b7 --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/283fc7f6-63a4-4018-8984-e07568cfe63b.json @@ -0,0 +1,114 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5.2-2025-12-11/1786164051.613399", + "evaluation_timestamp": "1782141731.0", + "retrieved_timestamp": "1786164051.613399", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "mean on hle/S-adaptive/+1ep/d4ee5c21 for scorer _llm_grader", + "source_data": { + "dataset_name": "hle", + "source_type": "hf_dataset", + "hf_repo": "cais/hle", + "samples_number": 3, + "sample_ids": [ + "hle_d7b7c3dc", + "hle_1dc141bb", + "hle_c2ff3a43" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1782141731.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 1.0, + "uncertainty": { + "num_samples": 2 + }, + "details": { + "total_matched_trajectories": "2", + "stopping_reason_count_repetition_guard": "2", + "avg_total_tokens_target_model": "53582.50", + "total_total_tokens_target_model": "107165.0", + "avg_total_tokens_other_models": "3407.50", + "total_total_tokens_other_models": "6815.0", + "avg_total_tokens_all_models": "56990.00", + "total_total_tokens_all_models": "113980.0", + "avg_turn_count": "5.00", + "total_turn_count": "10.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "16", + "max_tokens": "20000", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 5000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5.2-2025-12-11/283fc7f6-63a4-4018-8984-e07568cfe63b_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "1c8aa2346e8b4bf7dd68825e0f3a6918939192d0ed6c77019bf9ffad80fada7d", + "total_rows": 2 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5.2-2025-12-11/283fc7f6-63a4-4018-8984-e07568cfe63b_samples.jsonl b/data/hle/openai/gpt-5.2-2025-12-11/283fc7f6-63a4-4018-8984-e07568cfe63b_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..3ad246b54f40d8958d69a35d98f7c18879268f16 --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/283fc7f6-63a4-4018-8984-e07568cfe63b_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1c8aa2346e8b4bf7dd68825e0f3a6918939192d0ed6c77019bf9ffad80fada7d +size 35869 diff --git a/data/hle/openai/gpt-5.2-2025-12-11/32ce09e1-5209-4b8f-89ed-6192eafeda93.json b/data/hle/openai/gpt-5.2-2025-12-11/32ce09e1-5209-4b8f-89ed-6192eafeda93.json new file mode 100644 index 0000000000000000000000000000000000000000..f37be931e005f47e49c47010f856220cdcc313cf --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/32ce09e1-5209-4b8f-89ed-6192eafeda93.json @@ -0,0 +1,110 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5.2-2025-12-11/1786164047.600795", + "evaluation_timestamp": "1776803719.0", + "retrieved_timestamp": "1786164047.600795", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "mean on hle/S-adaptive/+1ep/9fd3632d for scorer _llm_grader", + "source_data": { + "dataset_name": "hle", + "source_type": "hf_dataset", + "hf_repo": "cais/hle", + "samples_number": 3, + "sample_ids": [ + "hle_ffb12d83", + "hle_37c97863", + "hle_0702accc" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1776803719.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.0, + "uncertainty": { + "num_samples": 3 + }, + "details": { + "total_matched_trajectories": "3", + "stopping_reason_count_repetition_guard": "3", + "avg_total_tokens_target_model": "22683.67", + "total_total_tokens_target_model": "68051.0", + "avg_total_tokens_other_models": "5197.67", + "total_total_tokens_other_models": "15593.0", + "avg_total_tokens_all_models": "27881.33", + "total_total_tokens_all_models": "83644.0", + "avg_turn_count": "5.00", + "total_turn_count": "15.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 5000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5.2-2025-12-11/32ce09e1-5209-4b8f-89ed-6192eafeda93_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "95c27cd071647b495001975f3a57785913e4c418a49553bd108e0a2894373f8a", + "total_rows": 3 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5.2-2025-12-11/32ce09e1-5209-4b8f-89ed-6192eafeda93_samples.jsonl b/data/hle/openai/gpt-5.2-2025-12-11/32ce09e1-5209-4b8f-89ed-6192eafeda93_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..89c477705140162dc62f057c5f4b755705b19a0d --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/32ce09e1-5209-4b8f-89ed-6192eafeda93_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:95c27cd071647b495001975f3a57785913e4c418a49553bd108e0a2894373f8a +size 59843 diff --git a/data/hle/openai/gpt-5.2-2025-12-11/342c41fe-a520-4b0f-8721-c5dde1228b3f.json b/data/hle/openai/gpt-5.2-2025-12-11/342c41fe-a520-4b0f-8721-c5dde1228b3f.json new file mode 100644 index 0000000000000000000000000000000000000000..0e8f93ba738207fada4add694f6ec3dacf470151 --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/342c41fe-a520-4b0f-8721-c5dde1228b3f.json @@ -0,0 +1,117 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5.2-2025-12-11/1786164132.404227", + "evaluation_timestamp": "1778975483.0", + "retrieved_timestamp": "1786164132.404227", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "mean on hle/S-adaptive/+4ep/455b48e6 for scorer _llm_grader", + "source_data": { + "dataset_name": "hle", + "source_type": "hf_dataset", + "hf_repo": "cais/hle", + "samples_number": 3, + "sample_ids": [ + "hle_a764505b", + "hle_e62ba103", + "hle_5f04b314" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1778975483.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.5, + "uncertainty": { + "standard_error": { + "value": 0.2886751345948129 + }, + "num_samples": 12 + }, + "details": { + "total_matched_trajectories": "12", + "stopping_reason_count_repetition_guard": "12", + "avg_total_tokens_target_model": "15880.50", + "total_total_tokens_target_model": "190566.0", + "avg_total_tokens_other_models": "3616.00", + "total_total_tokens_other_models": "43392.0", + "avg_total_tokens_all_models": "19496.50", + "total_total_tokens_all_models": "233958.0", + "avg_turn_count": "5.75", + "total_turn_count": "69.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "8", + "max_tokens": "32000", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 5000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "8", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5.2-2025-12-11/342c41fe-a520-4b0f-8721-c5dde1228b3f_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "91af9a3ddb16d3974e026538f68a4196f547c94a4292f05b1d9ebb632884f11a", + "total_rows": 12 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5.2-2025-12-11/342c41fe-a520-4b0f-8721-c5dde1228b3f_samples.jsonl b/data/hle/openai/gpt-5.2-2025-12-11/342c41fe-a520-4b0f-8721-c5dde1228b3f_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..155e18411b3df6ed98ba37f02305da900061bfde --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/342c41fe-a520-4b0f-8721-c5dde1228b3f_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:91af9a3ddb16d3974e026538f68a4196f547c94a4292f05b1d9ebb632884f11a +size 235621 diff --git a/data/hle/openai/gpt-5.2-2025-12-11/3c29096a-97d3-4895-bb2f-57731888b7e7.json b/data/hle/openai/gpt-5.2-2025-12-11/3c29096a-97d3-4895-bb2f-57731888b7e7.json new file mode 100644 index 0000000000000000000000000000000000000000..d19b75e3ca79a060676d1efc31ebe1c0be857021 --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/3c29096a-97d3-4895-bb2f-57731888b7e7.json @@ -0,0 +1,112 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5.2-2025-12-11/1786164083.451354", + "evaluation_timestamp": "1783341847.0", + "retrieved_timestamp": "1786164083.451354", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "mean on hle/S-adaptive/+2ep/9b8beab9 for scorer _llm_grader", + "source_data": { + "dataset_name": "hle", + "source_type": "hf_dataset", + "hf_repo": "cais/hle", + "samples_number": 1, + "sample_ids": [ + "hle_7808792d" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1783341847.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.5, + "uncertainty": { + "num_samples": 2 + }, + "details": { + "total_matched_trajectories": "2", + "stopping_reason_count_repetition_guard": "2", + "avg_total_tokens_target_model": "343870.00", + "total_total_tokens_target_model": "687740.0", + "avg_total_tokens_other_models": "15005.50", + "total_total_tokens_other_models": "30011.0", + "avg_total_tokens_all_models": "358875.50", + "total_total_tokens_all_models": "717751.0", + "avg_turn_count": "9.00", + "total_turn_count": "18.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "1", + "timeout": "1200", + "attempt_timeout": "3600", + "max_connections": "16", + "max_tokens": "64000", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 5000000 + }, + "sandbox": {}, + "max_attempts": 1 + }, + "additional_details": { + "max_retries": "1", + "timeout": "1200", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5.2-2025-12-11/3c29096a-97d3-4895-bb2f-57731888b7e7_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "f5cdb57a90f548b9ad847f331c2ad38c3b75138d8def98c18c5bbe5cc1e5782c", + "total_rows": 2 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5.2-2025-12-11/3c29096a-97d3-4895-bb2f-57731888b7e7_samples.jsonl b/data/hle/openai/gpt-5.2-2025-12-11/3c29096a-97d3-4895-bb2f-57731888b7e7_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..92d3853a2d099d07a8a106388f3aa7bcf663bd10 --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/3c29096a-97d3-4895-bb2f-57731888b7e7_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f5cdb57a90f548b9ad847f331c2ad38c3b75138d8def98c18c5bbe5cc1e5782c +size 68069 diff --git a/data/hle/openai/gpt-5.2-2025-12-11/53eae0f0-30b8-4225-8e68-465c40b9094f.json b/data/hle/openai/gpt-5.2-2025-12-11/53eae0f0-30b8-4225-8e68-465c40b9094f.json new file mode 100644 index 0000000000000000000000000000000000000000..96f7ab559449944266c4921009d57c43ed36149a --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/53eae0f0-30b8-4225-8e68-465c40b9094f.json @@ -0,0 +1,118 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5.2-2025-12-11/1786164111.107575", + "evaluation_timestamp": "1776430851.0", + "retrieved_timestamp": "1786164111.107575", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "mean on hle/S-adaptive/+4ep/b51cca68 for scorer _llm_grader", + "source_data": { + "dataset_name": "hle", + "source_type": "hf_dataset", + "hf_repo": "cais/hle", + "samples_number": 10, + "sample_ids": [ + "hle_d6b73839", + "hle_07662c9c", + "hle_4e3440f7", + "hle_baccd6da", + "hle_f46c2857", + "hle_86a18059", + "hle_8770ab31", + "hle_d1bd2767", + "hle_6499b500", + "hle_e1296346" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1776430851.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.3, + "uncertainty": { + "standard_error": { + "value": 0.13844373104863458 + }, + "num_samples": 40 + }, + "details": { + "total_matched_trajectories": "40", + "stopping_reason_count_repetition_guard": "40", + "avg_total_tokens_target_model": "22651.97", + "total_total_tokens_target_model": "906079.0", + "avg_total_tokens_other_models": "4597.23", + "total_total_tokens_other_models": "183889.0", + "avg_total_tokens_all_models": "27249.20", + "total_total_tokens_all_models": "1089968.0", + "avg_turn_count": "5.20", + "total_turn_count": "208.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 5000000 + }, + "sandbox": {} + }, + "additional_details": { + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5.2-2025-12-11/53eae0f0-30b8-4225-8e68-465c40b9094f_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "cef9eae96f38a4f155cd1391f4ba0ae5d04703bc0739383cd1d326e142b6d229", + "total_rows": 40 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5.2-2025-12-11/53eae0f0-30b8-4225-8e68-465c40b9094f_samples.jsonl b/data/hle/openai/gpt-5.2-2025-12-11/53eae0f0-30b8-4225-8e68-465c40b9094f_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..cb7554447610d5eff35eed8948fc4e505d853723 --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/53eae0f0-30b8-4225-8e68-465c40b9094f_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cef9eae96f38a4f155cd1391f4ba0ae5d04703bc0739383cd1d326e142b6d229 +size 789107 diff --git a/data/hle/openai/gpt-5.2-2025-12-11/5bb53ab6-da71-484d-b6a6-b8ef60eacd13.json b/data/hle/openai/gpt-5.2-2025-12-11/5bb53ab6-da71-484d-b6a6-b8ef60eacd13.json new file mode 100644 index 0000000000000000000000000000000000000000..f809eb854b59adfdf5137dfd5044b94069816c0e --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/5bb53ab6-da71-484d-b6a6-b8ef60eacd13.json @@ -0,0 +1,114 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5.2-2025-12-11/1786164049.538575", + "evaluation_timestamp": "1781905453.0", + "retrieved_timestamp": "1786164049.538575", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "mean on hle/S-adaptive/+1ep/cad52af3 for scorer _llm_grader", + "source_data": { + "dataset_name": "hle", + "source_type": "hf_dataset", + "hf_repo": "cais/hle", + "samples_number": 3, + "sample_ids": [ + "hle_1d580261", + "hle_b26bf8aa", + "hle_c2ff3a43" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1781905453.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.0, + "uncertainty": { + "num_samples": 2 + }, + "details": { + "total_matched_trajectories": "2", + "stopping_reason_count_repetition_guard": "2", + "avg_total_tokens_target_model": "37353.00", + "total_total_tokens_target_model": "74706.0", + "avg_total_tokens_other_models": "5500.50", + "total_total_tokens_other_models": "11001.0", + "avg_total_tokens_all_models": "42853.50", + "total_total_tokens_all_models": "85707.0", + "avg_turn_count": "5.00", + "total_turn_count": "10.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "16", + "max_tokens": "20000", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 5000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5.2-2025-12-11/5bb53ab6-da71-484d-b6a6-b8ef60eacd13_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "7aaf1bb0213915c6263a4f0d61cf9240f65efec34673dc8f4c7993f9bc04580a", + "total_rows": 2 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5.2-2025-12-11/5bb53ab6-da71-484d-b6a6-b8ef60eacd13_samples.jsonl b/data/hle/openai/gpt-5.2-2025-12-11/5bb53ab6-da71-484d-b6a6-b8ef60eacd13_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..9d49aad2fd428980c2b0c22c838180f12fb80eb3 --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/5bb53ab6-da71-484d-b6a6-b8ef60eacd13_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7aaf1bb0213915c6263a4f0d61cf9240f65efec34673dc8f4c7993f9bc04580a +size 40450 diff --git a/data/hle/openai/gpt-5.2-2025-12-11/5c517235-3a89-449b-b1a4-aacc93ee30c3.json b/data/hle/openai/gpt-5.2-2025-12-11/5c517235-3a89-449b-b1a4-aacc93ee30c3.json new file mode 100644 index 0000000000000000000000000000000000000000..b20daccd125afbaab404d902c4f7aecb72e977d5 --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/5c517235-3a89-449b-b1a4-aacc93ee30c3.json @@ -0,0 +1,112 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5.2-2025-12-11/1786164095.41234", + "evaluation_timestamp": "1780480226.0", + "retrieved_timestamp": "1786164095.41234", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "mean on hle/S-adaptive/+3ep/a97199c8 for scorer _llm_grader", + "source_data": { + "dataset_name": "hle", + "source_type": "hf_dataset", + "hf_repo": "cais/hle", + "samples_number": 1, + "sample_ids": [ + "hle_70e5aebc" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1780480226.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.0, + "uncertainty": { + "num_samples": 1 + }, + "details": { + "total_matched_trajectories": "1", + "stopping_reason_count_repetition_guard": "1", + "avg_total_tokens_target_model": "30256.00", + "total_total_tokens_target_model": "30256.0", + "avg_total_tokens_other_models": "3429.00", + "total_total_tokens_other_models": "3429.0", + "avg_total_tokens_all_models": "33685.00", + "total_total_tokens_all_models": "33685.0", + "avg_turn_count": "5.00", + "total_turn_count": "5.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "3", + "timeout": "1800", + "attempt_timeout": "1800", + "max_connections": "16", + "max_tokens": "20000", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 5000000 + }, + "sandbox": {}, + "max_attempts": 3 + }, + "additional_details": { + "max_retries": "3", + "timeout": "1800", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5.2-2025-12-11/5c517235-3a89-449b-b1a4-aacc93ee30c3_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "0c2622023a0aa99c8ffd0eddfc11fcb6044ff3fa4fcb3f02f4e6a6db254aff18", + "total_rows": 1 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5.2-2025-12-11/5c517235-3a89-449b-b1a4-aacc93ee30c3_samples.jsonl b/data/hle/openai/gpt-5.2-2025-12-11/5c517235-3a89-449b-b1a4-aacc93ee30c3_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..95fdac60a55d28607d7e34d781188fcc14b178f2 --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/5c517235-3a89-449b-b1a4-aacc93ee30c3_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0c2622023a0aa99c8ffd0eddfc11fcb6044ff3fa4fcb3f02f4e6a6db254aff18 +size 19018 diff --git a/data/hle/openai/gpt-5.2-2025-12-11/5f5732b4-bc7e-450c-bc8a-35cad8a10a8e.json b/data/hle/openai/gpt-5.2-2025-12-11/5f5732b4-bc7e-450c-bc8a-35cad8a10a8e.json new file mode 100644 index 0000000000000000000000000000000000000000..08d47896dc9e5e3f72dc706174b1ae26d2cc13cd --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/5f5732b4-bc7e-450c-bc8a-35cad8a10a8e.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5.2-2025-12-11/1786168717.333592", + "evaluation_timestamp": "1779017113.0", + "retrieved_timestamp": "1786168717.333592", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5.2-2025-12-11/5f5732b4-bc7e-450c-bc8a-35cad8a10a8e_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "eaee15dedc36d8b909b3845115b3c6e1a10fbb70607b37d6ef6ae021d1303343", + "total_rows": 2 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5.2-2025-12-11/5f5732b4-bc7e-450c-bc8a-35cad8a10a8e_samples.jsonl b/data/hle/openai/gpt-5.2-2025-12-11/5f5732b4-bc7e-450c-bc8a-35cad8a10a8e_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..80499c3189bacc90c20cc17e7637c73cf99dff41 --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/5f5732b4-bc7e-450c-bc8a-35cad8a10a8e_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:eaee15dedc36d8b909b3845115b3c6e1a10fbb70607b37d6ef6ae021d1303343 +size 38005 diff --git a/data/hle/openai/gpt-5.2-2025-12-11/6651f842-2a86-4346-8ed9-3fbbe8eacfc8.json b/data/hle/openai/gpt-5.2-2025-12-11/6651f842-2a86-4346-8ed9-3fbbe8eacfc8.json new file mode 100644 index 0000000000000000000000000000000000000000..4bc2f1fdaecc009bc4fa63636f2b923e05f15a34 --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/6651f842-2a86-4346-8ed9-3fbbe8eacfc8.json @@ -0,0 +1,115 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5.2-2025-12-11/1786164100.247041", + "evaluation_timestamp": "1778534692.0", + "retrieved_timestamp": "1786164100.247041", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "mean on hle/S-adaptive/+3ep/af1d3e97 for scorer _llm_grader", + "source_data": { + "dataset_name": "hle", + "source_type": "hf_dataset", + "hf_repo": "cais/hle", + "samples_number": 2, + "sample_ids": [ + "hle_ca6c436b", + "hle_28643173" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1778534692.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.5, + "uncertainty": { + "standard_error": { + "value": 0.5 + }, + "num_samples": 2 + }, + "details": { + "total_matched_trajectories": "2", + "stopping_reason_count_repetition_guard": "2", + "avg_total_tokens_target_model": "45416.50", + "total_total_tokens_target_model": "90833.0", + "avg_total_tokens_other_models": "7595.00", + "total_total_tokens_other_models": "15190.0", + "avg_total_tokens_all_models": "53011.50", + "total_total_tokens_all_models": "106023.0", + "avg_turn_count": "5.00", + "total_turn_count": "10.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 5000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5.2-2025-12-11/6651f842-2a86-4346-8ed9-3fbbe8eacfc8_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "8e2d1fe3b5f6d0ca094ae93188c0a82fe0219e3c7121be1c85f5b8d868a04188", + "total_rows": 2 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5.2-2025-12-11/6651f842-2a86-4346-8ed9-3fbbe8eacfc8_samples.jsonl b/data/hle/openai/gpt-5.2-2025-12-11/6651f842-2a86-4346-8ed9-3fbbe8eacfc8_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..a26e741f86e0819ca68ef22df96bd575ddbb4f56 --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/6651f842-2a86-4346-8ed9-3fbbe8eacfc8_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8e2d1fe3b5f6d0ca094ae93188c0a82fe0219e3c7121be1c85f5b8d868a04188 +size 48587 diff --git a/data/hle/openai/gpt-5.2-2025-12-11/682a9a3f-ea6d-4897-a931-4f5141da5b5e.json b/data/hle/openai/gpt-5.2-2025-12-11/682a9a3f-ea6d-4897-a931-4f5141da5b5e.json new file mode 100644 index 0000000000000000000000000000000000000000..8dfaa51a659d992d73e5591e8d09b8a6d61ccf90 --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/682a9a3f-ea6d-4897-a931-4f5141da5b5e.json @@ -0,0 +1,112 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5.2-2025-12-11/1786164059.370464", + "evaluation_timestamp": "1782153624.0", + "retrieved_timestamp": "1786164059.370464", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "mean on hle/S-adaptive/+2ep/6214441b for scorer _llm_grader", + "source_data": { + "dataset_name": "hle", + "source_type": "hf_dataset", + "hf_repo": "cais/hle", + "samples_number": 1, + "sample_ids": [ + "hle_34c69126" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1782153624.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 1.0, + "uncertainty": { + "num_samples": 1 + }, + "details": { + "total_matched_trajectories": "1", + "stopping_reason_count_repetition_guard": "1", + "avg_total_tokens_target_model": "35474.00", + "total_total_tokens_target_model": "35474.0", + "avg_total_tokens_other_models": "7684.00", + "total_total_tokens_other_models": "7684.0", + "avg_total_tokens_all_models": "43158.00", + "total_total_tokens_all_models": "43158.0", + "avg_turn_count": "5.00", + "total_turn_count": "5.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "16", + "max_tokens": "20000", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 5000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5.2-2025-12-11/682a9a3f-ea6d-4897-a931-4f5141da5b5e_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "2ed405e96f0fc83a3e7bf6d99f1b79af268e8335da760e2a69d902e8703d62fa", + "total_rows": 1 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5.2-2025-12-11/682a9a3f-ea6d-4897-a931-4f5141da5b5e_samples.jsonl b/data/hle/openai/gpt-5.2-2025-12-11/682a9a3f-ea6d-4897-a931-4f5141da5b5e_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..561588b309cc409bdacec31a5c75e6a38ac9c30d --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/682a9a3f-ea6d-4897-a931-4f5141da5b5e_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2ed405e96f0fc83a3e7bf6d99f1b79af268e8335da760e2a69d902e8703d62fa +size 21951 diff --git a/data/hle/openai/gpt-5.2-2025-12-11/6d430045-9b60-402d-a155-3c7afec9fc7d.json b/data/hle/openai/gpt-5.2-2025-12-11/6d430045-9b60-402d-a155-3c7afec9fc7d.json new file mode 100644 index 0000000000000000000000000000000000000000..59589adbc06483a6fb25e314a8f7a76aa7edbfe0 --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/6d430045-9b60-402d-a155-3c7afec9fc7d.json @@ -0,0 +1,113 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5.2-2025-12-11/1786164045.30817", + "evaluation_timestamp": "1781614424.0", + "retrieved_timestamp": "1786164045.30817", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "mean on hle/S-adaptive/+1ep/12554aeb for scorer _llm_grader", + "source_data": { + "dataset_name": "hle", + "source_type": "hf_dataset", + "hf_repo": "cais/hle", + "samples_number": 2, + "sample_ids": [ + "hle_ee938b26", + "hle_b26bf8aa" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1781614424.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.0, + "uncertainty": { + "num_samples": 1 + }, + "details": { + "total_matched_trajectories": "1", + "stopping_reason_count_repetition_guard": "1", + "avg_total_tokens_target_model": "31870.00", + "total_total_tokens_target_model": "31870.0", + "avg_total_tokens_other_models": "5896.00", + "total_total_tokens_other_models": "5896.0", + "avg_total_tokens_all_models": "37766.00", + "total_total_tokens_all_models": "37766.0", + "avg_turn_count": "5.00", + "total_turn_count": "5.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "16", + "max_tokens": "20000", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 5000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5.2-2025-12-11/6d430045-9b60-402d-a155-3c7afec9fc7d_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "a2fd9c76c4b3d513ad182cccae299e7ed0ec0e8693cb550a7cb2101073a20ec6", + "total_rows": 1 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5.2-2025-12-11/6d430045-9b60-402d-a155-3c7afec9fc7d_samples.jsonl b/data/hle/openai/gpt-5.2-2025-12-11/6d430045-9b60-402d-a155-3c7afec9fc7d_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..9f06629559669e3f732e779fbe37a05c83a754ab --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/6d430045-9b60-402d-a155-3c7afec9fc7d_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a2fd9c76c4b3d513ad182cccae299e7ed0ec0e8693cb550a7cb2101073a20ec6 +size 19669 diff --git a/data/hle/openai/gpt-5.2-2025-12-11/7320b37a-a23a-4306-882c-2a6b29300bf6.json b/data/hle/openai/gpt-5.2-2025-12-11/7320b37a-a23a-4306-882c-2a6b29300bf6.json new file mode 100644 index 0000000000000000000000000000000000000000..a9259f44e555d9cc931e0f6151b4e342abbc2a9b --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/7320b37a-a23a-4306-882c-2a6b29300bf6.json @@ -0,0 +1,108 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5.2-2025-12-11/1786164057.864049", + "evaluation_timestamp": "1777025785.0", + "retrieved_timestamp": "1786164057.864049", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "mean on hle/S-adaptive/+2ep/3fee0a63 for scorer _llm_grader", + "source_data": { + "dataset_name": "hle", + "source_type": "hf_dataset", + "hf_repo": "cais/hle", + "samples_number": 1, + "sample_ids": [ + "hle_25e0c71d" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1777025785.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.5, + "uncertainty": { + "num_samples": 2 + }, + "details": { + "total_matched_trajectories": "2", + "stopping_reason_count_repetition_guard": "2", + "avg_total_tokens_target_model": "25650.50", + "total_total_tokens_target_model": "51301.0", + "avg_total_tokens_other_models": "3719.00", + "total_total_tokens_other_models": "7438.0", + "avg_total_tokens_all_models": "29369.50", + "total_total_tokens_all_models": "58739.0", + "avg_turn_count": "5.00", + "total_turn_count": "10.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 5000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5.2-2025-12-11/7320b37a-a23a-4306-882c-2a6b29300bf6_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "c98b2ff26ed4768590db42ea39dae53cf00c88beeaaaea51c8389add1302d029", + "total_rows": 2 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5.2-2025-12-11/7320b37a-a23a-4306-882c-2a6b29300bf6_samples.jsonl b/data/hle/openai/gpt-5.2-2025-12-11/7320b37a-a23a-4306-882c-2a6b29300bf6_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..e3f115a89ee771a8641321c75794c90c8308956b --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/7320b37a-a23a-4306-882c-2a6b29300bf6_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c98b2ff26ed4768590db42ea39dae53cf00c88beeaaaea51c8389add1302d029 +size 38049 diff --git a/data/hle/openai/gpt-5.2-2025-12-11/799fb632-24db-43c2-a649-5f2cd71b1e3f.json b/data/hle/openai/gpt-5.2-2025-12-11/799fb632-24db-43c2-a649-5f2cd71b1e3f.json new file mode 100644 index 0000000000000000000000000000000000000000..907082104b3faec7e9393e4babf7297eba345634 --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/799fb632-24db-43c2-a649-5f2cd71b1e3f.json @@ -0,0 +1,112 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5.2-2025-12-11/1786164081.87326", + "evaluation_timestamp": "1782222650.0", + "retrieved_timestamp": "1786164081.87326", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "mean on hle/S-adaptive/+2ep/6214441b for scorer _llm_grader", + "source_data": { + "dataset_name": "hle", + "source_type": "hf_dataset", + "hf_repo": "cais/hle", + "samples_number": 1, + "sample_ids": [ + "hle_34c69126" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1782222650.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 1.0, + "uncertainty": { + "num_samples": 2 + }, + "details": { + "total_matched_trajectories": "1", + "stopping_reason_count_repetition_guard": "1", + "avg_total_tokens_target_model": "197453.00", + "total_total_tokens_target_model": "197453.0", + "avg_total_tokens_other_models": "6209.00", + "total_total_tokens_other_models": "6209.0", + "avg_total_tokens_all_models": "203662.00", + "total_total_tokens_all_models": "203662.0", + "avg_turn_count": "12.00", + "total_turn_count": "12.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "1200", + "max_connections": "16", + "max_tokens": "20000", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 5000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5.2-2025-12-11/799fb632-24db-43c2-a649-5f2cd71b1e3f_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "24bdf83d39b71c1d8e0760ccae02917582faa23e1ec46b07bf42add2d17873b1", + "total_rows": 2 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5.2-2025-12-11/799fb632-24db-43c2-a649-5f2cd71b1e3f_samples.jsonl b/data/hle/openai/gpt-5.2-2025-12-11/799fb632-24db-43c2-a649-5f2cd71b1e3f_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..ffadc273f812d3bd9225e016e458f0b193121762 --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/799fb632-24db-43c2-a649-5f2cd71b1e3f_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:24bdf83d39b71c1d8e0760ccae02917582faa23e1ec46b07bf42add2d17873b1 +size 45469 diff --git a/data/hle/openai/gpt-5.2-2025-12-11/7ee222aa-4c84-4151-af07-67fd75dcf0a0.json b/data/hle/openai/gpt-5.2-2025-12-11/7ee222aa-4c84-4151-af07-67fd75dcf0a0.json new file mode 100644 index 0000000000000000000000000000000000000000..3a2e3a54240f6108fae613555248d49a50fd1645 --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/7ee222aa-4c84-4151-af07-67fd75dcf0a0.json @@ -0,0 +1,112 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5.2-2025-12-11/1786164085.439402", + "evaluation_timestamp": "1776804373.0", + "retrieved_timestamp": "1786164085.439402", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "mean on hle/S-adaptive/+2ep/7aba9eed for scorer _llm_grader", + "source_data": { + "dataset_name": "hle", + "source_type": "hf_dataset", + "hf_repo": "cais/hle", + "samples_number": 2, + "sample_ids": [ + "hle_76cdecee", + "hle_142ffcff" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1776804373.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.5, + "uncertainty": { + "standard_error": { + "value": 0.5 + }, + "num_samples": 3 + }, + "details": { + "total_matched_trajectories": "3", + "stopping_reason_count_repetition_guard": "3", + "avg_total_tokens_target_model": "21683.00", + "total_total_tokens_target_model": "65049.0", + "avg_total_tokens_other_models": "3523.33", + "total_total_tokens_other_models": "10570.0", + "avg_total_tokens_all_models": "25206.33", + "total_total_tokens_all_models": "75619.0", + "avg_turn_count": "5.00", + "total_turn_count": "15.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 5000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5.2-2025-12-11/7ee222aa-4c84-4151-af07-67fd75dcf0a0_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "ba803c14fc55c6dd8de746674c9c280839b43c8494fab5cb63fd52e9998d4e1f", + "total_rows": 3 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5.2-2025-12-11/7ee222aa-4c84-4151-af07-67fd75dcf0a0_samples.jsonl b/data/hle/openai/gpt-5.2-2025-12-11/7ee222aa-4c84-4151-af07-67fd75dcf0a0_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..c907637c38550954d59c229152e26254cbe5d309 --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/7ee222aa-4c84-4151-af07-67fd75dcf0a0_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ba803c14fc55c6dd8de746674c9c280839b43c8494fab5cb63fd52e9998d4e1f +size 53387 diff --git a/data/hle/openai/gpt-5.2-2025-12-11/819f4fac-f465-4c50-b0e8-a1110d1afceb.json b/data/hle/openai/gpt-5.2-2025-12-11/819f4fac-f465-4c50-b0e8-a1110d1afceb.json new file mode 100644 index 0000000000000000000000000000000000000000..2119511c75ab75f046c8453a66de6ef1a98a8b3d --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/819f4fac-f465-4c50-b0e8-a1110d1afceb.json @@ -0,0 +1,125 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5.2-2025-12-11/1786164113.300613", + "evaluation_timestamp": "1776686674.0", + "retrieved_timestamp": "1786164113.300613", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "mean on hle/S-adaptive/+4ep/4db80b19 for scorer _llm_grader", + "source_data": { + "dataset_name": "hle", + "source_type": "hf_dataset", + "hf_repo": "cais/hle", + "samples_number": 15, + "sample_ids": [ + "hle_764ce81c", + "hle_ca6c436b", + "hle_a559277e", + "hle_8a419425", + "hle_85e95ef4", + "hle_28643173", + "hle_bc5fbce2", + "hle_25e0c71d", + "hle_76cdecee", + "hle_dfdb78f5", + "hle_142ffcff", + "hle_ffb12d83", + "hle_37c97863", + "hle_209b6b64", + "hle_0702accc" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1776686674.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.37037037037037035, + "uncertainty": { + "standard_error": { + "value": 0.14441179593039094 + }, + "num_samples": 25 + }, + "details": { + "total_matched_trajectories": "25", + "stopping_reason_count_repetition_guard": "25", + "avg_total_tokens_target_model": "23905.92", + "total_total_tokens_target_model": "597648.0", + "avg_total_tokens_other_models": "4427.16", + "total_total_tokens_other_models": "110679.0", + "avg_total_tokens_all_models": "28333.08", + "total_total_tokens_all_models": "708327.0", + "avg_turn_count": "5.00", + "total_turn_count": "125.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "3600", + "attempt_timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 5000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5.2-2025-12-11/819f4fac-f465-4c50-b0e8-a1110d1afceb_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "593b36d9accf0a8137b3f14db7c25ba7dfdb1fcf5bef4f6ecd87dd0a761a18ee", + "total_rows": 25 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5.2-2025-12-11/819f4fac-f465-4c50-b0e8-a1110d1afceb_samples.jsonl b/data/hle/openai/gpt-5.2-2025-12-11/819f4fac-f465-4c50-b0e8-a1110d1afceb_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..06c67f5c5dfe5a07d91df41b897d596187ca47c5 --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/819f4fac-f465-4c50-b0e8-a1110d1afceb_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:593b36d9accf0a8137b3f14db7c25ba7dfdb1fcf5bef4f6ecd87dd0a761a18ee +size 479514 diff --git a/data/hle/openai/gpt-5.2-2025-12-11/8272a09f-2e4e-4f59-8d79-f59aabe799f3.json b/data/hle/openai/gpt-5.2-2025-12-11/8272a09f-2e4e-4f59-8d79-f59aabe799f3.json new file mode 100644 index 0000000000000000000000000000000000000000..dcc98ba8458e1c3a56ca290949dc4696c8c682eb --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/8272a09f-2e4e-4f59-8d79-f59aabe799f3.json @@ -0,0 +1,112 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5.2-2025-12-11/1786164036.202895", + "evaluation_timestamp": "1779017113.0", + "retrieved_timestamp": "1786164036.202895", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "mean on hle/S-adaptive/+1ep/2364ca4e for scorer _llm_grader", + "source_data": { + "dataset_name": "hle", + "source_type": "hf_dataset", + "hf_repo": "cais/hle", + "samples_number": 1, + "sample_ids": [ + "hle_28643173" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1779017113.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.0, + "uncertainty": { + "num_samples": 1 + }, + "details": { + "total_matched_trajectories": "1", + "stopping_reason_count_repetition_guard": "1", + "avg_total_tokens_target_model": "55267.00", + "total_total_tokens_target_model": "55267.0", + "avg_total_tokens_other_models": "5536.00", + "total_total_tokens_other_models": "5536.0", + "avg_total_tokens_all_models": "60803.00", + "total_total_tokens_all_models": "60803.0", + "avg_turn_count": "5.00", + "total_turn_count": "5.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "8", + "max_tokens": "32000", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 5000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "8", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5.2-2025-12-11/8272a09f-2e4e-4f59-8d79-f59aabe799f3_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "d48cdd91dabc1b36bee3dabced4113ef14e5b84b5a9a0010f8d357684d918b17", + "total_rows": 1 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5.2-2025-12-11/8272a09f-2e4e-4f59-8d79-f59aabe799f3_samples.jsonl b/data/hle/openai/gpt-5.2-2025-12-11/8272a09f-2e4e-4f59-8d79-f59aabe799f3_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..f2189763b847e1452b2aa7de11662d783f579901 --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/8272a09f-2e4e-4f59-8d79-f59aabe799f3_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d48cdd91dabc1b36bee3dabced4113ef14e5b84b5a9a0010f8d357684d918b17 +size 20095 diff --git a/data/hle/openai/gpt-5.2-2025-12-11/82a30859-ee78-4a2e-8512-dc55f1103927.json b/data/hle/openai/gpt-5.2-2025-12-11/82a30859-ee78-4a2e-8512-dc55f1103927.json new file mode 100644 index 0000000000000000000000000000000000000000..1d6a034f25820a4bf7aa0a5a57b28738308dfbb4 --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/82a30859-ee78-4a2e-8512-dc55f1103927.json @@ -0,0 +1,108 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5.2-2025-12-11/1786164107.706816", + "evaluation_timestamp": "1777042503.0", + "retrieved_timestamp": "1786164107.706816", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "mean on hle/S-adaptive/+4ep/63638b66 for scorer _llm_grader", + "source_data": { + "dataset_name": "hle", + "source_type": "hf_dataset", + "hf_repo": "cais/hle", + "samples_number": 1, + "sample_ids": [ + "hle_ca6c436b" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1777042503.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.0, + "uncertainty": { + "num_samples": 1 + }, + "details": { + "total_matched_trajectories": "1", + "stopping_reason_count_repetition_guard": "1", + "avg_total_tokens_target_model": "44500.00", + "total_total_tokens_target_model": "44500.0", + "avg_total_tokens_other_models": "7918.00", + "total_total_tokens_other_models": "7918.0", + "avg_total_tokens_all_models": "52418.00", + "total_total_tokens_all_models": "52418.0", + "avg_turn_count": "5.00", + "total_turn_count": "5.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 5000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5.2-2025-12-11/82a30859-ee78-4a2e-8512-dc55f1103927_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "77f9f49b03f40358b0a760db4f0a6e3cd48e3b68a938545f71ca32e8a16b1b4d", + "total_rows": 1 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5.2-2025-12-11/82a30859-ee78-4a2e-8512-dc55f1103927_samples.jsonl b/data/hle/openai/gpt-5.2-2025-12-11/82a30859-ee78-4a2e-8512-dc55f1103927_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..3e05bbb0065fc286760a3988675a48b77e9b4f22 --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/82a30859-ee78-4a2e-8512-dc55f1103927_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:77f9f49b03f40358b0a760db4f0a6e3cd48e3b68a938545f71ca32e8a16b1b4d +size 27087 diff --git a/data/hle/openai/gpt-5.2-2025-12-11/890af2c0-c2c3-413c-ab57-3c3c15209241.json b/data/hle/openai/gpt-5.2-2025-12-11/890af2c0-c2c3-413c-ab57-3c3c15209241.json new file mode 100644 index 0000000000000000000000000000000000000000..96b119f78ca4181ad7d1270dece29b607b9bba3c --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/890af2c0-c2c3-413c-ab57-3c3c15209241.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5.2-2025-12-11/1786166767.143271", + "evaluation_timestamp": "1776686674.0", + "retrieved_timestamp": "1786166767.143271", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5.2-2025-12-11/890af2c0-c2c3-413c-ab57-3c3c15209241_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "af363644154a5403e932f008dd50171cd9ac7792bb940c3798aa34a6bc2dedf7", + "total_rows": 13 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5.2-2025-12-11/890af2c0-c2c3-413c-ab57-3c3c15209241_samples.jsonl b/data/hle/openai/gpt-5.2-2025-12-11/890af2c0-c2c3-413c-ab57-3c3c15209241_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..211eba302a7730da44f67ce737c8b08e9a1d41b0 --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/890af2c0-c2c3-413c-ab57-3c3c15209241_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:af363644154a5403e932f008dd50171cd9ac7792bb940c3798aa34a6bc2dedf7 +size 338971 diff --git a/data/hle/openai/gpt-5.2-2025-12-11/8aa4afa7-a846-4545-b9ce-807aaf790411.json b/data/hle/openai/gpt-5.2-2025-12-11/8aa4afa7-a846-4545-b9ce-807aaf790411.json new file mode 100644 index 0000000000000000000000000000000000000000..538ff69b70dabf007370a2d5840545d1c1b9b2cc --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/8aa4afa7-a846-4545-b9ce-807aaf790411.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5.2-2025-12-11/1786166768.787765", + "evaluation_timestamp": "1776955657.0", + "retrieved_timestamp": "1786166768.787765", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5.2-2025-12-11/8aa4afa7-a846-4545-b9ce-807aaf790411_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "68b567e45c285011b7bdc0f87ec04f47c78a905faee91cbc60e616d9b6bfc283", + "total_rows": 1 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5.2-2025-12-11/8aa4afa7-a846-4545-b9ce-807aaf790411_samples.jsonl b/data/hle/openai/gpt-5.2-2025-12-11/8aa4afa7-a846-4545-b9ce-807aaf790411_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..8cf5175c62e7a1fad1636b7bc6c83107d4b69beb --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/8aa4afa7-a846-4545-b9ce-807aaf790411_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68b567e45c285011b7bdc0f87ec04f47c78a905faee91cbc60e616d9b6bfc283 +size 42558 diff --git a/data/hle/openai/gpt-5.2-2025-12-11/9adffbe7-566e-46b9-a894-cbb332091fe4.json b/data/hle/openai/gpt-5.2-2025-12-11/9adffbe7-566e-46b9-a894-cbb332091fe4.json new file mode 100644 index 0000000000000000000000000000000000000000..87850719e2c536f18274609ab0b9c5de50b1527a --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/9adffbe7-566e-46b9-a894-cbb332091fe4.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5.2-2025-12-11/1786166777.350533", + "evaluation_timestamp": "1783049380.0", + "retrieved_timestamp": "1786166777.350533", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5.2-2025-12-11/9adffbe7-566e-46b9-a894-cbb332091fe4_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "74adcc5cffa349f5f9ebaee61b98795571b18ffc9655f83a6fe6c197a226b6b0", + "total_rows": 4 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5.2-2025-12-11/9adffbe7-566e-46b9-a894-cbb332091fe4_samples.jsonl b/data/hle/openai/gpt-5.2-2025-12-11/9adffbe7-566e-46b9-a894-cbb332091fe4_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..988edc21968803c8f941bd1cd51433a195ba4220 --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/9adffbe7-566e-46b9-a894-cbb332091fe4_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:74adcc5cffa349f5f9ebaee61b98795571b18ffc9655f83a6fe6c197a226b6b0 +size 82981 diff --git a/data/hle/openai/gpt-5.2-2025-12-11/9ca8e7e1-5b31-43e1-b6b2-c7a5073c6c9e.json b/data/hle/openai/gpt-5.2-2025-12-11/9ca8e7e1-5b31-43e1-b6b2-c7a5073c6c9e.json new file mode 100644 index 0000000000000000000000000000000000000000..ab328ead5d33fb8a168c0b2c81ca561ef81b21aa --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/9ca8e7e1-5b31-43e1-b6b2-c7a5073c6c9e.json @@ -0,0 +1,129 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5.2-2025-12-11/1786164138.661505", + "evaluation_timestamp": "1782789480.0", + "retrieved_timestamp": "1786164138.661505", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "mean on hle/S-adaptive/56c9740c for scorer _llm_grader", + "source_data": { + "dataset_name": "hle", + "source_type": "hf_dataset", + "hf_repo": "cais/hle", + "samples_number": 15, + "sample_ids": [ + "hle_764ce81c", + "hle_bc5fbce2", + "hle_dfdb78f5", + "hle_209b6b64", + "hle_4485802f", + "hle_44007b97", + "hle_7808792d", + "hle_e3b8f5c2", + "hle_9df5d371", + "hle_9255683e", + "hle_fcd56401", + "hle_5d49423c", + "hle_141cd3ef", + "hle_0ec247c5", + "hle_45d1c140" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1782789480.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.1, + "uncertainty": { + "standard_error": { + "value": 0.10000000000000002 + }, + "num_samples": 12 + }, + "details": { + "total_matched_trajectories": "12", + "stopping_reason_count_repetition_guard": "12", + "avg_total_tokens_target_model": "41671.92", + "total_total_tokens_target_model": "500063.0", + "avg_total_tokens_other_models": "3869.67", + "total_total_tokens_other_models": "46436.0", + "avg_total_tokens_all_models": "45541.58", + "total_total_tokens_all_models": "546499.0", + "avg_turn_count": "8.00", + "total_turn_count": "96.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "1200", + "max_connections": "16", + "max_tokens": "20000", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 5000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5.2-2025-12-11/9ca8e7e1-5b31-43e1-b6b2-c7a5073c6c9e_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "d753c9c6ae82eeefbab90ee6617ff2b8b79053896e1683148767fa27e82be1aa", + "total_rows": 12 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5.2-2025-12-11/9ca8e7e1-5b31-43e1-b6b2-c7a5073c6c9e_samples.jsonl b/data/hle/openai/gpt-5.2-2025-12-11/9ca8e7e1-5b31-43e1-b6b2-c7a5073c6c9e_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..43eaad88a54aab2b677ac6858d762e06fda29a71 --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/9ca8e7e1-5b31-43e1-b6b2-c7a5073c6c9e_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d753c9c6ae82eeefbab90ee6617ff2b8b79053896e1683148767fa27e82be1aa +size 265332 diff --git a/data/hle/openai/gpt-5.2-2025-12-11/9d35ebf3-f9b8-4e85-b16c-fdf8b32f7c27.json b/data/hle/openai/gpt-5.2-2025-12-11/9d35ebf3-f9b8-4e85-b16c-fdf8b32f7c27.json new file mode 100644 index 0000000000000000000000000000000000000000..e407da2636dc33c93c95edc3c2937cb21f4461e9 --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/9d35ebf3-f9b8-4e85-b16c-fdf8b32f7c27.json @@ -0,0 +1,113 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5.2-2025-12-11/1786164117.722708", + "evaluation_timestamp": "1782325142.0", + "retrieved_timestamp": "1786164117.722708", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "mean on hle/S-adaptive/+4ep/a7f8b3e3 for scorer _llm_grader", + "source_data": { + "dataset_name": "hle", + "source_type": "hf_dataset", + "hf_repo": "cais/hle", + "samples_number": 1, + "sample_ids": [ + "hle_ea41e785" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1782325142.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.25, + "uncertainty": { + "num_samples": 4 + }, + "details": { + "total_matched_trajectories": "4", + "stopping_reason_count_repetition_guard": "3", + "stopping_reason_count_token_limit": "1", + "avg_total_tokens_target_model": "1769001.75", + "total_total_tokens_target_model": "7076007.0", + "avg_total_tokens_other_models": "2300.75", + "total_total_tokens_other_models": "9203.0", + "avg_total_tokens_all_models": "1771302.50", + "total_total_tokens_all_models": "7085210.0", + "avg_turn_count": "78.25", + "total_turn_count": "313.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "1200", + "max_connections": "16", + "max_tokens": "20000", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 5000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5.2-2025-12-11/9d35ebf3-f9b8-4e85-b16c-fdf8b32f7c27_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "af97f1fb64b2cac550abd9e520aeec560f7c33e7071f6978bc840c2ae582978e", + "total_rows": 4 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5.2-2025-12-11/9d35ebf3-f9b8-4e85-b16c-fdf8b32f7c27_samples.jsonl b/data/hle/openai/gpt-5.2-2025-12-11/9d35ebf3-f9b8-4e85-b16c-fdf8b32f7c27_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..fc11bf0f482101dbdf6eedd34f6cee468b3b1ce9 --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/9d35ebf3-f9b8-4e85-b16c-fdf8b32f7c27_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:af97f1fb64b2cac550abd9e520aeec560f7c33e7071f6978bc840c2ae582978e +size 456481 diff --git a/data/hle/openai/gpt-5.2-2025-12-11/a1c70183-65a8-40e4-8983-3812bf3e4901.json b/data/hle/openai/gpt-5.2-2025-12-11/a1c70183-65a8-40e4-8983-3812bf3e4901.json new file mode 100644 index 0000000000000000000000000000000000000000..7e1a1eaf33b9647a0793dae6a270af30ce0f45d8 --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/a1c70183-65a8-40e4-8983-3812bf3e4901.json @@ -0,0 +1,118 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5.2-2025-12-11/1786164054.387681", + "evaluation_timestamp": "1783327709.0", + "retrieved_timestamp": "1786164054.387681", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "mean on hle/S-adaptive/+1ep/50045a97 for scorer _llm_grader", + "source_data": { + "dataset_name": "hle", + "source_type": "hf_dataset", + "hf_repo": "cais/hle", + "samples_number": 4, + "sample_ids": [ + "hle_dfdb78f5", + "hle_209b6b64", + "hle_4485802f", + "hle_9df5d371" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1783327709.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.5, + "uncertainty": { + "standard_error": { + "value": 0.28867513459481287 + }, + "num_samples": 4 + }, + "details": { + "total_matched_trajectories": "3", + "stopping_reason_count_repetition_guard": "3", + "avg_total_tokens_target_model": "474716.33", + "total_total_tokens_target_model": "1424149.0", + "avg_total_tokens_other_models": "3704.67", + "total_total_tokens_other_models": "11114.0", + "avg_total_tokens_all_models": "478421.00", + "total_total_tokens_all_models": "1435263.0", + "avg_turn_count": "8.67", + "total_turn_count": "26.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "1", + "timeout": "1200", + "attempt_timeout": "3600", + "max_connections": "16", + "max_tokens": "64000", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 5000000 + }, + "sandbox": {}, + "max_attempts": 1 + }, + "additional_details": { + "max_retries": "1", + "timeout": "1200", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5.2-2025-12-11/a1c70183-65a8-40e4-8983-3812bf3e4901_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "43e0ab69b1a057d73d012fff547e9f6e25a74f0f55c0184b09727c8d608a62e2", + "total_rows": 4 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5.2-2025-12-11/a1c70183-65a8-40e4-8983-3812bf3e4901_samples.jsonl b/data/hle/openai/gpt-5.2-2025-12-11/a1c70183-65a8-40e4-8983-3812bf3e4901_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..9c95ff940c5183aa4b107bd08dc1f0ee79e058b4 --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/a1c70183-65a8-40e4-8983-3812bf3e4901_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:43e0ab69b1a057d73d012fff547e9f6e25a74f0f55c0184b09727c8d608a62e2 +size 91189 diff --git a/data/hle/openai/gpt-5.2-2025-12-11/ae5e928f-8a6a-452c-b56b-5b5a62df54cc.json b/data/hle/openai/gpt-5.2-2025-12-11/ae5e928f-8a6a-452c-b56b-5b5a62df54cc.json new file mode 100644 index 0000000000000000000000000000000000000000..9ed96bc45114708d8a901868e9ffacff100c99fa --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/ae5e928f-8a6a-452c-b56b-5b5a62df54cc.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5.2-2025-12-11/1786168712.190554", + "evaluation_timestamp": "1776359245.0", + "retrieved_timestamp": "1786168712.190554", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5.2-2025-12-11/ae5e928f-8a6a-452c-b56b-5b5a62df54cc_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "a288c645af4c99745874e2e7bf8b788e314babfaa04641e0d7c2ecff157ebe33", + "total_rows": 10 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5.2-2025-12-11/ae5e928f-8a6a-452c-b56b-5b5a62df54cc_samples.jsonl b/data/hle/openai/gpt-5.2-2025-12-11/ae5e928f-8a6a-452c-b56b-5b5a62df54cc_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..85174e0f75e1879d06aa61a326b456fa316efa52 --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/ae5e928f-8a6a-452c-b56b-5b5a62df54cc_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a288c645af4c99745874e2e7bf8b788e314babfaa04641e0d7c2ecff157ebe33 +size 196204 diff --git a/data/hle/openai/gpt-5.2-2025-12-11/aefa406e-c0ae-4816-9531-668ec1a746a3.json b/data/hle/openai/gpt-5.2-2025-12-11/aefa406e-c0ae-4816-9531-668ec1a746a3.json new file mode 100644 index 0000000000000000000000000000000000000000..0538a98406d539e0761a128fee2717d6bfd14a0c --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/aefa406e-c0ae-4816-9531-668ec1a746a3.json @@ -0,0 +1,112 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5.2-2025-12-11/1786164096.961459", + "evaluation_timestamp": "1780662738.0", + "retrieved_timestamp": "1786164096.961459", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "mean on hle/S-adaptive/+3ep/a4f1cf4f for scorer _llm_grader", + "source_data": { + "dataset_name": "hle", + "source_type": "hf_dataset", + "hf_repo": "cais/hle", + "samples_number": 1, + "sample_ids": [ + "hle_1d580261" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1780662738.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.0, + "uncertainty": { + "num_samples": 1 + }, + "details": { + "total_matched_trajectories": "1", + "stopping_reason_count_repetition_guard": "1", + "avg_total_tokens_target_model": "34922.00", + "total_total_tokens_target_model": "34922.0", + "avg_total_tokens_other_models": "6589.00", + "total_total_tokens_other_models": "6589.0", + "avg_total_tokens_all_models": "41511.00", + "total_total_tokens_all_models": "41511.0", + "avg_turn_count": "5.00", + "total_turn_count": "5.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "3", + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "16", + "max_tokens": "20000", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 5000000 + }, + "sandbox": {}, + "max_attempts": 3 + }, + "additional_details": { + "max_retries": "3", + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5.2-2025-12-11/aefa406e-c0ae-4816-9531-668ec1a746a3_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "53aa7b4bbbeb359b40ab0a4af07961f53a6b559c872eefdb9d3d08cd9becc6db", + "total_rows": 1 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5.2-2025-12-11/aefa406e-c0ae-4816-9531-668ec1a746a3_samples.jsonl b/data/hle/openai/gpt-5.2-2025-12-11/aefa406e-c0ae-4816-9531-668ec1a746a3_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..b8662edf2b3d326414920f7738cb8ad510325a1d --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/aefa406e-c0ae-4816-9531-668ec1a746a3_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:53aa7b4bbbeb359b40ab0a4af07961f53a6b559c872eefdb9d3d08cd9becc6db +size 21186 diff --git a/data/hle/openai/gpt-5.2-2025-12-11/b298a499-91d2-4a4a-ab22-8c154869cf7b.json b/data/hle/openai/gpt-5.2-2025-12-11/b298a499-91d2-4a4a-ab22-8c154869cf7b.json new file mode 100644 index 0000000000000000000000000000000000000000..ca03334483467189aef1090c90dab4547773992f --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/b298a499-91d2-4a4a-ab22-8c154869cf7b.json @@ -0,0 +1,113 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5.2-2025-12-11/1786164128.535631", + "evaluation_timestamp": "1781648331.0", + "retrieved_timestamp": "1786164128.535631", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "mean on hle/S-adaptive/+4ep/2aae83db for scorer _llm_grader", + "source_data": { + "dataset_name": "hle", + "source_type": "hf_dataset", + "hf_repo": "cais/hle", + "samples_number": 2, + "sample_ids": [ + "hle_ea41e785", + "hle_d7b7c3dc" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1781648331.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 1.0, + "uncertainty": { + "num_samples": 1 + }, + "details": { + "total_matched_trajectories": "1", + "stopping_reason_count_repetition_guard": "1", + "avg_total_tokens_target_model": "56901.00", + "total_total_tokens_target_model": "56901.0", + "avg_total_tokens_other_models": "5935.00", + "total_total_tokens_other_models": "5935.0", + "avg_total_tokens_all_models": "62836.00", + "total_total_tokens_all_models": "62836.0", + "avg_turn_count": "5.00", + "total_turn_count": "5.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "16", + "max_tokens": "20000", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 5000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5.2-2025-12-11/b298a499-91d2-4a4a-ab22-8c154869cf7b_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "d49693cab62040e7dea82f236d73add78fd8a33b093e9b77864189901ab0676c", + "total_rows": 1 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5.2-2025-12-11/b298a499-91d2-4a4a-ab22-8c154869cf7b_samples.jsonl b/data/hle/openai/gpt-5.2-2025-12-11/b298a499-91d2-4a4a-ab22-8c154869cf7b_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..0ce64be6e0b4fed509d7ffb921e35eb8966e1c09 --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/b298a499-91d2-4a4a-ab22-8c154869cf7b_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d49693cab62040e7dea82f236d73add78fd8a33b093e9b77864189901ab0676c +size 20196 diff --git a/data/hle/openai/gpt-5.2-2025-12-11/b67c8835-22a2-4feb-bd75-186e691c2295.json b/data/hle/openai/gpt-5.2-2025-12-11/b67c8835-22a2-4feb-bd75-186e691c2295.json new file mode 100644 index 0000000000000000000000000000000000000000..cc315bad637aa703292beeb60dda7d8886363e40 --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/b67c8835-22a2-4feb-bd75-186e691c2295.json @@ -0,0 +1,116 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5.2-2025-12-11/1786164123.217151", + "evaluation_timestamp": "1779099713.0", + "retrieved_timestamp": "1786164123.217151", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "mean on hle/S-adaptive/+4ep/9afd8991 for scorer _llm_grader", + "source_data": { + "dataset_name": "hle", + "source_type": "hf_dataset", + "hf_repo": "cais/hle", + "samples_number": 2, + "sample_ids": [ + "hle_3cf4483d", + "hle_70e5aebc" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1779099713.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.375, + "uncertainty": { + "standard_error": { + "value": 0.37499999999999994 + }, + "num_samples": 5 + }, + "details": { + "total_matched_trajectories": "5", + "stopping_reason_count_repetition_guard": "5", + "avg_total_tokens_target_model": "19901.40", + "total_total_tokens_target_model": "99507.0", + "avg_total_tokens_other_models": "3413.20", + "total_total_tokens_other_models": "17066.0", + "avg_total_tokens_all_models": "23314.60", + "total_total_tokens_all_models": "116573.0", + "avg_turn_count": "5.00", + "total_turn_count": "25.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "8", + "max_tokens": "32000", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 5000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "8", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5.2-2025-12-11/b67c8835-22a2-4feb-bd75-186e691c2295_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "f1a1a61f9f5d2e90d60555952c792a58206b17b4dd46d20aa8fedfaed791fb9c", + "total_rows": 5 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5.2-2025-12-11/b67c8835-22a2-4feb-bd75-186e691c2295_samples.jsonl b/data/hle/openai/gpt-5.2-2025-12-11/b67c8835-22a2-4feb-bd75-186e691c2295_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..b804d5c56b786dfc36f0e29af3aa3e0b51f86e57 --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/b67c8835-22a2-4feb-bd75-186e691c2295_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f1a1a61f9f5d2e90d60555952c792a58206b17b4dd46d20aa8fedfaed791fb9c +size 93342 diff --git a/data/hle/openai/gpt-5.2-2025-12-11/b73b0b46-6d29-41bd-8d4c-4cb668d122b7.json b/data/hle/openai/gpt-5.2-2025-12-11/b73b0b46-6d29-41bd-8d4c-4cb668d122b7.json new file mode 100644 index 0000000000000000000000000000000000000000..5459a5563cdaeeac88ef731f466dacc0fa1d4a93 --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/b73b0b46-6d29-41bd-8d4c-4cb668d122b7.json @@ -0,0 +1,112 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5.2-2025-12-11/1786164037.813117", + "evaluation_timestamp": "1780651715.0", + "retrieved_timestamp": "1786164037.813117", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "mean on hle/S-adaptive/+1ep/a9c8a5bc for scorer _llm_grader", + "source_data": { + "dataset_name": "hle", + "source_type": "hf_dataset", + "hf_repo": "cais/hle", + "samples_number": 1, + "sample_ids": [ + "hle_6d963e58" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1780651715.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.0, + "uncertainty": { + "num_samples": 1 + }, + "details": { + "total_matched_trajectories": "1", + "stopping_reason_count_repetition_guard": "1", + "avg_total_tokens_target_model": "44254.00", + "total_total_tokens_target_model": "44254.0", + "avg_total_tokens_other_models": "7123.00", + "total_total_tokens_other_models": "7123.0", + "avg_total_tokens_all_models": "51377.00", + "total_total_tokens_all_models": "51377.0", + "avg_turn_count": "8.00", + "total_turn_count": "8.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "3", + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "16", + "max_tokens": "20000", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 5000000 + }, + "sandbox": {}, + "max_attempts": 3 + }, + "additional_details": { + "max_retries": "3", + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5.2-2025-12-11/b73b0b46-6d29-41bd-8d4c-4cb668d122b7_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "6e442c7344bd92450ebdbe5f8670fef6d80e724fc79174b3af91429d75fe01b3", + "total_rows": 1 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5.2-2025-12-11/b73b0b46-6d29-41bd-8d4c-4cb668d122b7_samples.jsonl b/data/hle/openai/gpt-5.2-2025-12-11/b73b0b46-6d29-41bd-8d4c-4cb668d122b7_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..67958ce339ecd3a6089f1db3951efe6a8b37205e --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/b73b0b46-6d29-41bd-8d4c-4cb668d122b7_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6e442c7344bd92450ebdbe5f8670fef6d80e724fc79174b3af91429d75fe01b3 +size 24983 diff --git a/data/hle/openai/gpt-5.2-2025-12-11/b77a6fd1-1124-41e0-ba3e-5749f9538491.json b/data/hle/openai/gpt-5.2-2025-12-11/b77a6fd1-1124-41e0-ba3e-5749f9538491.json new file mode 100644 index 0000000000000000000000000000000000000000..1b77ccb6ad437b3a63a880d062cb5a201deb352a --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/b77a6fd1-1124-41e0-ba3e-5749f9538491.json @@ -0,0 +1,113 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5.2-2025-12-11/1786164102.123956", + "evaluation_timestamp": "1781917022.0", + "retrieved_timestamp": "1786164102.123956", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "mean on hle/S-adaptive/+3ep/16d34158 for scorer _llm_grader", + "source_data": { + "dataset_name": "hle", + "source_type": "hf_dataset", + "hf_repo": "cais/hle", + "samples_number": 2, + "sample_ids": [ + "hle_d7b7c3dc", + "hle_34c69126" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1781917022.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 1.0, + "uncertainty": { + "num_samples": 3 + }, + "details": { + "total_matched_trajectories": "3", + "stopping_reason_count_repetition_guard": "3", + "avg_total_tokens_target_model": "34738.67", + "total_total_tokens_target_model": "104216.0", + "avg_total_tokens_other_models": "5630.67", + "total_total_tokens_other_models": "16892.0", + "avg_total_tokens_all_models": "40369.33", + "total_total_tokens_all_models": "121108.0", + "avg_turn_count": "5.00", + "total_turn_count": "15.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "16", + "max_tokens": "20000", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 5000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5.2-2025-12-11/b77a6fd1-1124-41e0-ba3e-5749f9538491_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "efeb2b68527cccec60a1b77600cc79f442a177da948967ed6fe0ba18591ea5f7", + "total_rows": 3 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5.2-2025-12-11/b77a6fd1-1124-41e0-ba3e-5749f9538491_samples.jsonl b/data/hle/openai/gpt-5.2-2025-12-11/b77a6fd1-1124-41e0-ba3e-5749f9538491_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..9cdcc9b151db0559bd2f80757bd20fe9c612b97f --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/b77a6fd1-1124-41e0-ba3e-5749f9538491_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:efeb2b68527cccec60a1b77600cc79f442a177da948967ed6fe0ba18591ea5f7 +size 60132 diff --git a/data/hle/openai/gpt-5.2-2025-12-11/b99fdc0e-15df-4845-93bc-1bcd60a07e4b.json b/data/hle/openai/gpt-5.2-2025-12-11/b99fdc0e-15df-4845-93bc-1bcd60a07e4b.json new file mode 100644 index 0000000000000000000000000000000000000000..c69e97f5e90a08573f989840edc620d8a4f27583 --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/b99fdc0e-15df-4845-93bc-1bcd60a07e4b.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5.2-2025-12-11/1786168721.942931", + "evaluation_timestamp": "1776686674.0", + "retrieved_timestamp": "1786168721.942931", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5.2-2025-12-11/b99fdc0e-15df-4845-93bc-1bcd60a07e4b_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "6f85ba16b2c5bf143610683c7e65374ba968be55c034830cfa7e2015b0cd27df", + "total_rows": 22 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5.2-2025-12-11/b99fdc0e-15df-4845-93bc-1bcd60a07e4b_samples.jsonl b/data/hle/openai/gpt-5.2-2025-12-11/b99fdc0e-15df-4845-93bc-1bcd60a07e4b_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..49d443c7eaa9a3ec62f33df22e2682912b21425c --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/b99fdc0e-15df-4845-93bc-1bcd60a07e4b_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6f85ba16b2c5bf143610683c7e65374ba968be55c034830cfa7e2015b0cd27df +size 423436 diff --git a/data/hle/openai/gpt-5.2-2025-12-11/bab3ccf8-9c25-479c-8ecb-40222389f43e.json b/data/hle/openai/gpt-5.2-2025-12-11/bab3ccf8-9c25-479c-8ecb-40222389f43e.json new file mode 100644 index 0000000000000000000000000000000000000000..6a5c9eb9fd0babd14118db79c3c24135682e40d2 --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/bab3ccf8-9c25-479c-8ecb-40222389f43e.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5.2-2025-12-11/1786166772.04895", + "evaluation_timestamp": "1776996193.0", + "retrieved_timestamp": "1786166772.04895", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5.2-2025-12-11/bab3ccf8-9c25-479c-8ecb-40222389f43e_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "d6694e0671300546536d61f6ddb936ebe0f9a944c68e7b5988ddb614fe67bc2a", + "total_rows": 2 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5.2-2025-12-11/bab3ccf8-9c25-479c-8ecb-40222389f43e_samples.jsonl b/data/hle/openai/gpt-5.2-2025-12-11/bab3ccf8-9c25-479c-8ecb-40222389f43e_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..e3905c0576ef4c8f182ac5dcf9caaec8e50441ff --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/bab3ccf8-9c25-479c-8ecb-40222389f43e_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d6694e0671300546536d61f6ddb936ebe0f9a944c68e7b5988ddb614fe67bc2a +size 85126 diff --git a/data/hle/openai/gpt-5.2-2025-12-11/bacb7071-0225-4519-9477-055020e9d0a6.json b/data/hle/openai/gpt-5.2-2025-12-11/bacb7071-0225-4519-9477-055020e9d0a6.json new file mode 100644 index 0000000000000000000000000000000000000000..35c2b324620a28531ceaddcb3282f35cbee12b72 --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/bacb7071-0225-4519-9477-055020e9d0a6.json @@ -0,0 +1,133 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5.2-2025-12-11/1786164126.607775", + "evaluation_timestamp": "1776825207.0", + "retrieved_timestamp": "1786164126.607775", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "mean on hle/S-adaptive/+4ep/31a44b4c for scorer _llm_grader", + "source_data": { + "dataset_name": "hle", + "source_type": "hf_dataset", + "hf_repo": "cais/hle", + "samples_number": 23, + "sample_ids": [ + "hle_28643173", + "hle_16388cc8", + "hle_4418bea4", + "hle_ebd62885", + "hle_9019c023", + "hle_9282e7ae", + "hle_f85a0fa1", + "hle_c3ca0bda", + "hle_376b8247", + "hle_31497469", + "hle_c43a27a3", + "hle_d0c8b07e", + "hle_0291c889", + "hle_b1a7811d", + "hle_dab7fbb1", + "hle_cf3d65bd", + "hle_5e235b48", + "hle_0ff76ff3", + "hle_aa2fed90", + "hle_5759e7d7", + "hle_ca59b32c", + "hle_33bc94e2", + "hle_e3817b7e" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1776825207.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.27380952380952384, + "uncertainty": { + "standard_error": { + "value": 0.09744467585562441 + }, + "num_samples": 79 + }, + "details": { + "total_matched_trajectories": "79", + "stopping_reason_count_repetition_guard": "79", + "avg_total_tokens_target_model": "18957.94", + "total_total_tokens_target_model": "1497677.0", + "avg_total_tokens_other_models": "3942.15", + "total_total_tokens_other_models": "311430.0", + "avg_total_tokens_all_models": "22900.09", + "total_total_tokens_all_models": "1809107.0", + "avg_turn_count": "5.35", + "total_turn_count": "423.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 5000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5.2-2025-12-11/bacb7071-0225-4519-9477-055020e9d0a6_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "8fac32a6e1ae5c745835a2ac16fbf0c042b8e20ab4ba1944251f848c28180e28", + "total_rows": 79 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5.2-2025-12-11/bacb7071-0225-4519-9477-055020e9d0a6_samples.jsonl b/data/hle/openai/gpt-5.2-2025-12-11/bacb7071-0225-4519-9477-055020e9d0a6_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..6d83a3d247c8d927a54deed124fe38c805e494e5 --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/bacb7071-0225-4519-9477-055020e9d0a6_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8fac32a6e1ae5c745835a2ac16fbf0c042b8e20ab4ba1944251f848c28180e28 +size 1522047 diff --git a/data/hle/openai/gpt-5.2-2025-12-11/c37c2546-aed4-4f2f-ba7a-ae49407a4831.json b/data/hle/openai/gpt-5.2-2025-12-11/c37c2546-aed4-4f2f-ba7a-ae49407a4831.json new file mode 100644 index 0000000000000000000000000000000000000000..c9a9219322f87f03df635e42e297cbc616f0723f --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/c37c2546-aed4-4f2f-ba7a-ae49407a4831.json @@ -0,0 +1,117 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5.2-2025-12-11/1786164106.124734", + "evaluation_timestamp": "1781630820.0", + "retrieved_timestamp": "1786164106.124734", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "mean on hle/S-adaptive/+3ep/3ddacda4 for scorer _llm_grader", + "source_data": { + "dataset_name": "hle", + "source_type": "hf_dataset", + "hf_repo": "cais/hle", + "samples_number": 3, + "sample_ids": [ + "hle_1dc141bb", + "hle_34c69126", + "hle_c2ff3a43" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1781630820.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.5, + "uncertainty": { + "standard_error": { + "value": 0.5 + }, + "num_samples": 3 + }, + "details": { + "total_matched_trajectories": "3", + "stopping_reason_count_repetition_guard": "3", + "avg_total_tokens_target_model": "45643.33", + "total_total_tokens_target_model": "136930.0", + "avg_total_tokens_other_models": "4104.67", + "total_total_tokens_other_models": "12314.0", + "avg_total_tokens_all_models": "49748.00", + "total_total_tokens_all_models": "149244.0", + "avg_turn_count": "5.00", + "total_turn_count": "15.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "16", + "max_tokens": "20000", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 5000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5.2-2025-12-11/c37c2546-aed4-4f2f-ba7a-ae49407a4831_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "a7d690a63668ffd169473c66f62a55c91eb00f13616760b4a136eb465d1726a3", + "total_rows": 3 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5.2-2025-12-11/c37c2546-aed4-4f2f-ba7a-ae49407a4831_samples.jsonl b/data/hle/openai/gpt-5.2-2025-12-11/c37c2546-aed4-4f2f-ba7a-ae49407a4831_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..29cafa5082de412526b0ee9f31f07aa82d115a56 --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/c37c2546-aed4-4f2f-ba7a-ae49407a4831_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a7d690a63668ffd169473c66f62a55c91eb00f13616760b4a136eb465d1726a3 +size 57027 diff --git a/data/hle/openai/gpt-5.2-2025-12-11/c576caa8-032b-45d9-9968-cb458f8d7918.json b/data/hle/openai/gpt-5.2-2025-12-11/c576caa8-032b-45d9-9968-cb458f8d7918.json new file mode 100644 index 0000000000000000000000000000000000000000..19c6aec5c3194ccd2670f93e95b4b334524ad53e --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/c576caa8-032b-45d9-9968-cb458f8d7918.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5.2-2025-12-11/1786166792.51664", + "evaluation_timestamp": "1777098978.0", + "retrieved_timestamp": "1786166792.51664", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5.2-2025-12-11/c576caa8-032b-45d9-9968-cb458f8d7918_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "7c01234256367794f08421c96c50312fae4fe1c9c3d464cc15e51cbe6156a929", + "total_rows": 145 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5.2-2025-12-11/c576caa8-032b-45d9-9968-cb458f8d7918_samples.jsonl b/data/hle/openai/gpt-5.2-2025-12-11/c576caa8-032b-45d9-9968-cb458f8d7918_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..bb87d72540cb2a2fbf168b43f26865132786393a --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/c576caa8-032b-45d9-9968-cb458f8d7918_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7c01234256367794f08421c96c50312fae4fe1c9c3d464cc15e51cbe6156a929 +size 4815027 diff --git a/data/hle/openai/gpt-5.2-2025-12-11/cd417b56-9329-4f16-8b8b-cbff3d59850c.json b/data/hle/openai/gpt-5.2-2025-12-11/cd417b56-9329-4f16-8b8b-cbff3d59850c.json new file mode 100644 index 0000000000000000000000000000000000000000..44bf0b73f536714fc6d24a0d3862aa928e9e9d0d --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/cd417b56-9329-4f16-8b8b-cbff3d59850c.json @@ -0,0 +1,112 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5.2-2025-12-11/1786164039.387662", + "evaluation_timestamp": "1782219554.0", + "retrieved_timestamp": "1786164039.387662", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "mean on hle/S-adaptive/+1ep/f59fd081 for scorer _llm_grader", + "source_data": { + "dataset_name": "hle", + "source_type": "hf_dataset", + "hf_repo": "cais/hle", + "samples_number": 1, + "sample_ids": [ + "hle_d7b7c3dc" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1782219554.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 1.0, + "uncertainty": { + "num_samples": 1 + }, + "details": { + "total_matched_trajectories": "1", + "stopping_reason_count_repetition_guard": "1", + "avg_total_tokens_target_model": "117497.00", + "total_total_tokens_target_model": "117497.0", + "avg_total_tokens_other_models": "5670.00", + "total_total_tokens_other_models": "5670.0", + "avg_total_tokens_all_models": "123167.00", + "total_total_tokens_all_models": "123167.0", + "avg_turn_count": "8.00", + "total_turn_count": "8.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "1200", + "max_connections": "16", + "max_tokens": "20000", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 5000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5.2-2025-12-11/cd417b56-9329-4f16-8b8b-cbff3d59850c_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "80ea6406186a9e782da5f54b9deb5625f7c0f072c18d34b5b8c1a9d1d54547b1", + "total_rows": 1 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5.2-2025-12-11/cd417b56-9329-4f16-8b8b-cbff3d59850c_samples.jsonl b/data/hle/openai/gpt-5.2-2025-12-11/cd417b56-9329-4f16-8b8b-cbff3d59850c_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..ca45ffd388926c5ebc2b287539f443454e1edde4 --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/cd417b56-9329-4f16-8b8b-cbff3d59850c_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:80ea6406186a9e782da5f54b9deb5625f7c0f072c18d34b5b8c1a9d1d54547b1 +size 24113 diff --git a/data/hle/openai/gpt-5.2-2025-12-11/cecc7458-25e8-492c-be22-a0acdf67e3d4.json b/data/hle/openai/gpt-5.2-2025-12-11/cecc7458-25e8-492c-be22-a0acdf67e3d4.json new file mode 100644 index 0000000000000000000000000000000000000000..7d89bf71646aa73a74c66a1df5be14f4c98da8a5 --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/cecc7458-25e8-492c-be22-a0acdf67e3d4.json @@ -0,0 +1,100 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5.2-2025-12-11/1786164041.284066", + "evaluation_timestamp": "1782322677.0", + "retrieved_timestamp": "1786164041.284066", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "mean on hle/S-adaptive/+1ep/6214441b for scorer _llm_grader", + "source_data": { + "dataset_name": "hle", + "source_type": "hf_dataset", + "hf_repo": "cais/hle", + "samples_number": 1, + "sample_ids": [ + "hle_34c69126" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1782322677.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 1.0, + "uncertainty": { + "num_samples": 1 + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "1200", + "max_connections": "16", + "max_tokens": "20000", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 5000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5.2-2025-12-11/cecc7458-25e8-492c-be22-a0acdf67e3d4_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "621bfb1291ac29fd4b71f6f2d6e8c0ab37499780fb8eaf7f0995fd12390a8e1d", + "total_rows": 1 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5.2-2025-12-11/cecc7458-25e8-492c-be22-a0acdf67e3d4_samples.jsonl b/data/hle/openai/gpt-5.2-2025-12-11/cecc7458-25e8-492c-be22-a0acdf67e3d4_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..9025f2fa1d4aca90ed626cc99d4ab8d3efea3f05 --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/cecc7458-25e8-492c-be22-a0acdf67e3d4_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:621bfb1291ac29fd4b71f6f2d6e8c0ab37499780fb8eaf7f0995fd12390a8e1d +size 22398 diff --git a/data/hle/openai/gpt-5.2-2025-12-11/cf4f8713-7372-41cb-b5c9-a9ef17fd28ba.json b/data/hle/openai/gpt-5.2-2025-12-11/cf4f8713-7372-41cb-b5c9-a9ef17fd28ba.json new file mode 100644 index 0000000000000000000000000000000000000000..81df73fc44635e4a5f142e06c2353b33a1b44021 --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/cf4f8713-7372-41cb-b5c9-a9ef17fd28ba.json @@ -0,0 +1,116 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5.2-2025-12-11/1786164087.170213", + "evaluation_timestamp": "1778931458.0", + "retrieved_timestamp": "1786164087.170213", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "mean on hle/S-adaptive/+2ep/af1d3e97 for scorer _llm_grader", + "source_data": { + "dataset_name": "hle", + "source_type": "hf_dataset", + "hf_repo": "cais/hle", + "samples_number": 2, + "sample_ids": [ + "hle_ca6c436b", + "hle_28643173" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1778931458.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.5, + "uncertainty": { + "standard_error": { + "value": 0.5 + }, + "num_samples": 3 + }, + "details": { + "total_matched_trajectories": "3", + "stopping_reason_count_repetition_guard": "3", + "avg_total_tokens_target_model": "56978.00", + "total_total_tokens_target_model": "170934.0", + "avg_total_tokens_other_models": "7232.67", + "total_total_tokens_other_models": "21698.0", + "avg_total_tokens_all_models": "64210.67", + "total_total_tokens_all_models": "192632.0", + "avg_turn_count": "5.00", + "total_turn_count": "15.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "8", + "max_tokens": "32000", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 5000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "8", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5.2-2025-12-11/cf4f8713-7372-41cb-b5c9-a9ef17fd28ba_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "718cbfc27f73ce4822ca210e5a7cefaffc96f0b35e33f042f732ca914c497d55", + "total_rows": 3 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5.2-2025-12-11/cf4f8713-7372-41cb-b5c9-a9ef17fd28ba_samples.jsonl b/data/hle/openai/gpt-5.2-2025-12-11/cf4f8713-7372-41cb-b5c9-a9ef17fd28ba_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..900770874e8f7a18ca111721eb5fe190f19222ec --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/cf4f8713-7372-41cb-b5c9-a9ef17fd28ba_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:718cbfc27f73ce4822ca210e5a7cefaffc96f0b35e33f042f732ca914c497d55 +size 74762 diff --git a/data/hle/openai/gpt-5.2-2025-12-11/d054d91b-62f7-443c-971d-d51ad181d733.json b/data/hle/openai/gpt-5.2-2025-12-11/d054d91b-62f7-443c-971d-d51ad181d733.json new file mode 100644 index 0000000000000000000000000000000000000000..0de813e3cfcb8f37fa751d4d671efdab7e548f6a --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/d054d91b-62f7-443c-971d-d51ad181d733.json @@ -0,0 +1,115 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5.2-2025-12-11/1786164093.747352", + "evaluation_timestamp": "1780652364.0", + "retrieved_timestamp": "1786164093.747352", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "mean on hle/S-adaptive/+2ep/15604978 for scorer _llm_grader", + "source_data": { + "dataset_name": "hle", + "source_type": "hf_dataset", + "hf_repo": "cais/hle", + "samples_number": 4, + "sample_ids": [ + "hle_70e5aebc", + "hle_b6cec502", + "hle_ee938b26", + "hle_b26bf8aa" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1780652364.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.0, + "uncertainty": { + "num_samples": 2 + }, + "details": { + "total_matched_trajectories": "2", + "stopping_reason_count_repetition_guard": "2", + "avg_total_tokens_target_model": "30790.50", + "total_total_tokens_target_model": "61581.0", + "avg_total_tokens_other_models": "4629.50", + "total_total_tokens_other_models": "9259.0", + "avg_total_tokens_all_models": "35420.00", + "total_total_tokens_all_models": "70840.0", + "avg_turn_count": "5.00", + "total_turn_count": "10.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "3", + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "16", + "max_tokens": "20000", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 5000000 + }, + "sandbox": {}, + "max_attempts": 3 + }, + "additional_details": { + "max_retries": "3", + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5.2-2025-12-11/d054d91b-62f7-443c-971d-d51ad181d733_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "241793c5137a6df8c3e8a024b7f65c4f6fc81ffe360c9938bc7e6a99bf8b6155", + "total_rows": 2 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5.2-2025-12-11/d054d91b-62f7-443c-971d-d51ad181d733_samples.jsonl b/data/hle/openai/gpt-5.2-2025-12-11/d054d91b-62f7-443c-971d-d51ad181d733_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..6d674ba513adb7abed52637a9179578ecbb0f18a --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/d054d91b-62f7-443c-971d-d51ad181d733_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:241793c5137a6df8c3e8a024b7f65c4f6fc81ffe360c9938bc7e6a99bf8b6155 +size 38177 diff --git a/data/hle/openai/gpt-5.2-2025-12-11/e3aaed43-3d3a-42b4-92e0-a38ac10d7431.json b/data/hle/openai/gpt-5.2-2025-12-11/e3aaed43-3d3a-42b4-92e0-a38ac10d7431.json new file mode 100644 index 0000000000000000000000000000000000000000..403170dfe996beca959f7ffb5d31e37271cbeda1 --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/e3aaed43-3d3a-42b4-92e0-a38ac10d7431.json @@ -0,0 +1,112 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5.2-2025-12-11/1786164098.615308", + "evaluation_timestamp": "1783017404.0", + "retrieved_timestamp": "1786164098.615308", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "mean on hle/S-adaptive/+3ep/90b2a41a for scorer _llm_grader", + "source_data": { + "dataset_name": "hle", + "source_type": "hf_dataset", + "hf_repo": "cais/hle", + "samples_number": 1, + "sample_ids": [ + "hle_4485802f" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1783017404.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.0, + "uncertainty": { + "num_samples": 2 + }, + "details": { + "total_matched_trajectories": "2", + "stopping_reason_count_repetition_guard": "2", + "avg_total_tokens_target_model": "45747.00", + "total_total_tokens_target_model": "91494.0", + "avg_total_tokens_other_models": "2937.00", + "total_total_tokens_other_models": "5874.0", + "avg_total_tokens_all_models": "48684.00", + "total_total_tokens_all_models": "97368.0", + "avg_turn_count": "6.00", + "total_turn_count": "12.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "1200", + "max_connections": "16", + "max_tokens": "20000", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 5000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5.2-2025-12-11/e3aaed43-3d3a-42b4-92e0-a38ac10d7431_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "5261d417cfba23b808c6cf279cbe86d5e137febbf5a76123aae1efd0a6e1fa44", + "total_rows": 2 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5.2-2025-12-11/e3aaed43-3d3a-42b4-92e0-a38ac10d7431_samples.jsonl b/data/hle/openai/gpt-5.2-2025-12-11/e3aaed43-3d3a-42b4-92e0-a38ac10d7431_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..1cddf2bf69af0d84af3a3a6294ae7039a89a3c60 --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/e3aaed43-3d3a-42b4-92e0-a38ac10d7431_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5261d417cfba23b808c6cf279cbe86d5e137febbf5a76123aae1efd0a6e1fa44 +size 38672 diff --git a/data/hle/openai/gpt-5.2-2025-12-11/e4f93563-d9fd-4d01-ad64-ffc89ed8087c.json b/data/hle/openai/gpt-5.2-2025-12-11/e4f93563-d9fd-4d01-ad64-ffc89ed8087c.json new file mode 100644 index 0000000000000000000000000000000000000000..4d968fbd21557f20b7f79565b1f900a00d784ca8 --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/e4f93563-d9fd-4d01-ad64-ffc89ed8087c.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5.2-2025-12-11/1786166775.056968", + "evaluation_timestamp": "1776430850.0", + "retrieved_timestamp": "1786166775.056968", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5.2-2025-12-11/e4f93563-d9fd-4d01-ad64-ffc89ed8087c_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "256313e40c4c1d72def425981ce7812e544682ba53235ae59a88ab2157ff6dac", + "total_rows": 30 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5.2-2025-12-11/e4f93563-d9fd-4d01-ad64-ffc89ed8087c_samples.jsonl b/data/hle/openai/gpt-5.2-2025-12-11/e4f93563-d9fd-4d01-ad64-ffc89ed8087c_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..e918e68aa7bc435a679a005ce843698f16b5db6f --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/e4f93563-d9fd-4d01-ad64-ffc89ed8087c_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:256313e40c4c1d72def425981ce7812e544682ba53235ae59a88ab2157ff6dac +size 1017274 diff --git a/data/hle/openai/gpt-5.2-2025-12-11/e80ca73a-2215-4f48-ae49-726b5eea6ad2.json b/data/hle/openai/gpt-5.2-2025-12-11/e80ca73a-2215-4f48-ae49-726b5eea6ad2.json new file mode 100644 index 0000000000000000000000000000000000000000..520bc8141bbaa434dad3094b4e18d292e92380a7 --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/e80ca73a-2215-4f48-ae49-726b5eea6ad2.json @@ -0,0 +1,112 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5.2-2025-12-11/1786164034.638626", + "evaluation_timestamp": "1783092534.0", + "retrieved_timestamp": "1786164034.638626", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "mean on hle/S-adaptive/+1ep/90b2a41a for scorer _llm_grader", + "source_data": { + "dataset_name": "hle", + "source_type": "hf_dataset", + "hf_repo": "cais/hle", + "samples_number": 1, + "sample_ids": [ + "hle_4485802f" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1783092534.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.0, + "uncertainty": { + "num_samples": 1 + }, + "details": { + "total_matched_trajectories": "1", + "stopping_reason_count_repetition_guard": "1", + "avg_total_tokens_target_model": "137889.00", + "total_total_tokens_target_model": "137889.0", + "avg_total_tokens_other_models": "3613.00", + "total_total_tokens_other_models": "3613.0", + "avg_total_tokens_all_models": "141502.00", + "total_total_tokens_all_models": "141502.0", + "avg_turn_count": "5.00", + "total_turn_count": "5.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "1", + "timeout": "1200", + "attempt_timeout": "3600", + "max_connections": "16", + "max_tokens": "64000", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 5000000 + }, + "sandbox": {}, + "max_attempts": 1 + }, + "additional_details": { + "max_retries": "1", + "timeout": "1200", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5.2-2025-12-11/e80ca73a-2215-4f48-ae49-726b5eea6ad2_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "ce7fd0e801d13e1e0c201f6137cb1594eb3c0b98d8c6e6359a582505a0b3940a", + "total_rows": 1 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5.2-2025-12-11/e80ca73a-2215-4f48-ae49-726b5eea6ad2_samples.jsonl b/data/hle/openai/gpt-5.2-2025-12-11/e80ca73a-2215-4f48-ae49-726b5eea6ad2_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..ac42eb5f3f5913919f8f5ffe10b942aafd178aaf --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/e80ca73a-2215-4f48-ae49-726b5eea6ad2_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ce7fd0e801d13e1e0c201f6137cb1594eb3c0b98d8c6e6359a582505a0b3940a +size 18689 diff --git a/data/hle/openai/gpt-5.2-2025-12-11/e825a2e5-d08b-4509-adfe-f3a3f55fa35c.json b/data/hle/openai/gpt-5.2-2025-12-11/e825a2e5-d08b-4509-adfe-f3a3f55fa35c.json new file mode 100644 index 0000000000000000000000000000000000000000..29b52a6b564fe37b6acbcbb950236641d0f17829 --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/e825a2e5-d08b-4509-adfe-f3a3f55fa35c.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5.2-2025-12-11/1786168724.107723", + "evaluation_timestamp": "1780669218.0", + "retrieved_timestamp": "1786168724.107723", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5.2-2025-12-11/e825a2e5-d08b-4509-adfe-f3a3f55fa35c_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "53bbd71783f66aa53665b0923d6e1df7f2e0254cda6dccca9cd281f66b8bb064", + "total_rows": 3 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5.2-2025-12-11/e825a2e5-d08b-4509-adfe-f3a3f55fa35c_samples.jsonl b/data/hle/openai/gpt-5.2-2025-12-11/e825a2e5-d08b-4509-adfe-f3a3f55fa35c_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..2aa4fc990e28dd34cb27480cab8f81153415366d --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/e825a2e5-d08b-4509-adfe-f3a3f55fa35c_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:53bbd71783f66aa53665b0923d6e1df7f2e0254cda6dccca9cd281f66b8bb064 +size 58036 diff --git a/data/hle/openai/gpt-5.2-2025-12-11/e922f976-86fb-487c-9034-8efe0ad0c6b8.json b/data/hle/openai/gpt-5.2-2025-12-11/e922f976-86fb-487c-9034-8efe0ad0c6b8.json new file mode 100644 index 0000000000000000000000000000000000000000..daf26cb3282bfb88e4a4d428dc06ab1275f07e6f --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/e922f976-86fb-487c-9034-8efe0ad0c6b8.json @@ -0,0 +1,112 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5.2-2025-12-11/1786164043.24143", + "evaluation_timestamp": "1782983153.0", + "retrieved_timestamp": "1786164043.24143", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "mean on hle/S-adaptive/+1ep/286511e3 for scorer _llm_grader", + "source_data": { + "dataset_name": "hle", + "source_type": "hf_dataset", + "hf_repo": "cais/hle", + "samples_number": 1, + "sample_ids": [ + "hle_141cd3ef" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1782983153.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.0, + "uncertainty": { + "num_samples": 1 + }, + "details": { + "total_matched_trajectories": "1", + "stopping_reason_count_repetition_guard": "1", + "avg_total_tokens_target_model": "24434.00", + "total_total_tokens_target_model": "24434.0", + "avg_total_tokens_other_models": "3011.00", + "total_total_tokens_other_models": "3011.0", + "avg_total_tokens_all_models": "27445.00", + "total_total_tokens_all_models": "27445.0", + "avg_turn_count": "5.00", + "total_turn_count": "5.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "1200", + "max_connections": "16", + "max_tokens": "20000", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 5000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5.2-2025-12-11/e922f976-86fb-487c-9034-8efe0ad0c6b8_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "d051575f7cced5e55fd6b903843db307412501ee29d16b47978d4549ef2662b1", + "total_rows": 1 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5.2-2025-12-11/e922f976-86fb-487c-9034-8efe0ad0c6b8_samples.jsonl b/data/hle/openai/gpt-5.2-2025-12-11/e922f976-86fb-487c-9034-8efe0ad0c6b8_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..f1d2610f812b917eafa356efc0c8b933894f119a --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/e922f976-86fb-487c-9034-8efe0ad0c6b8_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d051575f7cced5e55fd6b903843db307412501ee29d16b47978d4549ef2662b1 +size 17398 diff --git a/data/hle/openai/gpt-5.2-2025-12-11/f1c2403b-b2f2-45b5-a7c1-13d99ced91a2.json b/data/hle/openai/gpt-5.2-2025-12-11/f1c2403b-b2f2-45b5-a7c1-13d99ced91a2.json new file mode 100644 index 0000000000000000000000000000000000000000..81239623bbf9fef377406d1937dc4aaf519602c8 --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/f1c2403b-b2f2-45b5-a7c1-13d99ced91a2.json @@ -0,0 +1,113 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5.2-2025-12-11/1786164088.96196", + "evaluation_timestamp": "1781905453.0", + "retrieved_timestamp": "1786164088.96196", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "mean on hle/S-adaptive/+2ep/dbe7256f for scorer _llm_grader", + "source_data": { + "dataset_name": "hle", + "source_type": "hf_dataset", + "hf_repo": "cais/hle", + "samples_number": 2, + "sample_ids": [ + "hle_1dc141bb", + "hle_b6cec502" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1781905453.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.0, + "uncertainty": { + "num_samples": 3 + }, + "details": { + "total_matched_trajectories": "3", + "stopping_reason_count_repetition_guard": "3", + "avg_total_tokens_target_model": "42977.00", + "total_total_tokens_target_model": "128931.0", + "avg_total_tokens_other_models": "3930.00", + "total_total_tokens_other_models": "11790.0", + "avg_total_tokens_all_models": "46907.00", + "total_total_tokens_all_models": "140721.0", + "avg_turn_count": "5.00", + "total_turn_count": "15.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "16", + "max_tokens": "20000", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 5000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5.2-2025-12-11/f1c2403b-b2f2-45b5-a7c1-13d99ced91a2_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "19fd079149b1f17dc55458bfbdeb116d454077fef7bee5578be0a65db8c100af", + "total_rows": 3 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5.2-2025-12-11/f1c2403b-b2f2-45b5-a7c1-13d99ced91a2_samples.jsonl b/data/hle/openai/gpt-5.2-2025-12-11/f1c2403b-b2f2-45b5-a7c1-13d99ced91a2_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..ee38cda64ca3cb50e119a280d2edbdc95520d743 --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/f1c2403b-b2f2-45b5-a7c1-13d99ced91a2_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:19fd079149b1f17dc55458bfbdeb116d454077fef7bee5578be0a65db8c100af +size 55069 diff --git a/data/hle/openai/gpt-5.2-2025-12-11/f2adce90-9447-4bdc-a31d-19f60045d1f6.json b/data/hle/openai/gpt-5.2-2025-12-11/f2adce90-9447-4bdc-a31d-19f60045d1f6.json new file mode 100644 index 0000000000000000000000000000000000000000..6323c8934916ee8f19c3e24ff5c427a73a4f0334 --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/f2adce90-9447-4bdc-a31d-19f60045d1f6.json @@ -0,0 +1,116 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5.2-2025-12-11/1786164056.164675", + "evaluation_timestamp": "1776880341.0", + "retrieved_timestamp": "1786164056.164675", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "mean on hle/S-adaptive/+1ep/93e5a897 for scorer _llm_grader", + "source_data": { + "dataset_name": "hle", + "source_type": "hf_dataset", + "hf_repo": "cais/hle", + "samples_number": 6, + "sample_ids": [ + "hle_8a419425", + "hle_9282e7ae", + "hle_d0c8b07e", + "hle_cf3d65bd", + "hle_33bc94e2", + "hle_e3817b7e" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1776880341.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.4, + "uncertainty": { + "standard_error": { + "value": 0.24494897427831783 + }, + "num_samples": 5 + }, + "details": { + "total_matched_trajectories": "5", + "stopping_reason_count_repetition_guard": "5", + "avg_total_tokens_target_model": "31643.20", + "total_total_tokens_target_model": "158216.0", + "avg_total_tokens_other_models": "6418.80", + "total_total_tokens_other_models": "32094.0", + "avg_total_tokens_all_models": "38062.00", + "total_total_tokens_all_models": "190310.0", + "avg_turn_count": "5.00", + "total_turn_count": "25.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 5000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5.2-2025-12-11/f2adce90-9447-4bdc-a31d-19f60045d1f6_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "a6573160bc96db14c467fd2e7644b3cfb6bf9eb2fb8b59084039e5078816d58b", + "total_rows": 5 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5.2-2025-12-11/f2adce90-9447-4bdc-a31d-19f60045d1f6_samples.jsonl b/data/hle/openai/gpt-5.2-2025-12-11/f2adce90-9447-4bdc-a31d-19f60045d1f6_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..5d0786f90593e46d698214e3321ae7406e8a6252 --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/f2adce90-9447-4bdc-a31d-19f60045d1f6_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a6573160bc96db14c467fd2e7644b3cfb6bf9eb2fb8b59084039e5078816d58b +size 106305 diff --git a/data/hle/openai/gpt-5.2-2025-12-11/f5e5a6e3-03f0-40cc-b2c2-621a29dceca9.json b/data/hle/openai/gpt-5.2-2025-12-11/f5e5a6e3-03f0-40cc-b2c2-621a29dceca9.json new file mode 100644 index 0000000000000000000000000000000000000000..3d159255f165ff58c412d4994fdd46b23b08a10b --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/f5e5a6e3-03f0-40cc-b2c2-621a29dceca9.json @@ -0,0 +1,113 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5.2-2025-12-11/1786164104.133996", + "evaluation_timestamp": "1776819788.0", + "retrieved_timestamp": "1786164104.133996", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "mean on hle/S-adaptive/+3ep/f8ee45d3 for scorer _llm_grader", + "source_data": { + "dataset_name": "hle", + "source_type": "hf_dataset", + "hf_repo": "cais/hle", + "samples_number": 3, + "sample_ids": [ + "hle_a559277e", + "hle_8a419425", + "hle_85e95ef4" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1776819788.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.2222222222222222, + "uncertainty": { + "standard_error": { + "value": 0.22222222222222224 + }, + "num_samples": 7 + }, + "details": { + "total_matched_trajectories": "7", + "stopping_reason_count_repetition_guard": "7", + "avg_total_tokens_target_model": "29255.14", + "total_total_tokens_target_model": "204786.0", + "avg_total_tokens_other_models": "2791.57", + "total_total_tokens_other_models": "19541.0", + "avg_total_tokens_all_models": "32046.71", + "total_total_tokens_all_models": "224327.0", + "avg_turn_count": "5.00", + "total_turn_count": "35.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 5000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5.2-2025-12-11/f5e5a6e3-03f0-40cc-b2c2-621a29dceca9_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "50c0a62b751c9b3cb770f4a296c291dfcd47d540d83aaf3da06f1cf703f29120", + "total_rows": 7 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5.2-2025-12-11/f5e5a6e3-03f0-40cc-b2c2-621a29dceca9_samples.jsonl b/data/hle/openai/gpt-5.2-2025-12-11/f5e5a6e3-03f0-40cc-b2c2-621a29dceca9_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..df73020da36417a56d340be9c91f55a9570d1f21 --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/f5e5a6e3-03f0-40cc-b2c2-621a29dceca9_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:50c0a62b751c9b3cb770f4a296c291dfcd47d540d83aaf3da06f1cf703f29120 +size 122067 diff --git a/data/hle/openai/gpt-5.2-2025-12-11/f750c501-5548-4a7e-ac9a-8ebd3c4fa83a.json b/data/hle/openai/gpt-5.2-2025-12-11/f750c501-5548-4a7e-ac9a-8ebd3c4fa83a.json new file mode 100644 index 0000000000000000000000000000000000000000..ed9af3abb989d02d540d64caae4b396e061336c3 --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/f750c501-5548-4a7e-ac9a-8ebd3c4fa83a.json @@ -0,0 +1,112 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5.2-2025-12-11/1786164119.618784", + "evaluation_timestamp": "1783096473.0", + "retrieved_timestamp": "1786164119.618784", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "mean on hle/S-adaptive/+4ep/ea1d0e0e for scorer _llm_grader", + "source_data": { + "dataset_name": "hle", + "source_type": "hf_dataset", + "hf_repo": "cais/hle", + "samples_number": 1, + "sample_ids": [ + "hle_0ec247c5" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1783096473.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.0, + "uncertainty": { + "num_samples": 4 + }, + "details": { + "total_matched_trajectories": "4", + "stopping_reason_count_repetition_guard": "4", + "avg_total_tokens_target_model": "208016.25", + "total_total_tokens_target_model": "832065.0", + "avg_total_tokens_other_models": "4746.75", + "total_total_tokens_other_models": "18987.0", + "avg_total_tokens_all_models": "212763.00", + "total_total_tokens_all_models": "851052.0", + "avg_turn_count": "5.75", + "total_turn_count": "23.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "1", + "timeout": "1200", + "attempt_timeout": "3600", + "max_connections": "16", + "max_tokens": "64000", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 5000000 + }, + "sandbox": {}, + "max_attempts": 1 + }, + "additional_details": { + "max_retries": "1", + "timeout": "1200", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5.2-2025-12-11/f750c501-5548-4a7e-ac9a-8ebd3c4fa83a_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "e417a0b46a94048b686b1f557839df54b01feae403834be87eba106b0fcf600d", + "total_rows": 4 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5.2-2025-12-11/f750c501-5548-4a7e-ac9a-8ebd3c4fa83a_samples.jsonl b/data/hle/openai/gpt-5.2-2025-12-11/f750c501-5548-4a7e-ac9a-8ebd3c4fa83a_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..605f7950c09e8d0edb861cd2abcc5c42bc77cac6 --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/f750c501-5548-4a7e-ac9a-8ebd3c4fa83a_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e417a0b46a94048b686b1f557839df54b01feae403834be87eba106b0fcf600d +size 89469 diff --git a/data/hle/openai/gpt-5.2-2025-12-11/ff10aa73-f60a-45c9-aa45-642b614a2002.json b/data/hle/openai/gpt-5.2-2025-12-11/ff10aa73-f60a-45c9-aa45-642b614a2002.json new file mode 100644 index 0000000000000000000000000000000000000000..3df920ecd23b4216f45d5bd9edc7aa203004c29f --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/ff10aa73-f60a-45c9-aa45-642b614a2002.json @@ -0,0 +1,166 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5.2-2025-12-11/1786164151.722955", + "evaluation_timestamp": "1780489997.0", + "retrieved_timestamp": "1786164151.722955", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "mean on hle/S-adaptive/a22393ad for scorer _llm_grader", + "source_data": { + "dataset_name": "hle", + "source_type": "hf_dataset", + "hf_repo": "cais/hle", + "samples_number": 52, + "sample_ids": [ + "hle_764ce81c", + "hle_bc5fbce2", + "hle_dfdb78f5", + "hle_209b6b64", + "hle_4485802f", + "hle_44007b97", + "hle_7808792d", + "hle_e3b8f5c2", + "hle_ea41e785", + "hle_9df5d371", + "hle_d7b7c3dc", + "hle_1dc141bb", + "hle_b6cec502", + "hle_9255683e", + "hle_34c69126", + "hle_fcd56401", + "hle_b5c681d9", + "hle_d5a55900", + "hle_3e11b38a", + "hle_0a1eec3f", + "hle_5d49423c", + "hle_1d580261", + "hle_82f5c665", + "hle_cc7f25c0", + "hle_226e134b", + "hle_c1158fb8", + "hle_d0ea1019", + "hle_fcd42d0c", + "hle_b33012cf", + "hle_3697d2bc", + "hle_5e349785", + "hle_8b0384c4", + "hle_ee938b26", + "hle_d190d8c2", + "hle_b26bf8aa", + "hle_c2ff3a43", + "hle_141cd3ef", + "hle_5671e386", + "hle_a634e89e", + "hle_670287ad", + "hle_298b5616", + "hle_e14d8c23", + "hle_510b6040", + "hle_c6adc1b4", + "hle_0ec247c5", + "hle_aa343d51", + "hle_d52d3dae", + "hle_6d963e58", + "hle_35a40c52", + "hle_45d1c140", + "hle_c0ed25f7", + "hle_cd4aa5b3" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1780489997.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.25045045045045045, + "uncertainty": { + "standard_error": { + "value": 0.06623510996816717 + }, + "num_samples": 155 + }, + "details": { + "total_matched_trajectories": "155", + "stopping_reason_count_repetition_guard": "155", + "avg_total_tokens_target_model": "17181.49", + "total_total_tokens_target_model": "2663131.0", + "avg_total_tokens_other_models": "3817.53", + "total_total_tokens_other_models": "591717.0", + "avg_total_tokens_all_models": "20999.02", + "total_total_tokens_all_models": "3254848.0", + "avg_turn_count": "5.10", + "total_turn_count": "790.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "3", + "timeout": "1800", + "attempt_timeout": "1800", + "max_connections": "16", + "max_tokens": "20000", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 5000000 + }, + "sandbox": {}, + "max_attempts": 3 + }, + "additional_details": { + "max_retries": "3", + "timeout": "1800", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5.2-2025-12-11/ff10aa73-f60a-45c9-aa45-642b614a2002_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "68d839e62ff5a32ed09662841052cffa43d6dc0baedac1954c12ab3049c6d0d1", + "total_rows": 155 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5.2-2025-12-11/ff10aa73-f60a-45c9-aa45-642b614a2002_samples.jsonl b/data/hle/openai/gpt-5.2-2025-12-11/ff10aa73-f60a-45c9-aa45-642b614a2002_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..09fb00948bea900edc7ff521c86dac40ff3851ab --- /dev/null +++ b/data/hle/openai/gpt-5.2-2025-12-11/ff10aa73-f60a-45c9-aa45-642b614a2002_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68d839e62ff5a32ed09662841052cffa43d6dc0baedac1954c12ab3049c6d0d1 +size 2888828 diff --git a/data/hle/openai/gpt-5.4-2026-03-05/00435845-eb7b-4262-9ee1-18e2c1a06aa2.json b/data/hle/openai/gpt-5.4-2026-03-05/00435845-eb7b-4262-9ee1-18e2c1a06aa2.json new file mode 100644 index 0000000000000000000000000000000000000000..f6e67bb65f093fa288648f14d1181a6f496f9992 --- /dev/null +++ b/data/hle/openai/gpt-5.4-2026-03-05/00435845-eb7b-4262-9ee1-18e2c1a06aa2.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5.4-2026-03-05/1786168733.733655", + "evaluation_timestamp": "1776107987.0", + "retrieved_timestamp": "1786168733.733655", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5.4-2026-03-05/00435845-eb7b-4262-9ee1-18e2c1a06aa2_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "acf1ae9eeebadd30236b32ba1180a8bfd3e3aeba32a4334dd7c3e307f4beffca", + "total_rows": 20 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5.4-2026-03-05/00435845-eb7b-4262-9ee1-18e2c1a06aa2_samples.jsonl b/data/hle/openai/gpt-5.4-2026-03-05/00435845-eb7b-4262-9ee1-18e2c1a06aa2_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..545452e272edab775f0a57f8a6787abf9c6a6351 --- /dev/null +++ b/data/hle/openai/gpt-5.4-2026-03-05/00435845-eb7b-4262-9ee1-18e2c1a06aa2_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:acf1ae9eeebadd30236b32ba1180a8bfd3e3aeba32a4334dd7c3e307f4beffca +size 406425 diff --git a/data/hle/openai/gpt-5.4-2026-03-05/0330f59d-bd5a-48fa-b082-fec183a8133b.json b/data/hle/openai/gpt-5.4-2026-03-05/0330f59d-bd5a-48fa-b082-fec183a8133b.json new file mode 100644 index 0000000000000000000000000000000000000000..31d58b4c9482ea05943fae0106adebc5225a19a8 --- /dev/null +++ b/data/hle/openai/gpt-5.4-2026-03-05/0330f59d-bd5a-48fa-b082-fec183a8133b.json @@ -0,0 +1,109 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5.4-2026-03-05/1786164161.926087", + "evaluation_timestamp": "1777025747.0", + "retrieved_timestamp": "1786164161.926087", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "mean on hle/S-adaptive/+1ep/9f6dbad0 for scorer _llm_grader", + "source_data": { + "dataset_name": "hle", + "source_type": "hf_dataset", + "hf_repo": "cais/hle", + "samples_number": 2, + "sample_ids": [ + "hle_b6cec502", + "hle_c2ff3a43" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1777025747.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.0, + "uncertainty": { + "num_samples": 1 + }, + "details": { + "total_matched_trajectories": "1", + "stopping_reason_count_repetition_guard": "1", + "avg_total_tokens_target_model": "58964.00", + "total_total_tokens_target_model": "58964.0", + "avg_total_tokens_other_models": "3517.00", + "total_total_tokens_other_models": "3517.0", + "avg_total_tokens_all_models": "62481.00", + "total_total_tokens_all_models": "62481.0", + "avg_turn_count": "5.00", + "total_turn_count": "5.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 5000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5.4-2026-03-05/0330f59d-bd5a-48fa-b082-fec183a8133b_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "3894a69c4adbc723f47d25bc847e443da79d0c3c2a76dcd50af285ce23212b70", + "total_rows": 1 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5.4-2026-03-05/0330f59d-bd5a-48fa-b082-fec183a8133b_samples.jsonl b/data/hle/openai/gpt-5.4-2026-03-05/0330f59d-bd5a-48fa-b082-fec183a8133b_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..933a1bd4aad1cb3536478f1d5502056dc1401a9c --- /dev/null +++ b/data/hle/openai/gpt-5.4-2026-03-05/0330f59d-bd5a-48fa-b082-fec183a8133b_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3894a69c4adbc723f47d25bc847e443da79d0c3c2a76dcd50af285ce23212b70 +size 17867 diff --git a/data/hle/openai/gpt-5.4-2026-03-05/06dd6022-9983-4b51-b874-93a20cf5f6fe.json b/data/hle/openai/gpt-5.4-2026-03-05/06dd6022-9983-4b51-b874-93a20cf5f6fe.json new file mode 100644 index 0000000000000000000000000000000000000000..012fb0a663ecd95a4d930c68af7b1554f655d949 --- /dev/null +++ b/data/hle/openai/gpt-5.4-2026-03-05/06dd6022-9983-4b51-b874-93a20cf5f6fe.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5.4-2026-03-05/1786166802.379051", + "evaluation_timestamp": "1776519735.0", + "retrieved_timestamp": "1786166802.379051", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5.4-2026-03-05/06dd6022-9983-4b51-b874-93a20cf5f6fe_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "aed77fced8eccd9d8be1056584c16f319c57a849e9162e04093b9d272f354148", + "total_rows": 20 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5.4-2026-03-05/06dd6022-9983-4b51-b874-93a20cf5f6fe_samples.jsonl b/data/hle/openai/gpt-5.4-2026-03-05/06dd6022-9983-4b51-b874-93a20cf5f6fe_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..69f0ae04b568a956a04ab439f70596e0ab54b75f --- /dev/null +++ b/data/hle/openai/gpt-5.4-2026-03-05/06dd6022-9983-4b51-b874-93a20cf5f6fe_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:aed77fced8eccd9d8be1056584c16f319c57a849e9162e04093b9d272f354148 +size 807857 diff --git a/data/hle/openai/gpt-5.4-2026-03-05/1ad18d5d-0889-4b07-9cb0-6561888d6db7.json b/data/hle/openai/gpt-5.4-2026-03-05/1ad18d5d-0889-4b07-9cb0-6561888d6db7.json new file mode 100644 index 0000000000000000000000000000000000000000..bea0912a18bcfad7201e0b469765bfbcf027f9d1 --- /dev/null +++ b/data/hle/openai/gpt-5.4-2026-03-05/1ad18d5d-0889-4b07-9cb0-6561888d6db7.json @@ -0,0 +1,112 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5.4-2026-03-05/1786164159.734561", + "evaluation_timestamp": "1781905453.0", + "retrieved_timestamp": "1786164159.734561", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "mean on hle/S-adaptive/+1ep/9b8beab9 for scorer _llm_grader", + "source_data": { + "dataset_name": "hle", + "source_type": "hf_dataset", + "hf_repo": "cais/hle", + "samples_number": 1, + "sample_ids": [ + "hle_7808792d" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1781905453.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 1.0, + "uncertainty": { + "num_samples": 1 + }, + "details": { + "total_matched_trajectories": "1", + "stopping_reason_count_repetition_guard": "1", + "avg_total_tokens_target_model": "2095662.00", + "total_total_tokens_target_model": "2095662.0", + "avg_total_tokens_other_models": "17452.00", + "total_total_tokens_other_models": "17452.0", + "avg_total_tokens_all_models": "2113114.00", + "total_total_tokens_all_models": "2113114.0", + "avg_turn_count": "41.00", + "total_turn_count": "41.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "16", + "max_tokens": "20000", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 5000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5.4-2026-03-05/1ad18d5d-0889-4b07-9cb0-6561888d6db7_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "c8b9bbd2bc67f5ecd912a8c5c52c3e9acc6a6859e1913aabe1b62f36164d9e65", + "total_rows": 1 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5.4-2026-03-05/1ad18d5d-0889-4b07-9cb0-6561888d6db7_samples.jsonl b/data/hle/openai/gpt-5.4-2026-03-05/1ad18d5d-0889-4b07-9cb0-6561888d6db7_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..211c64d825b3bfc31956327b85cca9c7710c57bf --- /dev/null +++ b/data/hle/openai/gpt-5.4-2026-03-05/1ad18d5d-0889-4b07-9cb0-6561888d6db7_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c8b9bbd2bc67f5ecd912a8c5c52c3e9acc6a6859e1913aabe1b62f36164d9e65 +size 2781388 diff --git a/data/hle/openai/gpt-5.4-2026-03-05/1d49004f-6464-4b80-9cf6-c6b89c46c22d.json b/data/hle/openai/gpt-5.4-2026-03-05/1d49004f-6464-4b80-9cf6-c6b89c46c22d.json new file mode 100644 index 0000000000000000000000000000000000000000..17a09b1ca4f23a6ab7542d5838fa3a59843f0209 --- /dev/null +++ b/data/hle/openai/gpt-5.4-2026-03-05/1d49004f-6464-4b80-9cf6-c6b89c46c22d.json @@ -0,0 +1,112 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5.4-2026-03-05/1786164185.193253", + "evaluation_timestamp": "1778506822.0", + "retrieved_timestamp": "1786164185.193253", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "mean on hle/S-adaptive/+3ep/b30602d1 for scorer _llm_grader", + "source_data": { + "dataset_name": "hle", + "source_type": "hf_dataset", + "hf_repo": "cais/hle", + "samples_number": 2, + "sample_ids": [ + "hle_4485802f", + "hle_5d49423c" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1778506822.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 1.0, + "uncertainty": { + "num_samples": 1 + }, + "details": { + "total_matched_trajectories": "1", + "stopping_reason_count_repetition_guard": "1", + "avg_total_tokens_target_model": "87280.00", + "total_total_tokens_target_model": "87280.0", + "avg_total_tokens_other_models": "10105.00", + "total_total_tokens_other_models": "10105.0", + "avg_total_tokens_all_models": "97385.00", + "total_total_tokens_all_models": "97385.0", + "avg_turn_count": "5.00", + "total_turn_count": "5.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 5000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5.4-2026-03-05/1d49004f-6464-4b80-9cf6-c6b89c46c22d_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "4a28c4e2d5bc6c8dcfbeac2ebaeea7be5515d135fa2949e727a5e7ce0336f443", + "total_rows": 1 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5.4-2026-03-05/1d49004f-6464-4b80-9cf6-c6b89c46c22d_samples.jsonl b/data/hle/openai/gpt-5.4-2026-03-05/1d49004f-6464-4b80-9cf6-c6b89c46c22d_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..a71f76c06904c99002f0fa4c45965f1f85021974 --- /dev/null +++ b/data/hle/openai/gpt-5.4-2026-03-05/1d49004f-6464-4b80-9cf6-c6b89c46c22d_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4a28c4e2d5bc6c8dcfbeac2ebaeea7be5515d135fa2949e727a5e7ce0336f443 +size 21909 diff --git a/data/hle/openai/gpt-5.4-2026-03-05/1e3a9d3c-7b9c-4e13-b9a6-c218ca3274de.json b/data/hle/openai/gpt-5.4-2026-03-05/1e3a9d3c-7b9c-4e13-b9a6-c218ca3274de.json new file mode 100644 index 0000000000000000000000000000000000000000..df6610a348ff882d5c172d17f0abf69ac9565048 --- /dev/null +++ b/data/hle/openai/gpt-5.4-2026-03-05/1e3a9d3c-7b9c-4e13-b9a6-c218ca3274de.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5.4-2026-03-05/1786168825.615775", + "evaluation_timestamp": "1782983153.0", + "retrieved_timestamp": "1786168825.615775", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5.4-2026-03-05/1e3a9d3c-7b9c-4e13-b9a6-c218ca3274de_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "4f82ad83981cf16e17dd497332b8eb4b2f8b8955e1114f2f1d0b863d0bbd3b37", + "total_rows": 1 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5.4-2026-03-05/1e3a9d3c-7b9c-4e13-b9a6-c218ca3274de_samples.jsonl b/data/hle/openai/gpt-5.4-2026-03-05/1e3a9d3c-7b9c-4e13-b9a6-c218ca3274de_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..60d901edaedc05e6557fefc024c392646553d599 --- /dev/null +++ b/data/hle/openai/gpt-5.4-2026-03-05/1e3a9d3c-7b9c-4e13-b9a6-c218ca3274de_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f82ad83981cf16e17dd497332b8eb4b2f8b8955e1114f2f1d0b863d0bbd3b37 +size 4777057 diff --git a/data/hle/openai/gpt-5.4-2026-03-05/309f7226-1e28-46ef-aea4-c25ee1890b5d.json b/data/hle/openai/gpt-5.4-2026-03-05/309f7226-1e28-46ef-aea4-c25ee1890b5d.json new file mode 100644 index 0000000000000000000000000000000000000000..e6d3c61ed5b978f0140cef9bc3c06635d62ae8ab --- /dev/null +++ b/data/hle/openai/gpt-5.4-2026-03-05/309f7226-1e28-46ef-aea4-c25ee1890b5d.json @@ -0,0 +1,111 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5.4-2026-03-05/1786164153.888114", + "evaluation_timestamp": "1778489058.0", + "retrieved_timestamp": "1786164153.888114", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "mean on hle/S-adaptive/+1ep/4b073728 for scorer _llm_grader", + "source_data": { + "dataset_name": "hle", + "source_type": "hf_dataset", + "hf_repo": "cais/hle", + "samples_number": 1, + "sample_ids": [ + "hle_b6cec502" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1778489058.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.0, + "uncertainty": { + "num_samples": 1 + }, + "details": { + "total_matched_trajectories": "1", + "stopping_reason_count_repetition_guard": "1", + "avg_total_tokens_target_model": "67927.00", + "total_total_tokens_target_model": "67927.0", + "avg_total_tokens_other_models": "7034.00", + "total_total_tokens_other_models": "7034.0", + "avg_total_tokens_all_models": "74961.00", + "total_total_tokens_all_models": "74961.0", + "avg_turn_count": "5.00", + "total_turn_count": "5.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 5000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5.4-2026-03-05/309f7226-1e28-46ef-aea4-c25ee1890b5d_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "b70b09e2b7fdc85388afc5738e2ebc9d0579d6e5fb8adb13f3ff1664536b868a", + "total_rows": 1 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5.4-2026-03-05/309f7226-1e28-46ef-aea4-c25ee1890b5d_samples.jsonl b/data/hle/openai/gpt-5.4-2026-03-05/309f7226-1e28-46ef-aea4-c25ee1890b5d_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..cb7901da84a617d340f0296d5984d8ea6ddf5c02 --- /dev/null +++ b/data/hle/openai/gpt-5.4-2026-03-05/309f7226-1e28-46ef-aea4-c25ee1890b5d_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b70b09e2b7fdc85388afc5738e2ebc9d0579d6e5fb8adb13f3ff1664536b868a +size 19983 diff --git a/data/hle/openai/gpt-5.4-2026-03-05/310455ee-ad15-4b22-b9aa-7159c0693730.json b/data/hle/openai/gpt-5.4-2026-03-05/310455ee-ad15-4b22-b9aa-7159c0693730.json new file mode 100644 index 0000000000000000000000000000000000000000..ef623b1cd9143758b05b7dc8093a0df3cc81bfd3 --- /dev/null +++ b/data/hle/openai/gpt-5.4-2026-03-05/310455ee-ad15-4b22-b9aa-7159c0693730.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5.4-2026-03-05/1786168821.749228", + "evaluation_timestamp": "1776827375.0", + "retrieved_timestamp": "1786168821.749228", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5.4-2026-03-05/310455ee-ad15-4b22-b9aa-7159c0693730_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "a52c5eca3af7a933efd17a9b7cb2737feddfcc75c12e5dcc319e4df6cbb4c6d1", + "total_rows": 80 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5.4-2026-03-05/310455ee-ad15-4b22-b9aa-7159c0693730_samples.jsonl b/data/hle/openai/gpt-5.4-2026-03-05/310455ee-ad15-4b22-b9aa-7159c0693730_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..325f5ccfce92215722f3dfcf35b88914b9084add --- /dev/null +++ b/data/hle/openai/gpt-5.4-2026-03-05/310455ee-ad15-4b22-b9aa-7159c0693730_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a52c5eca3af7a933efd17a9b7cb2737feddfcc75c12e5dcc319e4df6cbb4c6d1 +size 1549897 diff --git a/data/hle/openai/gpt-5.4-2026-03-05/3f49b209-fa15-491e-a6bf-9cd0e2595056.json b/data/hle/openai/gpt-5.4-2026-03-05/3f49b209-fa15-491e-a6bf-9cd0e2595056.json new file mode 100644 index 0000000000000000000000000000000000000000..c3d6426d4dc001c6d4ecb095c4938ec9c273463d --- /dev/null +++ b/data/hle/openai/gpt-5.4-2026-03-05/3f49b209-fa15-491e-a6bf-9cd0e2595056.json @@ -0,0 +1,112 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5.4-2026-03-05/1786164169.981061", + "evaluation_timestamp": "1778937369.0", + "retrieved_timestamp": "1786164169.981061", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "mean on hle/S-adaptive/+2ep/70f636a7 for scorer _llm_grader", + "source_data": { + "dataset_name": "hle", + "source_type": "hf_dataset", + "hf_repo": "cais/hle", + "samples_number": 1, + "sample_ids": [ + "hle_5d49423c" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1778937369.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 1.0, + "uncertainty": { + "num_samples": 1 + }, + "details": { + "total_matched_trajectories": "1", + "stopping_reason_count_repetition_guard": "1", + "avg_total_tokens_target_model": "96925.00", + "total_total_tokens_target_model": "96925.0", + "avg_total_tokens_other_models": "9373.00", + "total_total_tokens_other_models": "9373.0", + "avg_total_tokens_all_models": "106298.00", + "total_total_tokens_all_models": "106298.0", + "avg_turn_count": "5.00", + "total_turn_count": "5.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "8", + "max_tokens": "32000", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 5000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "8", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5.4-2026-03-05/3f49b209-fa15-491e-a6bf-9cd0e2595056_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "1031b5f5447ff0944b79f5040353b2dc401835f217e93077db66d7e099aa990f", + "total_rows": 1 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5.4-2026-03-05/3f49b209-fa15-491e-a6bf-9cd0e2595056_samples.jsonl b/data/hle/openai/gpt-5.4-2026-03-05/3f49b209-fa15-491e-a6bf-9cd0e2595056_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..417ce9cd1796ec3667b0cc085718fcbfafc242ef --- /dev/null +++ b/data/hle/openai/gpt-5.4-2026-03-05/3f49b209-fa15-491e-a6bf-9cd0e2595056_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1031b5f5447ff0944b79f5040353b2dc401835f217e93077db66d7e099aa990f +size 21664 diff --git a/data/hle/openai/gpt-5.4-2026-03-05/452dff3c-19ac-4851-8802-dc55603a2f2d.json b/data/hle/openai/gpt-5.4-2026-03-05/452dff3c-19ac-4851-8802-dc55603a2f2d.json new file mode 100644 index 0000000000000000000000000000000000000000..839ee6e37d58be7063710e1f881f3f2155642b90 --- /dev/null +++ b/data/hle/openai/gpt-5.4-2026-03-05/452dff3c-19ac-4851-8802-dc55603a2f2d.json @@ -0,0 +1,112 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5.4-2026-03-05/1786164194.772103", + "evaluation_timestamp": "1783093658.0", + "retrieved_timestamp": "1786164194.772103", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "mean on hle/S-adaptive/+4ep/ea1d0e0e for scorer _llm_grader", + "source_data": { + "dataset_name": "hle", + "source_type": "hf_dataset", + "hf_repo": "cais/hle", + "samples_number": 1, + "sample_ids": [ + "hle_0ec247c5" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1783093658.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.5, + "uncertainty": { + "num_samples": 4 + }, + "details": { + "total_matched_trajectories": "4", + "stopping_reason_count_repetition_guard": "4", + "avg_total_tokens_target_model": "377231.25", + "total_total_tokens_target_model": "1508925.0", + "avg_total_tokens_other_models": "5328.25", + "total_total_tokens_other_models": "21313.0", + "avg_total_tokens_all_models": "382559.50", + "total_total_tokens_all_models": "1530238.0", + "avg_turn_count": "7.50", + "total_turn_count": "30.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "1", + "timeout": "1200", + "attempt_timeout": "3600", + "max_connections": "16", + "max_tokens": "64000", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 5000000 + }, + "sandbox": {}, + "max_attempts": 1 + }, + "additional_details": { + "max_retries": "1", + "timeout": "1200", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5.4-2026-03-05/452dff3c-19ac-4851-8802-dc55603a2f2d_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "fd8f6cd2d3d9bb1417af21a6030231e144b69b2a5d683e8bce7342337606141b", + "total_rows": 4 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5.4-2026-03-05/452dff3c-19ac-4851-8802-dc55603a2f2d_samples.jsonl b/data/hle/openai/gpt-5.4-2026-03-05/452dff3c-19ac-4851-8802-dc55603a2f2d_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..54c5f26461e46c7551b4edfb47acfb74f460935c --- /dev/null +++ b/data/hle/openai/gpt-5.4-2026-03-05/452dff3c-19ac-4851-8802-dc55603a2f2d_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fd8f6cd2d3d9bb1417af21a6030231e144b69b2a5d683e8bce7342337606141b +size 95606 diff --git a/data/hle/openai/gpt-5.4-2026-03-05/48db0c23-1cc2-4411-bedb-468bae4f87ae.json b/data/hle/openai/gpt-5.4-2026-03-05/48db0c23-1cc2-4411-bedb-468bae4f87ae.json new file mode 100644 index 0000000000000000000000000000000000000000..9f2dde7897c0c43ea03aef6701cda531953dfaff --- /dev/null +++ b/data/hle/openai/gpt-5.4-2026-03-05/48db0c23-1cc2-4411-bedb-468bae4f87ae.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5.4-2026-03-05/1786166828.46761", + "evaluation_timestamp": "1775078457.0", + "retrieved_timestamp": "1786166828.46761", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.202,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5.4-2026-03-05/48db0c23-1cc2-4411-bedb-468bae4f87ae_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "b0f7a036fca314ae7186a6a062c67f598734d3a12eb590c31b764f73412fb9b7", + "total_rows": 100 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5.4-2026-03-05/48db0c23-1cc2-4411-bedb-468bae4f87ae_samples.jsonl b/data/hle/openai/gpt-5.4-2026-03-05/48db0c23-1cc2-4411-bedb-468bae4f87ae_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..36b0a008d4fdfd31b54cb0506b34e6667c0de8f3 --- /dev/null +++ b/data/hle/openai/gpt-5.4-2026-03-05/48db0c23-1cc2-4411-bedb-468bae4f87ae_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b0f7a036fca314ae7186a6a062c67f598734d3a12eb590c31b764f73412fb9b7 +size 3701245 diff --git a/data/hle/openai/gpt-5.4-2026-03-05/5071767a-4ea9-4304-b2af-b07de8c5ff00.json b/data/hle/openai/gpt-5.4-2026-03-05/5071767a-4ea9-4304-b2af-b07de8c5ff00.json new file mode 100644 index 0000000000000000000000000000000000000000..283cee2ce8cf18c38c268ac7ad51252f17950afe --- /dev/null +++ b/data/hle/openai/gpt-5.4-2026-03-05/5071767a-4ea9-4304-b2af-b07de8c5ff00.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5.4-2026-03-05/1786168740.410677", + "evaluation_timestamp": "1776686673.0", + "retrieved_timestamp": "1786168740.410677", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5.4-2026-03-05/5071767a-4ea9-4304-b2af-b07de8c5ff00_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "3a7095956248ca437894249578615689e4fe5b4c8b8eb68506acf85e16e09233", + "total_rows": 25 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5.4-2026-03-05/5071767a-4ea9-4304-b2af-b07de8c5ff00_samples.jsonl b/data/hle/openai/gpt-5.4-2026-03-05/5071767a-4ea9-4304-b2af-b07de8c5ff00_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..0d3eb0fa587987af5532e79fdf94ae4ecd76e050 --- /dev/null +++ b/data/hle/openai/gpt-5.4-2026-03-05/5071767a-4ea9-4304-b2af-b07de8c5ff00_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3a7095956248ca437894249578615689e4fe5b4c8b8eb68506acf85e16e09233 +size 480672 diff --git a/data/hle/openai/gpt-5.4-2026-03-05/52ecf5dd-5ea1-4182-92e4-80503d2ad5bd.json b/data/hle/openai/gpt-5.4-2026-03-05/52ecf5dd-5ea1-4182-92e4-80503d2ad5bd.json new file mode 100644 index 0000000000000000000000000000000000000000..976421efca88a77c2f9846ae45c37666e74f5b1c --- /dev/null +++ b/data/hle/openai/gpt-5.4-2026-03-05/52ecf5dd-5ea1-4182-92e4-80503d2ad5bd.json @@ -0,0 +1,122 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5.4-2026-03-05/1786164178.461411", + "evaluation_timestamp": "1776686674.0", + "retrieved_timestamp": "1786164178.461411", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "mean on hle/S-adaptive/+3ep/727b518e for scorer _llm_grader", + "source_data": { + "dataset_name": "hle", + "source_type": "hf_dataset", + "hf_repo": "cais/hle", + "samples_number": 12, + "sample_ids": [ + "hle_dfdb78f5", + "hle_4485802f", + "hle_9019c023", + "hle_9282e7ae", + "hle_44007b97", + "hle_3cf4483d", + "hle_f85a0fa1", + "hle_c3ca0bda", + "hle_376b8247", + "hle_31497469", + "hle_c43a27a3", + "hle_ea41e785" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1776686674.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.4, + "uncertainty": { + "standard_error": { + "value": 0.1632993161855452 + }, + "num_samples": 26 + }, + "details": { + "total_matched_trajectories": "26", + "stopping_reason_count_repetition_guard": "26", + "avg_total_tokens_target_model": "41383.85", + "total_total_tokens_target_model": "1075980.0", + "avg_total_tokens_other_models": "4880.35", + "total_total_tokens_other_models": "126889.0", + "avg_total_tokens_all_models": "46264.19", + "total_total_tokens_all_models": "1202869.0", + "avg_turn_count": "5.58", + "total_turn_count": "145.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "3600", + "attempt_timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 5000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5.4-2026-03-05/52ecf5dd-5ea1-4182-92e4-80503d2ad5bd_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "f99f7023289596486ed8fd41e11591d5847d17853912061b225abab0facb965d", + "total_rows": 26 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5.4-2026-03-05/52ecf5dd-5ea1-4182-92e4-80503d2ad5bd_samples.jsonl b/data/hle/openai/gpt-5.4-2026-03-05/52ecf5dd-5ea1-4182-92e4-80503d2ad5bd_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..d79dea6b1d97d464d101a92128572497ad3aaf88 --- /dev/null +++ b/data/hle/openai/gpt-5.4-2026-03-05/52ecf5dd-5ea1-4182-92e4-80503d2ad5bd_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f99f7023289596486ed8fd41e11591d5847d17853912061b225abab0facb965d +size 515746 diff --git a/data/hle/openai/gpt-5.4-2026-03-05/537705ab-cbad-4eb3-a8d1-28be14d6ab03.json b/data/hle/openai/gpt-5.4-2026-03-05/537705ab-cbad-4eb3-a8d1-28be14d6ab03.json new file mode 100644 index 0000000000000000000000000000000000000000..6338cf69bc130a53af73b303d7425cb77aa52d85 --- /dev/null +++ b/data/hle/openai/gpt-5.4-2026-03-05/537705ab-cbad-4eb3-a8d1-28be14d6ab03.json @@ -0,0 +1,118 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5.4-2026-03-05/1786164156.024464", + "evaluation_timestamp": "1776430851.0", + "retrieved_timestamp": "1786164156.024464", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "mean on hle/S-adaptive/+1ep/37d87ffa for scorer _llm_grader", + "source_data": { + "dataset_name": "hle", + "source_type": "hf_dataset", + "hf_repo": "cais/hle", + "samples_number": 10, + "sample_ids": [ + "hle_e1296346", + "hle_ca6c436b", + "hle_142ffcff", + "hle_ffb12d83", + "hle_37c97863", + "hle_209b6b64", + "hle_0702accc", + "hle_16388cc8", + "hle_4418bea4", + "hle_ebd62885" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1776430851.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.1, + "uncertainty": { + "standard_error": { + "value": 0.09999999999999999 + }, + "num_samples": 10 + }, + "details": { + "total_matched_trajectories": "10", + "stopping_reason_count_repetition_guard": "10", + "avg_total_tokens_target_model": "41615.20", + "total_total_tokens_target_model": "416152.0", + "avg_total_tokens_other_models": "5157.10", + "total_total_tokens_other_models": "51571.0", + "avg_total_tokens_all_models": "46772.30", + "total_total_tokens_all_models": "467723.0", + "avg_turn_count": "6.10", + "total_turn_count": "61.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 5000000 + }, + "sandbox": {} + }, + "additional_details": { + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5.4-2026-03-05/537705ab-cbad-4eb3-a8d1-28be14d6ab03_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "d0f5df344de2caa1c1e02bd4f578cea8ba8df0db9b30f55bad044bfe01cc4d1d", + "total_rows": 10 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5.4-2026-03-05/537705ab-cbad-4eb3-a8d1-28be14d6ab03_samples.jsonl b/data/hle/openai/gpt-5.4-2026-03-05/537705ab-cbad-4eb3-a8d1-28be14d6ab03_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..d990efb0a2ae850467c3f4ab7af5bfc84e543e2c --- /dev/null +++ b/data/hle/openai/gpt-5.4-2026-03-05/537705ab-cbad-4eb3-a8d1-28be14d6ab03_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d0f5df344de2caa1c1e02bd4f578cea8ba8df0db9b30f55bad044bfe01cc4d1d +size 206862 diff --git a/data/hle/openai/gpt-5.4-2026-03-05/571e9780-4f86-4b6d-af8c-d1ce05168a03.json b/data/hle/openai/gpt-5.4-2026-03-05/571e9780-4f86-4b6d-af8c-d1ce05168a03.json new file mode 100644 index 0000000000000000000000000000000000000000..e41cf51753a6b51e5c7e17cabb0e67bed8a7e9ef --- /dev/null +++ b/data/hle/openai/gpt-5.4-2026-03-05/571e9780-4f86-4b6d-af8c-d1ce05168a03.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5.4-2026-03-05/1786168800.988302", + "evaluation_timestamp": "1776335800.0", + "retrieved_timestamp": "1786168800.988302", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5.4-2026-03-05/571e9780-4f86-4b6d-af8c-d1ce05168a03_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "14b0b0996178aa1adfa56d744c7085594af3ae97bf22ca896be95b6d1cedc559", + "total_rows": 20 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5.4-2026-03-05/571e9780-4f86-4b6d-af8c-d1ce05168a03_samples.jsonl b/data/hle/openai/gpt-5.4-2026-03-05/571e9780-4f86-4b6d-af8c-d1ce05168a03_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..981ede460503de49eb96e272da50b717f5bb4aca --- /dev/null +++ b/data/hle/openai/gpt-5.4-2026-03-05/571e9780-4f86-4b6d-af8c-d1ce05168a03_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:14b0b0996178aa1adfa56d744c7085594af3ae97bf22ca896be95b6d1cedc559 +size 340370 diff --git a/data/hle/openai/gpt-5.4-2026-03-05/5a58fe34-fa18-421b-ade0-ac218e046254.json b/data/hle/openai/gpt-5.4-2026-03-05/5a58fe34-fa18-421b-ade0-ac218e046254.json new file mode 100644 index 0000000000000000000000000000000000000000..667509a37960a8ef526bd6b07c6ba0ff56fe381d --- /dev/null +++ b/data/hle/openai/gpt-5.4-2026-03-05/5a58fe34-fa18-421b-ade0-ac218e046254.json @@ -0,0 +1,110 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5.4-2026-03-05/1786164166.46334", + "evaluation_timestamp": "1776955657.0", + "retrieved_timestamp": "1786164166.46334", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "mean on hle/S-adaptive/+1ep/2008043b for scorer _llm_grader", + "source_data": { + "dataset_name": "hle", + "source_type": "hf_dataset", + "hf_repo": "cais/hle", + "samples_number": 3, + "sample_ids": [ + "hle_b6cec502", + "hle_9255683e", + "hle_c2ff3a43" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1776955657.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 1.0, + "uncertainty": { + "num_samples": 1 + }, + "details": { + "total_matched_trajectories": "1", + "stopping_reason_count_repetition_guard": "1", + "avg_total_tokens_target_model": "73251.00", + "total_total_tokens_target_model": "73251.0", + "avg_total_tokens_other_models": "7242.00", + "total_total_tokens_other_models": "7242.0", + "avg_total_tokens_all_models": "80493.00", + "total_total_tokens_all_models": "80493.0", + "avg_turn_count": "5.00", + "total_turn_count": "5.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 5000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5.4-2026-03-05/5a58fe34-fa18-421b-ade0-ac218e046254_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "601f07880b1b3db5b3db0e43f323f6763d8d8f23805c164ccff22bcbff474d10", + "total_rows": 1 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5.4-2026-03-05/5a58fe34-fa18-421b-ade0-ac218e046254_samples.jsonl b/data/hle/openai/gpt-5.4-2026-03-05/5a58fe34-fa18-421b-ade0-ac218e046254_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..e37b62f504542e38e1f62cc22de7950ce3730090 --- /dev/null +++ b/data/hle/openai/gpt-5.4-2026-03-05/5a58fe34-fa18-421b-ade0-ac218e046254_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:601f07880b1b3db5b3db0e43f323f6763d8d8f23805c164ccff22bcbff474d10 +size 21811 diff --git a/data/hle/openai/gpt-5.4-2026-03-05/5fb4fc84-a39d-4859-984e-dea1c094cbb2.json b/data/hle/openai/gpt-5.4-2026-03-05/5fb4fc84-a39d-4859-984e-dea1c094cbb2.json new file mode 100644 index 0000000000000000000000000000000000000000..0b71ffea3b9fd8a6d70846ffa1ae1f821d70a90b --- /dev/null +++ b/data/hle/openai/gpt-5.4-2026-03-05/5fb4fc84-a39d-4859-984e-dea1c094cbb2.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5.4-2026-03-05/1786166806.376713", + "evaluation_timestamp": "1776955657.0", + "retrieved_timestamp": "1786166806.376713", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5.4-2026-03-05/5fb4fc84-a39d-4859-984e-dea1c094cbb2_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "fd530300806e7c15782278a1a57c423114aa53e5b1e6c8ff7f74d2893cf7c3a5", + "total_rows": 3 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5.4-2026-03-05/5fb4fc84-a39d-4859-984e-dea1c094cbb2_samples.jsonl b/data/hle/openai/gpt-5.4-2026-03-05/5fb4fc84-a39d-4859-984e-dea1c094cbb2_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..42133525ab5f03c2abdcda0f65564c31ef7fa964 --- /dev/null +++ b/data/hle/openai/gpt-5.4-2026-03-05/5fb4fc84-a39d-4859-984e-dea1c094cbb2_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fd530300806e7c15782278a1a57c423114aa53e5b1e6c8ff7f74d2893cf7c3a5 +size 119899 diff --git a/data/hle/openai/gpt-5.4-2026-03-05/60396f10-87cd-4e98-82ce-44de056c9c91.json b/data/hle/openai/gpt-5.4-2026-03-05/60396f10-87cd-4e98-82ce-44de056c9c91.json new file mode 100644 index 0000000000000000000000000000000000000000..53c8d6c21241634c65cf0b36346e8435d17530b1 --- /dev/null +++ b/data/hle/openai/gpt-5.4-2026-03-05/60396f10-87cd-4e98-82ce-44de056c9c91.json @@ -0,0 +1,112 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5.4-2026-03-05/1786164196.375925", + "evaluation_timestamp": "1778539230.0", + "retrieved_timestamp": "1786164196.375925", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "mean on hle/S-adaptive/+4ep/9410e543 for scorer _llm_grader", + "source_data": { + "dataset_name": "hle", + "source_type": "hf_dataset", + "hf_repo": "cais/hle", + "samples_number": 2, + "sample_ids": [ + "hle_dfdb78f5", + "hle_1d580261" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1778539230.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 1.0, + "uncertainty": { + "num_samples": 1 + }, + "details": { + "total_matched_trajectories": "1", + "stopping_reason_count_repetition_guard": "1", + "avg_total_tokens_target_model": "83362.00", + "total_total_tokens_target_model": "83362.0", + "avg_total_tokens_other_models": "9210.00", + "total_total_tokens_other_models": "9210.0", + "avg_total_tokens_all_models": "92572.00", + "total_total_tokens_all_models": "92572.0", + "avg_turn_count": "5.00", + "total_turn_count": "5.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 5000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5.4-2026-03-05/60396f10-87cd-4e98-82ce-44de056c9c91_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "f9b26a0aac7f517c1a965991b631c7647fb0127126a8478ed259d930ee6b5594", + "total_rows": 1 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5.4-2026-03-05/60396f10-87cd-4e98-82ce-44de056c9c91_samples.jsonl b/data/hle/openai/gpt-5.4-2026-03-05/60396f10-87cd-4e98-82ce-44de056c9c91_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..c94aeaad6a703dbab6e76d1b0b932a0b232f7a6b --- /dev/null +++ b/data/hle/openai/gpt-5.4-2026-03-05/60396f10-87cd-4e98-82ce-44de056c9c91_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f9b26a0aac7f517c1a965991b631c7647fb0127126a8478ed259d930ee6b5594 +size 22778 diff --git a/data/hle/openai/gpt-5.4-2026-03-05/661467b7-bc93-4dc8-96cb-6796b7a2ed40.json b/data/hle/openai/gpt-5.4-2026-03-05/661467b7-bc93-4dc8-96cb-6796b7a2ed40.json new file mode 100644 index 0000000000000000000000000000000000000000..cbeed7cf2545c90c04d5e5901ff79d7483245f92 --- /dev/null +++ b/data/hle/openai/gpt-5.4-2026-03-05/661467b7-bc93-4dc8-96cb-6796b7a2ed40.json @@ -0,0 +1,110 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5.4-2026-03-05/1786164188.269534", + "evaluation_timestamp": "1776885996.0", + "retrieved_timestamp": "1786164188.269534", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "mean on hle/S-adaptive/+3ep/795d2add for scorer _llm_grader", + "source_data": { + "dataset_name": "hle", + "source_type": "hf_dataset", + "hf_repo": "cais/hle", + "samples_number": 3, + "sample_ids": [ + "hle_4485802f", + "hle_b6cec502", + "hle_5d49423c" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1776885996.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.0, + "uncertainty": { + "num_samples": 2 + }, + "details": { + "total_matched_trajectories": "2", + "stopping_reason_count_repetition_guard": "2", + "avg_total_tokens_target_model": "72055.00", + "total_total_tokens_target_model": "144110.0", + "avg_total_tokens_other_models": "6241.00", + "total_total_tokens_other_models": "12482.0", + "avg_total_tokens_all_models": "78296.00", + "total_total_tokens_all_models": "156592.0", + "avg_turn_count": "6.50", + "total_turn_count": "13.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 5000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5.4-2026-03-05/661467b7-bc93-4dc8-96cb-6796b7a2ed40_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "ea9bb910a71aff1c20750408e953498ab80806af40eb537b1283a5355ed4a581", + "total_rows": 2 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5.4-2026-03-05/661467b7-bc93-4dc8-96cb-6796b7a2ed40_samples.jsonl b/data/hle/openai/gpt-5.4-2026-03-05/661467b7-bc93-4dc8-96cb-6796b7a2ed40_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..1eeabcc3cbe8e1543ac162f720150fd0b06a3532 --- /dev/null +++ b/data/hle/openai/gpt-5.4-2026-03-05/661467b7-bc93-4dc8-96cb-6796b7a2ed40_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ea9bb910a71aff1c20750408e953498ab80806af40eb537b1283a5355ed4a581 +size 44362 diff --git a/data/hle/openai/gpt-5.4-2026-03-05/6ab0c03f-7fcd-45b7-95e5-cbb6be9a5700.json b/data/hle/openai/gpt-5.4-2026-03-05/6ab0c03f-7fcd-45b7-95e5-cbb6be9a5700.json new file mode 100644 index 0000000000000000000000000000000000000000..5359f73c5b398450cb7b7e6a0abd9642b009e6b3 --- /dev/null +++ b/data/hle/openai/gpt-5.4-2026-03-05/6ab0c03f-7fcd-45b7-95e5-cbb6be9a5700.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5.4-2026-03-05/1786166797.067278", + "evaluation_timestamp": "1776677380.0", + "retrieved_timestamp": "1786166797.067278", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5.4-2026-03-05/6ab0c03f-7fcd-45b7-95e5-cbb6be9a5700_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "64ae6bebaa89e25355aa8c97de206b27add6fb4222990c0ab90f7059802f3b4c", + "total_rows": 2 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5.4-2026-03-05/6ab0c03f-7fcd-45b7-95e5-cbb6be9a5700_samples.jsonl b/data/hle/openai/gpt-5.4-2026-03-05/6ab0c03f-7fcd-45b7-95e5-cbb6be9a5700_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..0eb1e85a94e3754a15de2f055b8004f450051192 --- /dev/null +++ b/data/hle/openai/gpt-5.4-2026-03-05/6ab0c03f-7fcd-45b7-95e5-cbb6be9a5700_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:64ae6bebaa89e25355aa8c97de206b27add6fb4222990c0ab90f7059802f3b4c +size 119416 diff --git a/data/hle/openai/gpt-5.4-2026-03-05/6ae4090d-b98a-4fc6-b498-c4b6e9b26c29.json b/data/hle/openai/gpt-5.4-2026-03-05/6ae4090d-b98a-4fc6-b498-c4b6e9b26c29.json new file mode 100644 index 0000000000000000000000000000000000000000..f123be651135ac2987f8d6928957023aa85b6bf5 --- /dev/null +++ b/data/hle/openai/gpt-5.4-2026-03-05/6ae4090d-b98a-4fc6-b498-c4b6e9b26c29.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5.4-2026-03-05/1786166822.153239", + "evaluation_timestamp": "1775646127.0", + "retrieved_timestamp": "1786166822.153239", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5.4-2026-03-05/6ae4090d-b98a-4fc6-b498-c4b6e9b26c29_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "b635c8777abb7cf353af1a2687c91af78ef29e5178b1019d5688f244bca74aee", + "total_rows": 19 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5.4-2026-03-05/6ae4090d-b98a-4fc6-b498-c4b6e9b26c29_samples.jsonl b/data/hle/openai/gpt-5.4-2026-03-05/6ae4090d-b98a-4fc6-b498-c4b6e9b26c29_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..f041c16b6c0bd828274359fd48eb50755a0e05ca --- /dev/null +++ b/data/hle/openai/gpt-5.4-2026-03-05/6ae4090d-b98a-4fc6-b498-c4b6e9b26c29_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b635c8777abb7cf353af1a2687c91af78ef29e5178b1019d5688f244bca74aee +size 669628 diff --git a/data/hle/openai/gpt-5.4-2026-03-05/6b91623e-d0b7-49b2-8eca-785e4f94d81d.json b/data/hle/openai/gpt-5.4-2026-03-05/6b91623e-d0b7-49b2-8eca-785e4f94d81d.json new file mode 100644 index 0000000000000000000000000000000000000000..a438e8658ade314061e6ebc875252c29277c13c7 --- /dev/null +++ b/data/hle/openai/gpt-5.4-2026-03-05/6b91623e-d0b7-49b2-8eca-785e4f94d81d.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5.4-2026-03-05/1786166810.507392", + "evaluation_timestamp": "1782983153.0", + "retrieved_timestamp": "1786166810.507392", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5.4-2026-03-05/6b91623e-d0b7-49b2-8eca-785e4f94d81d_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "72b1ed42ef48c6d229752360a7b65eea8e68a4c3060316418b68ac2260515113", + "total_rows": 1 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5.4-2026-03-05/6b91623e-d0b7-49b2-8eca-785e4f94d81d_samples.jsonl b/data/hle/openai/gpt-5.4-2026-03-05/6b91623e-d0b7-49b2-8eca-785e4f94d81d_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..f002d6da05427ce81f86eb8e3e5aa9993ae29f8c --- /dev/null +++ b/data/hle/openai/gpt-5.4-2026-03-05/6b91623e-d0b7-49b2-8eca-785e4f94d81d_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:72b1ed42ef48c6d229752360a7b65eea8e68a4c3060316418b68ac2260515113 +size 1049869 diff --git a/data/hle/openai/gpt-5.4-2026-03-05/6d37d0d4-d1aa-428a-853e-d6506f3cc059.json b/data/hle/openai/gpt-5.4-2026-03-05/6d37d0d4-d1aa-428a-853e-d6506f3cc059.json new file mode 100644 index 0000000000000000000000000000000000000000..bffbe9b1d634d6929166641d9f860fd67cbd2ad7 --- /dev/null +++ b/data/hle/openai/gpt-5.4-2026-03-05/6d37d0d4-d1aa-428a-853e-d6506f3cc059.json @@ -0,0 +1,112 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5.4-2026-03-05/1786164172.079358", + "evaluation_timestamp": "1780489238.0", + "retrieved_timestamp": "1786164172.079358", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "mean on hle/S-adaptive/+2ep/a4f1cf4f for scorer _llm_grader", + "source_data": { + "dataset_name": "hle", + "source_type": "hf_dataset", + "hf_repo": "cais/hle", + "samples_number": 1, + "sample_ids": [ + "hle_1d580261" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1780489238.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.5, + "uncertainty": { + "num_samples": 2 + }, + "details": { + "total_matched_trajectories": "2", + "stopping_reason_count_repetition_guard": "2", + "avg_total_tokens_target_model": "1313902.00", + "total_total_tokens_target_model": "2627804.0", + "avg_total_tokens_other_models": "7158.50", + "total_total_tokens_other_models": "14317.0", + "avg_total_tokens_all_models": "1321060.50", + "total_total_tokens_all_models": "2642121.0", + "avg_turn_count": "26.50", + "total_turn_count": "53.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "3", + "timeout": "1800", + "attempt_timeout": "1800", + "max_connections": "16", + "max_tokens": "20000", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 5000000 + }, + "sandbox": {}, + "max_attempts": 3 + }, + "additional_details": { + "max_retries": "3", + "timeout": "1800", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5.4-2026-03-05/6d37d0d4-d1aa-428a-853e-d6506f3cc059_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "2a69318a11cb32c54f2e78b24f88b730e55e285ac453be5fb806c4d91d8bf945", + "total_rows": 2 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5.4-2026-03-05/6d37d0d4-d1aa-428a-853e-d6506f3cc059_samples.jsonl b/data/hle/openai/gpt-5.4-2026-03-05/6d37d0d4-d1aa-428a-853e-d6506f3cc059_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..892bc0b664144c38364a156c2718086470d76ead --- /dev/null +++ b/data/hle/openai/gpt-5.4-2026-03-05/6d37d0d4-d1aa-428a-853e-d6506f3cc059_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2a69318a11cb32c54f2e78b24f88b730e55e285ac453be5fb806c4d91d8bf945 +size 4049274 diff --git a/data/hle/openai/gpt-5.4-2026-03-05/6d70b13f-742d-4203-bb3a-91b82d4dfdc9.json b/data/hle/openai/gpt-5.4-2026-03-05/6d70b13f-742d-4203-bb3a-91b82d4dfdc9.json new file mode 100644 index 0000000000000000000000000000000000000000..d3fb6b9d3cb359777cddaa8a83dd9c4cc545ee91 --- /dev/null +++ b/data/hle/openai/gpt-5.4-2026-03-05/6d70b13f-742d-4203-bb3a-91b82d4dfdc9.json @@ -0,0 +1,133 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5.4-2026-03-05/1786164164.796807", + "evaluation_timestamp": "1776880341.0", + "retrieved_timestamp": "1786164164.796807", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "mean on hle/S-adaptive/+1ep/9fc2e053 for scorer _llm_grader", + "source_data": { + "dataset_name": "hle", + "source_type": "hf_dataset", + "hf_repo": "cais/hle", + "samples_number": 23, + "sample_ids": [ + "hle_d7b7c3dc", + "hle_9255683e", + "hle_34c69126", + "hle_226e134b", + "hle_c1158fb8", + "hle_d0ea1019", + "hle_fcd42d0c", + "hle_b33012cf", + "hle_3697d2bc", + "hle_5e349785", + "hle_8b0384c4", + "hle_ee938b26", + "hle_d190d8c2", + "hle_b26bf8aa", + "hle_141cd3ef", + "hle_5671e386", + "hle_a634e89e", + "hle_670287ad", + "hle_298b5616", + "hle_e14d8c23", + "hle_510b6040", + "hle_c6adc1b4", + "hle_aa343d51" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1776880341.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.3181818181818182, + "uncertainty": { + "standard_error": { + "value": 0.10163945352271771 + }, + "num_samples": 22 + }, + "details": { + "total_matched_trajectories": "22", + "stopping_reason_count_repetition_guard": "22", + "avg_total_tokens_target_model": "30838.73", + "total_total_tokens_target_model": "678452.0", + "avg_total_tokens_other_models": "5213.82", + "total_total_tokens_other_models": "114704.0", + "avg_total_tokens_all_models": "36052.55", + "total_total_tokens_all_models": "793156.0", + "avg_turn_count": "5.27", + "total_turn_count": "116.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 5000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5.4-2026-03-05/6d70b13f-742d-4203-bb3a-91b82d4dfdc9_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "54d3cabd110efb1306cd991de5ec70c7d75e8aa9ce3b40681baf97274d2048d6", + "total_rows": 22 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5.4-2026-03-05/6d70b13f-742d-4203-bb3a-91b82d4dfdc9_samples.jsonl b/data/hle/openai/gpt-5.4-2026-03-05/6d70b13f-742d-4203-bb3a-91b82d4dfdc9_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..2fc04908b8e81ba210f0f32d8a00c377f9568cf6 --- /dev/null +++ b/data/hle/openai/gpt-5.4-2026-03-05/6d70b13f-742d-4203-bb3a-91b82d4dfdc9_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:54d3cabd110efb1306cd991de5ec70c7d75e8aa9ce3b40681baf97274d2048d6 +size 425119 diff --git a/data/hle/openai/gpt-5.4-2026-03-05/6d93e563-0112-4fa6-bef1-4726b717409e.json b/data/hle/openai/gpt-5.4-2026-03-05/6d93e563-0112-4fa6-bef1-4726b717409e.json new file mode 100644 index 0000000000000000000000000000000000000000..4ddbb37917f6335e89249f33d586829066ff0a1b --- /dev/null +++ b/data/hle/openai/gpt-5.4-2026-03-05/6d93e563-0112-4fa6-bef1-4726b717409e.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5.4-2026-03-05/1786168727.607398", + "evaluation_timestamp": "1776359245.0", + "retrieved_timestamp": "1786168727.607398", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5.4-2026-03-05/6d93e563-0112-4fa6-bef1-4726b717409e_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "c3388fff92479ea0498aa2f0feb9f4bdac05ed4c01b290288335226ac85ff4d0", + "total_rows": 10 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5.4-2026-03-05/6d93e563-0112-4fa6-bef1-4726b717409e_samples.jsonl b/data/hle/openai/gpt-5.4-2026-03-05/6d93e563-0112-4fa6-bef1-4726b717409e_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..50dc9e09034807b12e13735c8ac64958b97930d8 --- /dev/null +++ b/data/hle/openai/gpt-5.4-2026-03-05/6d93e563-0112-4fa6-bef1-4726b717409e_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c3388fff92479ea0498aa2f0feb9f4bdac05ed4c01b290288335226ac85ff4d0 +size 195864 diff --git a/data/hle/openai/gpt-5.4-2026-03-05/6e1f0357-f09e-4fe5-9651-0b352a4b3468.json b/data/hle/openai/gpt-5.4-2026-03-05/6e1f0357-f09e-4fe5-9651-0b352a4b3468.json new file mode 100644 index 0000000000000000000000000000000000000000..d41cb030899324a618da9c3d1798bae744195763 --- /dev/null +++ b/data/hle/openai/gpt-5.4-2026-03-05/6e1f0357-f09e-4fe5-9651-0b352a4b3468.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5.4-2026-03-05/1786166799.779958", + "evaluation_timestamp": "1775558526.0", + "retrieved_timestamp": "1786166799.779958", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.202,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5.4-2026-03-05/6e1f0357-f09e-4fe5-9651-0b352a4b3468_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "72262d2fa247b2bbf1b26912b20ebac3333b3f944604826118cf26a282e10bed", + "total_rows": 20 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5.4-2026-03-05/6e1f0357-f09e-4fe5-9651-0b352a4b3468_samples.jsonl b/data/hle/openai/gpt-5.4-2026-03-05/6e1f0357-f09e-4fe5-9651-0b352a4b3468_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..ecc00d3e151df7bb6f966d7c781a7f54825a0883 --- /dev/null +++ b/data/hle/openai/gpt-5.4-2026-03-05/6e1f0357-f09e-4fe5-9651-0b352a4b3468_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:72262d2fa247b2bbf1b26912b20ebac3333b3f944604826118cf26a282e10bed +size 676777 diff --git a/data/hle/openai/gpt-5.4-2026-03-05/78794348-7913-4462-a041-4ae6cd04ea16.json b/data/hle/openai/gpt-5.4-2026-03-05/78794348-7913-4462-a041-4ae6cd04ea16.json new file mode 100644 index 0000000000000000000000000000000000000000..0b1e38440683f314da6c5e2a16821d4d742ea135 --- /dev/null +++ b/data/hle/openai/gpt-5.4-2026-03-05/78794348-7913-4462-a041-4ae6cd04ea16.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5.4-2026-03-05/1786166804.570664", + "evaluation_timestamp": "1776686674.0", + "retrieved_timestamp": "1786166804.570664", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5.4-2026-03-05/78794348-7913-4462-a041-4ae6cd04ea16_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "6e835262bed5a5086bff611e7c5ac67510a1232d21748abe81954c6ce27a4eed", + "total_rows": 18 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5.4-2026-03-05/78794348-7913-4462-a041-4ae6cd04ea16_samples.jsonl b/data/hle/openai/gpt-5.4-2026-03-05/78794348-7913-4462-a041-4ae6cd04ea16_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..badcf8ea95fa403311cc062662a17d43ccf4f38d --- /dev/null +++ b/data/hle/openai/gpt-5.4-2026-03-05/78794348-7913-4462-a041-4ae6cd04ea16_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6e835262bed5a5086bff611e7c5ac67510a1232d21748abe81954c6ce27a4eed +size 658646 diff --git a/data/hle/openai/gpt-5.4-2026-03-05/7d17cea0-8da2-4396-ae28-42efb472d85d.json b/data/hle/openai/gpt-5.4-2026-03-05/7d17cea0-8da2-4396-ae28-42efb472d85d.json new file mode 100644 index 0000000000000000000000000000000000000000..2b6661d7dbf479f54f1f7beb3b067f4952238cfb --- /dev/null +++ b/data/hle/openai/gpt-5.4-2026-03-05/7d17cea0-8da2-4396-ae28-42efb472d85d.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5.4-2026-03-05/1786166817.842689", + "evaluation_timestamp": "1775558526.0", + "retrieved_timestamp": "1786166817.842689", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.202,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5.4-2026-03-05/7d17cea0-8da2-4396-ae28-42efb472d85d_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "c5181bb872162dfcaf41702607361d92e75656b50361490f334a152361d1b931", + "total_rows": 20 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5.4-2026-03-05/7d17cea0-8da2-4396-ae28-42efb472d85d_samples.jsonl b/data/hle/openai/gpt-5.4-2026-03-05/7d17cea0-8da2-4396-ae28-42efb472d85d_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..e29171e5ad8711a6ef388b7f711d00d5752351d8 --- /dev/null +++ b/data/hle/openai/gpt-5.4-2026-03-05/7d17cea0-8da2-4396-ae28-42efb472d85d_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c5181bb872162dfcaf41702607361d92e75656b50361490f334a152361d1b931 +size 837026 diff --git a/data/hle/openai/gpt-5.4-2026-03-05/7dc94e38-f4d0-4ee6-ab2b-2d89dcb45371.json b/data/hle/openai/gpt-5.4-2026-03-05/7dc94e38-f4d0-4ee6-ab2b-2d89dcb45371.json new file mode 100644 index 0000000000000000000000000000000000000000..4cc04a3bee83650bd2f85bc9f2e9ceab9e3d9aa9 --- /dev/null +++ b/data/hle/openai/gpt-5.4-2026-03-05/7dc94e38-f4d0-4ee6-ab2b-2d89dcb45371.json @@ -0,0 +1,112 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5.4-2026-03-05/1786164180.220921", + "evaluation_timestamp": "1782325660.0", + "retrieved_timestamp": "1786164180.220921", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "mean on hle/S-adaptive/+3ep/90b2a41a for scorer _llm_grader", + "source_data": { + "dataset_name": "hle", + "source_type": "hf_dataset", + "hf_repo": "cais/hle", + "samples_number": 1, + "sample_ids": [ + "hle_4485802f" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1782325660.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.6666666666666666, + "uncertainty": { + "num_samples": 3 + }, + "details": { + "total_matched_trajectories": "3", + "stopping_reason_count_repetition_guard": "3", + "avg_total_tokens_target_model": "368256.00", + "total_total_tokens_target_model": "1104768.0", + "avg_total_tokens_other_models": "4069.00", + "total_total_tokens_other_models": "12207.0", + "avg_total_tokens_all_models": "372325.00", + "total_total_tokens_all_models": "1116975.0", + "avg_turn_count": "16.00", + "total_turn_count": "48.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "1200", + "max_connections": "16", + "max_tokens": "20000", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 5000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5.4-2026-03-05/7dc94e38-f4d0-4ee6-ab2b-2d89dcb45371_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "12bd083d822cbb2388861990848e5b9fc3f541f48a8158f4a8cb95730a4409cd", + "total_rows": 3 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5.4-2026-03-05/7dc94e38-f4d0-4ee6-ab2b-2d89dcb45371_samples.jsonl b/data/hle/openai/gpt-5.4-2026-03-05/7dc94e38-f4d0-4ee6-ab2b-2d89dcb45371_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..06909226b69f9d0feb379570e8d2deaa02e7388d --- /dev/null +++ b/data/hle/openai/gpt-5.4-2026-03-05/7dc94e38-f4d0-4ee6-ab2b-2d89dcb45371_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:12bd083d822cbb2388861990848e5b9fc3f541f48a8158f4a8cb95730a4409cd +size 99810 diff --git a/data/hle/openai/gpt-5.4-2026-03-05/7de92562-e6b3-4b5d-bbdb-1bf3dc4a0053.json b/data/hle/openai/gpt-5.4-2026-03-05/7de92562-e6b3-4b5d-bbdb-1bf3dc4a0053.json new file mode 100644 index 0000000000000000000000000000000000000000..5403ce5c8d7a33783784065ccefaa105502e0c9b --- /dev/null +++ b/data/hle/openai/gpt-5.4-2026-03-05/7de92562-e6b3-4b5d-bbdb-1bf3dc4a0053.json @@ -0,0 +1,117 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5.4-2026-03-05/1786164219.59328", + "evaluation_timestamp": "1781625228.0", + "retrieved_timestamp": "1786164219.59328", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "mean on hle/S-adaptive/ceb70202 for scorer _llm_grader", + "source_data": { + "dataset_name": "hle", + "source_type": "hf_dataset", + "hf_repo": "cais/hle", + "samples_number": 6, + "sample_ids": [ + "hle_bc5fbce2", + "hle_7808792d", + "hle_e3b8f5c2", + "hle_fcd56401", + "hle_0ec247c5", + "hle_45d1c140" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1781625228.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 1.0, + "uncertainty": { + "num_samples": 4 + }, + "details": { + "total_matched_trajectories": "4", + "stopping_reason_count_repetition_guard": "4", + "avg_total_tokens_target_model": "374269.00", + "total_total_tokens_target_model": "1497076.0", + "avg_total_tokens_other_models": "14019.00", + "total_total_tokens_other_models": "56076.0", + "avg_total_tokens_all_models": "388288.00", + "total_total_tokens_all_models": "1553152.0", + "avg_turn_count": "8.75", + "total_turn_count": "35.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "16", + "max_tokens": "20000", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 5000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5.4-2026-03-05/7de92562-e6b3-4b5d-bbdb-1bf3dc4a0053_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "e3c4b4d21683e6afab1b61c8d1f7fbfc98e772d206df53d7080204ed53317562", + "total_rows": 4 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5.4-2026-03-05/7de92562-e6b3-4b5d-bbdb-1bf3dc4a0053_samples.jsonl b/data/hle/openai/gpt-5.4-2026-03-05/7de92562-e6b3-4b5d-bbdb-1bf3dc4a0053_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..b41b65469d455418c3be99957c8a28de906b6a45 --- /dev/null +++ b/data/hle/openai/gpt-5.4-2026-03-05/7de92562-e6b3-4b5d-bbdb-1bf3dc4a0053_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e3c4b4d21683e6afab1b61c8d1f7fbfc98e772d206df53d7080204ed53317562 +size 1406842 diff --git a/data/hle/openai/gpt-5.4-2026-03-05/948336fe-ff0b-4e64-a610-e46d176e554a.json b/data/hle/openai/gpt-5.4-2026-03-05/948336fe-ff0b-4e64-a610-e46d176e554a.json new file mode 100644 index 0000000000000000000000000000000000000000..d02f361335d1b467f0e86827412d20d7259a3472 --- /dev/null +++ b/data/hle/openai/gpt-5.4-2026-03-05/948336fe-ff0b-4e64-a610-e46d176e554a.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5.4-2026-03-05/1786168787.868984", + "evaluation_timestamp": "1776268426.0", + "retrieved_timestamp": "1786168787.868984", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5.4-2026-03-05/948336fe-ff0b-4e64-a610-e46d176e554a_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "ebffa5acdb236f0708a4333e8f84cbbf1469ba2fd81304c85f58947f8ca3de12", + "total_rows": 30 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5.4-2026-03-05/948336fe-ff0b-4e64-a610-e46d176e554a_samples.jsonl b/data/hle/openai/gpt-5.4-2026-03-05/948336fe-ff0b-4e64-a610-e46d176e554a_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..fcb42f29b3e88204d716c31181844a4b03ae079c --- /dev/null +++ b/data/hle/openai/gpt-5.4-2026-03-05/948336fe-ff0b-4e64-a610-e46d176e554a_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ebffa5acdb236f0708a4333e8f84cbbf1469ba2fd81304c85f58947f8ca3de12 +size 606701 diff --git a/data/hle/openai/gpt-5.4-2026-03-05/9c097fba-645c-46c5-a3ab-2373a236d758.json b/data/hle/openai/gpt-5.4-2026-03-05/9c097fba-645c-46c5-a3ab-2373a236d758.json new file mode 100644 index 0000000000000000000000000000000000000000..c03078ba3d934ff9bf8f5c242350b1d52f4ecdfb --- /dev/null +++ b/data/hle/openai/gpt-5.4-2026-03-05/9c097fba-645c-46c5-a3ab-2373a236d758.json @@ -0,0 +1,112 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5.4-2026-03-05/1786164157.7221", + "evaluation_timestamp": "1779017113.0", + "retrieved_timestamp": "1786164157.7221", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "mean on hle/S-adaptive/+1ep/70f636a7 for scorer _llm_grader", + "source_data": { + "dataset_name": "hle", + "source_type": "hf_dataset", + "hf_repo": "cais/hle", + "samples_number": 1, + "sample_ids": [ + "hle_5d49423c" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1779017113.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 1.0, + "uncertainty": { + "num_samples": 1 + }, + "details": { + "total_matched_trajectories": "1", + "stopping_reason_count_repetition_guard": "1", + "avg_total_tokens_target_model": "91582.00", + "total_total_tokens_target_model": "91582.0", + "avg_total_tokens_other_models": "9020.00", + "total_total_tokens_other_models": "9020.0", + "avg_total_tokens_all_models": "100602.00", + "total_total_tokens_all_models": "100602.0", + "avg_turn_count": "5.00", + "total_turn_count": "5.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "8", + "max_tokens": "32000", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 5000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "8", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5.4-2026-03-05/9c097fba-645c-46c5-a3ab-2373a236d758_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "9da25328edb81b831435074fd37e563bd9cdbd5e851261bdf3122e56aa2db923", + "total_rows": 1 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5.4-2026-03-05/9c097fba-645c-46c5-a3ab-2373a236d758_samples.jsonl b/data/hle/openai/gpt-5.4-2026-03-05/9c097fba-645c-46c5-a3ab-2373a236d758_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..900ee161662cfb53e60255dd3598a694d2219c65 --- /dev/null +++ b/data/hle/openai/gpt-5.4-2026-03-05/9c097fba-645c-46c5-a3ab-2373a236d758_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9da25328edb81b831435074fd37e563bd9cdbd5e851261bdf3122e56aa2db923 +size 21064 diff --git a/data/hle/openai/gpt-5.4-2026-03-05/a52d8fee-0d75-4fb1-8456-59c53ac1d37d.json b/data/hle/openai/gpt-5.4-2026-03-05/a52d8fee-0d75-4fb1-8456-59c53ac1d37d.json new file mode 100644 index 0000000000000000000000000000000000000000..4c75dea6847bd56adf6bc8f57dbc620cf06c7976 --- /dev/null +++ b/data/hle/openai/gpt-5.4-2026-03-05/a52d8fee-0d75-4fb1-8456-59c53ac1d37d.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5.4-2026-03-05/1786166795.293738", + "evaluation_timestamp": "1776359245.0", + "retrieved_timestamp": "1786166795.293738", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5.4-2026-03-05/a52d8fee-0d75-4fb1-8456-59c53ac1d37d_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "b775e6ac418a94a19ff5579650dad5827172d97de910db4a5dbd7063dd7a8c43", + "total_rows": 10 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5.4-2026-03-05/a52d8fee-0d75-4fb1-8456-59c53ac1d37d_samples.jsonl b/data/hle/openai/gpt-5.4-2026-03-05/a52d8fee-0d75-4fb1-8456-59c53ac1d37d_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..2632998ff9b7281bcc8b01e93ce1998d8b30bb7e --- /dev/null +++ b/data/hle/openai/gpt-5.4-2026-03-05/a52d8fee-0d75-4fb1-8456-59c53ac1d37d_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b775e6ac418a94a19ff5579650dad5827172d97de910db4a5dbd7063dd7a8c43 +size 286165 diff --git a/data/hle/openai/gpt-5.4-2026-03-05/b8f6fa35-bc4d-4461-b3dd-045842bce87e.json b/data/hle/openai/gpt-5.4-2026-03-05/b8f6fa35-bc4d-4461-b3dd-045842bce87e.json new file mode 100644 index 0000000000000000000000000000000000000000..18038bae2b69533736a84d6f29b0d7b070d8a71a --- /dev/null +++ b/data/hle/openai/gpt-5.4-2026-03-05/b8f6fa35-bc4d-4461-b3dd-045842bce87e.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5.4-2026-03-05/1786166808.597773", + "evaluation_timestamp": "1776268426.0", + "retrieved_timestamp": "1786166808.597773", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5.4-2026-03-05/b8f6fa35-bc4d-4461-b3dd-045842bce87e_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "fe9a517099c3d1b7f47993b0461d7c7ceb511911f2d5e020eda75ae27f3920ad", + "total_rows": 20 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5.4-2026-03-05/b8f6fa35-bc4d-4461-b3dd-045842bce87e_samples.jsonl b/data/hle/openai/gpt-5.4-2026-03-05/b8f6fa35-bc4d-4461-b3dd-045842bce87e_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..df103db26ed33cc1c80befb976c067a28be4a0f2 --- /dev/null +++ b/data/hle/openai/gpt-5.4-2026-03-05/b8f6fa35-bc4d-4461-b3dd-045842bce87e_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fe9a517099c3d1b7f47993b0461d7c7ceb511911f2d5e020eda75ae27f3920ad +size 616445 diff --git a/data/hle/openai/gpt-5.4-2026-03-05/bf089add-750e-45bf-abdd-f610cc2d8c8d.json b/data/hle/openai/gpt-5.4-2026-03-05/bf089add-750e-45bf-abdd-f610cc2d8c8d.json new file mode 100644 index 0000000000000000000000000000000000000000..3808fdcd7f3f4ada4ece870a17a36a4b6d2ceeae --- /dev/null +++ b/data/hle/openai/gpt-5.4-2026-03-05/bf089add-750e-45bf-abdd-f610cc2d8c8d.json @@ -0,0 +1,113 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5.4-2026-03-05/1786164192.81602", + "evaluation_timestamp": "1782335122.0", + "retrieved_timestamp": "1786164192.81602", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "mean on hle/S-adaptive/+4ep/4f305b5d for scorer _llm_grader", + "source_data": { + "dataset_name": "hle", + "source_type": "hf_dataset", + "hf_repo": "cais/hle", + "samples_number": 1, + "sample_ids": [ + "hle_dfdb78f5" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1782335122.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 1.0, + "uncertainty": { + "num_samples": 4 + }, + "details": { + "total_matched_trajectories": "4", + "stopping_reason_count_repetition_guard": "2", + "stopping_reason_count_token_limit": "2", + "avg_total_tokens_target_model": "4486269.25", + "total_total_tokens_target_model": "17945077.0", + "avg_total_tokens_other_models": "4575.00", + "total_total_tokens_other_models": "18300.0", + "avg_total_tokens_all_models": "4490844.25", + "total_total_tokens_all_models": "17963377.0", + "avg_turn_count": "126.25", + "total_turn_count": "505.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "1200", + "max_connections": "16", + "max_tokens": "20000", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 5000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5.4-2026-03-05/bf089add-750e-45bf-abdd-f610cc2d8c8d_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "f7ad6f782078f61be2a7370e9a4cc786069161d8123bcd5f9f37207bc0651327", + "total_rows": 4 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5.4-2026-03-05/bf089add-750e-45bf-abdd-f610cc2d8c8d_samples.jsonl b/data/hle/openai/gpt-5.4-2026-03-05/bf089add-750e-45bf-abdd-f610cc2d8c8d_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..e36de5083d0343a52f1945475ced0e21442e68f3 --- /dev/null +++ b/data/hle/openai/gpt-5.4-2026-03-05/bf089add-750e-45bf-abdd-f610cc2d8c8d_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f7ad6f782078f61be2a7370e9a4cc786069161d8123bcd5f9f37207bc0651327 +size 915989 diff --git a/data/hle/openai/gpt-5.4-2026-03-05/c993dff3-2a2b-47d2-9f67-9ee11f9e5b0b.json b/data/hle/openai/gpt-5.4-2026-03-05/c993dff3-2a2b-47d2-9f67-9ee11f9e5b0b.json new file mode 100644 index 0000000000000000000000000000000000000000..087c24f8983c5947993d947853b5430366b28ef1 --- /dev/null +++ b/data/hle/openai/gpt-5.4-2026-03-05/c993dff3-2a2b-47d2-9f67-9ee11f9e5b0b.json @@ -0,0 +1,113 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5.4-2026-03-05/1786164186.702838", + "evaluation_timestamp": "1778948171.0", + "retrieved_timestamp": "1786164186.702838", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "mean on hle/S-adaptive/+3ep/5c4dbf1f for scorer _llm_grader", + "source_data": { + "dataset_name": "hle", + "source_type": "hf_dataset", + "hf_repo": "cais/hle", + "samples_number": 2, + "sample_ids": [ + "hle_4485802f", + "hle_1d580261" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1778948171.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.0, + "uncertainty": { + "num_samples": 1 + }, + "details": { + "total_matched_trajectories": "1", + "stopping_reason_count_repetition_guard": "1", + "avg_total_tokens_target_model": "87233.00", + "total_total_tokens_target_model": "87233.0", + "avg_total_tokens_other_models": "6224.00", + "total_total_tokens_other_models": "6224.0", + "avg_total_tokens_all_models": "93457.00", + "total_total_tokens_all_models": "93457.0", + "avg_turn_count": "5.00", + "total_turn_count": "5.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "8", + "max_tokens": "32000", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 5000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "8", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5.4-2026-03-05/c993dff3-2a2b-47d2-9f67-9ee11f9e5b0b_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "ab5ea5c53858913142f200c877a956ea594c1c49a9bc1cf1afe204484dae4d29", + "total_rows": 1 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5.4-2026-03-05/c993dff3-2a2b-47d2-9f67-9ee11f9e5b0b_samples.jsonl b/data/hle/openai/gpt-5.4-2026-03-05/c993dff3-2a2b-47d2-9f67-9ee11f9e5b0b_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..6b12ccf54061cb394236d8332e3a7803c58cbddc --- /dev/null +++ b/data/hle/openai/gpt-5.4-2026-03-05/c993dff3-2a2b-47d2-9f67-9ee11f9e5b0b_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ab5ea5c53858913142f200c877a956ea594c1c49a9bc1cf1afe204484dae4d29 +size 20367 diff --git a/data/hle/openai/gpt-5.4-2026-03-05/cc79edb5-2461-4252-85db-29b12d91fa94.json b/data/hle/openai/gpt-5.4-2026-03-05/cc79edb5-2461-4252-85db-29b12d91fa94.json new file mode 100644 index 0000000000000000000000000000000000000000..5f4c36cbc689f8bc6365e247c656d732c0b37d86 --- /dev/null +++ b/data/hle/openai/gpt-5.4-2026-03-05/cc79edb5-2461-4252-85db-29b12d91fa94.json @@ -0,0 +1,116 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5.4-2026-03-05/1786164175.92325", + "evaluation_timestamp": "1776880341.0", + "retrieved_timestamp": "1786164175.92325", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "mean on hle/S-adaptive/+2ep/08a4af9e for scorer _llm_grader", + "source_data": { + "dataset_name": "hle", + "source_type": "hf_dataset", + "hf_repo": "cais/hle", + "samples_number": 6, + "sample_ids": [ + "hle_c2ff3a43", + "hle_d52d3dae", + "hle_6d963e58", + "hle_35a40c52", + "hle_c0ed25f7", + "hle_cd4aa5b3" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1776880341.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.25, + "uncertainty": { + "standard_error": { + "value": 0.17078251276599332 + }, + "num_samples": 11 + }, + "details": { + "total_matched_trajectories": "11", + "stopping_reason_count_repetition_guard": "11", + "avg_total_tokens_target_model": "25383.64", + "total_total_tokens_target_model": "279220.0", + "avg_total_tokens_other_models": "3986.27", + "total_total_tokens_other_models": "43849.0", + "avg_total_tokens_all_models": "29369.91", + "total_total_tokens_all_models": "323069.0", + "avg_turn_count": "5.55", + "total_turn_count": "61.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 5000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5.4-2026-03-05/cc79edb5-2461-4252-85db-29b12d91fa94_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "305d99c212813cfda86aef05539cb573e06e66256a1d77cf0ae28fee375c23bd", + "total_rows": 11 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5.4-2026-03-05/cc79edb5-2461-4252-85db-29b12d91fa94_samples.jsonl b/data/hle/openai/gpt-5.4-2026-03-05/cc79edb5-2461-4252-85db-29b12d91fa94_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..bd357b0580bebb6309a034a3d03f153f71904ab8 --- /dev/null +++ b/data/hle/openai/gpt-5.4-2026-03-05/cc79edb5-2461-4252-85db-29b12d91fa94_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:305d99c212813cfda86aef05539cb573e06e66256a1d77cf0ae28fee375c23bd +size 209612 diff --git a/data/hle/openai/gpt-5.4-2026-03-05/d460bc30-6b61-4bcc-9927-2804b76dec01.json b/data/hle/openai/gpt-5.4-2026-03-05/d460bc30-6b61-4bcc-9927-2804b76dec01.json new file mode 100644 index 0000000000000000000000000000000000000000..f93698c067a5c9369f3a8283d48d5a774ac998be --- /dev/null +++ b/data/hle/openai/gpt-5.4-2026-03-05/d460bc30-6b61-4bcc-9927-2804b76dec01.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5.4-2026-03-05/1786168736.338501", + "evaluation_timestamp": "1776519735.0", + "retrieved_timestamp": "1786168736.338501", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5.4-2026-03-05/d460bc30-6b61-4bcc-9927-2804b76dec01_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "61691a1cc500990dc8af03fb95d46403aeed49c3484edac1654d78969b9cc060", + "total_rows": 20 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5.4-2026-03-05/d460bc30-6b61-4bcc-9927-2804b76dec01_samples.jsonl b/data/hle/openai/gpt-5.4-2026-03-05/d460bc30-6b61-4bcc-9927-2804b76dec01_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..858bd6f5727c35bc963c1709b0a9111cb350c9b9 --- /dev/null +++ b/data/hle/openai/gpt-5.4-2026-03-05/d460bc30-6b61-4bcc-9927-2804b76dec01_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:61691a1cc500990dc8af03fb95d46403aeed49c3484edac1654d78969b9cc060 +size 400245 diff --git a/data/hle/openai/gpt-5.4-2026-03-05/d749ee82-daa5-4e38-8bdf-c52fd23967ea.json b/data/hle/openai/gpt-5.4-2026-03-05/d749ee82-daa5-4e38-8bdf-c52fd23967ea.json new file mode 100644 index 0000000000000000000000000000000000000000..0aeebbb7c27a672de82e8a08557da9600e23d59f --- /dev/null +++ b/data/hle/openai/gpt-5.4-2026-03-05/d749ee82-daa5-4e38-8bdf-c52fd23967ea.json @@ -0,0 +1,112 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5.4-2026-03-05/1786164224.63255", + "evaluation_timestamp": "1776076028.0", + "retrieved_timestamp": "1786164224.63255", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "mean on hle/S-adaptive/+8ep for scorer _llm_grader", + "source_data": { + "dataset_name": "hle", + "source_type": "hf_dataset", + "hf_repo": "cais/hle", + "samples_number": 6, + "sample_ids": [ + "hle_d6b73839", + "hle_07662c9c", + "hle_764ce81c", + "hle_4e3440f7", + "hle_baccd6da", + "hle_f46c2857" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1776076028.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.625, + "uncertainty": { + "standard_error": { + "value": 0.19895560643855537 + }, + "num_samples": 48 + }, + "details": { + "total_matched_trajectories": "30", + "stopping_reason_count_repetition_guard": "30", + "avg_total_tokens_target_model": "36713.13", + "total_total_tokens_target_model": "1101394.0", + "avg_total_tokens_other_models": "6361.73", + "total_total_tokens_other_models": "190852.0", + "avg_total_tokens_all_models": "43074.87", + "total_total_tokens_all_models": "1292246.0", + "avg_turn_count": "6.23", + "total_turn_count": "187.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 5000000 + }, + "sandbox": {} + }, + "additional_details": { + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5.4-2026-03-05/d749ee82-daa5-4e38-8bdf-c52fd23967ea_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "7d0b3aef05fe9cbb371a1710527f4066fdad844e43fbf3362e9bb306c05af8a2", + "total_rows": 48 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5.4-2026-03-05/d749ee82-daa5-4e38-8bdf-c52fd23967ea_samples.jsonl b/data/hle/openai/gpt-5.4-2026-03-05/d749ee82-daa5-4e38-8bdf-c52fd23967ea_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..51fbdb61072ddc788791066f9a13cf1303e6caac --- /dev/null +++ b/data/hle/openai/gpt-5.4-2026-03-05/d749ee82-daa5-4e38-8bdf-c52fd23967ea_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7d0b3aef05fe9cbb371a1710527f4066fdad844e43fbf3362e9bb306c05af8a2 +size 944491 diff --git a/data/hle/openai/gpt-5.4-2026-03-05/de1c08f0-c835-4f04-99cb-b1511ac69c50.json b/data/hle/openai/gpt-5.4-2026-03-05/de1c08f0-c835-4f04-99cb-b1511ac69c50.json new file mode 100644 index 0000000000000000000000000000000000000000..fcda20bed57c83e4e04d47cf3cb4bdbce2c81a39 --- /dev/null +++ b/data/hle/openai/gpt-5.4-2026-03-05/de1c08f0-c835-4f04-99cb-b1511ac69c50.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5.4-2026-03-05/1786168796.255235", + "evaluation_timestamp": "1775733995.0", + "retrieved_timestamp": "1786168796.255235", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5.4-2026-03-05/de1c08f0-c835-4f04-99cb-b1511ac69c50_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "9cf78526e6232ed0d2adbfd6eb0f0ebabfbb52e8a468b5362ac38fbce8a9a941", + "total_rows": 100 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5.4-2026-03-05/de1c08f0-c835-4f04-99cb-b1511ac69c50_samples.jsonl b/data/hle/openai/gpt-5.4-2026-03-05/de1c08f0-c835-4f04-99cb-b1511ac69c50_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..459ca7fdb0b5a2ad83b8c552594664c5682325c1 --- /dev/null +++ b/data/hle/openai/gpt-5.4-2026-03-05/de1c08f0-c835-4f04-99cb-b1511ac69c50_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9cf78526e6232ed0d2adbfd6eb0f0ebabfbb52e8a468b5362ac38fbce8a9a941 +size 1979201 diff --git a/data/hle/openai/gpt-5.4-2026-03-05/e168f7b8-3e30-4064-bd7b-70f2adc0533b.json b/data/hle/openai/gpt-5.4-2026-03-05/e168f7b8-3e30-4064-bd7b-70f2adc0533b.json new file mode 100644 index 0000000000000000000000000000000000000000..3d3990af7aeb561ed4d34b173471fb6df7b11747 --- /dev/null +++ b/data/hle/openai/gpt-5.4-2026-03-05/e168f7b8-3e30-4064-bd7b-70f2adc0533b.json @@ -0,0 +1,117 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5.4-2026-03-05/1786164168.479557", + "evaluation_timestamp": "1776804136.0", + "retrieved_timestamp": "1786164168.479557", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "mean on hle/S-adaptive/+1ep/8b8ab4d2 for scorer _llm_grader", + "source_data": { + "dataset_name": "hle", + "source_type": "hf_dataset", + "hf_repo": "cais/hle", + "samples_number": 7, + "sample_ids": [ + "hle_9019c023", + "hle_9282e7ae", + "hle_44007b97", + "hle_f85a0fa1", + "hle_376b8247", + "hle_31497469", + "hle_c43a27a3" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1776804136.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.5714285714285714, + "uncertainty": { + "standard_error": { + "value": 0.20203050891044214 + }, + "num_samples": 7 + }, + "details": { + "total_matched_trajectories": "7", + "stopping_reason_count_repetition_guard": "7", + "avg_total_tokens_target_model": "36947.14", + "total_total_tokens_target_model": "258630.0", + "avg_total_tokens_other_models": "4867.86", + "total_total_tokens_other_models": "34075.0", + "avg_total_tokens_all_models": "41815.00", + "total_total_tokens_all_models": "292705.0", + "avg_turn_count": "5.00", + "total_turn_count": "35.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 5000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5.4-2026-03-05/e168f7b8-3e30-4064-bd7b-70f2adc0533b_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "b8e298881dc96765ff440541a24e875f5f7e70b7f012c31718cc0b5a33b01240", + "total_rows": 7 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5.4-2026-03-05/e168f7b8-3e30-4064-bd7b-70f2adc0533b_samples.jsonl b/data/hle/openai/gpt-5.4-2026-03-05/e168f7b8-3e30-4064-bd7b-70f2adc0533b_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..89eecc95b54850e1393a21ea6dac9d465bfcee38 --- /dev/null +++ b/data/hle/openai/gpt-5.4-2026-03-05/e168f7b8-3e30-4064-bd7b-70f2adc0533b_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b8e298881dc96765ff440541a24e875f5f7e70b7f012c31718cc0b5a33b01240 +size 134431 diff --git a/data/hle/openai/gpt-5.4-2026-03-05/ee9fe4cb-f419-4690-89dc-5c9d0bf62320.json b/data/hle/openai/gpt-5.4-2026-03-05/ee9fe4cb-f419-4690-89dc-5c9d0bf62320.json new file mode 100644 index 0000000000000000000000000000000000000000..de1305b165376e79fd5ee349505369e4050c5543 --- /dev/null +++ b/data/hle/openai/gpt-5.4-2026-03-05/ee9fe4cb-f419-4690-89dc-5c9d0bf62320.json @@ -0,0 +1,118 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5.4-2026-03-05/1786164198.550065", + "evaluation_timestamp": "1783108873.0", + "retrieved_timestamp": "1786164198.550065", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "mean on hle/S-adaptive/e28e6d55 for scorer _llm_grader", + "source_data": { + "dataset_name": "hle", + "source_type": "hf_dataset", + "hf_repo": "cais/hle", + "samples_number": 4, + "sample_ids": [ + "hle_bc5fbce2", + "hle_e3b8f5c2", + "hle_fcd56401", + "hle_45d1c140" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1783108873.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.15, + "uncertainty": { + "standard_error": { + "value": 0.15 + }, + "num_samples": 20 + }, + "details": { + "total_matched_trajectories": "20", + "stopping_reason_count_repetition_guard": "20", + "avg_total_tokens_target_model": "249040.55", + "total_total_tokens_target_model": "4980811.0", + "avg_total_tokens_other_models": "10095.85", + "total_total_tokens_other_models": "201917.0", + "avg_total_tokens_all_models": "259136.40", + "total_total_tokens_all_models": "5182728.0", + "avg_turn_count": "6.30", + "total_turn_count": "126.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "1", + "timeout": "1200", + "attempt_timeout": "3600", + "max_connections": "16", + "max_tokens": "64000", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 5000000 + }, + "sandbox": {}, + "max_attempts": 1 + }, + "additional_details": { + "max_retries": "1", + "timeout": "1200", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5.4-2026-03-05/ee9fe4cb-f419-4690-89dc-5c9d0bf62320_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "abcdad23cd67649b6fb0e31153dea215bda0264ddaa366f1f521bb38f52231eb", + "total_rows": 20 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5.4-2026-03-05/ee9fe4cb-f419-4690-89dc-5c9d0bf62320_samples.jsonl b/data/hle/openai/gpt-5.4-2026-03-05/ee9fe4cb-f419-4690-89dc-5c9d0bf62320_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..7ce1426d0b3cd1a25df3ab1c165c7d60adacffcd --- /dev/null +++ b/data/hle/openai/gpt-5.4-2026-03-05/ee9fe4cb-f419-4690-89dc-5c9d0bf62320_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:abcdad23cd67649b6fb0e31153dea215bda0264ddaa366f1f521bb38f52231eb +size 470514 diff --git a/data/hle/openai/gpt-5.4-2026-03-05/f27fcfb3-30b4-4ba1-9e59-3bf1921e8e25.json b/data/hle/openai/gpt-5.4-2026-03-05/f27fcfb3-30b4-4ba1-9e59-3bf1921e8e25.json new file mode 100644 index 0000000000000000000000000000000000000000..60c189a92ce5d8622ea8367d5b918170b7ff7d47 --- /dev/null +++ b/data/hle/openai/gpt-5.4-2026-03-05/f27fcfb3-30b4-4ba1-9e59-3bf1921e8e25.json @@ -0,0 +1,109 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5.4-2026-03-05/1786164174.019298", + "evaluation_timestamp": "1776810932.0", + "retrieved_timestamp": "1786164174.019298", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "mean on hle/S-adaptive/+2ep/b954d324 for scorer _llm_grader", + "source_data": { + "dataset_name": "hle", + "source_type": "hf_dataset", + "hf_repo": "cais/hle", + "samples_number": 2, + "sample_ids": [ + "hle_3cf4483d", + "hle_c3ca0bda" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1776810932.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.0, + "uncertainty": { + "num_samples": 4 + }, + "details": { + "total_matched_trajectories": "4", + "stopping_reason_count_repetition_guard": "4", + "avg_total_tokens_target_model": "36547.25", + "total_total_tokens_target_model": "146189.0", + "avg_total_tokens_other_models": "4299.25", + "total_total_tokens_other_models": "17197.0", + "avg_total_tokens_all_models": "40846.50", + "total_total_tokens_all_models": "163386.0", + "avg_turn_count": "5.75", + "total_turn_count": "23.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 5000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5.4-2026-03-05/f27fcfb3-30b4-4ba1-9e59-3bf1921e8e25_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "8e4399265d562f6fa4f0218703187cf8e59636b53c2f667d375c204a23ccd55f", + "total_rows": 4 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5.4-2026-03-05/f27fcfb3-30b4-4ba1-9e59-3bf1921e8e25_samples.jsonl b/data/hle/openai/gpt-5.4-2026-03-05/f27fcfb3-30b4-4ba1-9e59-3bf1921e8e25_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..8692317ee668f14ac390133fa2362ab9ee1d3c48 --- /dev/null +++ b/data/hle/openai/gpt-5.4-2026-03-05/f27fcfb3-30b4-4ba1-9e59-3bf1921e8e25_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8e4399265d562f6fa4f0218703187cf8e59636b53c2f667d375c204a23ccd55f +size 80518 diff --git a/data/hle/openai/gpt-5.4-2026-03-05/f7d7d6ad-1914-4cb0-bc24-a38aa225e777.json b/data/hle/openai/gpt-5.4-2026-03-05/f7d7d6ad-1914-4cb0-bc24-a38aa225e777.json new file mode 100644 index 0000000000000000000000000000000000000000..56ae752b9b96f2b072224be60dd3c3ec6a330124 --- /dev/null +++ b/data/hle/openai/gpt-5.4-2026-03-05/f7d7d6ad-1914-4cb0-bc24-a38aa225e777.json @@ -0,0 +1,136 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "hle/openai_gpt-5.4-2026-03-05/1786164183.219411", + "evaluation_timestamp": "1776813157.0", + "retrieved_timestamp": "1786164183.219411", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "mean on hle/S-adaptive/+3ep/f22cd2b4 for scorer _llm_grader", + "source_data": { + "dataset_name": "hle", + "source_type": "hf_dataset", + "hf_repo": "cais/hle", + "samples_number": 26, + "sample_ids": [ + "hle_ea41e785", + "hle_d0c8b07e", + "hle_0291c889", + "hle_b1a7811d", + "hle_dab7fbb1", + "hle_cf3d65bd", + "hle_5e235b48", + "hle_0ff76ff3", + "hle_70e5aebc", + "hle_aa2fed90", + "hle_9df5d371", + "hle_5759e7d7", + "hle_ca59b32c", + "hle_33bc94e2", + "hle_e3817b7e", + "hle_1dc141bb", + "hle_a764505b", + "hle_e62ba103", + "hle_9255683e", + "hle_5f04b314", + "hle_b5c681d9", + "hle_d5a55900", + "hle_3e11b38a", + "hle_0a1eec3f", + "hle_82f5c665", + "hle_cc7f25c0" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1776813157.0", + "metric_config": { + "evaluation_description": "mean", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.4102564102564103, + "uncertainty": { + "standard_error": { + "value": 0.09301721613957464 + }, + "num_samples": 75 + }, + "details": { + "total_matched_trajectories": "75", + "stopping_reason_count_repetition_guard": "75", + "avg_total_tokens_target_model": "35582.63", + "total_total_tokens_target_model": "2668697.0", + "avg_total_tokens_other_models": "5181.85", + "total_total_tokens_other_models": "388639.0", + "avg_total_tokens_all_models": "40764.48", + "total_total_tokens_all_models": "3057336.0", + "avg_turn_count": "5.24", + "total_turn_count": "393.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 5000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/hle/openai/gpt-5.4-2026-03-05/f7d7d6ad-1914-4cb0-bc24-a38aa225e777_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "807752cc901b9e807dc5bfbd9e96e48c03ba1af639bdeade7ccaa6ba1001d157", + "total_rows": 75 + } +} \ No newline at end of file diff --git a/data/hle/openai/gpt-5.4-2026-03-05/f7d7d6ad-1914-4cb0-bc24-a38aa225e777_samples.jsonl b/data/hle/openai/gpt-5.4-2026-03-05/f7d7d6ad-1914-4cb0-bc24-a38aa225e777_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..5c3ec121d244f60e6d6ec4bb4c64e3673139edea --- /dev/null +++ b/data/hle/openai/gpt-5.4-2026-03-05/f7d7d6ad-1914-4cb0-bc24-a38aa225e777_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:807752cc901b9e807dc5bfbd9e96e48c03ba1af639bdeade7ccaa6ba1001d157 +size 1465869 diff --git a/data/swebenchpro/anthropic/claude-opus-4-20250514/04e6e5c4-76c2-4649-a4b5-1ab82b01baed.json b/data/swebenchpro/anthropic/claude-opus-4-20250514/04e6e5c4-76c2-4649-a4b5-1ab82b01baed.json new file mode 100644 index 0000000000000000000000000000000000000000..a1f8aab18698ef52e85c1f0e356092dfe246b10e --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-20250514/04e6e5c4-76c2-4649-a4b5-1ab82b01baed.json @@ -0,0 +1,110 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/anthropic_claude-opus-4-20250514/1786165729.134354", + "evaluation_timestamp": "1778764801.0", + "retrieved_timestamp": "1786165729.134354", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-20250514", + "id": "anthropic/claude-opus-4-20250514", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on swebenchpro/S-adaptive/+2ep/e04ec096 for scorer harbor_scorer", + "source_data": { + "dataset_name": "swebenchpro", + "source_type": "hf_dataset", + "samples_number": 1, + "sample_ids": [ + "instance_ansible__ansible-de5858f48dc9e1ce9117034e0d7e76806f420ca8-v1055803c3a812189a1133297f7f5468579283f86" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1778764801.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.0, + "uncertainty": { + "num_samples": 1 + }, + "details": { + "total_matched_trajectories": "1", + "stopping_reason_count_repetition_guard": "1", + "avg_total_tokens_target_model": "4988281.00", + "total_total_tokens_target_model": "4988281.0", + "avg_total_tokens_other_models": "1677.00", + "total_total_tokens_other_models": "1677.0", + "avg_total_tokens_all_models": "4989958.00", + "total_total_tokens_all_models": "4989958.0", + "avg_turn_count": "125.00", + "total_turn_count": "125.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "8", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 30000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "8", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/anthropic/claude-opus-4-20250514/04e6e5c4-76c2-4649-a4b5-1ab82b01baed_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "4fbe23ffea084d441e5a4f0fbdcd0e0b26a0bfa82deb865a778de37d0570bc33", + "total_rows": 1 + } +} \ No newline at end of file diff --git a/data/swebenchpro/anthropic/claude-opus-4-20250514/04e6e5c4-76c2-4649-a4b5-1ab82b01baed_samples.jsonl b/data/swebenchpro/anthropic/claude-opus-4-20250514/04e6e5c4-76c2-4649-a4b5-1ab82b01baed_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..a5200ba619d884ce927a7b06c6a6c457592841e1 --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-20250514/04e6e5c4-76c2-4649-a4b5-1ab82b01baed_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4fbe23ffea084d441e5a4f0fbdcd0e0b26a0bfa82deb865a778de37d0570bc33 +size 485523 diff --git a/data/swebenchpro/anthropic/claude-opus-4-20250514/078c4359-bf96-43ec-b1c6-e5f0d2eef4a6.json b/data/swebenchpro/anthropic/claude-opus-4-20250514/078c4359-bf96-43ec-b1c6-e5f0d2eef4a6.json new file mode 100644 index 0000000000000000000000000000000000000000..64204be1289804dfc08e829743837e1cac817a12 --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-20250514/078c4359-bf96-43ec-b1c6-e5f0d2eef4a6.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/anthropic_claude-opus-4-20250514/1786167464.784447", + "evaluation_timestamp": "1778665507.0", + "retrieved_timestamp": "1786167464.784447", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-20250514", + "id": "anthropic/claude-opus-4-20250514", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/anthropic/claude-opus-4-20250514/078c4359-bf96-43ec-b1c6-e5f0d2eef4a6_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "3a4230b0dc02c5d34727eb375150cd71e4ce2c8164455c74412cf60c332acd60", + "total_rows": 2 + } +} \ No newline at end of file diff --git a/data/swebenchpro/anthropic/claude-opus-4-20250514/078c4359-bf96-43ec-b1c6-e5f0d2eef4a6_samples.jsonl b/data/swebenchpro/anthropic/claude-opus-4-20250514/078c4359-bf96-43ec-b1c6-e5f0d2eef4a6_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..d9bb98d67158e834e5ef56192b1f05078ae57afa --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-20250514/078c4359-bf96-43ec-b1c6-e5f0d2eef4a6_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3a4230b0dc02c5d34727eb375150cd71e4ce2c8164455c74412cf60c332acd60 +size 249393 diff --git a/data/swebenchpro/anthropic/claude-opus-4-20250514/08ba866d-cefa-461e-9114-3216a819cf51.json b/data/swebenchpro/anthropic/claude-opus-4-20250514/08ba866d-cefa-461e-9114-3216a819cf51.json new file mode 100644 index 0000000000000000000000000000000000000000..919dc39a4af8793fc97e792bdecde49a04e9cb35 --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-20250514/08ba866d-cefa-461e-9114-3216a819cf51.json @@ -0,0 +1,121 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/anthropic_claude-opus-4-20250514/1786165751.661707", + "evaluation_timestamp": "1778169019.0", + "retrieved_timestamp": "1786165751.661707", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-20250514", + "id": "anthropic/claude-opus-4-20250514", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on swebenchpro/S-adaptive/+3ep/292bdaf4 for scorer harbor_scorer", + "source_data": { + "dataset_name": "swebenchpro", + "source_type": "hf_dataset", + "samples_number": 9, + "sample_ids": [ + "instance_ansible__ansible-b2a289dcbb702003377221e25f62c8a3608f0e89-v173091e2e36d38c978002990795f66cfc0af30ad", + "instance_internetarchive__openlibrary-f8cc11d9c1575fdba5ac66aee0befca970da8d64-v13642507b4fc1f8d234172bf8129942da2c2ca26", + "instance_qutebrowser__qutebrowser-e70f5b03187bdd40e8bf70f5f3ead840f52d1f42-v02ad04386d5238fe2d1a1be450df257370de4b6a", + "instance_tutao__tutanota-d1aa0ecec288bfc800cfb9133b087c4f81ad8b38-vbc0d9ba8f0071fbe982809910959a6ff8884dbbf", + "instance_qutebrowser__qutebrowser-e64622cd2df5b521342cf4a62e0d4cb8f8c9ae5a-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_gravitational__teleport-ac2fb2f9b4fd1896b554d3011df23d3d71295779", + "instance_protonmail__webclients-cb8cc309c6968b0a2a5fe4288d0ae0a969ff31e1", + "instance_ansible__ansible-11c1777d56664b1acb56b387a1ad6aeadef1391d-v0f01c69f1e2528b935359cfe578530722bca2c59", + "instance_ansible__ansible-29aea9ff3466e4cd2ed00524b9e56738d568ce8b-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1778169019.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.6296296296296295, + "uncertainty": { + "standard_error": { + "value": 0.15158060688652686 + }, + "num_samples": 27 + }, + "details": { + "total_matched_trajectories": "27", + "stopping_reason_count_repetition_guard": "27", + "avg_total_tokens_target_model": "1302199.37", + "total_total_tokens_target_model": "35159383.0", + "avg_total_tokens_other_models": "1209.11", + "total_total_tokens_other_models": "32646.0", + "avg_total_tokens_all_models": "1303408.48", + "total_total_tokens_all_models": "35192029.0", + "avg_turn_count": "58.89", + "total_turn_count": "1590.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 30000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/anthropic/claude-opus-4-20250514/08ba866d-cefa-461e-9114-3216a819cf51_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "27e58aa53e2791751d785b9d998d5aa502ef9de2d72d427bd1703928e9c65180", + "total_rows": 27 + } +} \ No newline at end of file diff --git a/data/swebenchpro/anthropic/claude-opus-4-20250514/08ba866d-cefa-461e-9114-3216a819cf51_samples.jsonl b/data/swebenchpro/anthropic/claude-opus-4-20250514/08ba866d-cefa-461e-9114-3216a819cf51_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..9300f4009665939e6a186510afdb26280ccbde7f --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-20250514/08ba866d-cefa-461e-9114-3216a819cf51_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:27e58aa53e2791751d785b9d998d5aa502ef9de2d72d427bd1703928e9c65180 +size 6459300 diff --git a/data/swebenchpro/anthropic/claude-opus-4-20250514/2b7016a2-dce7-4989-89a4-4ab9cc112d43.json b/data/swebenchpro/anthropic/claude-opus-4-20250514/2b7016a2-dce7-4989-89a4-4ab9cc112d43.json new file mode 100644 index 0000000000000000000000000000000000000000..563a437c78c14fc5452e73df568d7158a9735f40 --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-20250514/2b7016a2-dce7-4989-89a4-4ab9cc112d43.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/anthropic_claude-opus-4-20250514/1786167474.448835", + "evaluation_timestamp": "1778595488.0", + "retrieved_timestamp": "1786167474.448835", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-20250514", + "id": "anthropic/claude-opus-4-20250514", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/anthropic/claude-opus-4-20250514/2b7016a2-dce7-4989-89a4-4ab9cc112d43_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "4bdf692c7bec5b88f7bb641ac724eba14f711087b71dd2d1ccdfdb07335ab7c4", + "total_rows": 44 + } +} \ No newline at end of file diff --git a/data/swebenchpro/anthropic/claude-opus-4-20250514/2b7016a2-dce7-4989-89a4-4ab9cc112d43_samples.jsonl b/data/swebenchpro/anthropic/claude-opus-4-20250514/2b7016a2-dce7-4989-89a4-4ab9cc112d43_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..9502f79fd1fd6b557e226ed9ceff1856ec092dd6 --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-20250514/2b7016a2-dce7-4989-89a4-4ab9cc112d43_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4bdf692c7bec5b88f7bb641ac724eba14f711087b71dd2d1ccdfdb07335ab7c4 +size 9389231 diff --git a/data/swebenchpro/anthropic/claude-opus-4-20250514/2b74906a-beea-47c8-9119-1a49edcec725.json b/data/swebenchpro/anthropic/claude-opus-4-20250514/2b74906a-beea-47c8-9119-1a49edcec725.json new file mode 100644 index 0000000000000000000000000000000000000000..93d628de90d22f063c45da35079e10ef23c49e44 --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-20250514/2b74906a-beea-47c8-9119-1a49edcec725.json @@ -0,0 +1,110 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/anthropic_claude-opus-4-20250514/1786163219.804591", + "evaluation_timestamp": "1778888168.0", + "retrieved_timestamp": "1786163219.804591", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-20250514", + "id": "anthropic/claude-opus-4-20250514", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on swebenchpro/S-adaptive+C/+4ep/e04ec096 for scorer harbor_scorer", + "source_data": { + "dataset_name": "swebenchpro", + "source_type": "hf_dataset", + "samples_number": 1, + "sample_ids": [ + "instance_ansible__ansible-de5858f48dc9e1ce9117034e0d7e76806f420ca8-v1055803c3a812189a1133297f7f5468579283f86" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1778888168.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.0, + "uncertainty": { + "num_samples": 2 + }, + "details": { + "total_matched_trajectories": "2", + "stopping_reason_count_repetition_guard": "2", + "avg_total_tokens_target_model": "3488500.50", + "total_total_tokens_target_model": "6977001.0", + "avg_total_tokens_other_models": "1755.00", + "total_total_tokens_other_models": "3510.0", + "avg_total_tokens_all_models": "3490255.50", + "total_total_tokens_all_models": "6980511.0", + "avg_turn_count": "86.00", + "total_turn_count": "172.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "8", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 30000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "8", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/anthropic/claude-opus-4-20250514/2b74906a-beea-47c8-9119-1a49edcec725_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "5e33eec07266d3d9c8177bc75f74d7216bde89a9524a40e0dee14aac47149637", + "total_rows": 2 + } +} \ No newline at end of file diff --git a/data/swebenchpro/anthropic/claude-opus-4-20250514/2b74906a-beea-47c8-9119-1a49edcec725_samples.jsonl b/data/swebenchpro/anthropic/claude-opus-4-20250514/2b74906a-beea-47c8-9119-1a49edcec725_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..40e5a978de1c017da6b89984805e25f1ea90e6ec --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-20250514/2b74906a-beea-47c8-9119-1a49edcec725_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5e33eec07266d3d9c8177bc75f74d7216bde89a9524a40e0dee14aac47149637 +size 823570 diff --git a/data/swebenchpro/anthropic/claude-opus-4-20250514/33b4de7d-2c68-441c-bdd2-edce80e1059b.json b/data/swebenchpro/anthropic/claude-opus-4-20250514/33b4de7d-2c68-441c-bdd2-edce80e1059b.json new file mode 100644 index 0000000000000000000000000000000000000000..73379512811428e5495b1419f208998d13fc0f1e --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-20250514/33b4de7d-2c68-441c-bdd2-edce80e1059b.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/anthropic_claude-opus-4-20250514/1786169496.785991", + "evaluation_timestamp": "1777420006.0", + "retrieved_timestamp": "1786169496.785991", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-20250514", + "id": "anthropic/claude-opus-4-20250514", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/anthropic/claude-opus-4-20250514/33b4de7d-2c68-441c-bdd2-edce80e1059b_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "8784cff826db826cadf77f8c2058d10fa4ce721262dfaf09be01a842761e1526", + "total_rows": 20 + } +} \ No newline at end of file diff --git a/data/swebenchpro/anthropic/claude-opus-4-20250514/33b4de7d-2c68-441c-bdd2-edce80e1059b_samples.jsonl b/data/swebenchpro/anthropic/claude-opus-4-20250514/33b4de7d-2c68-441c-bdd2-edce80e1059b_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..668c54e7772f643e81c61cdccf357d2f0fbcde1b --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-20250514/33b4de7d-2c68-441c-bdd2-edce80e1059b_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8784cff826db826cadf77f8c2058d10fa4ce721262dfaf09be01a842761e1526 +size 18742660 diff --git a/data/swebenchpro/anthropic/claude-opus-4-20250514/37a41783-54d5-40f2-916c-9ffb7089956b.json b/data/swebenchpro/anthropic/claude-opus-4-20250514/37a41783-54d5-40f2-916c-9ffb7089956b.json new file mode 100644 index 0000000000000000000000000000000000000000..d3f5e46f57a8de820bbd648d6210d6a2f0c7da8e --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-20250514/37a41783-54d5-40f2-916c-9ffb7089956b.json @@ -0,0 +1,133 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/anthropic_claude-opus-4-20250514/1786165738.005979", + "evaluation_timestamp": "1778545944.0", + "retrieved_timestamp": "1786165738.005979", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-20250514", + "id": "anthropic/claude-opus-4-20250514", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on swebenchpro/S-adaptive/+2ep/b2311a3e for scorer harbor_scorer", + "source_data": { + "dataset_name": "swebenchpro", + "source_type": "hf_dataset", + "samples_number": 21, + "sample_ids": [ + "instance_ansible__ansible-502270c804c33d3bc963930dc85e0f4ca359674d-v7eee2454f617569fd6889f2211f75bc02a35f9f8", + "instance_internetarchive__openlibrary-9c392b60e2c6fa1d68cb68084b4b4ff04d0cb35c-v2d9a6c849c60ed19fd0858ce9e40b7cc8e097e59", + "instance_element-hq__element-web-b007ea81b2ccd001b00f332bee65070aa7fc00f9-vnan", + "instance_ansible__ansible-811093f0225caa4dd33890933150a81c6a6d5226-v1055803c3a812189a1133297f7f5468579283f86", + "instance_flipt-io__flipt-cd2f3b0a9d4d8b8a6d3d56afab65851ecdc408e8", + "instance_internetarchive__openlibrary-7cbfb812ef0e1f9716e2d6e85d538a96fcb79d13-vfa6ff903cb27f336e17654595dd900fa943dcd91", + "instance_nodebb__nodebb-0f788b8eaa4bba3c142d171fd941d015c53b65fc-v0ec6d6c2baf3cb4797482ce4829bc25cd5716649", + "instance_flipt-io__flipt-0b119520afca1cf25c470ff4288c464d4510b944", + "instance_element-hq__element-web-f63160f38459fb552d00fcc60d4064977a9095a6-vnan", + "instance_ansible__ansible-f327e65d11bb905ed9f15996024f857a95592629-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_gravitational__teleport-73cc189b0e9636d418c4470ecce0d9af5dae2f02-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_future-architect__vuls-878c25bf5a9c9fd88ac32eb843f5636834d5712d", + "instance_ansible__ansible-42355d181a11b51ebfc56f6f4b3d9c74e01cb13b-v1055803c3a812189a1133297f7f5468579283f86", + "instance_tutao__tutanota-4b4e45949096bb288f2b522f657610e480efa3e8-vee878bb72091875e912c52fc32bc60ec3760227b", + "instance_flipt-io__flipt-05d7234fa582df632f70a7cd10194d61bd7043b9", + "instance_nodebb__nodebb-22368b996ee0e5f11a5189b400b33af3cc8d925a-v4fbcfae8b15e4ce5d132c408bca69ebb9cf146ed", + "instance_protonmail__webclients-2c3559cad02d1090985dba7e8eb5a129144d9811", + "instance_qutebrowser__qutebrowser-cf06f4e3708f886032d4d2a30108c2fddb042d81-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_internetarchive__openlibrary-3f580a5f244c299d936d73d9e327ba873b6401d9-v0f5aece3601a5b4419f7ccec1dbda2071be28ee4", + "instance_qutebrowser__qutebrowser-0aa57e4f7243024fa4bba8853306691b5dbd77b3-v5149fcda2a9a6fe1d35dfed1bade1444a11ef271", + "instance_future-architect__vuls-50580f6e98eeb36f53f27222f7f4fdfea0b21e8d" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1778545944.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.5714285714285714, + "uncertainty": { + "standard_error": { + "value": 0.10514001031342675 + }, + "num_samples": 42 + }, + "details": { + "total_matched_trajectories": "42", + "stopping_reason_count_repetition_guard": "42", + "avg_total_tokens_target_model": "3014145.50", + "total_total_tokens_target_model": "126594111.0", + "avg_total_tokens_other_models": "1315.31", + "total_total_tokens_other_models": "55243.0", + "avg_total_tokens_all_models": "3015460.81", + "total_total_tokens_all_models": "126649354.0", + "avg_turn_count": "84.10", + "total_turn_count": "3532.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 30000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/anthropic/claude-opus-4-20250514/37a41783-54d5-40f2-916c-9ffb7089956b_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "c372c3f6b7e7bd9afff8c990b40c44028870c322ee31ebb01d7599e53b5c8c3c", + "total_rows": 42 + } +} \ No newline at end of file diff --git a/data/swebenchpro/anthropic/claude-opus-4-20250514/37a41783-54d5-40f2-916c-9ffb7089956b_samples.jsonl b/data/swebenchpro/anthropic/claude-opus-4-20250514/37a41783-54d5-40f2-916c-9ffb7089956b_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..9fa94b4646b96f866ae212e4fff8b614b5fc0ccc --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-20250514/37a41783-54d5-40f2-916c-9ffb7089956b_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c372c3f6b7e7bd9afff8c990b40c44028870c322ee31ebb01d7599e53b5c8c3c +size 13250143 diff --git a/data/swebenchpro/anthropic/claude-opus-4-20250514/4362014d-0bd9-49cc-8b10-a1bc4ac9092f.json b/data/swebenchpro/anthropic/claude-opus-4-20250514/4362014d-0bd9-49cc-8b10-a1bc4ac9092f.json new file mode 100644 index 0000000000000000000000000000000000000000..6f9a842fd8bf04c6259a189f1460c9687c81991f --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-20250514/4362014d-0bd9-49cc-8b10-a1bc4ac9092f.json @@ -0,0 +1,192 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/anthropic_claude-opus-4-20250514/1786165817.755167", + "evaluation_timestamp": "1778274157.0", + "retrieved_timestamp": "1786165817.755167", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-20250514", + "id": "anthropic/claude-opus-4-20250514", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on swebenchpro/S-adaptive/9111ee97 for scorer harbor_scorer", + "source_data": { + "dataset_name": "swebenchpro", + "source_type": "hf_dataset", + "samples_number": 80, + "sample_ids": [ + "instance_nodebb__nodebb-70b4a0e2aebebe8f2f559de6680093d96a697b2f-vnan", + "instance_ansible__ansible-e22e103cdf8edc56ff7d9b848a58f94f1471a263-v1055803c3a812189a1133297f7f5468579283f86", + "instance_qutebrowser__qutebrowser-e15d26630934d0b6415ed2295ac42fd570a57620-va0fd88aac89cde702ec1ba84877234da33adce8a", + "instance_ansible__ansible-164881d871964aa64e0f911d03ae270acbad253c-v390e508d27db7a51eece36bb6d9698b63a5b638a", + "instance_ansible__ansible-502270c804c33d3bc963930dc85e0f4ca359674d-v7eee2454f617569fd6889f2211f75bc02a35f9f8", + "instance_internetarchive__openlibrary-9c392b60e2c6fa1d68cb68084b4b4ff04d0cb35c-v2d9a6c849c60ed19fd0858ce9e40b7cc8e097e59", + "instance_element-hq__element-web-b007ea81b2ccd001b00f332bee65070aa7fc00f9-vnan", + "instance_ansible__ansible-811093f0225caa4dd33890933150a81c6a6d5226-v1055803c3a812189a1133297f7f5468579283f86", + "instance_flipt-io__flipt-cd2f3b0a9d4d8b8a6d3d56afab65851ecdc408e8", + "instance_internetarchive__openlibrary-7cbfb812ef0e1f9716e2d6e85d538a96fcb79d13-vfa6ff903cb27f336e17654595dd900fa943dcd91", + "instance_nodebb__nodebb-0f788b8eaa4bba3c142d171fd941d015c53b65fc-v0ec6d6c2baf3cb4797482ce4829bc25cd5716649", + "instance_flipt-io__flipt-0b119520afca1cf25c470ff4288c464d4510b944", + "instance_element-hq__element-web-f63160f38459fb552d00fcc60d4064977a9095a6-vnan", + "instance_ansible__ansible-f327e65d11bb905ed9f15996024f857a95592629-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_gravitational__teleport-73cc189b0e9636d418c4470ecce0d9af5dae2f02-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_internetarchive__openlibrary-3c48b4bb782189e0858e6c3fc7956046cf3e1cfb-v2d9a6c849c60ed19fd0858ce9e40b7cc8e097e59", + "instance_future-architect__vuls-878c25bf5a9c9fd88ac32eb843f5636834d5712d", + "instance_element-hq__element-web-9bf77963ee5e036d54b2a3ca202fbf6378464a5e-vnan", + "instance_qutebrowser__qutebrowser-70248f256f93ed9b1984494d0a1a919ddd774892-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_future-architect__vuls-73f0adad95c4d227e2ccfa876c85cc95dd065e13", + "instance_protonmail__webclients-da91f084c0f532d9cc8ca385a701274d598057b8", + "instance_flipt-io__flipt-b2cd6a6dd73ca91b519015fd5924fde8d17f3f06", + "instance_qutebrowser__qutebrowser-cc360cd4a34a126274c7b51f3b63afbaf3e05a02-v5fc38aaf22415ab0b70567368332beee7955b367", + "instance_flipt-io__flipt-e2bd19dafa7166c96b082fb2a59eb54b4be0d778", + "instance_qutebrowser__qutebrowser-9ed748effa8f3bcd804612d9291da017b514e12f-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_ansible__ansible-1ee70fc272aff6bf3415357c6e13c5de5b928d9b-v1055803c3a812189a1133297f7f5468579283f86", + "instance_element-hq__element-web-75c2c1a572fa45d1ea1d1a96e9e36e303332ecaa-vnan", + "instance_qutebrowser__qutebrowser-f631cd4422744160d9dcf7a0455da532ce973315-v35616345bb8052ea303186706cec663146f0f184", + "instance_ansible__ansible-3db08adbb1cc6aa9941be5e0fc810132c6e1fa4b-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_qutebrowser__qutebrowser-e5340c449f23608803c286da0563b62f58ba25b0-v059c6fdc75567943479b23ebca7c07b5e9a7f34c", + "instance_ansible__ansible-42355d181a11b51ebfc56f6f4b3d9c74e01cb13b-v1055803c3a812189a1133297f7f5468579283f86", + "instance_internetarchive__openlibrary-09865f5fb549694d969f0a8e49b9d204ef1853ca-ve8c8d62a2b60610a3c4631f5f23ed866bada9818", + "instance_internetarchive__openlibrary-c506c1b0b678892af5cb22c1c1dbc35d96787a0a-v0f5aece3601a5b4419f7ccec1dbda2071be28ee4", + "instance_element-hq__element-web-2760bfc8369f1bee640d6d7a7e910783143d4c5f-vnan", + "instance_internetarchive__openlibrary-25858f9f0c165df25742acf8309ce909773f0cdd-v13642507b4fc1f8d234172bf8129942da2c2ca26", + "instance_protonmail__webclients-01ea5214d11e0df8b7170d91bafd34f23cb0f2b1", + "instance_tutao__tutanota-4b4e45949096bb288f2b522f657610e480efa3e8-vee878bb72091875e912c52fc32bc60ec3760227b", + "instance_flipt-io__flipt-05d7234fa582df632f70a7cd10194d61bd7043b9", + "instance_ansible__ansible-a02e22e902a69aeb465f16bf03f7f5a91b2cb828-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_protonmail__webclients-5f0745dd6993bb1430a951c62a49807c6635cd77", + "instance_tutao__tutanota-1ff82aa365763cee2d609c9d19360ad87fdf2ec7-vc4e41fd0029957297843cb9dec4a25c7c756f029", + "instance_internetarchive__openlibrary-30bc73a1395fba2300087c7f307e54bb5372b60a-v76304ecdb3a5954fcf13feb710e8c40fcf24b73c", + "instance_nodebb__nodebb-8ca65b0c78c67c1653487c02d1135e1b702185e1-vf2cf3cbd463b7ad942381f1c6d077626485a1e9e", + "instance_nodebb__nodebb-22368b996ee0e5f11a5189b400b33af3cc8d925a-v4fbcfae8b15e4ce5d132c408bca69ebb9cf146ed", + "instance_protonmail__webclients-2c3559cad02d1090985dba7e8eb5a129144d9811", + "instance_protonmail__webclients-01b519cd49e6a24d9a05d2eb97f54e420740072e", + "instance_ansible__ansible-77658704217d5f166404fc67997203c25381cb6e-v390e508d27db7a51eece36bb6d9698b63a5b638a", + "instance_qutebrowser__qutebrowser-2dd8966fdcf11972062c540b7a787e4d0de8d372-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_ansible__ansible-cd9c4eb5a6b2bfaf4a6709f001ce3d0c92c1eed2-v0f01c69f1e2528b935359cfe578530722bca2c59", + "instance_protonmail__webclients-ac23d1efa1a6ab7e62724779317ba44c28d78cfd", + "instance_future-architect__vuls-030b2e03525d68d74cb749959aac2d7f3fc0effa", + "instance_nodebb__nodebb-b1f9ad5534bb3a44dab5364f659876a4b7fe34c1-vnan", + "instance_flipt-io__flipt-292fdaca9be39e6a921aaa8874c011d0fdd3e874", + "instance_qutebrowser__qutebrowser-ef5ba1a0360b39f9eff027fbdc57f363597c3c3b-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_ansible__ansible-622a493ae03bd5e5cf517d336fc426e9d12208c7-v906c969b551b346ef54a2c0b41e04f632b7b73c2", + "instance_gravitational__teleport-82185f232ae8974258397e121b3bc2ed0c3729ed-v626ec2a48416b10a88641359a169d99e935ff037", + "instance_ansible__ansible-9a21e247786ebd294dafafca1105fcd770ff46c6-v67cdaa49f89b34e42b69d5b7830b3c3ad3d8803f", + "instance_protonmail__webclients-8afd9ce04c8dde9e150e1c2b50d32e7ee2efa3e7", + "instance_flipt-io__flipt-b22f5f02e40b225b6b93fff472914973422e97c6", + "instance_gravitational__teleport-2bb3bbbd8aff1164a2353381cb79e1dc93b90d28-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_ansible__ansible-de5858f48dc9e1ce9117034e0d7e76806f420ca8-v1055803c3a812189a1133297f7f5468579283f86", + "instance_gravitational__teleport-46aa81b1ce96ebb4ebed2ae53fd78cd44a05da6c-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_tutao__tutanota-8513a9e8114a8b42e64f4348335e0f23efa054c4-vee878bb72091875e912c52fc32bc60ec3760227b", + "instance_internetarchive__openlibrary-630221ab686c64e75a2ce253c893c033e4814b2e-v93c53c13d5f9b383ebb411ee7750b49dcd1a34c6", + "instance_internetarchive__openlibrary-7f6b722a10f822171501d027cad60afe53337732-ve8c8d62a2b60610a3c4631f5f23ed866bada9818", + "instance_nodebb__nodebb-a917210c5b2c20637094545401f85783905c074c-vf2cf3cbd463b7ad942381f1c6d077626485a1e9e", + "instance_internetarchive__openlibrary-1351c59fd43689753de1fca32c78d539a116ffc1-v29f82c9cf21d57b242f8d8b0e541525d259e2d63", + "instance_qutebrowser__qutebrowser-a25e8a09873838ca9efefd36ea8a45170bbeb95c-vc2f56a753b62a190ddb23cd330c257b9cf560d12", + "instance_protonmail__webclients-281a6b3f190f323ec2c0630999354fafb84b2880", + "instance_future-architect__vuls-999529a05b202b0fd29c6fca5039a4c47a3766bb", + "instance_element-hq__element-web-1077729a19c0ce902e713cf6fab42c91fb7907f1-vnan", + "instance_future-architect__vuls-be7b9114cc9545e68fb0ee7bc63d7ec53d1a00ad", + "instance_qutebrowser__qutebrowser-c580ebf0801e5a3ecabc54f327498bb753c6d5f2-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_gravitational__teleport-3ff19cf7c41f396ae468797d3aeb61515517edc9-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_future-architect__vuls-aaea15e516ece43978cf98e09e52080478b1d39f", + "instance_protonmail__webclients-f161c10cf7d31abf82e8d64d7a99c9fac5acfa18", + "instance_qutebrowser__qutebrowser-cf06f4e3708f886032d4d2a30108c2fddb042d81-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_internetarchive__openlibrary-3f580a5f244c299d936d73d9e327ba873b6401d9-v0f5aece3601a5b4419f7ccec1dbda2071be28ee4", + "instance_qutebrowser__qutebrowser-0aa57e4f7243024fa4bba8853306691b5dbd77b3-v5149fcda2a9a6fe1d35dfed1bade1444a11ef271", + "instance_future-architect__vuls-50580f6e98eeb36f53f27222f7f4fdfea0b21e8d" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1778274157.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.5010416666666667, + "uncertainty": { + "standard_error": { + "value": 0.04786432543042873 + }, + "num_samples": 296 + }, + "details": { + "total_matched_trajectories": "296", + "stopping_reason_count_repetition_guard": "296", + "avg_total_tokens_target_model": "1607824.13", + "total_total_tokens_target_model": "475915943.0", + "avg_total_tokens_other_models": "1310.08", + "total_total_tokens_other_models": "387784.0", + "avg_total_tokens_all_models": "1609134.21", + "total_total_tokens_all_models": "476303727.0", + "avg_turn_count": "63.36", + "total_turn_count": "18755.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 30000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/anthropic/claude-opus-4-20250514/4362014d-0bd9-49cc-8b10-a1bc4ac9092f_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "c77c4ef42899f228f59866d5ffd3ae9fd1245af7795f3d71a3aecccb0923546e", + "total_rows": 296 + } +} \ No newline at end of file diff --git a/data/swebenchpro/anthropic/claude-opus-4-20250514/4362014d-0bd9-49cc-8b10-a1bc4ac9092f_samples.jsonl b/data/swebenchpro/anthropic/claude-opus-4-20250514/4362014d-0bd9-49cc-8b10-a1bc4ac9092f_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..4adc2ed7680cdaa517a35311e762a5a3ebb8b118 --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-20250514/4362014d-0bd9-49cc-8b10-a1bc4ac9092f_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c77c4ef42899f228f59866d5ffd3ae9fd1245af7795f3d71a3aecccb0923546e +size 75591738 diff --git a/data/swebenchpro/anthropic/claude-opus-4-20250514/4c5021d6-18f1-4cb9-851d-25b58432f1dd.json b/data/swebenchpro/anthropic/claude-opus-4-20250514/4c5021d6-18f1-4cb9-851d-25b58432f1dd.json new file mode 100644 index 0000000000000000000000000000000000000000..ea9853e7aa55c1002e0678ac72ec6bae4e8826ff --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-20250514/4c5021d6-18f1-4cb9-851d-25b58432f1dd.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/anthropic_claude-opus-4-20250514/1786169526.880115", + "evaluation_timestamp": "1777772817.0", + "retrieved_timestamp": "1786169526.880115", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-20250514", + "id": "anthropic/claude-opus-4-20250514", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/anthropic/claude-opus-4-20250514/4c5021d6-18f1-4cb9-851d-25b58432f1dd_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "785898e28af17252709225ff5298d372478d30db87393a6a80fb97dec284e923", + "total_rows": 31 + } +} \ No newline at end of file diff --git a/data/swebenchpro/anthropic/claude-opus-4-20250514/4c5021d6-18f1-4cb9-851d-25b58432f1dd_samples.jsonl b/data/swebenchpro/anthropic/claude-opus-4-20250514/4c5021d6-18f1-4cb9-851d-25b58432f1dd_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..d0245fbaac3d09f3d1a3c389fae3885f9f068607 --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-20250514/4c5021d6-18f1-4cb9-851d-25b58432f1dd_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:785898e28af17252709225ff5298d372478d30db87393a6a80fb97dec284e923 +size 24974388 diff --git a/data/swebenchpro/anthropic/claude-opus-4-20250514/735476bd-ef1e-4213-8886-3738d6cf7e11.json b/data/swebenchpro/anthropic/claude-opus-4-20250514/735476bd-ef1e-4213-8886-3738d6cf7e11.json new file mode 100644 index 0000000000000000000000000000000000000000..0dffae3e314c1a53cd1f5b89423e9ad6e01bb88b --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-20250514/735476bd-ef1e-4213-8886-3738d6cf7e11.json @@ -0,0 +1,112 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/anthropic_claude-opus-4-20250514/1786163194.934347", + "evaluation_timestamp": "1778703021.0", + "retrieved_timestamp": "1786163194.934347", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-20250514", + "id": "anthropic/claude-opus-4-20250514", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on swebenchpro/S-adaptive+C/+2ep/e7886a1f for scorer harbor_scorer", + "source_data": { + "dataset_name": "swebenchpro", + "source_type": "hf_dataset", + "samples_number": 3, + "sample_ids": [ + "instance_internetarchive__openlibrary-f8cc11d9c1575fdba5ac66aee0befca970da8d64-v13642507b4fc1f8d234172bf8129942da2c2ca26", + "instance_qutebrowser__qutebrowser-e70f5b03187bdd40e8bf70f5f3ead840f52d1f42-v02ad04386d5238fe2d1a1be450df257370de4b6a", + "instance_element-hq__element-web-9bf77963ee5e036d54b2a3ca202fbf6378464a5e-vnan" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1778703021.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 1.0, + "uncertainty": { + "num_samples": 6 + }, + "details": { + "total_matched_trajectories": "6", + "stopping_reason_count_completed_on_successful_submit": "6", + "avg_total_tokens_target_model": "500951.17", + "total_total_tokens_target_model": "3005707.0", + "avg_total_tokens_other_models": "0.00", + "total_total_tokens_other_models": "0.0", + "avg_total_tokens_all_models": "500951.17", + "total_total_tokens_all_models": "3005707.0", + "avg_turn_count": "32.00", + "total_turn_count": "192.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "8", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 30000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "8", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/anthropic/claude-opus-4-20250514/735476bd-ef1e-4213-8886-3738d6cf7e11_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "7610c0ccead2015d45a40f84d87e284e640156d48159fa4cdfa989eec70dc4d8", + "total_rows": 6 + } +} \ No newline at end of file diff --git a/data/swebenchpro/anthropic/claude-opus-4-20250514/735476bd-ef1e-4213-8886-3738d6cf7e11_samples.jsonl b/data/swebenchpro/anthropic/claude-opus-4-20250514/735476bd-ef1e-4213-8886-3738d6cf7e11_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..b5a4df51bf5cf90dac514968b7ecfbb83d83c7f3 --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-20250514/735476bd-ef1e-4213-8886-3738d6cf7e11_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7610c0ccead2015d45a40f84d87e284e640156d48159fa4cdfa989eec70dc4d8 +size 896865 diff --git a/data/swebenchpro/anthropic/claude-opus-4-20250514/7d65fd44-7584-4f04-895c-d6354b83303f.json b/data/swebenchpro/anthropic/claude-opus-4-20250514/7d65fd44-7584-4f04-895c-d6354b83303f.json new file mode 100644 index 0000000000000000000000000000000000000000..7908ee82347bd66d57068a8d6f5293bf96bdfa28 --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-20250514/7d65fd44-7584-4f04-895c-d6354b83303f.json @@ -0,0 +1,111 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/anthropic_claude-opus-4-20250514/1786163207.761319", + "evaluation_timestamp": "1778704753.0", + "retrieved_timestamp": "1786163207.761319", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-20250514", + "id": "anthropic/claude-opus-4-20250514", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on swebenchpro/S-adaptive+C/+3ep/12fd67db for scorer harbor_scorer", + "source_data": { + "dataset_name": "swebenchpro", + "source_type": "hf_dataset", + "samples_number": 1, + "sample_ids": [ + "instance_element-hq__element-web-75c2c1a572fa45d1ea1d1a96e9e36e303332ecaa-vnan" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1778704753.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.3333333333333333, + "uncertainty": { + "num_samples": 3 + }, + "details": { + "total_matched_trajectories": "3", + "stopping_reason_count_completed_on_successful_submit": "1", + "stopping_reason_count_repetition_guard": "2", + "avg_total_tokens_target_model": "1330209.33", + "total_total_tokens_target_model": "3990628.0", + "avg_total_tokens_other_models": "858.00", + "total_total_tokens_other_models": "2574.0", + "avg_total_tokens_all_models": "1331067.33", + "total_total_tokens_all_models": "3993202.0", + "avg_turn_count": "55.33", + "total_turn_count": "166.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "8", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 30000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "8", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/anthropic/claude-opus-4-20250514/7d65fd44-7584-4f04-895c-d6354b83303f_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "373f1891a7459f24ba04af48c13c92d68fdfc823cbbf078a1a0ef22fcd019757", + "total_rows": 3 + } +} \ No newline at end of file diff --git a/data/swebenchpro/anthropic/claude-opus-4-20250514/7d65fd44-7584-4f04-895c-d6354b83303f_samples.jsonl b/data/swebenchpro/anthropic/claude-opus-4-20250514/7d65fd44-7584-4f04-895c-d6354b83303f_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..453f6a84be898c75b36aa5ea265a71c9e25b4592 --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-20250514/7d65fd44-7584-4f04-895c-d6354b83303f_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:373f1891a7459f24ba04af48c13c92d68fdfc823cbbf078a1a0ef22fcd019757 +size 780365 diff --git a/data/swebenchpro/anthropic/claude-opus-4-20250514/95a1ea16-3df0-4b07-9a10-a0612d02221b.json b/data/swebenchpro/anthropic/claude-opus-4-20250514/95a1ea16-3df0-4b07-9a10-a0612d02221b.json new file mode 100644 index 0000000000000000000000000000000000000000..44bb3192be7ececa1f977544b7acf89c7a1ce830 --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-20250514/95a1ea16-3df0-4b07-9a10-a0612d02221b.json @@ -0,0 +1,132 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/anthropic_claude-opus-4-20250514/1786163215.575933", + "evaluation_timestamp": "1778858244.0", + "retrieved_timestamp": "1786163215.575933", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-20250514", + "id": "anthropic/claude-opus-4-20250514", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on swebenchpro/S-adaptive+C/+3ep/7f13f3ab for scorer harbor_scorer", + "source_data": { + "dataset_name": "swebenchpro", + "source_type": "hf_dataset", + "samples_number": 19, + "sample_ids": [ + "instance_ansible__ansible-811093f0225caa4dd33890933150a81c6a6d5226-v1055803c3a812189a1133297f7f5468579283f86", + "instance_flipt-io__flipt-b22f5f02e40b225b6b93fff472914973422e97c6", + "instance_gravitational__teleport-2bb3bbbd8aff1164a2353381cb79e1dc93b90d28-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_gravitational__teleport-46aa81b1ce96ebb4ebed2ae53fd78cd44a05da6c-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_tutao__tutanota-8513a9e8114a8b42e64f4348335e0f23efa054c4-vee878bb72091875e912c52fc32bc60ec3760227b", + "instance_internetarchive__openlibrary-630221ab686c64e75a2ce253c893c033e4814b2e-v93c53c13d5f9b383ebb411ee7750b49dcd1a34c6", + "instance_internetarchive__openlibrary-7f6b722a10f822171501d027cad60afe53337732-ve8c8d62a2b60610a3c4631f5f23ed866bada9818", + "instance_nodebb__nodebb-a917210c5b2c20637094545401f85783905c074c-vf2cf3cbd463b7ad942381f1c6d077626485a1e9e", + "instance_internetarchive__openlibrary-1351c59fd43689753de1fca32c78d539a116ffc1-v29f82c9cf21d57b242f8d8b0e541525d259e2d63", + "instance_protonmail__webclients-281a6b3f190f323ec2c0630999354fafb84b2880", + "instance_future-architect__vuls-999529a05b202b0fd29c6fca5039a4c47a3766bb", + "instance_future-architect__vuls-be7b9114cc9545e68fb0ee7bc63d7ec53d1a00ad", + "instance_qutebrowser__qutebrowser-c580ebf0801e5a3ecabc54f327498bb753c6d5f2-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_gravitational__teleport-3ff19cf7c41f396ae468797d3aeb61515517edc9-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_future-architect__vuls-aaea15e516ece43978cf98e09e52080478b1d39f", + "instance_qutebrowser__qutebrowser-cf06f4e3708f886032d4d2a30108c2fddb042d81-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_internetarchive__openlibrary-3f580a5f244c299d936d73d9e327ba873b6401d9-v0f5aece3601a5b4419f7ccec1dbda2071be28ee4", + "instance_qutebrowser__qutebrowser-0aa57e4f7243024fa4bba8853306691b5dbd77b3-v5149fcda2a9a6fe1d35dfed1bade1444a11ef271", + "instance_future-architect__vuls-50580f6e98eeb36f53f27222f7f4fdfea0b21e8d" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1778858244.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.45614035087719296, + "uncertainty": { + "standard_error": { + "value": 0.1087782177645513 + }, + "num_samples": 56 + }, + "details": { + "total_matched_trajectories": "56", + "stopping_reason_count_repetition_guard": "31", + "stopping_reason_count_completed_on_successful_submit": "25", + "avg_total_tokens_target_model": "2667723.91", + "total_total_tokens_target_model": "149392539.0", + "avg_total_tokens_other_models": "830.14", + "total_total_tokens_other_models": "46488.0", + "avg_total_tokens_all_models": "2668554.05", + "total_total_tokens_all_models": "149439027.0", + "avg_turn_count": "81.62", + "total_turn_count": "4571.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "8", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 30000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "8", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/anthropic/claude-opus-4-20250514/95a1ea16-3df0-4b07-9a10-a0612d02221b_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "19e23741c3586334f062ccca09873851590a495153841306b72a9e44faca1ce7", + "total_rows": 56 + } +} \ No newline at end of file diff --git a/data/swebenchpro/anthropic/claude-opus-4-20250514/95a1ea16-3df0-4b07-9a10-a0612d02221b_samples.jsonl b/data/swebenchpro/anthropic/claude-opus-4-20250514/95a1ea16-3df0-4b07-9a10-a0612d02221b_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..000501d8236f793ee21950aa74a3e518fc2127bd --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-20250514/95a1ea16-3df0-4b07-9a10-a0612d02221b_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:19e23741c3586334f062ccca09873851590a495153841306b72a9e44faca1ce7 +size 17954849 diff --git a/data/swebenchpro/anthropic/claude-opus-4-20250514/97b5746b-cd8d-4052-ab81-f791cdd9cd19.json b/data/swebenchpro/anthropic/claude-opus-4-20250514/97b5746b-cd8d-4052-ab81-f791cdd9cd19.json new file mode 100644 index 0000000000000000000000000000000000000000..7ece01890eed6dbe4e30a621410a09ec088fdf1a --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-20250514/97b5746b-cd8d-4052-ab81-f791cdd9cd19.json @@ -0,0 +1,153 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/anthropic_claude-opus-4-20250514/1786163203.390802", + "evaluation_timestamp": "1778854176.0", + "retrieved_timestamp": "1786163203.390802", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-20250514", + "id": "anthropic/claude-opus-4-20250514", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on swebenchpro/S-adaptive+C/+2ep/43258232 for scorer harbor_scorer", + "source_data": { + "dataset_name": "swebenchpro", + "source_type": "hf_dataset", + "samples_number": 40, + "sample_ids": [ + "instance_qutebrowser__qutebrowser-2e961080a85d660148937ee8f0f6b3445a8f2c01-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_flipt-io__flipt-b68b8960b8a08540d5198d78c665a7eb0bea4008", + "instance_qutebrowser__qutebrowser-01d1d1494411380d97cac14614a829d3a69cecaf-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_future-architect__vuls-78b52d6a7f480bd610b692de9bf0c86f57332f23", + "instance_ansible__ansible-e22e103cdf8edc56ff7d9b848a58f94f1471a263-v1055803c3a812189a1133297f7f5468579283f86", + "instance_flipt-io__flipt-cd2f3b0a9d4d8b8a6d3d56afab65851ecdc408e8", + "instance_future-architect__vuls-878c25bf5a9c9fd88ac32eb843f5636834d5712d", + "instance_qutebrowser__qutebrowser-70248f256f93ed9b1984494d0a1a919ddd774892-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_future-architect__vuls-73f0adad95c4d227e2ccfa876c85cc95dd065e13", + "instance_protonmail__webclients-da91f084c0f532d9cc8ca385a701274d598057b8", + "instance_flipt-io__flipt-b2cd6a6dd73ca91b519015fd5924fde8d17f3f06", + "instance_qutebrowser__qutebrowser-cc360cd4a34a126274c7b51f3b63afbaf3e05a02-v5fc38aaf22415ab0b70567368332beee7955b367", + "instance_flipt-io__flipt-e2bd19dafa7166c96b082fb2a59eb54b4be0d778", + "instance_qutebrowser__qutebrowser-9ed748effa8f3bcd804612d9291da017b514e12f-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_ansible__ansible-3db08adbb1cc6aa9941be5e0fc810132c6e1fa4b-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_ansible__ansible-42355d181a11b51ebfc56f6f4b3d9c74e01cb13b-v1055803c3a812189a1133297f7f5468579283f86", + "instance_internetarchive__openlibrary-09865f5fb549694d969f0a8e49b9d204ef1853ca-ve8c8d62a2b60610a3c4631f5f23ed866bada9818", + "instance_internetarchive__openlibrary-c506c1b0b678892af5cb22c1c1dbc35d96787a0a-v0f5aece3601a5b4419f7ccec1dbda2071be28ee4", + "instance_element-hq__element-web-2760bfc8369f1bee640d6d7a7e910783143d4c5f-vnan", + "instance_internetarchive__openlibrary-25858f9f0c165df25742acf8309ce909773f0cdd-v13642507b4fc1f8d234172bf8129942da2c2ca26", + "instance_protonmail__webclients-01ea5214d11e0df8b7170d91bafd34f23cb0f2b1", + "instance_tutao__tutanota-4b4e45949096bb288f2b522f657610e480efa3e8-vee878bb72091875e912c52fc32bc60ec3760227b", + "instance_flipt-io__flipt-05d7234fa582df632f70a7cd10194d61bd7043b9", + "instance_ansible__ansible-a02e22e902a69aeb465f16bf03f7f5a91b2cb828-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_protonmail__webclients-5f0745dd6993bb1430a951c62a49807c6635cd77", + "instance_tutao__tutanota-1ff82aa365763cee2d609c9d19360ad87fdf2ec7-vc4e41fd0029957297843cb9dec4a25c7c756f029", + "instance_internetarchive__openlibrary-30bc73a1395fba2300087c7f307e54bb5372b60a-v76304ecdb3a5954fcf13feb710e8c40fcf24b73c", + "instance_nodebb__nodebb-8ca65b0c78c67c1653487c02d1135e1b702185e1-vf2cf3cbd463b7ad942381f1c6d077626485a1e9e", + "instance_nodebb__nodebb-22368b996ee0e5f11a5189b400b33af3cc8d925a-v4fbcfae8b15e4ce5d132c408bca69ebb9cf146ed", + "instance_protonmail__webclients-01b519cd49e6a24d9a05d2eb97f54e420740072e", + "instance_ansible__ansible-77658704217d5f166404fc67997203c25381cb6e-v390e508d27db7a51eece36bb6d9698b63a5b638a", + "instance_qutebrowser__qutebrowser-2dd8966fdcf11972062c540b7a787e4d0de8d372-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_ansible__ansible-cd9c4eb5a6b2bfaf4a6709f001ce3d0c92c1eed2-v0f01c69f1e2528b935359cfe578530722bca2c59", + "instance_protonmail__webclients-ac23d1efa1a6ab7e62724779317ba44c28d78cfd", + "instance_future-architect__vuls-030b2e03525d68d74cb749959aac2d7f3fc0effa", + "instance_qutebrowser__qutebrowser-ef5ba1a0360b39f9eff027fbdc57f363597c3c3b-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_ansible__ansible-622a493ae03bd5e5cf517d336fc426e9d12208c7-v906c969b551b346ef54a2c0b41e04f632b7b73c2", + "instance_gravitational__teleport-82185f232ae8974258397e121b3bc2ed0c3729ed-v626ec2a48416b10a88641359a169d99e935ff037", + "instance_ansible__ansible-9a21e247786ebd294dafafca1105fcd770ff46c6-v67cdaa49f89b34e42b69d5b7830b3c3ad3d8803f", + "instance_protonmail__webclients-8afd9ce04c8dde9e150e1c2b50d32e7ee2efa3e7" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1778854176.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.575, + "uncertainty": { + "standard_error": { + "value": 0.07283191044787428 + }, + "num_samples": 78 + }, + "details": { + "total_matched_trajectories": "78", + "stopping_reason_count_completed_on_successful_submit": "45", + "stopping_reason_count_repetition_guard": "33", + "avg_total_tokens_target_model": "1703346.35", + "total_total_tokens_target_model": "132861015.0", + "avg_total_tokens_other_models": "633.74", + "total_total_tokens_other_models": "49432.0", + "avg_total_tokens_all_models": "1703980.09", + "total_total_tokens_all_models": "132910447.0", + "avg_turn_count": "62.31", + "total_turn_count": "4860.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "8", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 30000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "8", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/anthropic/claude-opus-4-20250514/97b5746b-cd8d-4052-ab81-f791cdd9cd19_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "77a9661d78c37895222634762b379b0967e08bd01027b017344c42e3b7854f95", + "total_rows": 78 + } +} \ No newline at end of file diff --git a/data/swebenchpro/anthropic/claude-opus-4-20250514/97b5746b-cd8d-4052-ab81-f791cdd9cd19_samples.jsonl b/data/swebenchpro/anthropic/claude-opus-4-20250514/97b5746b-cd8d-4052-ab81-f791cdd9cd19_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..48b5ad724cca10e3f2a4eb6ecca3ad406c8c9e92 --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-20250514/97b5746b-cd8d-4052-ab81-f791cdd9cd19_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:77a9661d78c37895222634762b379b0967e08bd01027b017344c42e3b7854f95 +size 19744552 diff --git a/data/swebenchpro/anthropic/claude-opus-4-20250514/a3d749fd-9a4e-4dc9-9632-955f4c207261.json b/data/swebenchpro/anthropic/claude-opus-4-20250514/a3d749fd-9a4e-4dc9-9632-955f4c207261.json new file mode 100644 index 0000000000000000000000000000000000000000..3cd8ec8294bfdc062c346a83bb5d2bfc86b5d821 --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-20250514/a3d749fd-9a4e-4dc9-9632-955f4c207261.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/anthropic_claude-opus-4-20250514/1786167467.258899", + "evaluation_timestamp": "1778675421.0", + "retrieved_timestamp": "1786167467.258899", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-20250514", + "id": "anthropic/claude-opus-4-20250514", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/anthropic/claude-opus-4-20250514/a3d749fd-9a4e-4dc9-9632-955f4c207261_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "9dd39a9bd358efb40bf60902e959ed962e03a8800761b361aaa062293a648495", + "total_rows": 1 + } +} \ No newline at end of file diff --git a/data/swebenchpro/anthropic/claude-opus-4-20250514/a3d749fd-9a4e-4dc9-9632-955f4c207261_samples.jsonl b/data/swebenchpro/anthropic/claude-opus-4-20250514/a3d749fd-9a4e-4dc9-9632-955f4c207261_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..2c31e02887a46ac9140fedbe4479da4b0ebd5abd --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-20250514/a3d749fd-9a4e-4dc9-9632-955f4c207261_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9dd39a9bd358efb40bf60902e959ed962e03a8800761b361aaa062293a648495 +size 131107 diff --git a/data/swebenchpro/anthropic/claude-opus-4-20250514/ba498d65-792a-4017-8584-5a90866a36d9.json b/data/swebenchpro/anthropic/claude-opus-4-20250514/ba498d65-792a-4017-8584-5a90866a36d9.json new file mode 100644 index 0000000000000000000000000000000000000000..b9129e440e122cd545ff80b81635d120fedaa774 --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-20250514/ba498d65-792a-4017-8584-5a90866a36d9.json @@ -0,0 +1,110 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/anthropic_claude-opus-4-20250514/1786165759.551444", + "evaluation_timestamp": "1778591254.0", + "retrieved_timestamp": "1786165759.551444", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-20250514", + "id": "anthropic/claude-opus-4-20250514", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on swebenchpro/S-adaptive/+4ep/e04ec096 for scorer harbor_scorer", + "source_data": { + "dataset_name": "swebenchpro", + "source_type": "hf_dataset", + "samples_number": 1, + "sample_ids": [ + "instance_ansible__ansible-de5858f48dc9e1ce9117034e0d7e76806f420ca8-v1055803c3a812189a1133297f7f5468579283f86" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1778591254.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.0, + "uncertainty": { + "num_samples": 2 + }, + "details": { + "total_matched_trajectories": "2", + "stopping_reason_count_repetition_guard": "2", + "avg_total_tokens_target_model": "2846830.50", + "total_total_tokens_target_model": "5693661.0", + "avg_total_tokens_other_models": "1734.50", + "total_total_tokens_other_models": "3469.0", + "avg_total_tokens_all_models": "2848565.00", + "total_total_tokens_all_models": "5697130.0", + "avg_turn_count": "92.00", + "total_turn_count": "184.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 30000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/anthropic/claude-opus-4-20250514/ba498d65-792a-4017-8584-5a90866a36d9_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "b74ee16693706a1d6321406ee11a548164b4af20529c30b9f605239df8f46d75", + "total_rows": 2 + } +} \ No newline at end of file diff --git a/data/swebenchpro/anthropic/claude-opus-4-20250514/ba498d65-792a-4017-8584-5a90866a36d9_samples.jsonl b/data/swebenchpro/anthropic/claude-opus-4-20250514/ba498d65-792a-4017-8584-5a90866a36d9_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..8c01881761a5738365f60b1f9f2ba24293da0884 --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-20250514/ba498d65-792a-4017-8584-5a90866a36d9_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b74ee16693706a1d6321406ee11a548164b4af20529c30b9f605239df8f46d75 +size 788402 diff --git a/data/swebenchpro/anthropic/claude-opus-4-20250514/c3b1c5c9-c32f-4029-8721-0202e26bb8da.json b/data/swebenchpro/anthropic/claude-opus-4-20250514/c3b1c5c9-c32f-4029-8721-0202e26bb8da.json new file mode 100644 index 0000000000000000000000000000000000000000..09abb4e671c664aa60b6f01dbd0eb1976a31a2b9 --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-20250514/c3b1c5c9-c32f-4029-8721-0202e26bb8da.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/anthropic_claude-opus-4-20250514/1786167508.033498", + "evaluation_timestamp": "1778763061.0", + "retrieved_timestamp": "1786167508.033498", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-20250514", + "id": "anthropic/claude-opus-4-20250514", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/anthropic/claude-opus-4-20250514/c3b1c5c9-c32f-4029-8721-0202e26bb8da_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "9da87140a612cf522928dbfbd235aa3ef9a72191953c855c5ebfbffaf6c17c6e", + "total_rows": 159 + } +} \ No newline at end of file diff --git a/data/swebenchpro/anthropic/claude-opus-4-20250514/c3b1c5c9-c32f-4029-8721-0202e26bb8da_samples.jsonl b/data/swebenchpro/anthropic/claude-opus-4-20250514/c3b1c5c9-c32f-4029-8721-0202e26bb8da_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..6608414b5695d1d5ad47d77760ad09635c430cf1 --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-20250514/c3b1c5c9-c32f-4029-8721-0202e26bb8da_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9da87140a612cf522928dbfbd235aa3ef9a72191953c855c5ebfbffaf6c17c6e +size 44862938 diff --git a/data/swebenchpro/anthropic/claude-opus-4-20250514/cece1192-b954-4a31-8464-7afdd0b2cfad.json b/data/swebenchpro/anthropic/claude-opus-4-20250514/cece1192-b954-4a31-8464-7afdd0b2cfad.json new file mode 100644 index 0000000000000000000000000000000000000000..2b111abe6b6a9bf5d212a358beaf31c028ea6c49 --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-20250514/cece1192-b954-4a31-8464-7afdd0b2cfad.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/anthropic_claude-opus-4-20250514/1786169514.16688", + "evaluation_timestamp": "1778017874.0", + "retrieved_timestamp": "1786169514.16688", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-20250514", + "id": "anthropic/claude-opus-4-20250514", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/anthropic/claude-opus-4-20250514/cece1192-b954-4a31-8464-7afdd0b2cfad_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "875c1245b0929b57ea063faa39cde4c72c305665a75c402a59bcc6fc0f4857ee", + "total_rows": 20 + } +} \ No newline at end of file diff --git a/data/swebenchpro/anthropic/claude-opus-4-20250514/cece1192-b954-4a31-8464-7afdd0b2cfad_samples.jsonl b/data/swebenchpro/anthropic/claude-opus-4-20250514/cece1192-b954-4a31-8464-7afdd0b2cfad_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..b4d6eb61d88d52bea77eec50f4f71798ef53ca7e --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-20250514/cece1192-b954-4a31-8464-7afdd0b2cfad_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:875c1245b0929b57ea063faa39cde4c72c305665a75c402a59bcc6fc0f4857ee +size 10699826 diff --git a/data/swebenchpro/anthropic/claude-opus-4-20250514/d3caff69-cf5c-4b6d-9b8b-f9655286aa4b.json b/data/swebenchpro/anthropic/claude-opus-4-20250514/d3caff69-cf5c-4b6d-9b8b-f9655286aa4b.json new file mode 100644 index 0000000000000000000000000000000000000000..88bae4478cbe715c9447db0b631a8996fc1dc5f6 --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-20250514/d3caff69-cf5c-4b6d-9b8b-f9655286aa4b.json @@ -0,0 +1,116 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/anthropic_claude-opus-4-20250514/1786163192.771527", + "evaluation_timestamp": "1780423668.0", + "retrieved_timestamp": "1786163192.771527", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-20250514", + "id": "anthropic/claude-opus-4-20250514", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on swebenchpro/S-adaptive+C/+1ep/6538475d for scorer harbor_scorer", + "source_data": { + "dataset_name": "swebenchpro", + "source_type": "hf_dataset", + "samples_number": 3, + "sample_ids": [ + "instance_flipt-io__flipt-cd2f3b0a9d4d8b8a6d3d56afab65851ecdc408e8", + "instance_ansible__ansible-42355d181a11b51ebfc56f6f4b3d9c74e01cb13b-v1055803c3a812189a1133297f7f5468579283f86", + "instance_qutebrowser__qutebrowser-cf06f4e3708f886032d4d2a30108c2fddb042d81-v2ef375ac784985212b1805e1d0431dc8f1b3c171" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1780423668.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.3333333333333333, + "uncertainty": { + "standard_error": { + "value": 0.33333333333333337 + }, + "num_samples": 3 + }, + "details": { + "total_matched_trajectories": "3", + "stopping_reason_count_repetition_guard": "2", + "stopping_reason_count_completed_on_successful_submit": "1", + "avg_total_tokens_target_model": "3164006.67", + "total_total_tokens_target_model": "9492020.0", + "avg_total_tokens_other_models": "941.00", + "total_total_tokens_other_models": "2823.0", + "avg_total_tokens_all_models": "3164947.67", + "total_total_tokens_all_models": "9494843.0", + "avg_turn_count": "99.67", + "total_turn_count": "299.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "3", + "timeout": "1800", + "attempt_timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 30000000 + }, + "sandbox": {}, + "max_attempts": 3 + }, + "additional_details": { + "max_retries": "3", + "timeout": "1800", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/anthropic/claude-opus-4-20250514/d3caff69-cf5c-4b6d-9b8b-f9655286aa4b_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "4f30a361721c646d3b5b369562440a015700b4b925f6d66cee2f9d239fb9f137", + "total_rows": 3 + } +} \ No newline at end of file diff --git a/data/swebenchpro/anthropic/claude-opus-4-20250514/d3caff69-cf5c-4b6d-9b8b-f9655286aa4b_samples.jsonl b/data/swebenchpro/anthropic/claude-opus-4-20250514/d3caff69-cf5c-4b6d-9b8b-f9655286aa4b_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..78771976564f5eb40e3cb6a182f081545fb3a251 --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-20250514/d3caff69-cf5c-4b6d-9b8b-f9655286aa4b_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f30a361721c646d3b5b369562440a015700b4b925f6d66cee2f9d239fb9f137 +size 1113060 diff --git a/data/swebenchpro/anthropic/claude-opus-4-20250514/de0d0d73-cb41-4345-b7f2-96a855d1dd9d.json b/data/swebenchpro/anthropic/claude-opus-4-20250514/de0d0d73-cb41-4345-b7f2-96a855d1dd9d.json new file mode 100644 index 0000000000000000000000000000000000000000..9b108d5ae18096ff3e0d16b1cbe9df8da099e0ac --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-20250514/de0d0d73-cb41-4345-b7f2-96a855d1dd9d.json @@ -0,0 +1,111 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/anthropic_claude-opus-4-20250514/1786165703.191274", + "evaluation_timestamp": "1778168188.0", + "retrieved_timestamp": "1786165703.191274", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-20250514", + "id": "anthropic/claude-opus-4-20250514", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on swebenchpro/S-adaptive/+1ep/54266d45 for scorer harbor_scorer", + "source_data": { + "dataset_name": "swebenchpro", + "source_type": "hf_dataset", + "samples_number": 2, + "sample_ids": [ + "instance_flipt-io__flipt-b68b8960b8a08540d5198d78c665a7eb0bea4008", + "instance_navidrome__navidrome-d8e794317f788198227e10fb667e10496b3eb99a" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1778168188.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.0, + "uncertainty": { + "num_samples": 2 + }, + "details": { + "total_matched_trajectories": "2", + "stopping_reason_count_repetition_guard": "2", + "avg_total_tokens_target_model": "1801672.50", + "total_total_tokens_target_model": "3603345.0", + "avg_total_tokens_other_models": "1360.50", + "total_total_tokens_other_models": "2721.0", + "avg_total_tokens_all_models": "1803033.00", + "total_total_tokens_all_models": "3606066.0", + "avg_turn_count": "68.50", + "total_turn_count": "137.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 30000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/anthropic/claude-opus-4-20250514/de0d0d73-cb41-4345-b7f2-96a855d1dd9d_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "dd6ba401065bd2758c77715c58d1bf4bfa5fba94a8be2526f79760690b9cc707", + "total_rows": 2 + } +} \ No newline at end of file diff --git a/data/swebenchpro/anthropic/claude-opus-4-20250514/de0d0d73-cb41-4345-b7f2-96a855d1dd9d_samples.jsonl b/data/swebenchpro/anthropic/claude-opus-4-20250514/de0d0d73-cb41-4345-b7f2-96a855d1dd9d_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..f0ebd3f94de27715534071d59117516dbf19b601 --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-20250514/de0d0d73-cb41-4345-b7f2-96a855d1dd9d_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dd6ba401065bd2758c77715c58d1bf4bfa5fba94a8be2526f79760690b9cc707 +size 515551 diff --git a/data/swebenchpro/anthropic/claude-opus-4-20250514/eb858e41-6a4e-44a9-acf3-4509cc35f6b5.json b/data/swebenchpro/anthropic/claude-opus-4-20250514/eb858e41-6a4e-44a9-acf3-4509cc35f6b5.json new file mode 100644 index 0000000000000000000000000000000000000000..e958bf40e7148d0216c18bed2003e6e9754f5c83 --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-20250514/eb858e41-6a4e-44a9-acf3-4509cc35f6b5.json @@ -0,0 +1,149 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/anthropic_claude-opus-4-20250514/1786163233.356671", + "evaluation_timestamp": "1778706574.0", + "retrieved_timestamp": "1786163233.356671", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-20250514", + "id": "anthropic/claude-opus-4-20250514", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on swebenchpro/S-adaptive+C/+4ep/d5aa280f for scorer harbor_scorer", + "source_data": { + "dataset_name": "swebenchpro", + "source_type": "hf_dataset", + "samples_number": 36, + "sample_ids": [ + "instance_flipt-io__flipt-e42da21a07a5ae35835ec54f74004ebd58713874", + "instance_nodebb__nodebb-84e065752f6d7fbe5c08cbf50cb173ffb866b8fa-vf2cf3cbd463b7ad942381f1c6d077626485a1e9e", + "instance_element-hq__element-web-f3534b42df3dcfe36dc48bddbf14034085af6d30-vnan", + "instance_navidrome__navidrome-d8e794317f788198227e10fb667e10496b3eb99a", + "instance_future-architect__vuls-f0b3a8b1db98eb1bd32685f1c36c41a99c3452ed", + "instance_future-architect__vuls-6eff6a9329a65cc412e79b8f82444dfa3d0f0b5a", + "instance_internetarchive__openlibrary-fdbc0d8f418333c7e575c40b661b582c301ef7ac-v13642507b4fc1f8d234172bf8129942da2c2ca26", + "instance_ansible__ansible-b2a289dcbb702003377221e25f62c8a3608f0e89-v173091e2e36d38c978002990795f66cfc0af30ad", + "instance_tutao__tutanota-d1aa0ecec288bfc800cfb9133b087c4f81ad8b38-vbc0d9ba8f0071fbe982809910959a6ff8884dbbf", + "instance_qutebrowser__qutebrowser-e64622cd2df5b521342cf4a62e0d4cb8f8c9ae5a-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_gravitational__teleport-ac2fb2f9b4fd1896b554d3011df23d3d71295779", + "instance_protonmail__webclients-cb8cc309c6968b0a2a5fe4288d0ae0a969ff31e1", + "instance_ansible__ansible-11c1777d56664b1acb56b387a1ad6aeadef1391d-v0f01c69f1e2528b935359cfe578530722bca2c59", + "instance_ansible__ansible-29aea9ff3466e4cd2ed00524b9e56738d568ce8b-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_nodebb__nodebb-70b4a0e2aebebe8f2f559de6680093d96a697b2f-vnan", + "instance_qutebrowser__qutebrowser-e15d26630934d0b6415ed2295ac42fd570a57620-va0fd88aac89cde702ec1ba84877234da33adce8a", + "instance_ansible__ansible-164881d871964aa64e0f911d03ae270acbad253c-v390e508d27db7a51eece36bb6d9698b63a5b638a", + "instance_ansible__ansible-502270c804c33d3bc963930dc85e0f4ca359674d-v7eee2454f617569fd6889f2211f75bc02a35f9f8", + "instance_internetarchive__openlibrary-9c392b60e2c6fa1d68cb68084b4b4ff04d0cb35c-v2d9a6c849c60ed19fd0858ce9e40b7cc8e097e59", + "instance_element-hq__element-web-b007ea81b2ccd001b00f332bee65070aa7fc00f9-vnan", + "instance_internetarchive__openlibrary-7cbfb812ef0e1f9716e2d6e85d538a96fcb79d13-vfa6ff903cb27f336e17654595dd900fa943dcd91", + "instance_nodebb__nodebb-0f788b8eaa4bba3c142d171fd941d015c53b65fc-v0ec6d6c2baf3cb4797482ce4829bc25cd5716649", + "instance_flipt-io__flipt-0b119520afca1cf25c470ff4288c464d4510b944", + "instance_element-hq__element-web-f63160f38459fb552d00fcc60d4064977a9095a6-vnan", + "instance_ansible__ansible-f327e65d11bb905ed9f15996024f857a95592629-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_gravitational__teleport-73cc189b0e9636d418c4470ecce0d9af5dae2f02-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_internetarchive__openlibrary-3c48b4bb782189e0858e6c3fc7956046cf3e1cfb-v2d9a6c849c60ed19fd0858ce9e40b7cc8e097e59", + "instance_ansible__ansible-1ee70fc272aff6bf3415357c6e13c5de5b928d9b-v1055803c3a812189a1133297f7f5468579283f86", + "instance_qutebrowser__qutebrowser-f631cd4422744160d9dcf7a0455da532ce973315-v35616345bb8052ea303186706cec663146f0f184", + "instance_qutebrowser__qutebrowser-e5340c449f23608803c286da0563b62f58ba25b0-v059c6fdc75567943479b23ebca7c07b5e9a7f34c", + "instance_protonmail__webclients-2c3559cad02d1090985dba7e8eb5a129144d9811", + "instance_nodebb__nodebb-b1f9ad5534bb3a44dab5364f659876a4b7fe34c1-vnan", + "instance_flipt-io__flipt-292fdaca9be39e6a921aaa8874c011d0fdd3e874", + "instance_qutebrowser__qutebrowser-a25e8a09873838ca9efefd36ea8a45170bbeb95c-vc2f56a753b62a190ddb23cd330c257b9cf560d12", + "instance_element-hq__element-web-1077729a19c0ce902e713cf6fab42c91fb7907f1-vnan", + "instance_protonmail__webclients-f161c10cf7d31abf82e8d64d7a99c9fac5acfa18" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1778706574.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.6875, + "uncertainty": { + "standard_error": { + "value": 0.06998795247800504 + }, + "num_samples": 144 + }, + "details": { + "total_matched_trajectories": "144", + "stopping_reason_count_completed_on_successful_submit": "99", + "stopping_reason_count_repetition_guard": "45", + "avg_total_tokens_target_model": "1357528.49", + "total_total_tokens_target_model": "195484103.0", + "avg_total_tokens_other_models": "420.70", + "total_total_tokens_other_models": "60581.0", + "avg_total_tokens_all_models": "1357949.19", + "total_total_tokens_all_models": "195544684.0", + "avg_turn_count": "58.55", + "total_turn_count": "8431.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "8", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 30000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "8", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/anthropic/claude-opus-4-20250514/eb858e41-6a4e-44a9-acf3-4509cc35f6b5_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "9ae5b7a35906138f9548003514c19ccc7bfbb3e1a5c3c2733c2660fc0517bed1", + "total_rows": 144 + } +} \ No newline at end of file diff --git a/data/swebenchpro/anthropic/claude-opus-4-20250514/eb858e41-6a4e-44a9-acf3-4509cc35f6b5_samples.jsonl b/data/swebenchpro/anthropic/claude-opus-4-20250514/eb858e41-6a4e-44a9-acf3-4509cc35f6b5_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..e5e14ffc5d651d140638602e83281f942c9ee2c3 --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-20250514/eb858e41-6a4e-44a9-acf3-4509cc35f6b5_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9ae5b7a35906138f9548003514c19ccc7bfbb3e1a5c3c2733c2660fc0517bed1 +size 32485465 diff --git a/data/swebenchpro/anthropic/claude-opus-4-20250514/f1a4bf58-b4a0-4e67-8f31-542422d3044a.json b/data/swebenchpro/anthropic/claude-opus-4-20250514/f1a4bf58-b4a0-4e67-8f31-542422d3044a.json new file mode 100644 index 0000000000000000000000000000000000000000..dfa02c469d9c4d021e26c171a8c55a1d33ab26dc --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-20250514/f1a4bf58-b4a0-4e67-8f31-542422d3044a.json @@ -0,0 +1,170 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/anthropic_claude-opus-4-20250514/1786165720.572873", + "evaluation_timestamp": "1778513006.0", + "retrieved_timestamp": "1786165720.572873", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-20250514", + "id": "anthropic/claude-opus-4-20250514", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on swebenchpro/S-adaptive/+1ep/e2df6d1e for scorer harbor_scorer", + "source_data": { + "dataset_name": "swebenchpro", + "source_type": "hf_dataset", + "samples_number": 58, + "sample_ids": [ + "instance_nodebb__nodebb-70b4a0e2aebebe8f2f559de6680093d96a697b2f-vnan", + "instance_ansible__ansible-e22e103cdf8edc56ff7d9b848a58f94f1471a263-v1055803c3a812189a1133297f7f5468579283f86", + "instance_qutebrowser__qutebrowser-e15d26630934d0b6415ed2295ac42fd570a57620-va0fd88aac89cde702ec1ba84877234da33adce8a", + "instance_ansible__ansible-164881d871964aa64e0f911d03ae270acbad253c-v390e508d27db7a51eece36bb6d9698b63a5b638a", + "instance_internetarchive__openlibrary-3c48b4bb782189e0858e6c3fc7956046cf3e1cfb-v2d9a6c849c60ed19fd0858ce9e40b7cc8e097e59", + "instance_element-hq__element-web-9bf77963ee5e036d54b2a3ca202fbf6378464a5e-vnan", + "instance_qutebrowser__qutebrowser-70248f256f93ed9b1984494d0a1a919ddd774892-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_future-architect__vuls-73f0adad95c4d227e2ccfa876c85cc95dd065e13", + "instance_protonmail__webclients-da91f084c0f532d9cc8ca385a701274d598057b8", + "instance_flipt-io__flipt-b2cd6a6dd73ca91b519015fd5924fde8d17f3f06", + "instance_qutebrowser__qutebrowser-cc360cd4a34a126274c7b51f3b63afbaf3e05a02-v5fc38aaf22415ab0b70567368332beee7955b367", + "instance_flipt-io__flipt-e2bd19dafa7166c96b082fb2a59eb54b4be0d778", + "instance_qutebrowser__qutebrowser-9ed748effa8f3bcd804612d9291da017b514e12f-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_ansible__ansible-1ee70fc272aff6bf3415357c6e13c5de5b928d9b-v1055803c3a812189a1133297f7f5468579283f86", + "instance_element-hq__element-web-75c2c1a572fa45d1ea1d1a96e9e36e303332ecaa-vnan", + "instance_qutebrowser__qutebrowser-f631cd4422744160d9dcf7a0455da532ce973315-v35616345bb8052ea303186706cec663146f0f184", + "instance_ansible__ansible-3db08adbb1cc6aa9941be5e0fc810132c6e1fa4b-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_qutebrowser__qutebrowser-e5340c449f23608803c286da0563b62f58ba25b0-v059c6fdc75567943479b23ebca7c07b5e9a7f34c", + "instance_internetarchive__openlibrary-09865f5fb549694d969f0a8e49b9d204ef1853ca-ve8c8d62a2b60610a3c4631f5f23ed866bada9818", + "instance_internetarchive__openlibrary-c506c1b0b678892af5cb22c1c1dbc35d96787a0a-v0f5aece3601a5b4419f7ccec1dbda2071be28ee4", + "instance_element-hq__element-web-2760bfc8369f1bee640d6d7a7e910783143d4c5f-vnan", + "instance_internetarchive__openlibrary-25858f9f0c165df25742acf8309ce909773f0cdd-v13642507b4fc1f8d234172bf8129942da2c2ca26", + "instance_protonmail__webclients-01ea5214d11e0df8b7170d91bafd34f23cb0f2b1", + "instance_ansible__ansible-a02e22e902a69aeb465f16bf03f7f5a91b2cb828-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_protonmail__webclients-5f0745dd6993bb1430a951c62a49807c6635cd77", + "instance_tutao__tutanota-1ff82aa365763cee2d609c9d19360ad87fdf2ec7-vc4e41fd0029957297843cb9dec4a25c7c756f029", + "instance_internetarchive__openlibrary-30bc73a1395fba2300087c7f307e54bb5372b60a-v76304ecdb3a5954fcf13feb710e8c40fcf24b73c", + "instance_nodebb__nodebb-8ca65b0c78c67c1653487c02d1135e1b702185e1-vf2cf3cbd463b7ad942381f1c6d077626485a1e9e", + "instance_protonmail__webclients-01b519cd49e6a24d9a05d2eb97f54e420740072e", + "instance_ansible__ansible-77658704217d5f166404fc67997203c25381cb6e-v390e508d27db7a51eece36bb6d9698b63a5b638a", + "instance_qutebrowser__qutebrowser-2dd8966fdcf11972062c540b7a787e4d0de8d372-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_ansible__ansible-cd9c4eb5a6b2bfaf4a6709f001ce3d0c92c1eed2-v0f01c69f1e2528b935359cfe578530722bca2c59", + "instance_protonmail__webclients-ac23d1efa1a6ab7e62724779317ba44c28d78cfd", + "instance_future-architect__vuls-030b2e03525d68d74cb749959aac2d7f3fc0effa", + "instance_nodebb__nodebb-b1f9ad5534bb3a44dab5364f659876a4b7fe34c1-vnan", + "instance_flipt-io__flipt-292fdaca9be39e6a921aaa8874c011d0fdd3e874", + "instance_qutebrowser__qutebrowser-ef5ba1a0360b39f9eff027fbdc57f363597c3c3b-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_ansible__ansible-622a493ae03bd5e5cf517d336fc426e9d12208c7-v906c969b551b346ef54a2c0b41e04f632b7b73c2", + "instance_gravitational__teleport-82185f232ae8974258397e121b3bc2ed0c3729ed-v626ec2a48416b10a88641359a169d99e935ff037", + "instance_ansible__ansible-9a21e247786ebd294dafafca1105fcd770ff46c6-v67cdaa49f89b34e42b69d5b7830b3c3ad3d8803f", + "instance_protonmail__webclients-8afd9ce04c8dde9e150e1c2b50d32e7ee2efa3e7", + "instance_flipt-io__flipt-b22f5f02e40b225b6b93fff472914973422e97c6", + "instance_gravitational__teleport-2bb3bbbd8aff1164a2353381cb79e1dc93b90d28-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_gravitational__teleport-46aa81b1ce96ebb4ebed2ae53fd78cd44a05da6c-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_tutao__tutanota-8513a9e8114a8b42e64f4348335e0f23efa054c4-vee878bb72091875e912c52fc32bc60ec3760227b", + "instance_internetarchive__openlibrary-630221ab686c64e75a2ce253c893c033e4814b2e-v93c53c13d5f9b383ebb411ee7750b49dcd1a34c6", + "instance_internetarchive__openlibrary-7f6b722a10f822171501d027cad60afe53337732-ve8c8d62a2b60610a3c4631f5f23ed866bada9818", + "instance_nodebb__nodebb-a917210c5b2c20637094545401f85783905c074c-vf2cf3cbd463b7ad942381f1c6d077626485a1e9e", + "instance_internetarchive__openlibrary-1351c59fd43689753de1fca32c78d539a116ffc1-v29f82c9cf21d57b242f8d8b0e541525d259e2d63", + "instance_qutebrowser__qutebrowser-a25e8a09873838ca9efefd36ea8a45170bbeb95c-vc2f56a753b62a190ddb23cd330c257b9cf560d12", + "instance_protonmail__webclients-281a6b3f190f323ec2c0630999354fafb84b2880", + "instance_future-architect__vuls-999529a05b202b0fd29c6fca5039a4c47a3766bb", + "instance_element-hq__element-web-1077729a19c0ce902e713cf6fab42c91fb7907f1-vnan", + "instance_future-architect__vuls-be7b9114cc9545e68fb0ee7bc63d7ec53d1a00ad", + "instance_qutebrowser__qutebrowser-c580ebf0801e5a3ecabc54f327498bb753c6d5f2-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_gravitational__teleport-3ff19cf7c41f396ae468797d3aeb61515517edc9-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_future-architect__vuls-aaea15e516ece43978cf98e09e52080478b1d39f", + "instance_protonmail__webclients-f161c10cf7d31abf82e8d64d7a99c9fac5acfa18" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1778513006.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.48214285714285715, + "uncertainty": { + "standard_error": { + "value": 0.06737697508644648 + }, + "num_samples": 56 + }, + "details": { + "total_matched_trajectories": "56", + "stopping_reason_count_repetition_guard": "56", + "avg_total_tokens_target_model": "1409848.77", + "total_total_tokens_target_model": "78951531.0", + "avg_total_tokens_other_models": "1358.61", + "total_total_tokens_other_models": "76082.0", + "avg_total_tokens_all_models": "1411207.38", + "total_total_tokens_all_models": "79027613.0", + "avg_turn_count": "60.21", + "total_turn_count": "3372.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 30000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/anthropic/claude-opus-4-20250514/f1a4bf58-b4a0-4e67-8f31-542422d3044a_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "bfeac015bd978a9e89da1b41f4cf2f6a4118e13c4882646dcdb0570eca36aa17", + "total_rows": 56 + } +} \ No newline at end of file diff --git a/data/swebenchpro/anthropic/claude-opus-4-20250514/f1a4bf58-b4a0-4e67-8f31-542422d3044a_samples.jsonl b/data/swebenchpro/anthropic/claude-opus-4-20250514/f1a4bf58-b4a0-4e67-8f31-542422d3044a_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..10aa687ae72a15c170bd417ace8f208b9d2e12ab --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-20250514/f1a4bf58-b4a0-4e67-8f31-542422d3044a_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bfeac015bd978a9e89da1b41f4cf2f6a4118e13c4882646dcdb0570eca36aa17 +size 13619367 diff --git a/data/swebenchpro/anthropic/claude-opus-4-20250514/f28e8d25-f0d4-4b16-97fb-34d9c462e302.json b/data/swebenchpro/anthropic/claude-opus-4-20250514/f28e8d25-f0d4-4b16-97fb-34d9c462e302.json new file mode 100644 index 0000000000000000000000000000000000000000..ccbd3b9eb0bddc653f5eb02fddb865c72cafc302 --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-20250514/f28e8d25-f0d4-4b16-97fb-34d9c462e302.json @@ -0,0 +1,111 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/anthropic_claude-opus-4-20250514/1786165706.127064", + "evaluation_timestamp": "1778590202.0", + "retrieved_timestamp": "1786165706.127064", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-20250514", + "id": "anthropic/claude-opus-4-20250514", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on swebenchpro/S-adaptive/+1ep/d8b5f4d5 for scorer harbor_scorer", + "source_data": { + "dataset_name": "swebenchpro", + "source_type": "hf_dataset", + "samples_number": 2, + "sample_ids": [ + "instance_qutebrowser__qutebrowser-9ed748effa8f3bcd804612d9291da017b514e12f-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_ansible__ansible-1ee70fc272aff6bf3415357c6e13c5de5b928d9b-v1055803c3a812189a1133297f7f5468579283f86" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1778590202.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 1.0, + "uncertainty": { + "num_samples": 2 + }, + "details": { + "total_matched_trajectories": "2", + "stopping_reason_count_repetition_guard": "2", + "avg_total_tokens_target_model": "952435.00", + "total_total_tokens_target_model": "1904870.0", + "avg_total_tokens_other_models": "1037.00", + "total_total_tokens_other_models": "2074.0", + "avg_total_tokens_all_models": "953472.00", + "total_total_tokens_all_models": "1906944.0", + "avg_turn_count": "51.50", + "total_turn_count": "103.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 30000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/anthropic/claude-opus-4-20250514/f28e8d25-f0d4-4b16-97fb-34d9c462e302_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "4891372dc3b52dd8d6b7bbfa8d6a8107f378c523e10bcce69a27f0a9eb7fe728", + "total_rows": 2 + } +} \ No newline at end of file diff --git a/data/swebenchpro/anthropic/claude-opus-4-20250514/f28e8d25-f0d4-4b16-97fb-34d9c462e302_samples.jsonl b/data/swebenchpro/anthropic/claude-opus-4-20250514/f28e8d25-f0d4-4b16-97fb-34d9c462e302_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..655d3ec491d0c7378e45fd96993c1896e36495eb --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-20250514/f28e8d25-f0d4-4b16-97fb-34d9c462e302_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4891372dc3b52dd8d6b7bbfa8d6a8107f378c523e10bcce69a27f0a9eb7fe728 +size 430209 diff --git a/data/swebenchpro/anthropic/claude-opus-4-5-20251101/00e58287-4b09-4a43-a990-fd7e67721c4f.json b/data/swebenchpro/anthropic/claude-opus-4-5-20251101/00e58287-4b09-4a43-a990-fd7e67721c4f.json new file mode 100644 index 0000000000000000000000000000000000000000..6c9892a4ece0a7d2346d841080668f7bf13ee6ee --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-5-20251101/00e58287-4b09-4a43-a990-fd7e67721c4f.json @@ -0,0 +1,105 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/anthropic_claude-opus-4-5-20251101/1786163090.260744", + "evaluation_timestamp": "1776444232.0", + "retrieved_timestamp": "1786163090.260744", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-5-20251101", + "id": "anthropic/claude-opus-4-5-20251101", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on swebenchpro/S-adaptive+C/+4ep/dae38cd0 for scorer harbor_scorer", + "source_data": { + "dataset_name": "swebenchpro", + "source_type": "hf_dataset", + "samples_number": 2, + "sample_ids": [ + "instance_gravitational__teleport-82185f232ae8974258397e121b3bc2ed0c3729ed-v626ec2a48416b10a88641359a169d99e935ff037", + "instance_future-architect__vuls-50580f6e98eeb36f53f27222f7f4fdfea0b21e8d" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1776444232.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.0, + "uncertainty": { + "num_samples": 8 + }, + "details": { + "total_matched_trajectories": "8", + "stopping_reason_count_repetition_guard": "8", + "avg_total_tokens_target_model": "5573506.00", + "total_total_tokens_target_model": "44588048.0", + "avg_total_tokens_other_models": "0.00", + "total_total_tokens_other_models": "0.0", + "avg_total_tokens_all_models": "5573506.00", + "total_total_tokens_all_models": "44588048.0", + "avg_turn_count": "93.25", + "total_turn_count": "746.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "max_connections": "10", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 30000000 + }, + "sandbox": {} + }, + "additional_details": { + "max_connections": "10" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/anthropic/claude-opus-4-5-20251101/00e58287-4b09-4a43-a990-fd7e67721c4f_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "262093036934b2a91501ba997c9861db0bf7009394f0a186e739d3539ee30e31", + "total_rows": 8 + } +} \ No newline at end of file diff --git a/data/swebenchpro/anthropic/claude-opus-4-5-20251101/00e58287-4b09-4a43-a990-fd7e67721c4f_samples.jsonl b/data/swebenchpro/anthropic/claude-opus-4-5-20251101/00e58287-4b09-4a43-a990-fd7e67721c4f_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..8af3bd281f86dcc0556339bbf1307df4e225c356 --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-5-20251101/00e58287-4b09-4a43-a990-fd7e67721c4f_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:262093036934b2a91501ba997c9861db0bf7009394f0a186e739d3539ee30e31 +size 13680020 diff --git a/data/swebenchpro/anthropic/claude-opus-4-5-20251101/0b3db2ac-2d3b-4ab6-bbb2-ed971646512c.json b/data/swebenchpro/anthropic/claude-opus-4-5-20251101/0b3db2ac-2d3b-4ab6-bbb2-ed971646512c.json new file mode 100644 index 0000000000000000000000000000000000000000..4c7e572f97dd1488c0e009d96e2c9ee3a79ea056 --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-5-20251101/0b3db2ac-2d3b-4ab6-bbb2-ed971646512c.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/anthropic_claude-opus-4-5-20251101/1786167415.341246", + "evaluation_timestamp": "1776417565.0", + "retrieved_timestamp": "1786167415.341246", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-5-20251101", + "id": "anthropic/claude-opus-4-5-20251101", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/anthropic/claude-opus-4-5-20251101/0b3db2ac-2d3b-4ab6-bbb2-ed971646512c_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "e4d8c9b0618c00a95b72915fd8c5e793ca81f6f51cda704252e5da0587fbd348", + "total_rows": 7 + } +} \ No newline at end of file diff --git a/data/swebenchpro/anthropic/claude-opus-4-5-20251101/0b3db2ac-2d3b-4ab6-bbb2-ed971646512c_samples.jsonl b/data/swebenchpro/anthropic/claude-opus-4-5-20251101/0b3db2ac-2d3b-4ab6-bbb2-ed971646512c_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..c5204b4d29976a0891dffe72b1036ea9430556b3 --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-5-20251101/0b3db2ac-2d3b-4ab6-bbb2-ed971646512c_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e4d8c9b0618c00a95b72915fd8c5e793ca81f6f51cda704252e5da0587fbd348 +size 2102022 diff --git a/data/swebenchpro/anthropic/claude-opus-4-5-20251101/0deefadd-a71b-4438-9d53-0f73d670c0e0.json b/data/swebenchpro/anthropic/claude-opus-4-5-20251101/0deefadd-a71b-4438-9d53-0f73d670c0e0.json new file mode 100644 index 0000000000000000000000000000000000000000..9d541d3d16ce54013272feba20128b100eb2412a --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-5-20251101/0deefadd-a71b-4438-9d53-0f73d670c0e0.json @@ -0,0 +1,142 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/anthropic_claude-opus-4-5-20251101/1786165511.922792", + "evaluation_timestamp": "1777656314.0", + "retrieved_timestamp": "1786165511.922792", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-5-20251101", + "id": "anthropic/claude-opus-4-5-20251101", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on swebenchpro/S-adaptive/+1ep/01059fca for scorer harbor_scorer", + "source_data": { + "dataset_name": "swebenchpro", + "source_type": "hf_dataset", + "samples_number": 32, + "sample_ids": [ + "instance_protonmail__webclients-01ea5214d11e0df8b7170d91bafd34f23cb0f2b1", + "instance_flipt-io__flipt-05d7234fa582df632f70a7cd10194d61bd7043b9", + "instance_ansible__ansible-a02e22e902a69aeb465f16bf03f7f5a91b2cb828-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_protonmail__webclients-5f0745dd6993bb1430a951c62a49807c6635cd77", + "instance_internetarchive__openlibrary-30bc73a1395fba2300087c7f307e54bb5372b60a-v76304ecdb3a5954fcf13feb710e8c40fcf24b73c", + "instance_nodebb__nodebb-8ca65b0c78c67c1653487c02d1135e1b702185e1-vf2cf3cbd463b7ad942381f1c6d077626485a1e9e", + "instance_nodebb__nodebb-22368b996ee0e5f11a5189b400b33af3cc8d925a-v4fbcfae8b15e4ce5d132c408bca69ebb9cf146ed", + "instance_protonmail__webclients-2c3559cad02d1090985dba7e8eb5a129144d9811", + "instance_ansible__ansible-77658704217d5f166404fc67997203c25381cb6e-v390e508d27db7a51eece36bb6d9698b63a5b638a", + "instance_qutebrowser__qutebrowser-2dd8966fdcf11972062c540b7a787e4d0de8d372-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_ansible__ansible-cd9c4eb5a6b2bfaf4a6709f001ce3d0c92c1eed2-v0f01c69f1e2528b935359cfe578530722bca2c59", + "instance_protonmail__webclients-ac23d1efa1a6ab7e62724779317ba44c28d78cfd", + "instance_future-architect__vuls-030b2e03525d68d74cb749959aac2d7f3fc0effa", + "instance_nodebb__nodebb-b1f9ad5534bb3a44dab5364f659876a4b7fe34c1-vnan", + "instance_flipt-io__flipt-292fdaca9be39e6a921aaa8874c011d0fdd3e874", + "instance_qutebrowser__qutebrowser-ef5ba1a0360b39f9eff027fbdc57f363597c3c3b-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_ansible__ansible-622a493ae03bd5e5cf517d336fc426e9d12208c7-v906c969b551b346ef54a2c0b41e04f632b7b73c2", + "instance_gravitational__teleport-82185f232ae8974258397e121b3bc2ed0c3729ed-v626ec2a48416b10a88641359a169d99e935ff037", + "instance_ansible__ansible-9a21e247786ebd294dafafca1105fcd770ff46c6-v67cdaa49f89b34e42b69d5b7830b3c3ad3d8803f", + "instance_protonmail__webclients-8afd9ce04c8dde9e150e1c2b50d32e7ee2efa3e7", + "instance_ansible__ansible-de5858f48dc9e1ce9117034e0d7e76806f420ca8-v1055803c3a812189a1133297f7f5468579283f86", + "instance_gravitational__teleport-46aa81b1ce96ebb4ebed2ae53fd78cd44a05da6c-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_internetarchive__openlibrary-630221ab686c64e75a2ce253c893c033e4814b2e-v93c53c13d5f9b383ebb411ee7750b49dcd1a34c6", + "instance_internetarchive__openlibrary-7f6b722a10f822171501d027cad60afe53337732-ve8c8d62a2b60610a3c4631f5f23ed866bada9818", + "instance_nodebb__nodebb-a917210c5b2c20637094545401f85783905c074c-vf2cf3cbd463b7ad942381f1c6d077626485a1e9e", + "instance_internetarchive__openlibrary-1351c59fd43689753de1fca32c78d539a116ffc1-v29f82c9cf21d57b242f8d8b0e541525d259e2d63", + "instance_qutebrowser__qutebrowser-a25e8a09873838ca9efefd36ea8a45170bbeb95c-vc2f56a753b62a190ddb23cd330c257b9cf560d12", + "instance_future-architect__vuls-999529a05b202b0fd29c6fca5039a4c47a3766bb", + "instance_element-hq__element-web-1077729a19c0ce902e713cf6fab42c91fb7907f1-vnan", + "instance_future-architect__vuls-be7b9114cc9545e68fb0ee7bc63d7ec53d1a00ad", + "instance_qutebrowser__qutebrowser-c580ebf0801e5a3ecabc54f327498bb753c6d5f2-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_gravitational__teleport-3ff19cf7c41f396ae468797d3aeb61515517edc9-vee9b09fb20c43af7e520f57e9239bbcf46b7113d" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1777656314.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.65625, + "uncertainty": { + "standard_error": { + "value": 0.08530513305661301 + }, + "num_samples": 32 + }, + "details": { + "total_matched_trajectories": "32", + "stopping_reason_count_token_limit": "13", + "stopping_reason_count_repetition_guard": "19", + "avg_total_tokens_target_model": "19551450.41", + "total_total_tokens_target_model": "625646413.0", + "avg_total_tokens_other_models": "390.88", + "total_total_tokens_other_models": "12508.0", + "avg_total_tokens_all_models": "19551841.28", + "total_total_tokens_all_models": "625658921.0", + "avg_turn_count": "277.94", + "total_turn_count": "8894.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 30000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/anthropic/claude-opus-4-5-20251101/0deefadd-a71b-4438-9d53-0f73d670c0e0_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "5933a0815bccbdfdd42275d76e5dec40bd4cc81332f2302073dbdb2fc78e3387", + "total_rows": 32 + } +} \ No newline at end of file diff --git a/data/swebenchpro/anthropic/claude-opus-4-5-20251101/0deefadd-a71b-4438-9d53-0f73d670c0e0_samples.jsonl b/data/swebenchpro/anthropic/claude-opus-4-5-20251101/0deefadd-a71b-4438-9d53-0f73d670c0e0_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..0672cdc2d8c01a6c668fc25fc9a76167e7bf65a4 --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-5-20251101/0deefadd-a71b-4438-9d53-0f73d670c0e0_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5933a0815bccbdfdd42275d76e5dec40bd4cc81332f2302073dbdb2fc78e3387 +size 36167268 diff --git a/data/swebenchpro/anthropic/claude-opus-4-5-20251101/0f30a8a8-89ae-4fa4-867a-2856dca81ab0.json b/data/swebenchpro/anthropic/claude-opus-4-5-20251101/0f30a8a8-89ae-4fa4-867a-2856dca81ab0.json new file mode 100644 index 0000000000000000000000000000000000000000..996007b84daceb23053a117a506845edf233fd59 --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-5-20251101/0f30a8a8-89ae-4fa4-867a-2856dca81ab0.json @@ -0,0 +1,304 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/anthropic_claude-opus-4-5-20251101/1786182352.622346", + "evaluation_timestamp": "1774867708.0", + "retrieved_timestamp": "1786182352.622346", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.196.dev25+gb973b316,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-5-20251101", + "id": "anthropic/claude-opus-4-5-20251101", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on eval_task for scorer harbor_scorer", + "source_data": { + "dataset_name": "swebenchpro", + "source_type": "hf_dataset", + "samples_number": 731, + "sample_ids": [ + "instance_flipt-io__flipt-e91615cf07966da41756017a7d571f9fc0fdbe80", + "instance_ansible__ansible-8127abbc298cabf04aaa89a478fc5e5e3432a6fc-v30a923fb5c164d6cd18280c02422f75e611e8fb2", + "instance_protonmail__webclients-da91f084c0f532d9cc8ca385a701274d598057b8", + "instance_qutebrowser__qutebrowser-deeb15d6f009b3ca0c3bd503a7cef07462bd16b4-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_internetarchive__openlibrary-5069b09e5f64428dce59b33455c8bb17fe577070-v8717e18970bcdc4e0d2cea3b1527752b21e74866", + "instance_protonmail__webclients-c5a2089ca2bfe9aa1d85a664b8ad87ef843a1c9c", + "instance_qutebrowser__qutebrowser-fd6790fe8c02b144ab2464f1fc8ab3d02ce3c476-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_navidrome__navidrome-eebfbc5381a1e506ff17b5f1371d1ad83d5fd642", + "instance_navidrome__navidrome-812dc2090f20ac4f8ac271b6ed95be5889d1a3ca", + "instance_nodebb__nodebb-0c81642997ea1d827dbd02c311db9d4976112cd4-vf2cf3cbd463b7ad942381f1c6d077626485a1e9e", + "instance_internetarchive__openlibrary-c506c1b0b678892af5cb22c1c1dbc35d96787a0a-v0f5aece3601a5b4419f7ccec1dbda2071be28ee4", + "instance_flipt-io__flipt-5c7037ececb0bead0a8eb56054e224bcd7ac5922", + "instance_qutebrowser__qutebrowser-01d1d1494411380d97cac14614a829d3a69cecaf-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_internetarchive__openlibrary-7f7e53aa4cf74a4f8549a5bcd4810c527e2f6d7e-v13642507b4fc1f8d234172bf8129942da2c2ca26", + "instance_internetarchive__openlibrary-53e02a22972e9253aeded0e1981e6845e1e521fe-vfa6ff903cb27f336e17654595dd900fa943dcd91", + "instance_ansible__ansible-34db57a47f875d11c4068567b9ec7ace174ec4cf-v1055803c3a812189a1133297f7f5468579283f86", + "instance_qutebrowser__qutebrowser-996487c43e4fcc265b541f9eca1e7930e3c5cf05-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_gravitational__teleport-96019ce0be7a2c8e36363f359eb7c943b41dde70", + "instance_protonmail__webclients-df60460f163fd5c34e844ab9015e3176f1ab1ac0", + "instance_future-architect__vuls-ad2edbb8448e2c41a097f1c0b52696c0f6c5924d", + "instance_ansible__ansible-9142be2f6cabbe6597c9254c5bb9186d17036d55-v0f01c69f1e2528b935359cfe578530722bca2c59", + "instance_element-hq__element-web-33299af5c9b7a7ec5a9c31d578d4ec5b18088fb7-vnan", + "instance_element-hq__element-web-772df3021201d9c73835a626df8dcb6334ad9a3e-vnan", + "instance_internetarchive__openlibrary-1be7de788a444f6255e89c10ef6aa608550604a8-v29f82c9cf21d57b242f8d8b0e541525d259e2d63", + "instance_internetarchive__openlibrary-630221ab686c64e75a2ce253c893c033e4814b2e-v93c53c13d5f9b383ebb411ee7750b49dcd1a34c6", + "instance_gravitational__teleport-f432a71a13e698b6e1c4672a2e9e9c1f32d35c12", + "instance_protonmail__webclients-0d0267c4438cf378bda90bc85eed3a3615871ac4", + "instance_internetarchive__openlibrary-dbbd9d539c6d4fd45d5be9662aa19b6d664b5137-v08d8e8889ec945ab821fb156c04c7d2e2810debb", + "instance_flipt-io__flipt-b2170346dc37cf42fda1386cd630f24821ad2ac5", + "instance_element-hq__element-web-aec454dd6feeb93000380523cbb0b3681c0275fd-vnan", + "instance_internetarchive__openlibrary-4a5d2a7d24c9e4c11d3069220c0685b736d5ecde-v13642507b4fc1f8d234172bf8129942da2c2ca26", + "instance_qutebrowser__qutebrowser-ec2dcfce9eee9f808efc17a1b99e227fc4421dea-v5149fcda2a9a6fe1d35dfed1bade1444a11ef271", + "instance_gravitational__teleport-1330415d33a27594c948a36d9d7701f496229e9f", + "instance_protonmail__webclients-3a6790f480309130b5d6332dce6c9d5ccca13ee3", + "instance_flipt-io__flipt-492cc0b158200089dceede3b1aba0ed28df3fb1d", + "instance_future-architect__vuls-edb324c3d9ec3b107bf947f00e38af99d05b3e16", + "instance_ansible__ansible-bec27fb4c0a40c5f8bbcf26a475704227d65ee73-v30a923fb5c164d6cd18280c02422f75e611e8fb2", + "instance_gravitational__teleport-c1b1c6a1541c478d7777a48fca993cc8206c73b9", + "instance_tutao__tutanota-1e516e989b3c0221f4af6b297d9c0e4c43e4adc3-vbc0d9ba8f0071fbe982809910959a6ff8884dbbf", + "instance_flipt-io__flipt-406f9396ad65696d58865b3a6283109cd4eaf40e", + "instance_future-architect__vuls-139f3a81b66c47e6d8f70ce6c4afe7a9196a6ea8", + "instance_flipt-io__flipt-3b2c25ee8a3ac247c3fad13ad8d64ace34ec8ee7", + "instance_gravitational__teleport-0ac7334939981cf85b9591ac295c3816954e287e", + "instance_qutebrowser__qutebrowser-e15d26630934d0b6415ed2295ac42fd570a57620-va0fd88aac89cde702ec1ba84877234da33adce8a", + "instance_ansible__ansible-5e88cd9972f10b66dd97e1ee684c910c6a2dd25e-v906c969b551b346ef54a2c0b41e04f632b7b73c2", + "instance_element-hq__element-web-923ad4323b2006b2b180544429455ffe7d4a6cc3-vnan", + "instance_protonmail__webclients-4feccbc9990980aee26ea29035f8f931d6089895", + "instance_flipt-io__flipt-a0cbc0cb65ae601270bdbe3f5313e2dfd49c80e4", + "instance_gravitational__teleport-b4e7cd3a5e246736d3fe8d6886af55030b232277", + "instance_ansible__ansible-6cc97447aac5816745278f3735af128afb255c81-v0f01c69f1e2528b935359cfe578530722bca2c59", + "instance_ansible__ansible-f327e65d11bb905ed9f15996024f857a95592629-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_qutebrowser__qutebrowser-5fdc83e5da6222fe61163395baaad7ae57fa2cb4-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_navidrome__navidrome-3982ba725883e71d4e3e618c61d5140eeb8d850a", + "instance_qutebrowser__qutebrowser-ebfe9b7aa0c4ba9d451f993e08955004aaec4345-v059c6fdc75567943479b23ebca7c07b5e9a7f34c", + "instance_ansible__ansible-709484969c8a4ffd74b839a673431a8c5caa6457-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_gravitational__teleport-ad41b3c15414b28a6cec8c25424a19bfa7abd0e9-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_protonmail__webclients-2c3559cad02d1090985dba7e8eb5a129144d9811", + "instance_flipt-io__flipt-5af0757e96dec4962a076376d1bedc79de0d4249", + "instance_element-hq__element-web-56c7fc1948923b4b3f3507799e725ac16bcf8018-vnan", + "instance_flipt-io__flipt-381b90f718435c4694380b5fcd0d5cf8e3b5a25a", + "instance_flipt-io__flipt-6fd0f9e2587f14ac1fdd1c229f0bcae0468c8daa", + "instance_ansible__ansible-4c5ce5a1a9e79a845aff4978cfeb72a0d4ecf7d6-v1055803c3a812189a1133297f7f5468579283f86", + "instance_navidrome__navidrome-7073d18b54da7e53274d11c9e2baef1242e8769e", + "instance_internetarchive__openlibrary-0a90f9f0256e4f933523e9842799e39f95ae29ce-v76304ecdb3a5954fcf13feb710e8c40fcf24b73c", + "instance_nodebb__nodebb-be43cd25974681c9743d424238b7536c357dc8d3-vf2cf3cbd463b7ad942381f1c6d077626485a1e9e", + "instance_flipt-io__flipt-05d7234fa582df632f70a7cd10194d61bd7043b9", + "instance_nodebb__nodebb-76c6e30282906ac664f2c9278fc90999b27b1f48-vd59a5728dfc977f44533186ace531248c2917516", + "instance_qutebrowser__qutebrowser-6b320dc18662580e1313d2548fdd6231d2a97e6d-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_future-architect__vuls-01441351c3407abfc21c48a38e28828e1b504e0c", + "instance_ansible__ansible-1a4644ff15355fd696ac5b9d074a566a80fe7ca3-v30a923fb5c164d6cd18280c02422f75e611e8fb2", + "instance_navidrome__navidrome-09ae41a2da66264c60ef307882362d2e2d8d8b89", + "instance_internetarchive__openlibrary-5c6c22f3d2edf2f1b10f5dc335e32cb6a5f40341-v76304ecdb3a5954fcf13feb710e8c40fcf24b73c", + "instance_navidrome__navidrome-bf2bcb12799b21069f137749e0c331f761d1f693", + "instance_flipt-io__flipt-2ce8a0331e8a8f63f2c1b555db8277ffe5aa2e63", + "instance_ansible__ansible-a1569ea4ca6af5480cf0b7b3135f5e12add28a44-v0f01c69f1e2528b935359cfe578530722bca2c59", + "instance_gravitational__teleport-82185f232ae8974258397e121b3bc2ed0c3729ed-v626ec2a48416b10a88641359a169d99e935ff037", + "instance_flipt-io__flipt-72d06db14d58692bfb4d07b1aa745a37b35956f3", + "instance_ansible__ansible-b6290e1d156af608bd79118d209a64a051c55001-v390e508d27db7a51eece36bb6d9698b63a5b638a", + "instance_gravitational__teleport-32bcd71591c234f0d8b091ec01f1f5cbfdc0f13c-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_ansible__ansible-a02e22e902a69aeb465f16bf03f7f5a91b2cb828-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_gravitational__teleport-eda668c30d9d3b56d9c69197b120b01013611186", + "instance_nodebb__nodebb-0f788b8eaa4bba3c142d171fd941d015c53b65fc-v0ec6d6c2baf3cb4797482ce4829bc25cd5716649", + "instance_gravitational__teleport-47530e1fd8bfb84ec096ebcbbc29990f30829655-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_qutebrowser__qutebrowser-e64622cd2df5b521342cf4a62e0d4cb8f8c9ae5a-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_element-hq__element-web-494d9de6f0a94ffb491e74744d2735bce02dc0ab-vnan", + "instance_element-hq__element-web-9bf77963ee5e036d54b2a3ca202fbf6378464a5e-vnan", + "instance_nodebb__nodebb-9c576a0758690f45a6ca03b5884c601e473bf2c1-vd59a5728dfc977f44533186ace531248c2917516", + "instance_internetarchive__openlibrary-25858f9f0c165df25742acf8309ce909773f0cdd-v13642507b4fc1f8d234172bf8129942da2c2ca26", + "instance_protonmail__webclients-f161c10cf7d31abf82e8d64d7a99c9fac5acfa18", + "instance_gravitational__teleport-cb712e3f0b06dadc679f895daef8072cae400c26-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_gravitational__teleport-46a13210519461c7cec8d643bfbe750265775b41", + "instance_ansible__ansible-e40889e7112ae00a21a2c74312b330e67a766cc0-v1055803c3a812189a1133297f7f5468579283f86", + "instance_gravitational__teleport-6eaaf3a27e64f4ef4ef855bd35d7ec338cf17460-v626ec2a48416b10a88641359a169d99e935ff037", + "instance_internetarchive__openlibrary-0d13e6b4bf80bced6c0946b969b9a1b6963f6bce-v76304ecdb3a5954fcf13feb710e8c40fcf24b73c", + "instance_protonmail__webclients-2f2f6c311c6128fe86976950d3c0c2db07b03921", + "instance_ansible__ansible-5260527c4a71bfed99d803e687dd19619423b134-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_qutebrowser__qutebrowser-8d05f0282a271bfd45e614238bd1b555c58b3fc1-v35616345bb8052ea303186706cec663146f0f184", + "instance_qutebrowser__qutebrowser-50efac08f623644a85441bbe02ab9347d2b71a9d-v9f8e9d96c85c85a605e382f1510bd08563afc566", + "instance_flipt-io__flipt-dae029cba7cdb98dfb1a6b416c00d324241e6063", + "instance_nodebb__nodebb-18c45b44613aecd53e9f60457b9812049ab2998d-v0495b863a912fbff5749c67e860612b91825407c", + "instance_future-architect__vuls-e4728e388120b311c4ed469e4f942e0347a2689b-v264a82e2f4818e30f5a25e4da53b27ba119f62b5", + "instance_qutebrowser__qutebrowser-5e0d6dc1483cb3336ea0e3dcbd4fe4aa00fc1742-v5149fcda2a9a6fe1d35dfed1bade1444a11ef271", + "instance_flipt-io__flipt-c12967bc73fdf02054cf3ef8498c05e25f0a18c0", + "instance_qutebrowser__qutebrowser-ff1c025ad3210506fc76e1f604d8c8c27637d88e-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_ansible__ansible-83909bfa22573777e3db5688773bda59721962ad-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_qutebrowser__qutebrowser-1a9e74bfaf9a9db2a510dc14572d33ded6040a57-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_gravitational__teleport-24cafecd8721891092210afc55f6413ab46ca211-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_internetarchive__openlibrary-4b7ea2977be2747496ba792a678940baa985f7ea-v0f5aece3601a5b4419f7ccec1dbda2071be28ee4", + "instance_internetarchive__openlibrary-fad4a40acf5ff5f06cd7441a5c7baf41a7d81fe4-vfa6ff903cb27f336e17654595dd900fa943dcd91", + "instance_qutebrowser__qutebrowser-3fd8e12949b8feda401930574facf09dd4180bba", + "instance_ansible__ansible-164881d871964aa64e0f911d03ae270acbad253c-v390e508d27db7a51eece36bb6d9698b63a5b638a", + "instance_ansible__ansible-fb144c44144f8bd3542e71f5db62b6d322c7bd85-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_element-hq__element-web-582a1b093fc0b77538052f45cbb9c7295f991b51-vnan", + "instance_element-hq__element-web-ec0f940ef0e8e3b61078f145f34dc40d1938e6c5-vnan", + "instance_internetarchive__openlibrary-6fdbbeee4c0a7e976ff3e46fb1d36f4eb110c428-v08d8e8889ec945ab821fb156c04c7d2e2810debb", + "instance_flipt-io__flipt-7161f7b876773a911afdd804b281e52681cb7321", + "instance_navidrome__navidrome-28389fb05e1523564dfc61fa43ed8eb8a10f938c", + "instance_ansible__ansible-5f4e332e3762999d94af27746db29ff1729252c1-v0f01c69f1e2528b935359cfe578530722bca2c59", + "instance_nodebb__nodebb-0e07f3c9bace416cbab078a30eae972868c0a8a3-vf2cf3cbd463b7ad942381f1c6d077626485a1e9e", + "instance_nodebb__nodebb-da0211b1a001d45d73b4c84c6417a4f1b0312575-vf2cf3cbd463b7ad942381f1c6d077626485a1e9e", + "instance_navidrome__navidrome-31799662706fedddf5bcc1a76b50409d1f91d327", + "instance_future-architect__vuls-50580f6e98eeb36f53f27222f7f4fdfea0b21e8d", + "instance_internetarchive__openlibrary-60725705782832a2cb22e17c49697948a42a9d03-v298a7a812ceed28c4c18355a091f1b268fe56d86", + "instance_ansible__ansible-3db08adbb1cc6aa9941be5e0fc810132c6e1fa4b-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_internetarchive__openlibrary-8a5a63af6e0be406aa6c8c9b6d5f28b2f1b6af5a-v0f5aece3601a5b4419f7ccec1dbda2071be28ee4", + "instance_future-architect__vuls-e1df74cbc1a1d1889428b3333a3b2405c4651993", + "instance_internetarchive__openlibrary-123e6e5e1c85b9c07d1e98f70bfc480bc8016890-v2733ff199fb72f0d033a30dc62cb0a4742e3a7f4", + "instance_protonmail__webclients-ae36cb23a1682dcfd69587c1b311ae0227e28f39", + "instance_protonmail__webclients-8be4f6cb9380fcd2e67bcb18cef931ae0d4b869c", + "instance_qutebrowser__qutebrowser-a25e8a09873838ca9efefd36ea8a45170bbeb95c-vc2f56a753b62a190ddb23cd330c257b9cf560d12", + "instance_qutebrowser__qutebrowser-44e64199ed38003253f0296badd4a447645067b6-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_gravitational__teleport-7744f72c6eb631791434b648ba41083b5f6d2278-vce94f93ad1030e3136852817f2423c1b3ac37bc4", + "instance_tutao__tutanota-40e94dee2bcec2b63f362da283123e9df1874cc1-vc4e41fd0029957297843cb9dec4a25c7c756f029", + "instance_future-architect__vuls-c11ba27509f733d7d280bdf661cbbe2e7a99df4c", + "instance_flipt-io__flipt-b2cd6a6dd73ca91b519015fd5924fde8d17f3f06", + "instance_gravitational__teleport-6a14edcf1ff010172fdbac622d0a474ed6af46de", + "instance_flipt-io__flipt-36e62baffae2132f78f9d34dc300a9baa2d7ae0e", + "instance_navidrome__navidrome-c90468b895f6171e33e937ff20dc915c995274f0", + "instance_ansible__ansible-ecea15c508f0e081525be036cf76bbb56dbcdd9d-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_ansible__ansible-e22e103cdf8edc56ff7d9b848a58f94f1471a263-v1055803c3a812189a1133297f7f5468579283f86", + "instance_future-architect__vuls-dc496468b9e9fb73371f9606cdcdb0f8e12e70ca", + "instance_gravitational__teleport-c782838c3a174fdff80cafd8cd3b1aa4dae8beb2", + "instance_internetarchive__openlibrary-91efee627df01e32007abf2d6ebf73f9d9053076-vbee42ad1b72fb23c6a1c874868a720b370983ed2", + "instance_qutebrowser__qutebrowser-52708364b5f91e198defb022d1a5b4b3ebd9b563-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_ansible__ansible-eea46a0d1b99a6dadedbb6a3502d599235fa7ec3-v390e508d27db7a51eece36bb6d9698b63a5b638a", + "instance_protonmail__webclients-c6f65d205c401350a226bb005f42fac1754b0b5b", + "instance_ansible__ansible-cb94c0cc550df9e98f1247bc71d8c2b861c75049-v1055803c3a812189a1133297f7f5468579283f86", + "instance_nodebb__nodebb-05f2236193f407cf8e2072757fbd6bb170bc13f0-vf2cf3cbd463b7ad942381f1c6d077626485a1e9e", + "instance_navidrome__navidrome-55730514ea59d5f1d0b8e3f8745569c29bdbf7b4", + "instance_navidrome__navidrome-0130c6dc13438b48cf0fdfab08a89e357b5517c9", + "instance_flipt-io__flipt-507170da0f7f4da330f6732bffdf11c4df7fc192", + "instance_tutao__tutanota-fe240cbf7f0fdd6744ef7bef8cb61676bcdbb621-vc4e41fd0029957297843cb9dec4a25c7c756f029", + "instance_qutebrowser__qutebrowser-233cb1cc48635130e5602549856a6fa4ab4c087f-v35616345bb8052ea303186706cec663146f0f184", + "instance_internetarchive__openlibrary-f8cc11d9c1575fdba5ac66aee0befca970da8d64-v13642507b4fc1f8d234172bf8129942da2c2ca26", + "instance_qutebrowser__qutebrowser-fec187c2cb53d769c2682b35ca77858a811414a8-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_tutao__tutanota-da4edb7375c10f47f4ed3860a591c5e6557f7b5c-vbc0d9ba8f0071fbe982809910959a6ff8884dbbf", + "instance_ansible__ansible-be2c376ab87e3e872ca21697508f12c6909cf85a-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_future-architect__vuls-457a3a9627fb9a0800d0aecf1d4713fb634a9011", + "instance_internetarchive__openlibrary-d40ec88713dc95ea791b252f92d2f7b75e107440-v13642507b4fc1f8d234172bf8129942da2c2ca26", + "instance_protonmail__webclients-863d524b5717b9d33ce08a0f0535e3fd8e8d1ed8", + "instance_internetarchive__openlibrary-308a35d6999427c02b1dbf5211c033ad3b352556-ve8c8d62a2b60610a3c4631f5f23ed866bada9818", + "instance_flipt-io__flipt-15b76cada1ef29cfa56b0fba36754be36243dded", + "instance_future-architect__vuls-61c39637f2f3809e1b5dad05f0c57c799dce1587", + "instance_ansible__ansible-5d253a13807e884b7ce0b6b57a963a45e2f0322c-v1055803c3a812189a1133297f7f5468579283f86", + "instance_element-hq__element-web-44b98896a79ede48f5ad7ff22619a39d5f6ff03c-vnan", + "instance_qutebrowser__qutebrowser-f91ace96223cac8161c16dd061907e138fe85111-v059c6fdc75567943479b23ebca7c07b5e9a7f34c", + "instance_gravitational__teleport-bb562408da4adeae16e025be65e170959d1ec492-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_gravitational__teleport-78b0d8c72637df1129fb6ff84fc49ef4b5ab1288", + "instance_internetarchive__openlibrary-3f7db6bbbcc7c418b3db72d157c6aed1d45b2ccf-v430f20c722405e462d9ef44dee7d34c41e76fe7a", + "instance_future-architect__vuls-e52fa8d6ed1d23e36f2a86e5d3efe9aa057a1b0d", + "instance_navidrome__navidrome-3bc9e75b2843f91f6a1e9b604e321c2bd4fd442a", + "instance_tutao__tutanota-fb32e5f9d9fc152a00144d56dd0af01760a2d4dc-vc4e41fd0029957297843cb9dec4a25c7c756f029", + "instance_gravitational__teleport-af5e2517de7d18406b614e413aca61c319312171-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_ansible__ansible-a20a52701402a12f91396549df04ac55809f68e9-v1055803c3a812189a1133297f7f5468579283f86", + "instance_protonmail__webclients-1501eb765873b2884b6f1944fd242ecfc9d6b103", + "instance_gravitational__teleport-db89206db6c2969266e664c7c0fb51b70e958b64", + "instance_ansible__ansible-be59caa59bf47ca78a4760eb7ff38568372a8260-v1055803c3a812189a1133297f7f5468579283f86", + "instance_flipt-io__flipt-65581fef4aa807540cb933753d085feb0d7e736f", + "instance_gravitational__teleport-4e1c39639edf1ab494dd7562844c8b277b5cfa18-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_future-architect__vuls-f0b3a8b1db98eb1bd32685f1c36c41a99c3452ed", + "instance_navidrome__navidrome-b65e76293a917ee2dfc5d4b373b1c62e054d0dca", + "instance_element-hq__element-web-33e8edb3d508d6eefb354819ca693b7accc695e7", + "instance_flipt-io__flipt-e50808c03e4b9d25a6a78af9c61a3b1616ea356b", + "instance_future-architect__vuls-e3c27e1817d68248043bd09d63cc31f3344a6f2c", + "instance_element-hq__element-web-ad26925bb6628260cfe0fcf90ec0a8cba381f4a4-vnan", + "instance_nodebb__nodebb-1ea9481af6125ffd6da0592ed439aa62af0bca11-vd59a5728dfc977f44533186ace531248c2917516", + "instance_ansible__ansible-489156378c8e97374a75a544c7c9c2c0dd8146d1-v390e508d27db7a51eece36bb6d9698b63a5b638a", + "instance_internetarchive__openlibrary-bb152d23c004f3d68986877143bb0f83531fe401-ve8c8d62a2b60610a3c4631f5f23ed866bada9818", + "instance_qutebrowser__qutebrowser-0b621cb0ce2b54d3f93d8d41d8ff4257888a87e5-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_qutebrowser__qutebrowser-1af602b258b97aaba69d2585ed499d95e2303ac2-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_future-architect__vuls-e1fab805afcfc92a2a615371d0ec1e667503c254-v264a82e2f4818e30f5a25e4da53b27ba119f62b5", + "instance_internetarchive__openlibrary-7edd1ef09d91fe0b435707633c5cc9af41dedddf-v76304ecdb3a5954fcf13feb710e8c40fcf24b73c", + "instance_qutebrowser__qutebrowser-99029144b5109bb1b2a53964a7c129e009980cd9-va0fd88aac89cde702ec1ba84877234da33adce8a", + "instance_ansible__ansible-12734fa21c08a0ce8c84e533abdc560db2eb1955-v7eee2454f617569fd6889f2211f75bc02a35f9f8", + "instance_protonmail__webclients-7e54526774e577c0ebb58ced7ba8bef349a69fec", + "instance_gravitational__teleport-1a77b7945a022ab86858029d30ac7ad0d5239d00-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_nodebb__nodebb-8ca65b0c78c67c1653487c02d1135e1b702185e1-vf2cf3cbd463b7ad942381f1c6d077626485a1e9e", + "instance_flipt-io__flipt-9f8127f225a86245fa35dca4885c2daef824ee55", + "instance_gravitational__teleport-ac2fb2f9b4fd1896b554d3011df23d3d71295779", + "instance_ansible__ansible-f02a62db509dc7463fab642c9c3458b9bc3476cc-v390e508d27db7a51eece36bb6d9698b63a5b638a" + ], + "additional_details": { + "shuffled": "True" + } + }, + "evaluation_timestamp": "1774867708.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.565, + "uncertainty": { + "standard_error": { + "value": 0.03514328173714403 + }, + "num_samples": 200 + }, + "details": { + "total_matched_trajectories": "25", + "stopping_reason_count_completed_on_successful_submit": "16", + "stopping_reason_count_repetition_guard": "9", + "avg_total_tokens_target_model": "2712832.40", + "total_total_tokens_target_model": "67820810.0", + "avg_total_tokens_other_models": "0.00", + "total_total_tokens_other_models": "0.0", + "avg_total_tokens_all_models": "2712832.40", + "total_total_tokens_all_models": "67820810.0", + "avg_turn_count": "55.76", + "total_turn_count": "1394.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"If you do not submit an answer, you will receive a score of zero.\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"Your submission was incorrect. Please review your work, identify what went wrong, and try again.\", \"to_float\"], \"submit\": null, \"on_continue\": null, \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\"}, \"params_passed\": {\"prompt\": [\"If you do not submit an answer, you will receive a score of zero.\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"Your submission was incorrect. Please review your work, identify what went wrong, and try again.\", \"to_float\"], \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 30000000 + }, + "sandbox": { + "type": "k8s" + } + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/anthropic/claude-opus-4-5-20251101/0f30a8a8-89ae-4fa4-867a-2856dca81ab0_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "dcc4280606fe7a1cef29e48c3c182a4452f4214375cdcd3563ef419804c8945e", + "total_rows": 200 + } +} \ No newline at end of file diff --git a/data/swebenchpro/anthropic/claude-opus-4-5-20251101/0f30a8a8-89ae-4fa4-867a-2856dca81ab0_samples.jsonl b/data/swebenchpro/anthropic/claude-opus-4-5-20251101/0f30a8a8-89ae-4fa4-867a-2856dca81ab0_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..95e92349df26eceba3ffc7a6ca4c824265b1ea58 --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-5-20251101/0f30a8a8-89ae-4fa4-867a-2856dca81ab0_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dcc4280606fe7a1cef29e48c3c182a4452f4214375cdcd3563ef419804c8945e +size 164690748 diff --git a/data/swebenchpro/anthropic/claude-opus-4-5-20251101/1ef18b24-2034-44c2-aa3c-713cde058425.json b/data/swebenchpro/anthropic/claude-opus-4-5-20251101/1ef18b24-2034-44c2-aa3c-713cde058425.json new file mode 100644 index 0000000000000000000000000000000000000000..2d9b4d87611ed3cbd48ba1ea6185927973095e63 --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-5-20251101/1ef18b24-2034-44c2-aa3c-713cde058425.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/anthropic_claude-opus-4-5-20251101/1786169480.622756", + "evaluation_timestamp": "1777303755.0", + "retrieved_timestamp": "1786169480.622756", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-5-20251101", + "id": "anthropic/claude-opus-4-5-20251101", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/anthropic/claude-opus-4-5-20251101/1ef18b24-2034-44c2-aa3c-713cde058425_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "8eb7abb57ef84d5330c8a4d8a36b227d80bbee5c589567b081cb5db5d1edf57b", + "total_rows": 63 + } +} \ No newline at end of file diff --git a/data/swebenchpro/anthropic/claude-opus-4-5-20251101/1ef18b24-2034-44c2-aa3c-713cde058425_samples.jsonl b/data/swebenchpro/anthropic/claude-opus-4-5-20251101/1ef18b24-2034-44c2-aa3c-713cde058425_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..084d32339d6b2d4dd0a520a962113a6f02dac388 --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-5-20251101/1ef18b24-2034-44c2-aa3c-713cde058425_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8eb7abb57ef84d5330c8a4d8a36b227d80bbee5c589567b081cb5db5d1edf57b +size 73492939 diff --git a/data/swebenchpro/anthropic/claude-opus-4-5-20251101/1f83b1ed-dbb1-436d-b4d3-dc0173b58e14.json b/data/swebenchpro/anthropic/claude-opus-4-5-20251101/1f83b1ed-dbb1-436d-b4d3-dc0173b58e14.json new file mode 100644 index 0000000000000000000000000000000000000000..7470d23fd8b959b224e057c6316fb2c585fe3caa --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-5-20251101/1f83b1ed-dbb1-436d-b4d3-dc0173b58e14.json @@ -0,0 +1,110 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/anthropic_claude-opus-4-5-20251101/1786165484.274739", + "evaluation_timestamp": "1778055257.0", + "retrieved_timestamp": "1786165484.274739", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-5-20251101", + "id": "anthropic/claude-opus-4-5-20251101", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on swebenchpro/S-adaptive/+1ep/38bb1c85 for scorer harbor_scorer", + "source_data": { + "dataset_name": "swebenchpro", + "source_type": "hf_dataset", + "samples_number": 1, + "sample_ids": [ + "instance_tutao__tutanota-8513a9e8114a8b42e64f4348335e0f23efa054c4-vee878bb72091875e912c52fc32bc60ec3760227b" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1778055257.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 1.0, + "uncertainty": { + "num_samples": 1 + }, + "details": { + "total_matched_trajectories": "1", + "stopping_reason_count_repetition_guard": "1", + "avg_total_tokens_target_model": "1646153.00", + "total_total_tokens_target_model": "1646153.0", + "avg_total_tokens_other_models": "1379.00", + "total_total_tokens_other_models": "1379.0", + "avg_total_tokens_all_models": "1647532.00", + "total_total_tokens_all_models": "1647532.0", + "avg_turn_count": "48.00", + "total_turn_count": "48.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 30000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/anthropic/claude-opus-4-5-20251101/1f83b1ed-dbb1-436d-b4d3-dc0173b58e14_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "6e8d0d04e205af4d948d6186ce0ec631691d0c50d6d5af7f8e040a54d962328e", + "total_rows": 1 + } +} \ No newline at end of file diff --git a/data/swebenchpro/anthropic/claude-opus-4-5-20251101/1f83b1ed-dbb1-436d-b4d3-dc0173b58e14_samples.jsonl b/data/swebenchpro/anthropic/claude-opus-4-5-20251101/1f83b1ed-dbb1-436d-b4d3-dc0173b58e14_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..ffc0821a8da2a5b129cead16ae12ecf959fd134f --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-5-20251101/1f83b1ed-dbb1-436d-b4d3-dc0173b58e14_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6e8d0d04e205af4d948d6186ce0ec631691d0c50d6d5af7f8e040a54d962328e +size 305142 diff --git a/data/swebenchpro/anthropic/claude-opus-4-5-20251101/2afdc58e-7508-443d-af57-2d2fc4cc6ba4.json b/data/swebenchpro/anthropic/claude-opus-4-5-20251101/2afdc58e-7508-443d-af57-2d2fc4cc6ba4.json new file mode 100644 index 0000000000000000000000000000000000000000..736c7368d20e82454f104802155d854e7138b906 --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-5-20251101/2afdc58e-7508-443d-af57-2d2fc4cc6ba4.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/anthropic_claude-opus-4-5-20251101/1786169229.198038", + "evaluation_timestamp": "1777078567.0", + "retrieved_timestamp": "1786169229.198038", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-5-20251101", + "id": "anthropic/claude-opus-4-5-20251101", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/anthropic/claude-opus-4-5-20251101/2afdc58e-7508-443d-af57-2d2fc4cc6ba4_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "caa031c7b99abd39cbdb0ec9c690c5ce38b1a6be0212f899bacdcde42ebc6e1b", + "total_rows": 43 + } +} \ No newline at end of file diff --git a/data/swebenchpro/anthropic/claude-opus-4-5-20251101/2afdc58e-7508-443d-af57-2d2fc4cc6ba4_samples.jsonl b/data/swebenchpro/anthropic/claude-opus-4-5-20251101/2afdc58e-7508-443d-af57-2d2fc4cc6ba4_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..9675fc086e6b010bee044e8786435d7367ecd5ef --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-5-20251101/2afdc58e-7508-443d-af57-2d2fc4cc6ba4_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:caa031c7b99abd39cbdb0ec9c690c5ce38b1a6be0212f899bacdcde42ebc6e1b +size 50706524 diff --git a/data/swebenchpro/anthropic/claude-opus-4-5-20251101/3ea8dc68-4d2e-4373-b1cd-2fa15bc2a3ab.json b/data/swebenchpro/anthropic/claude-opus-4-5-20251101/3ea8dc68-4d2e-4373-b1cd-2fa15bc2a3ab.json new file mode 100644 index 0000000000000000000000000000000000000000..01c74b27e7449742889cc6efaab79ea39e33962c --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-5-20251101/3ea8dc68-4d2e-4373-b1cd-2fa15bc2a3ab.json @@ -0,0 +1,110 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/anthropic_claude-opus-4-5-20251101/1786163045.002672", + "evaluation_timestamp": "1776433475.0", + "retrieved_timestamp": "1786163045.002672", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-5-20251101", + "id": "anthropic/claude-opus-4-5-20251101", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on swebenchpro/S-adaptive+C/+2ep/779be3d6 for scorer harbor_scorer", + "source_data": { + "dataset_name": "swebenchpro", + "source_type": "hf_dataset", + "samples_number": 7, + "sample_ids": [ + "instance_qutebrowser__qutebrowser-01d1d1494411380d97cac14614a829d3a69cecaf-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_internetarchive__openlibrary-f8cc11d9c1575fdba5ac66aee0befca970da8d64-v13642507b4fc1f8d234172bf8129942da2c2ca26", + "instance_qutebrowser__qutebrowser-e64622cd2df5b521342cf4a62e0d4cb8f8c9ae5a-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_ansible__ansible-e22e103cdf8edc56ff7d9b848a58f94f1471a263-v1055803c3a812189a1133297f7f5468579283f86", + "instance_qutebrowser__qutebrowser-e15d26630934d0b6415ed2295ac42fd570a57620-va0fd88aac89cde702ec1ba84877234da33adce8a", + "instance_ansible__ansible-164881d871964aa64e0f911d03ae270acbad253c-v390e508d27db7a51eece36bb6d9698b63a5b638a", + "instance_ansible__ansible-f327e65d11bb905ed9f15996024f857a95592629-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1776433475.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 1.0, + "uncertainty": { + "num_samples": 14 + }, + "details": { + "total_matched_trajectories": "14", + "stopping_reason_count_completed_on_successful_submit": "14", + "avg_total_tokens_target_model": "2678416.43", + "total_total_tokens_target_model": "37497830.0", + "avg_total_tokens_other_models": "0.00", + "total_total_tokens_other_models": "0.0", + "avg_total_tokens_all_models": "2678416.43", + "total_total_tokens_all_models": "37497830.0", + "avg_turn_count": "68.93", + "total_turn_count": "965.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "max_connections": "10", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 30000000 + }, + "sandbox": {} + }, + "additional_details": { + "max_connections": "10" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/anthropic/claude-opus-4-5-20251101/3ea8dc68-4d2e-4373-b1cd-2fa15bc2a3ab_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "a40f57ec66dda0271eb669603822753519f7c0e8d325fdbc9d52a0d824bed1f3", + "total_rows": 14 + } +} \ No newline at end of file diff --git a/data/swebenchpro/anthropic/claude-opus-4-5-20251101/3ea8dc68-4d2e-4373-b1cd-2fa15bc2a3ab_samples.jsonl b/data/swebenchpro/anthropic/claude-opus-4-5-20251101/3ea8dc68-4d2e-4373-b1cd-2fa15bc2a3ab_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..8e60a8e9aca759db0c54596da5abb5c623f02129 --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-5-20251101/3ea8dc68-4d2e-4373-b1cd-2fa15bc2a3ab_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a40f57ec66dda0271eb669603822753519f7c0e8d325fdbc9d52a0d824bed1f3 +size 4594155 diff --git a/data/swebenchpro/anthropic/claude-opus-4-5-20251101/45002255-0a58-4e02-b2fd-24610f5931b9.json b/data/swebenchpro/anthropic/claude-opus-4-5-20251101/45002255-0a58-4e02-b2fd-24610f5931b9.json new file mode 100644 index 0000000000000000000000000000000000000000..b4ce8667384cd4a0f3d9a4a8c7c16bf0c00f608d --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-5-20251101/45002255-0a58-4e02-b2fd-24610f5931b9.json @@ -0,0 +1,133 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/anthropic_claude-opus-4-5-20251101/1786163041.506745", + "evaluation_timestamp": "1778228784.0", + "retrieved_timestamp": "1786163041.506745", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-5-20251101", + "id": "anthropic/claude-opus-4-5-20251101", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on swebenchpro/S-adaptive+C/+1ep/85524ce9 for scorer harbor_scorer", + "source_data": { + "dataset_name": "swebenchpro", + "source_type": "hf_dataset", + "samples_number": 20, + "sample_ids": [ + "instance_tutao__tutanota-d1aa0ecec288bfc800cfb9133b087c4f81ad8b38-vbc0d9ba8f0071fbe982809910959a6ff8884dbbf", + "instance_protonmail__webclients-cb8cc309c6968b0a2a5fe4288d0ae0a969ff31e1", + "instance_ansible__ansible-11c1777d56664b1acb56b387a1ad6aeadef1391d-v0f01c69f1e2528b935359cfe578530722bca2c59", + "instance_nodebb__nodebb-70b4a0e2aebebe8f2f559de6680093d96a697b2f-vnan", + "instance_element-hq__element-web-b007ea81b2ccd001b00f332bee65070aa7fc00f9-vnan", + "instance_flipt-io__flipt-cd2f3b0a9d4d8b8a6d3d56afab65851ecdc408e8", + "instance_element-hq__element-web-2760bfc8369f1bee640d6d7a7e910783143d4c5f-vnan", + "instance_protonmail__webclients-5f0745dd6993bb1430a951c62a49807c6635cd77", + "instance_gravitational__teleport-46aa81b1ce96ebb4ebed2ae53fd78cd44a05da6c-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_tutao__tutanota-8513a9e8114a8b42e64f4348335e0f23efa054c4-vee878bb72091875e912c52fc32bc60ec3760227b", + "instance_internetarchive__openlibrary-7f6b722a10f822171501d027cad60afe53337732-ve8c8d62a2b60610a3c4631f5f23ed866bada9818", + "instance_nodebb__nodebb-a917210c5b2c20637094545401f85783905c074c-vf2cf3cbd463b7ad942381f1c6d077626485a1e9e", + "instance_internetarchive__openlibrary-1351c59fd43689753de1fca32c78d539a116ffc1-v29f82c9cf21d57b242f8d8b0e541525d259e2d63", + "instance_protonmail__webclients-281a6b3f190f323ec2c0630999354fafb84b2880", + "instance_future-architect__vuls-999529a05b202b0fd29c6fca5039a4c47a3766bb", + "instance_element-hq__element-web-1077729a19c0ce902e713cf6fab42c91fb7907f1-vnan", + "instance_future-architect__vuls-be7b9114cc9545e68fb0ee7bc63d7ec53d1a00ad", + "instance_qutebrowser__qutebrowser-c580ebf0801e5a3ecabc54f327498bb753c6d5f2-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_internetarchive__openlibrary-3f580a5f244c299d936d73d9e327ba873b6401d9-v0f5aece3601a5b4419f7ccec1dbda2071be28ee4", + "instance_qutebrowser__qutebrowser-0aa57e4f7243024fa4bba8853306691b5dbd77b3-v5149fcda2a9a6fe1d35dfed1bade1444a11ef271" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1778228784.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.8, + "uncertainty": { + "standard_error": { + "value": 0.0917662935482247 + }, + "num_samples": 20 + }, + "details": { + "total_matched_trajectories": "20", + "stopping_reason_count_completed_on_successful_submit": "16", + "stopping_reason_count_repetition_guard": "4", + "avg_total_tokens_target_model": "3745022.85", + "total_total_tokens_target_model": "74900457.0", + "avg_total_tokens_other_models": "292.60", + "total_total_tokens_other_models": "5852.0", + "avg_total_tokens_all_models": "3745315.45", + "total_total_tokens_all_models": "74906309.0", + "avg_turn_count": "75.60", + "total_turn_count": "1512.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 30000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/anthropic/claude-opus-4-5-20251101/45002255-0a58-4e02-b2fd-24610f5931b9_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "030427a207eb15f077b081d581f1b7004e476d015882625aebcf3967d050ca59", + "total_rows": 20 + } +} \ No newline at end of file diff --git a/data/swebenchpro/anthropic/claude-opus-4-5-20251101/45002255-0a58-4e02-b2fd-24610f5931b9_samples.jsonl b/data/swebenchpro/anthropic/claude-opus-4-5-20251101/45002255-0a58-4e02-b2fd-24610f5931b9_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..a564080aac78ccd7bb3e1a4749db441a94b8b4ca --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-5-20251101/45002255-0a58-4e02-b2fd-24610f5931b9_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:030427a207eb15f077b081d581f1b7004e476d015882625aebcf3967d050ca59 +size 7583716 diff --git a/data/swebenchpro/anthropic/claude-opus-4-5-20251101/58194ac4-41ec-4bc5-a654-5d075d47ba37.json b/data/swebenchpro/anthropic/claude-opus-4-5-20251101/58194ac4-41ec-4bc5-a654-5d075d47ba37.json new file mode 100644 index 0000000000000000000000000000000000000000..54dbdbb26169fe8ad009acac506e78e807967632 --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-5-20251101/58194ac4-41ec-4bc5-a654-5d075d47ba37.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/anthropic_claude-opus-4-5-20251101/1786167418.519097", + "evaluation_timestamp": "1776417565.0", + "retrieved_timestamp": "1786167418.519097", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-5-20251101", + "id": "anthropic/claude-opus-4-5-20251101", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/anthropic/claude-opus-4-5-20251101/58194ac4-41ec-4bc5-a654-5d075d47ba37_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "5a1709b1cdb6bc963bbd7eb4da07a169200a1c39e5712fc6ecf97a402ed8b975", + "total_rows": 3 + } +} \ No newline at end of file diff --git a/data/swebenchpro/anthropic/claude-opus-4-5-20251101/58194ac4-41ec-4bc5-a654-5d075d47ba37_samples.jsonl b/data/swebenchpro/anthropic/claude-opus-4-5-20251101/58194ac4-41ec-4bc5-a654-5d075d47ba37_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..c69c490d83acff018e7a1bddc6c63052002eff4c --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-5-20251101/58194ac4-41ec-4bc5-a654-5d075d47ba37_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5a1709b1cdb6bc963bbd7eb4da07a169200a1c39e5712fc6ecf97a402ed8b975 +size 1072992 diff --git a/data/swebenchpro/anthropic/claude-opus-4-5-20251101/5d6c3a74-3d11-4c29-b92c-59dca2f69b01.json b/data/swebenchpro/anthropic/claude-opus-4-5-20251101/5d6c3a74-3d11-4c29-b92c-59dca2f69b01.json new file mode 100644 index 0000000000000000000000000000000000000000..f5a4e561c1af5fe4dddea93050662364a08a9ba3 --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-5-20251101/5d6c3a74-3d11-4c29-b92c-59dca2f69b01.json @@ -0,0 +1,126 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/anthropic_claude-opus-4-5-20251101/1786165493.319434", + "evaluation_timestamp": "1777292858.0", + "retrieved_timestamp": "1786165493.319434", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-5-20251101", + "id": "anthropic/claude-opus-4-5-20251101", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on swebenchpro/S-adaptive/+1ep/f4ac0573 for scorer harbor_scorer", + "source_data": { + "dataset_name": "swebenchpro", + "source_type": "hf_dataset", + "samples_number": 16, + "sample_ids": [ + "instance_internetarchive__openlibrary-fdbc0d8f418333c7e575c40b661b582c301ef7ac-v13642507b4fc1f8d234172bf8129942da2c2ca26", + "instance_ansible__ansible-b2a289dcbb702003377221e25f62c8a3608f0e89-v173091e2e36d38c978002990795f66cfc0af30ad", + "instance_internetarchive__openlibrary-f8cc11d9c1575fdba5ac66aee0befca970da8d64-v13642507b4fc1f8d234172bf8129942da2c2ca26", + "instance_qutebrowser__qutebrowser-e70f5b03187bdd40e8bf70f5f3ead840f52d1f42-v02ad04386d5238fe2d1a1be450df257370de4b6a", + "instance_tutao__tutanota-d1aa0ecec288bfc800cfb9133b087c4f81ad8b38-vbc0d9ba8f0071fbe982809910959a6ff8884dbbf", + "instance_qutebrowser__qutebrowser-e64622cd2df5b521342cf4a62e0d4cb8f8c9ae5a-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_gravitational__teleport-ac2fb2f9b4fd1896b554d3011df23d3d71295779", + "instance_protonmail__webclients-cb8cc309c6968b0a2a5fe4288d0ae0a969ff31e1", + "instance_ansible__ansible-11c1777d56664b1acb56b387a1ad6aeadef1391d-v0f01c69f1e2528b935359cfe578530722bca2c59", + "instance_ansible__ansible-29aea9ff3466e4cd2ed00524b9e56738d568ce8b-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_nodebb__nodebb-70b4a0e2aebebe8f2f559de6680093d96a697b2f-vnan", + "instance_ansible__ansible-e22e103cdf8edc56ff7d9b848a58f94f1471a263-v1055803c3a812189a1133297f7f5468579283f86", + "instance_qutebrowser__qutebrowser-e15d26630934d0b6415ed2295ac42fd570a57620-va0fd88aac89cde702ec1ba84877234da33adce8a", + "instance_ansible__ansible-164881d871964aa64e0f911d03ae270acbad253c-v390e508d27db7a51eece36bb6d9698b63a5b638a", + "instance_ansible__ansible-502270c804c33d3bc963930dc85e0f4ca359674d-v7eee2454f617569fd6889f2211f75bc02a35f9f8", + "instance_internetarchive__openlibrary-9c392b60e2c6fa1d68cb68084b4b4ff04d0cb35c-v2d9a6c849c60ed19fd0858ce9e40b7cc8e097e59" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1777292858.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.9375, + "uncertainty": { + "standard_error": { + "value": 0.0625 + }, + "num_samples": 16 + }, + "details": { + "total_matched_trajectories": "16", + "stopping_reason_count_repetition_guard": "9", + "stopping_reason_count_token_limit": "7", + "avg_total_tokens_target_model": "20780280.94", + "total_total_tokens_target_model": "332484495.0", + "avg_total_tokens_other_models": "310.19", + "total_total_tokens_other_models": "4963.0", + "avg_total_tokens_all_models": "20780591.12", + "total_total_tokens_all_models": "332489458.0", + "avg_turn_count": "303.50", + "total_turn_count": "4856.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 30000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/anthropic/claude-opus-4-5-20251101/5d6c3a74-3d11-4c29-b92c-59dca2f69b01_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "8a9772f8ea058366182f538dd116cce9aa4a132509922c6a63412363fbcdd939", + "total_rows": 16 + } +} \ No newline at end of file diff --git a/data/swebenchpro/anthropic/claude-opus-4-5-20251101/5d6c3a74-3d11-4c29-b92c-59dca2f69b01_samples.jsonl b/data/swebenchpro/anthropic/claude-opus-4-5-20251101/5d6c3a74-3d11-4c29-b92c-59dca2f69b01_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..a106642e8c5a8dd15311ee2706db39fefc5725de --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-5-20251101/5d6c3a74-3d11-4c29-b92c-59dca2f69b01_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8a9772f8ea058366182f538dd116cce9aa4a132509922c6a63412363fbcdd939 +size 20182443 diff --git a/data/swebenchpro/anthropic/claude-opus-4-5-20251101/6c50ea8c-703c-4d1b-98d7-b892b50a0f98.json b/data/swebenchpro/anthropic/claude-opus-4-5-20251101/6c50ea8c-703c-4d1b-98d7-b892b50a0f98.json new file mode 100644 index 0000000000000000000000000000000000000000..cc2354c81ad0917560fb665ef48366229c351727 --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-5-20251101/6c50ea8c-703c-4d1b-98d7-b892b50a0f98.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/anthropic_claude-opus-4-5-20251101/1786169203.200411", + "evaluation_timestamp": "1777586285.0", + "retrieved_timestamp": "1786169203.200411", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-5-20251101", + "id": "anthropic/claude-opus-4-5-20251101", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/anthropic/claude-opus-4-5-20251101/6c50ea8c-703c-4d1b-98d7-b892b50a0f98_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "9c5beaa2a5a1a081c187495446edfbde1df6bf6813370d07fe057b1a27a6fa7e", + "total_rows": 4 + } +} \ No newline at end of file diff --git a/data/swebenchpro/anthropic/claude-opus-4-5-20251101/6c50ea8c-703c-4d1b-98d7-b892b50a0f98_samples.jsonl b/data/swebenchpro/anthropic/claude-opus-4-5-20251101/6c50ea8c-703c-4d1b-98d7-b892b50a0f98_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..0e8acb0e292c818ad0ba86e537a1256555eb3d9f --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-5-20251101/6c50ea8c-703c-4d1b-98d7-b892b50a0f98_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9c5beaa2a5a1a081c187495446edfbde1df6bf6813370d07fe057b1a27a6fa7e +size 4990903 diff --git a/data/swebenchpro/anthropic/claude-opus-4-5-20251101/6fa6e07b-2b68-440d-80eb-475b34d8e50e.json b/data/swebenchpro/anthropic/claude-opus-4-5-20251101/6fa6e07b-2b68-440d-80eb-475b34d8e50e.json new file mode 100644 index 0000000000000000000000000000000000000000..6f148eb44c1863ce13719436abecba3a38d30b82 --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-5-20251101/6fa6e07b-2b68-440d-80eb-475b34d8e50e.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/anthropic_claude-opus-4-5-20251101/1786169182.431887", + "evaluation_timestamp": "1777645894.0", + "retrieved_timestamp": "1786169182.431887", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-5-20251101", + "id": "anthropic/claude-opus-4-5-20251101", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/anthropic/claude-opus-4-5-20251101/6fa6e07b-2b68-440d-80eb-475b34d8e50e_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "50102f16b163b3c112d8231f37dc395ba16dcf96abad3cd3b7f9c13c6df6aab6", + "total_rows": 1 + } +} \ No newline at end of file diff --git a/data/swebenchpro/anthropic/claude-opus-4-5-20251101/6fa6e07b-2b68-440d-80eb-475b34d8e50e_samples.jsonl b/data/swebenchpro/anthropic/claude-opus-4-5-20251101/6fa6e07b-2b68-440d-80eb-475b34d8e50e_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..6d225cfeb6329fb8da3b584708a588bb1c833fc5 --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-5-20251101/6fa6e07b-2b68-440d-80eb-475b34d8e50e_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:50102f16b163b3c112d8231f37dc395ba16dcf96abad3cd3b7f9c13c6df6aab6 +size 1066700 diff --git a/data/swebenchpro/anthropic/claude-opus-4-5-20251101/70617389-cabd-4ba1-9882-b8ad4dfebbeb.json b/data/swebenchpro/anthropic/claude-opus-4-5-20251101/70617389-cabd-4ba1-9882-b8ad4dfebbeb.json new file mode 100644 index 0000000000000000000000000000000000000000..026d2e2208a182666070477bedfcc1245a2494b0 --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-5-20251101/70617389-cabd-4ba1-9882-b8ad4dfebbeb.json @@ -0,0 +1,178 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/anthropic_claude-opus-4-5-20251101/1786163134.827185", + "evaluation_timestamp": "1778056929.0", + "retrieved_timestamp": "1786163134.827185", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-5-20251101", + "id": "anthropic/claude-opus-4-5-20251101", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on swebenchpro/S-adaptive+C/c54f2dbe for scorer harbor_scorer", + "source_data": { + "dataset_name": "swebenchpro", + "source_type": "hf_dataset", + "samples_number": 64, + "sample_ids": [ + "instance_tutao__tutanota-d1aa0ecec288bfc800cfb9133b087c4f81ad8b38-vbc0d9ba8f0071fbe982809910959a6ff8884dbbf", + "instance_protonmail__webclients-cb8cc309c6968b0a2a5fe4288d0ae0a969ff31e1", + "instance_ansible__ansible-11c1777d56664b1acb56b387a1ad6aeadef1391d-v0f01c69f1e2528b935359cfe578530722bca2c59", + "instance_ansible__ansible-29aea9ff3466e4cd2ed00524b9e56738d568ce8b-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_nodebb__nodebb-70b4a0e2aebebe8f2f559de6680093d96a697b2f-vnan", + "instance_ansible__ansible-502270c804c33d3bc963930dc85e0f4ca359674d-v7eee2454f617569fd6889f2211f75bc02a35f9f8", + "instance_internetarchive__openlibrary-9c392b60e2c6fa1d68cb68084b4b4ff04d0cb35c-v2d9a6c849c60ed19fd0858ce9e40b7cc8e097e59", + "instance_element-hq__element-web-b007ea81b2ccd001b00f332bee65070aa7fc00f9-vnan", + "instance_ansible__ansible-811093f0225caa4dd33890933150a81c6a6d5226-v1055803c3a812189a1133297f7f5468579283f86", + "instance_flipt-io__flipt-cd2f3b0a9d4d8b8a6d3d56afab65851ecdc408e8", + "instance_internetarchive__openlibrary-7cbfb812ef0e1f9716e2d6e85d538a96fcb79d13-vfa6ff903cb27f336e17654595dd900fa943dcd91", + "instance_flipt-io__flipt-0b119520afca1cf25c470ff4288c464d4510b944", + "instance_element-hq__element-web-f63160f38459fb552d00fcc60d4064977a9095a6-vnan", + "instance_gravitational__teleport-73cc189b0e9636d418c4470ecce0d9af5dae2f02-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_internetarchive__openlibrary-3c48b4bb782189e0858e6c3fc7956046cf3e1cfb-v2d9a6c849c60ed19fd0858ce9e40b7cc8e097e59", + "instance_future-architect__vuls-878c25bf5a9c9fd88ac32eb843f5636834d5712d", + "instance_qutebrowser__qutebrowser-70248f256f93ed9b1984494d0a1a919ddd774892-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_future-architect__vuls-73f0adad95c4d227e2ccfa876c85cc95dd065e13", + "instance_qutebrowser__qutebrowser-cc360cd4a34a126274c7b51f3b63afbaf3e05a02-v5fc38aaf22415ab0b70567368332beee7955b367", + "instance_flipt-io__flipt-e2bd19dafa7166c96b082fb2a59eb54b4be0d778", + "instance_qutebrowser__qutebrowser-9ed748effa8f3bcd804612d9291da017b514e12f-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_ansible__ansible-1ee70fc272aff6bf3415357c6e13c5de5b928d9b-v1055803c3a812189a1133297f7f5468579283f86", + "instance_element-hq__element-web-75c2c1a572fa45d1ea1d1a96e9e36e303332ecaa-vnan", + "instance_qutebrowser__qutebrowser-f631cd4422744160d9dcf7a0455da532ce973315-v35616345bb8052ea303186706cec663146f0f184", + "instance_qutebrowser__qutebrowser-e5340c449f23608803c286da0563b62f58ba25b0-v059c6fdc75567943479b23ebca7c07b5e9a7f34c", + "instance_ansible__ansible-42355d181a11b51ebfc56f6f4b3d9c74e01cb13b-v1055803c3a812189a1133297f7f5468579283f86", + "instance_internetarchive__openlibrary-09865f5fb549694d969f0a8e49b9d204ef1853ca-ve8c8d62a2b60610a3c4631f5f23ed866bada9818", + "instance_element-hq__element-web-2760bfc8369f1bee640d6d7a7e910783143d4c5f-vnan", + "instance_protonmail__webclients-01ea5214d11e0df8b7170d91bafd34f23cb0f2b1", + "instance_tutao__tutanota-4b4e45949096bb288f2b522f657610e480efa3e8-vee878bb72091875e912c52fc32bc60ec3760227b", + "instance_protonmail__webclients-5f0745dd6993bb1430a951c62a49807c6635cd77", + "instance_tutao__tutanota-1ff82aa365763cee2d609c9d19360ad87fdf2ec7-vc4e41fd0029957297843cb9dec4a25c7c756f029", + "instance_internetarchive__openlibrary-30bc73a1395fba2300087c7f307e54bb5372b60a-v76304ecdb3a5954fcf13feb710e8c40fcf24b73c", + "instance_nodebb__nodebb-22368b996ee0e5f11a5189b400b33af3cc8d925a-v4fbcfae8b15e4ce5d132c408bca69ebb9cf146ed", + "instance_protonmail__webclients-01b519cd49e6a24d9a05d2eb97f54e420740072e", + "instance_ansible__ansible-77658704217d5f166404fc67997203c25381cb6e-v390e508d27db7a51eece36bb6d9698b63a5b638a", + "instance_qutebrowser__qutebrowser-2dd8966fdcf11972062c540b7a787e4d0de8d372-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_ansible__ansible-cd9c4eb5a6b2bfaf4a6709f001ce3d0c92c1eed2-v0f01c69f1e2528b935359cfe578530722bca2c59", + "instance_protonmail__webclients-ac23d1efa1a6ab7e62724779317ba44c28d78cfd", + "instance_future-architect__vuls-030b2e03525d68d74cb749959aac2d7f3fc0effa", + "instance_nodebb__nodebb-b1f9ad5534bb3a44dab5364f659876a4b7fe34c1-vnan", + "instance_flipt-io__flipt-292fdaca9be39e6a921aaa8874c011d0fdd3e874", + "instance_qutebrowser__qutebrowser-ef5ba1a0360b39f9eff027fbdc57f363597c3c3b-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_ansible__ansible-622a493ae03bd5e5cf517d336fc426e9d12208c7-v906c969b551b346ef54a2c0b41e04f632b7b73c2", + "instance_ansible__ansible-9a21e247786ebd294dafafca1105fcd770ff46c6-v67cdaa49f89b34e42b69d5b7830b3c3ad3d8803f", + "instance_protonmail__webclients-8afd9ce04c8dde9e150e1c2b50d32e7ee2efa3e7", + "instance_flipt-io__flipt-b22f5f02e40b225b6b93fff472914973422e97c6", + "instance_gravitational__teleport-2bb3bbbd8aff1164a2353381cb79e1dc93b90d28-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_ansible__ansible-de5858f48dc9e1ce9117034e0d7e76806f420ca8-v1055803c3a812189a1133297f7f5468579283f86", + "instance_gravitational__teleport-46aa81b1ce96ebb4ebed2ae53fd78cd44a05da6c-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_tutao__tutanota-8513a9e8114a8b42e64f4348335e0f23efa054c4-vee878bb72091875e912c52fc32bc60ec3760227b", + "instance_internetarchive__openlibrary-7f6b722a10f822171501d027cad60afe53337732-ve8c8d62a2b60610a3c4631f5f23ed866bada9818", + "instance_nodebb__nodebb-a917210c5b2c20637094545401f85783905c074c-vf2cf3cbd463b7ad942381f1c6d077626485a1e9e", + "instance_internetarchive__openlibrary-1351c59fd43689753de1fca32c78d539a116ffc1-v29f82c9cf21d57b242f8d8b0e541525d259e2d63", + "instance_protonmail__webclients-281a6b3f190f323ec2c0630999354fafb84b2880", + "instance_future-architect__vuls-999529a05b202b0fd29c6fca5039a4c47a3766bb", + "instance_element-hq__element-web-1077729a19c0ce902e713cf6fab42c91fb7907f1-vnan", + "instance_future-architect__vuls-be7b9114cc9545e68fb0ee7bc63d7ec53d1a00ad", + "instance_qutebrowser__qutebrowser-c580ebf0801e5a3ecabc54f327498bb753c6d5f2-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_gravitational__teleport-3ff19cf7c41f396ae468797d3aeb61515517edc9-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_future-architect__vuls-aaea15e516ece43978cf98e09e52080478b1d39f", + "instance_qutebrowser__qutebrowser-cf06f4e3708f886032d4d2a30108c2fddb042d81-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_internetarchive__openlibrary-3f580a5f244c299d936d73d9e327ba873b6401d9-v0f5aece3601a5b4419f7ccec1dbda2071be28ee4", + "instance_qutebrowser__qutebrowser-0aa57e4f7243024fa4bba8853306691b5dbd77b3-v5149fcda2a9a6fe1d35dfed1bade1444a11ef271" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1778056929.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.746875, + "uncertainty": { + "standard_error": { + "value": 0.054416491194204074 + }, + "num_samples": 300 + }, + "details": { + "total_matched_trajectories": "300", + "stopping_reason_count_completed_on_successful_submit": "223", + "stopping_reason_count_repetition_guard": "76", + "stopping_reason_count_completed_with_submit": "1", + "avg_total_tokens_target_model": "3597106.09", + "total_total_tokens_target_model": "1079131828.0", + "avg_total_tokens_other_models": "397.00", + "total_total_tokens_other_models": "119100.0", + "avg_total_tokens_all_models": "3597503.09", + "total_total_tokens_all_models": "1079250928.0", + "avg_turn_count": "75.03", + "total_turn_count": "22509.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 30000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/anthropic/claude-opus-4-5-20251101/70617389-cabd-4ba1-9882-b8ad4dfebbeb_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "86e6b6b18d757d601ce36cae50a28ab45dd95e17be9dd6efe691958693e6e870", + "total_rows": 300 + } +} \ No newline at end of file diff --git a/data/swebenchpro/anthropic/claude-opus-4-5-20251101/70617389-cabd-4ba1-9882-b8ad4dfebbeb_samples.jsonl b/data/swebenchpro/anthropic/claude-opus-4-5-20251101/70617389-cabd-4ba1-9882-b8ad4dfebbeb_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..b178244fe457cbf76d7c92cd27d82185840200d4 --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-5-20251101/70617389-cabd-4ba1-9882-b8ad4dfebbeb_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:86e6b6b18d757d601ce36cae50a28ab45dd95e17be9dd6efe691958693e6e870 +size 113804707 diff --git a/data/swebenchpro/anthropic/claude-opus-4-5-20251101/7ef458a8-4124-4aab-9fc6-876424ab8de2.json b/data/swebenchpro/anthropic/claude-opus-4-5-20251101/7ef458a8-4124-4aab-9fc6-876424ab8de2.json new file mode 100644 index 0000000000000000000000000000000000000000..dc8308af642f15d55d64ac7526769c8ae0653cbf --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-5-20251101/7ef458a8-4124-4aab-9fc6-876424ab8de2.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/anthropic_claude-opus-4-5-20251101/1786169374.388476", + "evaluation_timestamp": "1777369892.0", + "retrieved_timestamp": "1786169374.388476", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-5-20251101", + "id": "anthropic/claude-opus-4-5-20251101", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/anthropic/claude-opus-4-5-20251101/7ef458a8-4124-4aab-9fc6-876424ab8de2_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "1f29646cdc825ffb08b8d66ff9fbee1949a9ad5c6ceddd71d9adbb63f0a0354e", + "total_rows": 63 + } +} \ No newline at end of file diff --git a/data/swebenchpro/anthropic/claude-opus-4-5-20251101/7ef458a8-4124-4aab-9fc6-876424ab8de2_samples.jsonl b/data/swebenchpro/anthropic/claude-opus-4-5-20251101/7ef458a8-4124-4aab-9fc6-876424ab8de2_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..e08b5ddbff20a6e3192963eae1c7fef3509ad23e --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-5-20251101/7ef458a8-4124-4aab-9fc6-876424ab8de2_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1f29646cdc825ffb08b8d66ff9fbee1949a9ad5c6ceddd71d9adbb63f0a0354e +size 68988233 diff --git a/data/swebenchpro/anthropic/claude-opus-4-5-20251101/82e216d6-4ae3-40f7-a3af-a3afdcb2603e.json b/data/swebenchpro/anthropic/claude-opus-4-5-20251101/82e216d6-4ae3-40f7-a3af-a3afdcb2603e.json new file mode 100644 index 0000000000000000000000000000000000000000..e22f4576e37321c7a0fc1453bcd15bb13c6b694a --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-5-20251101/82e216d6-4ae3-40f7-a3af-a3afdcb2603e.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/anthropic_claude-opus-4-5-20251101/1786167411.153298", + "evaluation_timestamp": "1776359446.0", + "retrieved_timestamp": "1786167411.153298", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-5-20251101", + "id": "anthropic/claude-opus-4-5-20251101", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/anthropic/claude-opus-4-5-20251101/82e216d6-4ae3-40f7-a3af-a3afdcb2603e_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "6fdb4495a5c30b13706322d470917c58c7b17bccbe4bed4ee2cb2ba6d552c6ea", + "total_rows": 20 + } +} \ No newline at end of file diff --git a/data/swebenchpro/anthropic/claude-opus-4-5-20251101/82e216d6-4ae3-40f7-a3af-a3afdcb2603e_samples.jsonl b/data/swebenchpro/anthropic/claude-opus-4-5-20251101/82e216d6-4ae3-40f7-a3af-a3afdcb2603e_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..f8d93a6e3e2da3aa2a1a8ca7d6a4ce52c8549019 --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-5-20251101/82e216d6-4ae3-40f7-a3af-a3afdcb2603e_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6fdb4495a5c30b13706322d470917c58c7b17bccbe4bed4ee2cb2ba6d552c6ea +size 10427681 diff --git a/data/swebenchpro/anthropic/claude-opus-4-5-20251101/a930c53a-1c2f-430c-88db-377ba0bcc427.json b/data/swebenchpro/anthropic/claude-opus-4-5-20251101/a930c53a-1c2f-430c-88db-377ba0bcc427.json new file mode 100644 index 0000000000000000000000000000000000000000..772e4e69326de8ea7141516a350796dbb1f51051 --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-5-20251101/a930c53a-1c2f-430c-88db-377ba0bcc427.json @@ -0,0 +1,122 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/anthropic_claude-opus-4-5-20251101/1786165525.643199", + "evaluation_timestamp": "1777719650.0", + "retrieved_timestamp": "1786165525.643199", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-5-20251101", + "id": "anthropic/claude-opus-4-5-20251101", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on swebenchpro/S-adaptive/+2ep/d24ea760 for scorer harbor_scorer", + "source_data": { + "dataset_name": "swebenchpro", + "source_type": "hf_dataset", + "samples_number": 12, + "sample_ids": [ + "instance_tutao__tutanota-4b4e45949096bb288f2b522f657610e480efa3e8-vee878bb72091875e912c52fc32bc60ec3760227b", + "instance_tutao__tutanota-1ff82aa365763cee2d609c9d19360ad87fdf2ec7-vc4e41fd0029957297843cb9dec4a25c7c756f029", + "instance_protonmail__webclients-01b519cd49e6a24d9a05d2eb97f54e420740072e", + "instance_flipt-io__flipt-b22f5f02e40b225b6b93fff472914973422e97c6", + "instance_gravitational__teleport-2bb3bbbd8aff1164a2353381cb79e1dc93b90d28-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_protonmail__webclients-281a6b3f190f323ec2c0630999354fafb84b2880", + "instance_future-architect__vuls-aaea15e516ece43978cf98e09e52080478b1d39f", + "instance_protonmail__webclients-f161c10cf7d31abf82e8d64d7a99c9fac5acfa18", + "instance_qutebrowser__qutebrowser-cf06f4e3708f886032d4d2a30108c2fddb042d81-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_internetarchive__openlibrary-3f580a5f244c299d936d73d9e327ba873b6401d9-v0f5aece3601a5b4419f7ccec1dbda2071be28ee4", + "instance_qutebrowser__qutebrowser-0aa57e4f7243024fa4bba8853306691b5dbd77b3-v5149fcda2a9a6fe1d35dfed1bade1444a11ef271", + "instance_future-architect__vuls-50580f6e98eeb36f53f27222f7f4fdfea0b21e8d" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1777719650.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.6666666666666666, + "uncertainty": { + "standard_error": { + "value": 0.1421338109037403 + }, + "num_samples": 24 + }, + "details": { + "total_matched_trajectories": "24", + "stopping_reason_count_token_limit": "10", + "stopping_reason_count_repetition_guard": "14", + "avg_total_tokens_target_model": "18479169.67", + "total_total_tokens_target_model": "443500072.0", + "avg_total_tokens_other_models": "417.58", + "total_total_tokens_other_models": "10022.0", + "avg_total_tokens_all_models": "18479587.25", + "total_total_tokens_all_models": "443510094.0", + "avg_turn_count": "262.42", + "total_turn_count": "6298.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 30000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/anthropic/claude-opus-4-5-20251101/a930c53a-1c2f-430c-88db-377ba0bcc427_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "c3f66d6b2fb9dfe425513c006dc2184ebf5e59c40e1aa0b43fe5b11645ab2dba", + "total_rows": 24 + } +} \ No newline at end of file diff --git a/data/swebenchpro/anthropic/claude-opus-4-5-20251101/a930c53a-1c2f-430c-88db-377ba0bcc427_samples.jsonl b/data/swebenchpro/anthropic/claude-opus-4-5-20251101/a930c53a-1c2f-430c-88db-377ba0bcc427_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..82a59ac1a0e7e80ec35783625b5cdf30dbc87519 --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-5-20251101/a930c53a-1c2f-430c-88db-377ba0bcc427_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c3f66d6b2fb9dfe425513c006dc2184ebf5e59c40e1aa0b43fe5b11645ab2dba +size 24502272 diff --git a/data/swebenchpro/anthropic/claude-opus-4-5-20251101/b81fab29-13f7-429c-8b79-299ad8608b68.json b/data/swebenchpro/anthropic/claude-opus-4-5-20251101/b81fab29-13f7-429c-8b79-299ad8608b68.json new file mode 100644 index 0000000000000000000000000000000000000000..ac8c88a573dd6be1579a6d1b7264f4ec2bb7322f --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-5-20251101/b81fab29-13f7-429c-8b79-299ad8608b68.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/anthropic_claude-opus-4-5-20251101/1786169195.83336", + "evaluation_timestamp": "1776893158.0", + "retrieved_timestamp": "1786169195.83336", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-5-20251101", + "id": "anthropic/claude-opus-4-5-20251101", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/anthropic/claude-opus-4-5-20251101/b81fab29-13f7-429c-8b79-299ad8608b68_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "ab5f5ab1fe2b8483be522d87b442d3d8fa653af04b805ff35c3e714568e2f2fa", + "total_rows": 26 + } +} \ No newline at end of file diff --git a/data/swebenchpro/anthropic/claude-opus-4-5-20251101/b81fab29-13f7-429c-8b79-299ad8608b68_samples.jsonl b/data/swebenchpro/anthropic/claude-opus-4-5-20251101/b81fab29-13f7-429c-8b79-299ad8608b68_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..0318f306b6d3618355b9770734cfff33c3c04f2d --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-5-20251101/b81fab29-13f7-429c-8b79-299ad8608b68_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ab5f5ab1fe2b8483be522d87b442d3d8fa653af04b805ff35c3e714568e2f2fa +size 26838317 diff --git a/data/swebenchpro/anthropic/claude-opus-4-5-20251101/b9d27d09-7736-49b7-bf86-ad8ddd2f05e5.json b/data/swebenchpro/anthropic/claude-opus-4-5-20251101/b9d27d09-7736-49b7-bf86-ad8ddd2f05e5.json new file mode 100644 index 0000000000000000000000000000000000000000..a559a2933168936ec424f204bd245ef24782d98f --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-5-20251101/b9d27d09-7736-49b7-bf86-ad8ddd2f05e5.json @@ -0,0 +1,121 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/anthropic_claude-opus-4-5-20251101/1786163079.975465", + "evaluation_timestamp": "1777025747.0", + "retrieved_timestamp": "1786163079.975465", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-5-20251101", + "id": "anthropic/claude-opus-4-5-20251101", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on swebenchpro/S-adaptive+C/+4ep/e7811166 for scorer harbor_scorer", + "source_data": { + "dataset_name": "swebenchpro", + "source_type": "hf_dataset", + "samples_number": 11, + "sample_ids": [ + "instance_qutebrowser__qutebrowser-2e961080a85d660148937ee8f0f6b3445a8f2c01-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_flipt-io__flipt-b68b8960b8a08540d5198d78c665a7eb0bea4008", + "instance_future-architect__vuls-78b52d6a7f480bd610b692de9bf0c86f57332f23", + "instance_flipt-io__flipt-e42da21a07a5ae35835ec54f74004ebd58713874", + "instance_nodebb__nodebb-84e065752f6d7fbe5c08cbf50cb173ffb866b8fa-vf2cf3cbd463b7ad942381f1c6d077626485a1e9e", + "instance_element-hq__element-web-f3534b42df3dcfe36dc48bddbf14034085af6d30-vnan", + "instance_navidrome__navidrome-d8e794317f788198227e10fb667e10496b3eb99a", + "instance_future-architect__vuls-6eff6a9329a65cc412e79b8f82444dfa3d0f0b5a", + "instance_internetarchive__openlibrary-fdbc0d8f418333c7e575c40b661b582c301ef7ac-v13642507b4fc1f8d234172bf8129942da2c2ca26", + "instance_ansible__ansible-b2a289dcbb702003377221e25f62c8a3608f0e89-v173091e2e36d38c978002990795f66cfc0af30ad", + "instance_qutebrowser__qutebrowser-e70f5b03187bdd40e8bf70f5f3ead840f52d1f42-v02ad04386d5238fe2d1a1be450df257370de4b6a" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1777025747.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.5454545454545454, + "uncertainty": { + "standard_error": { + "value": 0.1574591643244434 + }, + "num_samples": 44 + }, + "details": { + "total_matched_trajectories": "44", + "stopping_reason_count_completed_on_successful_submit": "24", + "stopping_reason_count_repetition_guard": "20", + "avg_total_tokens_target_model": "5767140.25", + "total_total_tokens_target_model": "253754171.0", + "avg_total_tokens_other_models": "0.00", + "total_total_tokens_other_models": "0.0", + "avg_total_tokens_all_models": "5767140.25", + "total_total_tokens_all_models": "253754171.0", + "avg_turn_count": "102.93", + "total_turn_count": "4529.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 30000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/anthropic/claude-opus-4-5-20251101/b9d27d09-7736-49b7-bf86-ad8ddd2f05e5_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "815feed9044fb20d173af36df7cc01060fa56273ddb9713a4d4103f1d44e177c", + "total_rows": 44 + } +} \ No newline at end of file diff --git a/data/swebenchpro/anthropic/claude-opus-4-5-20251101/b9d27d09-7736-49b7-bf86-ad8ddd2f05e5_samples.jsonl b/data/swebenchpro/anthropic/claude-opus-4-5-20251101/b9d27d09-7736-49b7-bf86-ad8ddd2f05e5_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..da450ecbf50a8d9c1f2bca51e75c2366d56986b3 --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-5-20251101/b9d27d09-7736-49b7-bf86-ad8ddd2f05e5_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:815feed9044fb20d173af36df7cc01060fa56273ddb9713a4d4103f1d44e177c +size 36858027 diff --git a/data/swebenchpro/anthropic/claude-opus-4-5-20251101/bf94b048-27ce-4e0a-b2a6-1898af7b8546.json b/data/swebenchpro/anthropic/claude-opus-4-5-20251101/bf94b048-27ce-4e0a-b2a6-1898af7b8546.json new file mode 100644 index 0000000000000000000000000000000000000000..49acdda78178abec1657e7e2c3e1642ac3b86914 --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-5-20251101/bf94b048-27ce-4e0a-b2a6-1898af7b8546.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/anthropic_claude-opus-4-5-20251101/1786167398.281914", + "evaluation_timestamp": "1776918738.0", + "retrieved_timestamp": "1786167398.281914", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-5-20251101", + "id": "anthropic/claude-opus-4-5-20251101", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/anthropic/claude-opus-4-5-20251101/bf94b048-27ce-4e0a-b2a6-1898af7b8546_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "8d2cd60734e3417a75496b80428f4a2cbd588289b69dbad1967fccabf3dbe2ca", + "total_rows": 11 + } +} \ No newline at end of file diff --git a/data/swebenchpro/anthropic/claude-opus-4-5-20251101/bf94b048-27ce-4e0a-b2a6-1898af7b8546_samples.jsonl b/data/swebenchpro/anthropic/claude-opus-4-5-20251101/bf94b048-27ce-4e0a-b2a6-1898af7b8546_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..dde16f10802c4b5b918171430f51a91b7c891d60 --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-5-20251101/bf94b048-27ce-4e0a-b2a6-1898af7b8546_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8d2cd60734e3417a75496b80428f4a2cbd588289b69dbad1967fccabf3dbe2ca +size 9420890 diff --git a/data/swebenchpro/anthropic/claude-opus-4-5-20251101/c3b703cb-b218-4c6f-b95e-f79bcca4b55e.json b/data/swebenchpro/anthropic/claude-opus-4-5-20251101/c3b703cb-b218-4c6f-b95e-f79bcca4b55e.json new file mode 100644 index 0000000000000000000000000000000000000000..5467dea4c2274bd09785025dcf851f5099dda2c7 --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-5-20251101/c3b703cb-b218-4c6f-b95e-f79bcca4b55e.json @@ -0,0 +1,291 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/anthropic_claude-opus-4-5-20251101/1786182846.707106", + "evaluation_timestamp": "1775602861.0", + "retrieved_timestamp": "1786182846.707106", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.196.dev25+gb973b316,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-5-20251101", + "id": "anthropic/claude-opus-4-5-20251101", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on eval_task for scorer harbor_scorer", + "source_data": { + "dataset_name": "swebenchpro", + "source_type": "hf_dataset", + "samples_number": 731, + "sample_ids": [ + "instance_flipt-io__flipt-e91615cf07966da41756017a7d571f9fc0fdbe80", + "instance_ansible__ansible-8127abbc298cabf04aaa89a478fc5e5e3432a6fc-v30a923fb5c164d6cd18280c02422f75e611e8fb2", + "instance_protonmail__webclients-da91f084c0f532d9cc8ca385a701274d598057b8", + "instance_qutebrowser__qutebrowser-deeb15d6f009b3ca0c3bd503a7cef07462bd16b4-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_internetarchive__openlibrary-5069b09e5f64428dce59b33455c8bb17fe577070-v8717e18970bcdc4e0d2cea3b1527752b21e74866", + "instance_protonmail__webclients-c5a2089ca2bfe9aa1d85a664b8ad87ef843a1c9c", + "instance_qutebrowser__qutebrowser-fd6790fe8c02b144ab2464f1fc8ab3d02ce3c476-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_navidrome__navidrome-eebfbc5381a1e506ff17b5f1371d1ad83d5fd642", + "instance_navidrome__navidrome-812dc2090f20ac4f8ac271b6ed95be5889d1a3ca", + "instance_nodebb__nodebb-0c81642997ea1d827dbd02c311db9d4976112cd4-vf2cf3cbd463b7ad942381f1c6d077626485a1e9e", + "instance_internetarchive__openlibrary-c506c1b0b678892af5cb22c1c1dbc35d96787a0a-v0f5aece3601a5b4419f7ccec1dbda2071be28ee4", + "instance_flipt-io__flipt-5c7037ececb0bead0a8eb56054e224bcd7ac5922", + "instance_qutebrowser__qutebrowser-01d1d1494411380d97cac14614a829d3a69cecaf-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_internetarchive__openlibrary-7f7e53aa4cf74a4f8549a5bcd4810c527e2f6d7e-v13642507b4fc1f8d234172bf8129942da2c2ca26", + "instance_internetarchive__openlibrary-53e02a22972e9253aeded0e1981e6845e1e521fe-vfa6ff903cb27f336e17654595dd900fa943dcd91", + "instance_ansible__ansible-34db57a47f875d11c4068567b9ec7ace174ec4cf-v1055803c3a812189a1133297f7f5468579283f86", + "instance_qutebrowser__qutebrowser-996487c43e4fcc265b541f9eca1e7930e3c5cf05-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_gravitational__teleport-96019ce0be7a2c8e36363f359eb7c943b41dde70", + "instance_protonmail__webclients-df60460f163fd5c34e844ab9015e3176f1ab1ac0", + "instance_future-architect__vuls-ad2edbb8448e2c41a097f1c0b52696c0f6c5924d", + "instance_ansible__ansible-9142be2f6cabbe6597c9254c5bb9186d17036d55-v0f01c69f1e2528b935359cfe578530722bca2c59", + "instance_element-hq__element-web-33299af5c9b7a7ec5a9c31d578d4ec5b18088fb7-vnan", + "instance_element-hq__element-web-772df3021201d9c73835a626df8dcb6334ad9a3e-vnan", + "instance_internetarchive__openlibrary-1be7de788a444f6255e89c10ef6aa608550604a8-v29f82c9cf21d57b242f8d8b0e541525d259e2d63", + "instance_internetarchive__openlibrary-630221ab686c64e75a2ce253c893c033e4814b2e-v93c53c13d5f9b383ebb411ee7750b49dcd1a34c6", + "instance_gravitational__teleport-f432a71a13e698b6e1c4672a2e9e9c1f32d35c12", + "instance_protonmail__webclients-0d0267c4438cf378bda90bc85eed3a3615871ac4", + "instance_internetarchive__openlibrary-dbbd9d539c6d4fd45d5be9662aa19b6d664b5137-v08d8e8889ec945ab821fb156c04c7d2e2810debb", + "instance_flipt-io__flipt-b2170346dc37cf42fda1386cd630f24821ad2ac5", + "instance_element-hq__element-web-aec454dd6feeb93000380523cbb0b3681c0275fd-vnan", + "instance_internetarchive__openlibrary-4a5d2a7d24c9e4c11d3069220c0685b736d5ecde-v13642507b4fc1f8d234172bf8129942da2c2ca26", + "instance_qutebrowser__qutebrowser-ec2dcfce9eee9f808efc17a1b99e227fc4421dea-v5149fcda2a9a6fe1d35dfed1bade1444a11ef271", + "instance_gravitational__teleport-1330415d33a27594c948a36d9d7701f496229e9f", + "instance_protonmail__webclients-3a6790f480309130b5d6332dce6c9d5ccca13ee3", + "instance_flipt-io__flipt-492cc0b158200089dceede3b1aba0ed28df3fb1d", + "instance_future-architect__vuls-edb324c3d9ec3b107bf947f00e38af99d05b3e16", + "instance_ansible__ansible-bec27fb4c0a40c5f8bbcf26a475704227d65ee73-v30a923fb5c164d6cd18280c02422f75e611e8fb2", + "instance_gravitational__teleport-c1b1c6a1541c478d7777a48fca993cc8206c73b9", + "instance_tutao__tutanota-1e516e989b3c0221f4af6b297d9c0e4c43e4adc3-vbc0d9ba8f0071fbe982809910959a6ff8884dbbf", + "instance_flipt-io__flipt-406f9396ad65696d58865b3a6283109cd4eaf40e", + "instance_future-architect__vuls-139f3a81b66c47e6d8f70ce6c4afe7a9196a6ea8", + "instance_flipt-io__flipt-3b2c25ee8a3ac247c3fad13ad8d64ace34ec8ee7", + "instance_gravitational__teleport-0ac7334939981cf85b9591ac295c3816954e287e", + "instance_qutebrowser__qutebrowser-e15d26630934d0b6415ed2295ac42fd570a57620-va0fd88aac89cde702ec1ba84877234da33adce8a", + "instance_ansible__ansible-5e88cd9972f10b66dd97e1ee684c910c6a2dd25e-v906c969b551b346ef54a2c0b41e04f632b7b73c2", + "instance_element-hq__element-web-923ad4323b2006b2b180544429455ffe7d4a6cc3-vnan", + "instance_protonmail__webclients-4feccbc9990980aee26ea29035f8f931d6089895", + "instance_flipt-io__flipt-a0cbc0cb65ae601270bdbe3f5313e2dfd49c80e4", + "instance_gravitational__teleport-b4e7cd3a5e246736d3fe8d6886af55030b232277", + "instance_ansible__ansible-6cc97447aac5816745278f3735af128afb255c81-v0f01c69f1e2528b935359cfe578530722bca2c59", + "instance_ansible__ansible-f327e65d11bb905ed9f15996024f857a95592629-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_qutebrowser__qutebrowser-5fdc83e5da6222fe61163395baaad7ae57fa2cb4-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_navidrome__navidrome-3982ba725883e71d4e3e618c61d5140eeb8d850a", + "instance_qutebrowser__qutebrowser-ebfe9b7aa0c4ba9d451f993e08955004aaec4345-v059c6fdc75567943479b23ebca7c07b5e9a7f34c", + "instance_ansible__ansible-709484969c8a4ffd74b839a673431a8c5caa6457-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_gravitational__teleport-ad41b3c15414b28a6cec8c25424a19bfa7abd0e9-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_protonmail__webclients-2c3559cad02d1090985dba7e8eb5a129144d9811", + "instance_flipt-io__flipt-5af0757e96dec4962a076376d1bedc79de0d4249", + "instance_element-hq__element-web-56c7fc1948923b4b3f3507799e725ac16bcf8018-vnan", + "instance_flipt-io__flipt-381b90f718435c4694380b5fcd0d5cf8e3b5a25a", + "instance_flipt-io__flipt-6fd0f9e2587f14ac1fdd1c229f0bcae0468c8daa", + "instance_ansible__ansible-4c5ce5a1a9e79a845aff4978cfeb72a0d4ecf7d6-v1055803c3a812189a1133297f7f5468579283f86", + "instance_navidrome__navidrome-7073d18b54da7e53274d11c9e2baef1242e8769e", + "instance_internetarchive__openlibrary-0a90f9f0256e4f933523e9842799e39f95ae29ce-v76304ecdb3a5954fcf13feb710e8c40fcf24b73c", + "instance_nodebb__nodebb-be43cd25974681c9743d424238b7536c357dc8d3-vf2cf3cbd463b7ad942381f1c6d077626485a1e9e", + "instance_flipt-io__flipt-05d7234fa582df632f70a7cd10194d61bd7043b9", + "instance_nodebb__nodebb-76c6e30282906ac664f2c9278fc90999b27b1f48-vd59a5728dfc977f44533186ace531248c2917516", + "instance_qutebrowser__qutebrowser-6b320dc18662580e1313d2548fdd6231d2a97e6d-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_future-architect__vuls-01441351c3407abfc21c48a38e28828e1b504e0c", + "instance_ansible__ansible-1a4644ff15355fd696ac5b9d074a566a80fe7ca3-v30a923fb5c164d6cd18280c02422f75e611e8fb2", + "instance_navidrome__navidrome-09ae41a2da66264c60ef307882362d2e2d8d8b89", + "instance_internetarchive__openlibrary-5c6c22f3d2edf2f1b10f5dc335e32cb6a5f40341-v76304ecdb3a5954fcf13feb710e8c40fcf24b73c", + "instance_navidrome__navidrome-bf2bcb12799b21069f137749e0c331f761d1f693", + "instance_flipt-io__flipt-2ce8a0331e8a8f63f2c1b555db8277ffe5aa2e63", + "instance_ansible__ansible-a1569ea4ca6af5480cf0b7b3135f5e12add28a44-v0f01c69f1e2528b935359cfe578530722bca2c59", + "instance_gravitational__teleport-82185f232ae8974258397e121b3bc2ed0c3729ed-v626ec2a48416b10a88641359a169d99e935ff037", + "instance_flipt-io__flipt-72d06db14d58692bfb4d07b1aa745a37b35956f3", + "instance_ansible__ansible-b6290e1d156af608bd79118d209a64a051c55001-v390e508d27db7a51eece36bb6d9698b63a5b638a", + "instance_gravitational__teleport-32bcd71591c234f0d8b091ec01f1f5cbfdc0f13c-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_ansible__ansible-a02e22e902a69aeb465f16bf03f7f5a91b2cb828-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_gravitational__teleport-eda668c30d9d3b56d9c69197b120b01013611186", + "instance_nodebb__nodebb-0f788b8eaa4bba3c142d171fd941d015c53b65fc-v0ec6d6c2baf3cb4797482ce4829bc25cd5716649", + "instance_gravitational__teleport-47530e1fd8bfb84ec096ebcbbc29990f30829655-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_qutebrowser__qutebrowser-e64622cd2df5b521342cf4a62e0d4cb8f8c9ae5a-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_element-hq__element-web-494d9de6f0a94ffb491e74744d2735bce02dc0ab-vnan", + "instance_element-hq__element-web-9bf77963ee5e036d54b2a3ca202fbf6378464a5e-vnan", + "instance_nodebb__nodebb-9c576a0758690f45a6ca03b5884c601e473bf2c1-vd59a5728dfc977f44533186ace531248c2917516", + "instance_internetarchive__openlibrary-25858f9f0c165df25742acf8309ce909773f0cdd-v13642507b4fc1f8d234172bf8129942da2c2ca26", + "instance_protonmail__webclients-f161c10cf7d31abf82e8d64d7a99c9fac5acfa18", + "instance_gravitational__teleport-cb712e3f0b06dadc679f895daef8072cae400c26-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_gravitational__teleport-46a13210519461c7cec8d643bfbe750265775b41", + "instance_ansible__ansible-e40889e7112ae00a21a2c74312b330e67a766cc0-v1055803c3a812189a1133297f7f5468579283f86", + "instance_gravitational__teleport-6eaaf3a27e64f4ef4ef855bd35d7ec338cf17460-v626ec2a48416b10a88641359a169d99e935ff037", + "instance_internetarchive__openlibrary-0d13e6b4bf80bced6c0946b969b9a1b6963f6bce-v76304ecdb3a5954fcf13feb710e8c40fcf24b73c", + "instance_protonmail__webclients-2f2f6c311c6128fe86976950d3c0c2db07b03921", + "instance_ansible__ansible-5260527c4a71bfed99d803e687dd19619423b134-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_qutebrowser__qutebrowser-8d05f0282a271bfd45e614238bd1b555c58b3fc1-v35616345bb8052ea303186706cec663146f0f184", + "instance_qutebrowser__qutebrowser-50efac08f623644a85441bbe02ab9347d2b71a9d-v9f8e9d96c85c85a605e382f1510bd08563afc566", + "instance_flipt-io__flipt-dae029cba7cdb98dfb1a6b416c00d324241e6063", + "instance_nodebb__nodebb-18c45b44613aecd53e9f60457b9812049ab2998d-v0495b863a912fbff5749c67e860612b91825407c", + "instance_future-architect__vuls-e4728e388120b311c4ed469e4f942e0347a2689b-v264a82e2f4818e30f5a25e4da53b27ba119f62b5", + "instance_qutebrowser__qutebrowser-5e0d6dc1483cb3336ea0e3dcbd4fe4aa00fc1742-v5149fcda2a9a6fe1d35dfed1bade1444a11ef271", + "instance_flipt-io__flipt-c12967bc73fdf02054cf3ef8498c05e25f0a18c0", + "instance_qutebrowser__qutebrowser-ff1c025ad3210506fc76e1f604d8c8c27637d88e-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_ansible__ansible-83909bfa22573777e3db5688773bda59721962ad-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_qutebrowser__qutebrowser-1a9e74bfaf9a9db2a510dc14572d33ded6040a57-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_gravitational__teleport-24cafecd8721891092210afc55f6413ab46ca211-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_internetarchive__openlibrary-4b7ea2977be2747496ba792a678940baa985f7ea-v0f5aece3601a5b4419f7ccec1dbda2071be28ee4", + "instance_internetarchive__openlibrary-fad4a40acf5ff5f06cd7441a5c7baf41a7d81fe4-vfa6ff903cb27f336e17654595dd900fa943dcd91", + "instance_qutebrowser__qutebrowser-3fd8e12949b8feda401930574facf09dd4180bba", + "instance_ansible__ansible-164881d871964aa64e0f911d03ae270acbad253c-v390e508d27db7a51eece36bb6d9698b63a5b638a", + "instance_ansible__ansible-fb144c44144f8bd3542e71f5db62b6d322c7bd85-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_element-hq__element-web-582a1b093fc0b77538052f45cbb9c7295f991b51-vnan", + "instance_element-hq__element-web-ec0f940ef0e8e3b61078f145f34dc40d1938e6c5-vnan", + "instance_internetarchive__openlibrary-6fdbbeee4c0a7e976ff3e46fb1d36f4eb110c428-v08d8e8889ec945ab821fb156c04c7d2e2810debb", + "instance_flipt-io__flipt-7161f7b876773a911afdd804b281e52681cb7321", + "instance_navidrome__navidrome-28389fb05e1523564dfc61fa43ed8eb8a10f938c", + "instance_ansible__ansible-5f4e332e3762999d94af27746db29ff1729252c1-v0f01c69f1e2528b935359cfe578530722bca2c59", + "instance_nodebb__nodebb-0e07f3c9bace416cbab078a30eae972868c0a8a3-vf2cf3cbd463b7ad942381f1c6d077626485a1e9e", + "instance_nodebb__nodebb-da0211b1a001d45d73b4c84c6417a4f1b0312575-vf2cf3cbd463b7ad942381f1c6d077626485a1e9e", + "instance_navidrome__navidrome-31799662706fedddf5bcc1a76b50409d1f91d327", + "instance_future-architect__vuls-50580f6e98eeb36f53f27222f7f4fdfea0b21e8d", + "instance_internetarchive__openlibrary-60725705782832a2cb22e17c49697948a42a9d03-v298a7a812ceed28c4c18355a091f1b268fe56d86", + "instance_ansible__ansible-3db08adbb1cc6aa9941be5e0fc810132c6e1fa4b-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_internetarchive__openlibrary-8a5a63af6e0be406aa6c8c9b6d5f28b2f1b6af5a-v0f5aece3601a5b4419f7ccec1dbda2071be28ee4", + "instance_future-architect__vuls-e1df74cbc1a1d1889428b3333a3b2405c4651993", + "instance_internetarchive__openlibrary-123e6e5e1c85b9c07d1e98f70bfc480bc8016890-v2733ff199fb72f0d033a30dc62cb0a4742e3a7f4", + "instance_protonmail__webclients-ae36cb23a1682dcfd69587c1b311ae0227e28f39", + "instance_protonmail__webclients-8be4f6cb9380fcd2e67bcb18cef931ae0d4b869c", + "instance_qutebrowser__qutebrowser-a25e8a09873838ca9efefd36ea8a45170bbeb95c-vc2f56a753b62a190ddb23cd330c257b9cf560d12", + "instance_qutebrowser__qutebrowser-44e64199ed38003253f0296badd4a447645067b6-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_gravitational__teleport-7744f72c6eb631791434b648ba41083b5f6d2278-vce94f93ad1030e3136852817f2423c1b3ac37bc4", + "instance_tutao__tutanota-40e94dee2bcec2b63f362da283123e9df1874cc1-vc4e41fd0029957297843cb9dec4a25c7c756f029", + "instance_future-architect__vuls-c11ba27509f733d7d280bdf661cbbe2e7a99df4c", + "instance_flipt-io__flipt-b2cd6a6dd73ca91b519015fd5924fde8d17f3f06", + "instance_gravitational__teleport-6a14edcf1ff010172fdbac622d0a474ed6af46de", + "instance_flipt-io__flipt-36e62baffae2132f78f9d34dc300a9baa2d7ae0e", + "instance_navidrome__navidrome-c90468b895f6171e33e937ff20dc915c995274f0", + "instance_ansible__ansible-ecea15c508f0e081525be036cf76bbb56dbcdd9d-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_ansible__ansible-e22e103cdf8edc56ff7d9b848a58f94f1471a263-v1055803c3a812189a1133297f7f5468579283f86", + "instance_future-architect__vuls-dc496468b9e9fb73371f9606cdcdb0f8e12e70ca", + "instance_gravitational__teleport-c782838c3a174fdff80cafd8cd3b1aa4dae8beb2", + "instance_internetarchive__openlibrary-91efee627df01e32007abf2d6ebf73f9d9053076-vbee42ad1b72fb23c6a1c874868a720b370983ed2", + "instance_qutebrowser__qutebrowser-52708364b5f91e198defb022d1a5b4b3ebd9b563-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_ansible__ansible-eea46a0d1b99a6dadedbb6a3502d599235fa7ec3-v390e508d27db7a51eece36bb6d9698b63a5b638a", + "instance_protonmail__webclients-c6f65d205c401350a226bb005f42fac1754b0b5b", + "instance_ansible__ansible-cb94c0cc550df9e98f1247bc71d8c2b861c75049-v1055803c3a812189a1133297f7f5468579283f86", + "instance_nodebb__nodebb-05f2236193f407cf8e2072757fbd6bb170bc13f0-vf2cf3cbd463b7ad942381f1c6d077626485a1e9e", + "instance_navidrome__navidrome-55730514ea59d5f1d0b8e3f8745569c29bdbf7b4", + "instance_navidrome__navidrome-0130c6dc13438b48cf0fdfab08a89e357b5517c9", + "instance_flipt-io__flipt-507170da0f7f4da330f6732bffdf11c4df7fc192", + "instance_tutao__tutanota-fe240cbf7f0fdd6744ef7bef8cb61676bcdbb621-vc4e41fd0029957297843cb9dec4a25c7c756f029", + "instance_qutebrowser__qutebrowser-233cb1cc48635130e5602549856a6fa4ab4c087f-v35616345bb8052ea303186706cec663146f0f184", + "instance_internetarchive__openlibrary-f8cc11d9c1575fdba5ac66aee0befca970da8d64-v13642507b4fc1f8d234172bf8129942da2c2ca26", + "instance_qutebrowser__qutebrowser-fec187c2cb53d769c2682b35ca77858a811414a8-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_tutao__tutanota-da4edb7375c10f47f4ed3860a591c5e6557f7b5c-vbc0d9ba8f0071fbe982809910959a6ff8884dbbf", + "instance_ansible__ansible-be2c376ab87e3e872ca21697508f12c6909cf85a-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_future-architect__vuls-457a3a9627fb9a0800d0aecf1d4713fb634a9011", + "instance_internetarchive__openlibrary-d40ec88713dc95ea791b252f92d2f7b75e107440-v13642507b4fc1f8d234172bf8129942da2c2ca26", + "instance_protonmail__webclients-863d524b5717b9d33ce08a0f0535e3fd8e8d1ed8", + "instance_internetarchive__openlibrary-308a35d6999427c02b1dbf5211c033ad3b352556-ve8c8d62a2b60610a3c4631f5f23ed866bada9818", + "instance_flipt-io__flipt-15b76cada1ef29cfa56b0fba36754be36243dded", + "instance_future-architect__vuls-61c39637f2f3809e1b5dad05f0c57c799dce1587", + "instance_ansible__ansible-5d253a13807e884b7ce0b6b57a963a45e2f0322c-v1055803c3a812189a1133297f7f5468579283f86", + "instance_element-hq__element-web-44b98896a79ede48f5ad7ff22619a39d5f6ff03c-vnan", + "instance_qutebrowser__qutebrowser-f91ace96223cac8161c16dd061907e138fe85111-v059c6fdc75567943479b23ebca7c07b5e9a7f34c", + "instance_gravitational__teleport-bb562408da4adeae16e025be65e170959d1ec492-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_gravitational__teleport-78b0d8c72637df1129fb6ff84fc49ef4b5ab1288", + "instance_internetarchive__openlibrary-3f7db6bbbcc7c418b3db72d157c6aed1d45b2ccf-v430f20c722405e462d9ef44dee7d34c41e76fe7a", + "instance_future-architect__vuls-e52fa8d6ed1d23e36f2a86e5d3efe9aa057a1b0d", + "instance_navidrome__navidrome-3bc9e75b2843f91f6a1e9b604e321c2bd4fd442a", + "instance_tutao__tutanota-fb32e5f9d9fc152a00144d56dd0af01760a2d4dc-vc4e41fd0029957297843cb9dec4a25c7c756f029", + "instance_gravitational__teleport-af5e2517de7d18406b614e413aca61c319312171-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_ansible__ansible-a20a52701402a12f91396549df04ac55809f68e9-v1055803c3a812189a1133297f7f5468579283f86", + "instance_protonmail__webclients-1501eb765873b2884b6f1944fd242ecfc9d6b103", + "instance_gravitational__teleport-db89206db6c2969266e664c7c0fb51b70e958b64", + "instance_ansible__ansible-be59caa59bf47ca78a4760eb7ff38568372a8260-v1055803c3a812189a1133297f7f5468579283f86", + "instance_flipt-io__flipt-65581fef4aa807540cb933753d085feb0d7e736f", + "instance_gravitational__teleport-4e1c39639edf1ab494dd7562844c8b277b5cfa18-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_future-architect__vuls-f0b3a8b1db98eb1bd32685f1c36c41a99c3452ed", + "instance_navidrome__navidrome-b65e76293a917ee2dfc5d4b373b1c62e054d0dca", + "instance_element-hq__element-web-33e8edb3d508d6eefb354819ca693b7accc695e7", + "instance_flipt-io__flipt-e50808c03e4b9d25a6a78af9c61a3b1616ea356b", + "instance_future-architect__vuls-e3c27e1817d68248043bd09d63cc31f3344a6f2c", + "instance_element-hq__element-web-ad26925bb6628260cfe0fcf90ec0a8cba381f4a4-vnan", + "instance_nodebb__nodebb-1ea9481af6125ffd6da0592ed439aa62af0bca11-vd59a5728dfc977f44533186ace531248c2917516", + "instance_ansible__ansible-489156378c8e97374a75a544c7c9c2c0dd8146d1-v390e508d27db7a51eece36bb6d9698b63a5b638a", + "instance_internetarchive__openlibrary-bb152d23c004f3d68986877143bb0f83531fe401-ve8c8d62a2b60610a3c4631f5f23ed866bada9818", + "instance_qutebrowser__qutebrowser-0b621cb0ce2b54d3f93d8d41d8ff4257888a87e5-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_qutebrowser__qutebrowser-1af602b258b97aaba69d2585ed499d95e2303ac2-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_future-architect__vuls-e1fab805afcfc92a2a615371d0ec1e667503c254-v264a82e2f4818e30f5a25e4da53b27ba119f62b5", + "instance_internetarchive__openlibrary-7edd1ef09d91fe0b435707633c5cc9af41dedddf-v76304ecdb3a5954fcf13feb710e8c40fcf24b73c", + "instance_qutebrowser__qutebrowser-99029144b5109bb1b2a53964a7c129e009980cd9-va0fd88aac89cde702ec1ba84877234da33adce8a", + "instance_ansible__ansible-12734fa21c08a0ce8c84e533abdc560db2eb1955-v7eee2454f617569fd6889f2211f75bc02a35f9f8", + "instance_protonmail__webclients-7e54526774e577c0ebb58ced7ba8bef349a69fec", + "instance_gravitational__teleport-1a77b7945a022ab86858029d30ac7ad0d5239d00-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_nodebb__nodebb-8ca65b0c78c67c1653487c02d1135e1b702185e1-vf2cf3cbd463b7ad942381f1c6d077626485a1e9e", + "instance_flipt-io__flipt-9f8127f225a86245fa35dca4885c2daef824ee55", + "instance_gravitational__teleport-ac2fb2f9b4fd1896b554d3011df23d3d71295779", + "instance_ansible__ansible-f02a62db509dc7463fab642c9c3458b9bc3476cc-v390e508d27db7a51eece36bb6d9698b63a5b638a" + ], + "additional_details": { + "shuffled": "True" + } + }, + "evaluation_timestamp": "1775602861.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.5, + "uncertainty": { + "standard_error": { + "value": 0.0354440602504168 + }, + "num_samples": 200 + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"If you do not submit an answer, you will receive a score of zero.\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [10, \"Your answer has been saved. Consider whether your previous approach is likely correct or whether a different approach would be more promising. Then either refine your previous answer or try a substantially different approach, and submit when ready.\", \"\"], \"submit\": null, \"on_continue\": null, \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\"}, \"params_passed\": {\"prompt\": [\"If you do not submit an answer, you will receive a score of zero.\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [10, \"Your answer has been saved. Consider whether your previous approach is likely correct or whether a different approach would be more promising. Then either refine your previous answer or try a substantially different approach, and submit when ready.\", \"\"], \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "reasoning_effort": "\"high\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 20000000 + }, + "sandbox": { + "type": "k8s" + } + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/anthropic/claude-opus-4-5-20251101/c3b703cb-b218-4c6f-b95e-f79bcca4b55e_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "4ef81b38402f4bdb7ad0e617c2f7d71587eff302091b50522c1cd14b8b7f26c0", + "total_rows": 200 + } +} \ No newline at end of file diff --git a/data/swebenchpro/anthropic/claude-opus-4-5-20251101/c3b703cb-b218-4c6f-b95e-f79bcca4b55e_samples.jsonl b/data/swebenchpro/anthropic/claude-opus-4-5-20251101/c3b703cb-b218-4c6f-b95e-f79bcca4b55e_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..ae819d3f0388469fd2c503fc860630a530f6e040 --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-5-20251101/c3b703cb-b218-4c6f-b95e-f79bcca4b55e_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4ef81b38402f4bdb7ad0e617c2f7d71587eff302091b50522c1cd14b8b7f26c0 +size 90628427 diff --git a/data/swebenchpro/anthropic/claude-opus-4-5-20251101/d4329097-d5f8-4481-8db3-2c9dfe382156.json b/data/swebenchpro/anthropic/claude-opus-4-5-20251101/d4329097-d5f8-4481-8db3-2c9dfe382156.json new file mode 100644 index 0000000000000000000000000000000000000000..0b09eabbc33c98165e1909975950b0c13787ef63 --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-5-20251101/d4329097-d5f8-4481-8db3-2c9dfe382156.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/anthropic_claude-opus-4-5-20251101/1786169441.31341", + "evaluation_timestamp": "1777292858.0", + "retrieved_timestamp": "1786169441.31341", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-5-20251101", + "id": "anthropic/claude-opus-4-5-20251101", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/anthropic/claude-opus-4-5-20251101/d4329097-d5f8-4481-8db3-2c9dfe382156_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "4f369751d9f475b853d20c2266936cacbcadb981973555b39d449cc21a2b15e7", + "total_rows": 79 + } +} \ No newline at end of file diff --git a/data/swebenchpro/anthropic/claude-opus-4-5-20251101/d4329097-d5f8-4481-8db3-2c9dfe382156_samples.jsonl b/data/swebenchpro/anthropic/claude-opus-4-5-20251101/d4329097-d5f8-4481-8db3-2c9dfe382156_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..edc5f9a332deafd87c0f46009608e208ad34ad70 --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-5-20251101/d4329097-d5f8-4481-8db3-2c9dfe382156_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f369751d9f475b853d20c2266936cacbcadb981973555b39d449cc21a2b15e7 +size 96016753 diff --git a/data/swebenchpro/anthropic/claude-opus-4-5-20251101/dae4d305-f0cc-4762-85b2-c92668e40d4c.json b/data/swebenchpro/anthropic/claude-opus-4-5-20251101/dae4d305-f0cc-4762-85b2-c92668e40d4c.json new file mode 100644 index 0000000000000000000000000000000000000000..387170e3b1845a039dc8144f696d8d535423fc39 --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-5-20251101/dae4d305-f0cc-4762-85b2-c92668e40d4c.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/anthropic_claude-opus-4-5-20251101/1786169269.538694", + "evaluation_timestamp": "1777369892.0", + "retrieved_timestamp": "1786169269.538694", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-5-20251101", + "id": "anthropic/claude-opus-4-5-20251101", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/anthropic/claude-opus-4-5-20251101/dae4d305-f0cc-4762-85b2-c92668e40d4c_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "4891f7faad25c6054b3219546066fee22c953c86d31738b6b6ff8c6b6bbb446e", + "total_rows": 58 + } +} \ No newline at end of file diff --git a/data/swebenchpro/anthropic/claude-opus-4-5-20251101/dae4d305-f0cc-4762-85b2-c92668e40d4c_samples.jsonl b/data/swebenchpro/anthropic/claude-opus-4-5-20251101/dae4d305-f0cc-4762-85b2-c92668e40d4c_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..1bb00e3ad145273b8f68ee21faf4c914162bed26 --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-5-20251101/dae4d305-f0cc-4762-85b2-c92668e40d4c_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4891f7faad25c6054b3219546066fee22c953c86d31738b6b6ff8c6b6bbb446e +size 59162273 diff --git a/data/swebenchpro/anthropic/claude-opus-4-5-20251101/e6e37cd8-011e-4402-a0f7-248411195a4c.json b/data/swebenchpro/anthropic/claude-opus-4-5-20251101/e6e37cd8-011e-4402-a0f7-248411195a4c.json new file mode 100644 index 0000000000000000000000000000000000000000..b5f0a18085a56514a7e3806c2f51c3f02000f088 --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-5-20251101/e6e37cd8-011e-4402-a0f7-248411195a4c.json @@ -0,0 +1,136 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/anthropic_claude-opus-4-5-20251101/1786165607.650335", + "evaluation_timestamp": "1777025747.0", + "retrieved_timestamp": "1786165607.650335", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-5-20251101", + "id": "anthropic/claude-opus-4-5-20251101", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on swebenchpro/S-adaptive/+4ep/91322b10 for scorer harbor_scorer", + "source_data": { + "dataset_name": "swebenchpro", + "source_type": "hf_dataset", + "samples_number": 26, + "sample_ids": [ + "instance_qutebrowser__qutebrowser-2e961080a85d660148937ee8f0f6b3445a8f2c01-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_flipt-io__flipt-b68b8960b8a08540d5198d78c665a7eb0bea4008", + "instance_qutebrowser__qutebrowser-01d1d1494411380d97cac14614a829d3a69cecaf-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_future-architect__vuls-78b52d6a7f480bd610b692de9bf0c86f57332f23", + "instance_flipt-io__flipt-e42da21a07a5ae35835ec54f74004ebd58713874", + "instance_nodebb__nodebb-84e065752f6d7fbe5c08cbf50cb173ffb866b8fa-vf2cf3cbd463b7ad942381f1c6d077626485a1e9e", + "instance_element-hq__element-web-f3534b42df3dcfe36dc48bddbf14034085af6d30-vnan", + "instance_navidrome__navidrome-d8e794317f788198227e10fb667e10496b3eb99a", + "instance_future-architect__vuls-f0b3a8b1db98eb1bd32685f1c36c41a99c3452ed", + "instance_future-architect__vuls-6eff6a9329a65cc412e79b8f82444dfa3d0f0b5a", + "instance_internetarchive__openlibrary-fdbc0d8f418333c7e575c40b661b582c301ef7ac-v13642507b4fc1f8d234172bf8129942da2c2ca26", + "instance_ansible__ansible-b2a289dcbb702003377221e25f62c8a3608f0e89-v173091e2e36d38c978002990795f66cfc0af30ad", + "instance_internetarchive__openlibrary-f8cc11d9c1575fdba5ac66aee0befca970da8d64-v13642507b4fc1f8d234172bf8129942da2c2ca26", + "instance_qutebrowser__qutebrowser-e70f5b03187bdd40e8bf70f5f3ead840f52d1f42-v02ad04386d5238fe2d1a1be450df257370de4b6a", + "instance_tutao__tutanota-d1aa0ecec288bfc800cfb9133b087c4f81ad8b38-vbc0d9ba8f0071fbe982809910959a6ff8884dbbf", + "instance_qutebrowser__qutebrowser-e64622cd2df5b521342cf4a62e0d4cb8f8c9ae5a-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_gravitational__teleport-ac2fb2f9b4fd1896b554d3011df23d3d71295779", + "instance_protonmail__webclients-cb8cc309c6968b0a2a5fe4288d0ae0a969ff31e1", + "instance_ansible__ansible-11c1777d56664b1acb56b387a1ad6aeadef1391d-v0f01c69f1e2528b935359cfe578530722bca2c59", + "instance_ansible__ansible-29aea9ff3466e4cd2ed00524b9e56738d568ce8b-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_nodebb__nodebb-70b4a0e2aebebe8f2f559de6680093d96a697b2f-vnan", + "instance_ansible__ansible-e22e103cdf8edc56ff7d9b848a58f94f1471a263-v1055803c3a812189a1133297f7f5468579283f86", + "instance_qutebrowser__qutebrowser-e15d26630934d0b6415ed2295ac42fd570a57620-va0fd88aac89cde702ec1ba84877234da33adce8a", + "instance_ansible__ansible-164881d871964aa64e0f911d03ae270acbad253c-v390e508d27db7a51eece36bb6d9698b63a5b638a", + "instance_ansible__ansible-502270c804c33d3bc963930dc85e0f4ca359674d-v7eee2454f617569fd6889f2211f75bc02a35f9f8", + "instance_internetarchive__openlibrary-9c392b60e2c6fa1d68cb68084b4b4ff04d0cb35c-v2d9a6c849c60ed19fd0858ce9e40b7cc8e097e59" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1777025747.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.7307692307692307, + "uncertainty": { + "standard_error": { + "value": 0.08871201995900614 + }, + "num_samples": 88 + }, + "details": { + "total_matched_trajectories": "88", + "stopping_reason_count_repetition_guard": "53", + "stopping_reason_count_token_limit": "35", + "avg_total_tokens_target_model": "18483757.43", + "total_total_tokens_target_model": "1626570654.0", + "avg_total_tokens_other_models": "442.93", + "total_total_tokens_other_models": "38978.0", + "avg_total_tokens_all_models": "18484200.36", + "total_total_tokens_all_models": "1626609632.0", + "avg_turn_count": "267.36", + "total_turn_count": "23528.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 30000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/anthropic/claude-opus-4-5-20251101/e6e37cd8-011e-4402-a0f7-248411195a4c_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "1d09e2fb9e0589918d07745e3f2e12d51a7677298dd92366914895af5d2ed396", + "total_rows": 88 + } +} \ No newline at end of file diff --git a/data/swebenchpro/anthropic/claude-opus-4-5-20251101/e6e37cd8-011e-4402-a0f7-248411195a4c_samples.jsonl b/data/swebenchpro/anthropic/claude-opus-4-5-20251101/e6e37cd8-011e-4402-a0f7-248411195a4c_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..06f62436e482b04c09f146c6af39310aaa87d94a --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-5-20251101/e6e37cd8-011e-4402-a0f7-248411195a4c_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1d09e2fb9e0589918d07745e3f2e12d51a7677298dd92366914895af5d2ed396 +size 91690112 diff --git a/data/swebenchpro/anthropic/claude-opus-4-5-20251101/eaabc11e-382e-4cdd-95bb-800118e465b1.json b/data/swebenchpro/anthropic/claude-opus-4-5-20251101/eaabc11e-382e-4cdd-95bb-800118e465b1.json new file mode 100644 index 0000000000000000000000000000000000000000..4d4f2ea92328dea8179dc81267e15663372ee0bf --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-5-20251101/eaabc11e-382e-4cdd-95bb-800118e465b1.json @@ -0,0 +1,108 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/anthropic_claude-opus-4-5-20251101/1786165535.295577", + "evaluation_timestamp": "1777787601.0", + "retrieved_timestamp": "1786165535.295577", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-5-20251101", + "id": "anthropic/claude-opus-4-5-20251101", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on swebenchpro/S-adaptive/+4ep/38bb1c85 for scorer harbor_scorer", + "source_data": { + "dataset_name": "swebenchpro", + "source_type": "hf_dataset", + "samples_number": 1, + "sample_ids": [ + "instance_tutao__tutanota-8513a9e8114a8b42e64f4348335e0f23efa054c4-vee878bb72091875e912c52fc32bc60ec3760227b" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1777787601.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 1.0, + "uncertainty": { + "num_samples": 3 + }, + "details": { + "total_matched_trajectories": "3", + "stopping_reason_count_repetition_guard": "2", + "stopping_reason_count_token_limit": "1", + "avg_total_tokens_target_model": "18226756.00", + "total_total_tokens_target_model": "54680268.0", + "avg_total_tokens_other_models": "115.67", + "total_total_tokens_other_models": "347.0", + "avg_total_tokens_all_models": "18226871.67", + "total_total_tokens_all_models": "54680615.0", + "avg_turn_count": "249.00", + "total_turn_count": "747.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 30000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/anthropic/claude-opus-4-5-20251101/eaabc11e-382e-4cdd-95bb-800118e465b1_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "95ef9be03bb1bb33aef1910cfff603c2ef335821fb6562e24681b7727c47eec8", + "total_rows": 3 + } +} \ No newline at end of file diff --git a/data/swebenchpro/anthropic/claude-opus-4-5-20251101/eaabc11e-382e-4cdd-95bb-800118e465b1_samples.jsonl b/data/swebenchpro/anthropic/claude-opus-4-5-20251101/eaabc11e-382e-4cdd-95bb-800118e465b1_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..d0c3cfa765277fa66cdc5a237f999407f7fa9231 --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-5-20251101/eaabc11e-382e-4cdd-95bb-800118e465b1_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:95ef9be03bb1bb33aef1910cfff603c2ef335821fb6562e24681b7727c47eec8 +size 3109042 diff --git a/data/swebenchpro/anthropic/claude-opus-4-5-20251101/f5acf456-8e90-4e38-b11e-daac7ee3f4d0.json b/data/swebenchpro/anthropic/claude-opus-4-5-20251101/f5acf456-8e90-4e38-b11e-daac7ee3f4d0.json new file mode 100644 index 0000000000000000000000000000000000000000..6b227e84807c4ab1c264e3c0faee1f761d050fcb --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-5-20251101/f5acf456-8e90-4e38-b11e-daac7ee3f4d0.json @@ -0,0 +1,123 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/anthropic_claude-opus-4-5-20251101/1786163058.763405", + "evaluation_timestamp": "1776433475.0", + "retrieved_timestamp": "1786163058.763405", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-5-20251101", + "id": "anthropic/claude-opus-4-5-20251101", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on swebenchpro/S-adaptive+C/+3ep/bfa59b09 for scorer harbor_scorer", + "source_data": { + "dataset_name": "swebenchpro", + "source_type": "hf_dataset", + "samples_number": 16, + "sample_ids": [ + "instance_future-architect__vuls-f0b3a8b1db98eb1bd32685f1c36c41a99c3452ed", + "instance_gravitational__teleport-ac2fb2f9b4fd1896b554d3011df23d3d71295779", + "instance_nodebb__nodebb-0f788b8eaa4bba3c142d171fd941d015c53b65fc-v0ec6d6c2baf3cb4797482ce4829bc25cd5716649", + "instance_element-hq__element-web-9bf77963ee5e036d54b2a3ca202fbf6378464a5e-vnan", + "instance_protonmail__webclients-da91f084c0f532d9cc8ca385a701274d598057b8", + "instance_flipt-io__flipt-b2cd6a6dd73ca91b519015fd5924fde8d17f3f06", + "instance_ansible__ansible-3db08adbb1cc6aa9941be5e0fc810132c6e1fa4b-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_internetarchive__openlibrary-c506c1b0b678892af5cb22c1c1dbc35d96787a0a-v0f5aece3601a5b4419f7ccec1dbda2071be28ee4", + "instance_internetarchive__openlibrary-25858f9f0c165df25742acf8309ce909773f0cdd-v13642507b4fc1f8d234172bf8129942da2c2ca26", + "instance_flipt-io__flipt-05d7234fa582df632f70a7cd10194d61bd7043b9", + "instance_ansible__ansible-a02e22e902a69aeb465f16bf03f7f5a91b2cb828-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_nodebb__nodebb-8ca65b0c78c67c1653487c02d1135e1b702185e1-vf2cf3cbd463b7ad942381f1c6d077626485a1e9e", + "instance_protonmail__webclients-2c3559cad02d1090985dba7e8eb5a129144d9811", + "instance_internetarchive__openlibrary-630221ab686c64e75a2ce253c893c033e4814b2e-v93c53c13d5f9b383ebb411ee7750b49dcd1a34c6", + "instance_qutebrowser__qutebrowser-a25e8a09873838ca9efefd36ea8a45170bbeb95c-vc2f56a753b62a190ddb23cd330c257b9cf560d12", + "instance_protonmail__webclients-f161c10cf7d31abf82e8d64d7a99c9fac5acfa18" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1776433475.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.75, + "uncertainty": { + "standard_error": { + "value": 0.11180339887498948 + }, + "num_samples": 48 + }, + "details": { + "total_matched_trajectories": "48", + "stopping_reason_count_completed_on_successful_submit": "35", + "stopping_reason_count_repetition_guard": "13", + "avg_total_tokens_target_model": "3766788.77", + "total_total_tokens_target_model": "180805861.0", + "avg_total_tokens_other_models": "0.00", + "total_total_tokens_other_models": "0.0", + "avg_total_tokens_all_models": "3766788.77", + "total_total_tokens_all_models": "180805861.0", + "avg_turn_count": "77.69", + "total_turn_count": "3729.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "max_connections": "10", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 30000000 + }, + "sandbox": {} + }, + "additional_details": { + "max_connections": "10" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/anthropic/claude-opus-4-5-20251101/f5acf456-8e90-4e38-b11e-daac7ee3f4d0_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "427b046317687830c1d5fd576fdbc090ccb7dd304802d351d6ad16a5dc61ecaa", + "total_rows": 48 + } +} \ No newline at end of file diff --git a/data/swebenchpro/anthropic/claude-opus-4-5-20251101/f5acf456-8e90-4e38-b11e-daac7ee3f4d0_samples.jsonl b/data/swebenchpro/anthropic/claude-opus-4-5-20251101/f5acf456-8e90-4e38-b11e-daac7ee3f4d0_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..3932b15ff205530bacaa3f38dc19d07460e0e660 --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-5-20251101/f5acf456-8e90-4e38-b11e-daac7ee3f4d0_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:427b046317687830c1d5fd576fdbc090ccb7dd304802d351d6ad16a5dc61ecaa +size 29013986 diff --git a/data/swebenchpro/anthropic/claude-opus-4-6/179733e3-d347-4acc-a6df-d5abae5cad4b.json b/data/swebenchpro/anthropic/claude-opus-4-6/179733e3-d347-4acc-a6df-d5abae5cad4b.json new file mode 100644 index 0000000000000000000000000000000000000000..ddf76506ca541250d9fd74afe6c59c749bb9a1e4 --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-6/179733e3-d347-4acc-a6df-d5abae5cad4b.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/anthropic_claude-opus-4-6/1786167420.430653", + "evaluation_timestamp": "1778676015.0", + "retrieved_timestamp": "1786167420.430653", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-6", + "id": "anthropic/claude-opus-4-6", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/anthropic/claude-opus-4-6/179733e3-d347-4acc-a6df-d5abae5cad4b_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "4adda30c21c5975174fedec28514c099cbb28d903c3986fed8a32aa063e7d272", + "total_rows": 1 + } +} \ No newline at end of file diff --git a/data/swebenchpro/anthropic/claude-opus-4-6/179733e3-d347-4acc-a6df-d5abae5cad4b_samples.jsonl b/data/swebenchpro/anthropic/claude-opus-4-6/179733e3-d347-4acc-a6df-d5abae5cad4b_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..e99463d175a2694f30d25eb28a948a22ad905e42 --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-6/179733e3-d347-4acc-a6df-d5abae5cad4b_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4adda30c21c5975174fedec28514c099cbb28d903c3986fed8a32aa063e7d272 +size 122423 diff --git a/data/swebenchpro/anthropic/claude-opus-4-6/203094b7-060e-41dd-9650-3eee5da88ed1.json b/data/swebenchpro/anthropic/claude-opus-4-6/203094b7-060e-41dd-9650-3eee5da88ed1.json new file mode 100644 index 0000000000000000000000000000000000000000..edf4c6c20960852bd261e5590293dec9b9531451 --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-6/203094b7-060e-41dd-9650-3eee5da88ed1.json @@ -0,0 +1,212 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/anthropic_claude-opus-4-6/1786165689.27021", + "evaluation_timestamp": "1778280802.0", + "retrieved_timestamp": "1786165689.27021", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-6", + "id": "anthropic/claude-opus-4-6", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on swebenchpro/S-adaptive/c7cac3d9 for scorer harbor_scorer", + "source_data": { + "dataset_name": "swebenchpro", + "source_type": "hf_dataset", + "samples_number": 100, + "sample_ids": [ + "instance_qutebrowser__qutebrowser-2e961080a85d660148937ee8f0f6b3445a8f2c01-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_flipt-io__flipt-b68b8960b8a08540d5198d78c665a7eb0bea4008", + "instance_qutebrowser__qutebrowser-01d1d1494411380d97cac14614a829d3a69cecaf-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_future-architect__vuls-78b52d6a7f480bd610b692de9bf0c86f57332f23", + "instance_flipt-io__flipt-e42da21a07a5ae35835ec54f74004ebd58713874", + "instance_nodebb__nodebb-84e065752f6d7fbe5c08cbf50cb173ffb866b8fa-vf2cf3cbd463b7ad942381f1c6d077626485a1e9e", + "instance_element-hq__element-web-f3534b42df3dcfe36dc48bddbf14034085af6d30-vnan", + "instance_navidrome__navidrome-d8e794317f788198227e10fb667e10496b3eb99a", + "instance_future-architect__vuls-f0b3a8b1db98eb1bd32685f1c36c41a99c3452ed", + "instance_future-architect__vuls-6eff6a9329a65cc412e79b8f82444dfa3d0f0b5a", + "instance_internetarchive__openlibrary-fdbc0d8f418333c7e575c40b661b582c301ef7ac-v13642507b4fc1f8d234172bf8129942da2c2ca26", + "instance_ansible__ansible-b2a289dcbb702003377221e25f62c8a3608f0e89-v173091e2e36d38c978002990795f66cfc0af30ad", + "instance_internetarchive__openlibrary-f8cc11d9c1575fdba5ac66aee0befca970da8d64-v13642507b4fc1f8d234172bf8129942da2c2ca26", + "instance_qutebrowser__qutebrowser-e70f5b03187bdd40e8bf70f5f3ead840f52d1f42-v02ad04386d5238fe2d1a1be450df257370de4b6a", + "instance_tutao__tutanota-d1aa0ecec288bfc800cfb9133b087c4f81ad8b38-vbc0d9ba8f0071fbe982809910959a6ff8884dbbf", + "instance_qutebrowser__qutebrowser-e64622cd2df5b521342cf4a62e0d4cb8f8c9ae5a-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_gravitational__teleport-ac2fb2f9b4fd1896b554d3011df23d3d71295779", + "instance_protonmail__webclients-cb8cc309c6968b0a2a5fe4288d0ae0a969ff31e1", + "instance_ansible__ansible-11c1777d56664b1acb56b387a1ad6aeadef1391d-v0f01c69f1e2528b935359cfe578530722bca2c59", + "instance_ansible__ansible-29aea9ff3466e4cd2ed00524b9e56738d568ce8b-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_nodebb__nodebb-70b4a0e2aebebe8f2f559de6680093d96a697b2f-vnan", + "instance_ansible__ansible-e22e103cdf8edc56ff7d9b848a58f94f1471a263-v1055803c3a812189a1133297f7f5468579283f86", + "instance_qutebrowser__qutebrowser-e15d26630934d0b6415ed2295ac42fd570a57620-va0fd88aac89cde702ec1ba84877234da33adce8a", + "instance_ansible__ansible-164881d871964aa64e0f911d03ae270acbad253c-v390e508d27db7a51eece36bb6d9698b63a5b638a", + "instance_ansible__ansible-502270c804c33d3bc963930dc85e0f4ca359674d-v7eee2454f617569fd6889f2211f75bc02a35f9f8", + "instance_internetarchive__openlibrary-9c392b60e2c6fa1d68cb68084b4b4ff04d0cb35c-v2d9a6c849c60ed19fd0858ce9e40b7cc8e097e59", + "instance_element-hq__element-web-b007ea81b2ccd001b00f332bee65070aa7fc00f9-vnan", + "instance_ansible__ansible-811093f0225caa4dd33890933150a81c6a6d5226-v1055803c3a812189a1133297f7f5468579283f86", + "instance_flipt-io__flipt-cd2f3b0a9d4d8b8a6d3d56afab65851ecdc408e8", + "instance_internetarchive__openlibrary-7cbfb812ef0e1f9716e2d6e85d538a96fcb79d13-vfa6ff903cb27f336e17654595dd900fa943dcd91", + "instance_nodebb__nodebb-0f788b8eaa4bba3c142d171fd941d015c53b65fc-v0ec6d6c2baf3cb4797482ce4829bc25cd5716649", + "instance_flipt-io__flipt-0b119520afca1cf25c470ff4288c464d4510b944", + "instance_element-hq__element-web-f63160f38459fb552d00fcc60d4064977a9095a6-vnan", + "instance_ansible__ansible-f327e65d11bb905ed9f15996024f857a95592629-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_gravitational__teleport-73cc189b0e9636d418c4470ecce0d9af5dae2f02-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_internetarchive__openlibrary-3c48b4bb782189e0858e6c3fc7956046cf3e1cfb-v2d9a6c849c60ed19fd0858ce9e40b7cc8e097e59", + "instance_future-architect__vuls-878c25bf5a9c9fd88ac32eb843f5636834d5712d", + "instance_element-hq__element-web-9bf77963ee5e036d54b2a3ca202fbf6378464a5e-vnan", + "instance_qutebrowser__qutebrowser-70248f256f93ed9b1984494d0a1a919ddd774892-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_future-architect__vuls-73f0adad95c4d227e2ccfa876c85cc95dd065e13", + "instance_protonmail__webclients-da91f084c0f532d9cc8ca385a701274d598057b8", + "instance_flipt-io__flipt-b2cd6a6dd73ca91b519015fd5924fde8d17f3f06", + "instance_qutebrowser__qutebrowser-cc360cd4a34a126274c7b51f3b63afbaf3e05a02-v5fc38aaf22415ab0b70567368332beee7955b367", + "instance_flipt-io__flipt-e2bd19dafa7166c96b082fb2a59eb54b4be0d778", + "instance_qutebrowser__qutebrowser-9ed748effa8f3bcd804612d9291da017b514e12f-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_ansible__ansible-1ee70fc272aff6bf3415357c6e13c5de5b928d9b-v1055803c3a812189a1133297f7f5468579283f86", + "instance_element-hq__element-web-75c2c1a572fa45d1ea1d1a96e9e36e303332ecaa-vnan", + "instance_qutebrowser__qutebrowser-f631cd4422744160d9dcf7a0455da532ce973315-v35616345bb8052ea303186706cec663146f0f184", + "instance_ansible__ansible-3db08adbb1cc6aa9941be5e0fc810132c6e1fa4b-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_qutebrowser__qutebrowser-e5340c449f23608803c286da0563b62f58ba25b0-v059c6fdc75567943479b23ebca7c07b5e9a7f34c", + "instance_ansible__ansible-42355d181a11b51ebfc56f6f4b3d9c74e01cb13b-v1055803c3a812189a1133297f7f5468579283f86", + "instance_internetarchive__openlibrary-09865f5fb549694d969f0a8e49b9d204ef1853ca-ve8c8d62a2b60610a3c4631f5f23ed866bada9818", + "instance_internetarchive__openlibrary-c506c1b0b678892af5cb22c1c1dbc35d96787a0a-v0f5aece3601a5b4419f7ccec1dbda2071be28ee4", + "instance_element-hq__element-web-2760bfc8369f1bee640d6d7a7e910783143d4c5f-vnan", + "instance_internetarchive__openlibrary-25858f9f0c165df25742acf8309ce909773f0cdd-v13642507b4fc1f8d234172bf8129942da2c2ca26", + "instance_protonmail__webclients-01ea5214d11e0df8b7170d91bafd34f23cb0f2b1", + "instance_tutao__tutanota-4b4e45949096bb288f2b522f657610e480efa3e8-vee878bb72091875e912c52fc32bc60ec3760227b", + "instance_flipt-io__flipt-05d7234fa582df632f70a7cd10194d61bd7043b9", + "instance_ansible__ansible-a02e22e902a69aeb465f16bf03f7f5a91b2cb828-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_protonmail__webclients-5f0745dd6993bb1430a951c62a49807c6635cd77", + "instance_tutao__tutanota-1ff82aa365763cee2d609c9d19360ad87fdf2ec7-vc4e41fd0029957297843cb9dec4a25c7c756f029", + "instance_internetarchive__openlibrary-30bc73a1395fba2300087c7f307e54bb5372b60a-v76304ecdb3a5954fcf13feb710e8c40fcf24b73c", + "instance_nodebb__nodebb-8ca65b0c78c67c1653487c02d1135e1b702185e1-vf2cf3cbd463b7ad942381f1c6d077626485a1e9e", + "instance_nodebb__nodebb-22368b996ee0e5f11a5189b400b33af3cc8d925a-v4fbcfae8b15e4ce5d132c408bca69ebb9cf146ed", + "instance_protonmail__webclients-2c3559cad02d1090985dba7e8eb5a129144d9811", + "instance_protonmail__webclients-01b519cd49e6a24d9a05d2eb97f54e420740072e", + "instance_ansible__ansible-77658704217d5f166404fc67997203c25381cb6e-v390e508d27db7a51eece36bb6d9698b63a5b638a", + "instance_qutebrowser__qutebrowser-2dd8966fdcf11972062c540b7a787e4d0de8d372-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_ansible__ansible-cd9c4eb5a6b2bfaf4a6709f001ce3d0c92c1eed2-v0f01c69f1e2528b935359cfe578530722bca2c59", + "instance_protonmail__webclients-ac23d1efa1a6ab7e62724779317ba44c28d78cfd", + "instance_future-architect__vuls-030b2e03525d68d74cb749959aac2d7f3fc0effa", + "instance_nodebb__nodebb-b1f9ad5534bb3a44dab5364f659876a4b7fe34c1-vnan", + "instance_flipt-io__flipt-292fdaca9be39e6a921aaa8874c011d0fdd3e874", + "instance_qutebrowser__qutebrowser-ef5ba1a0360b39f9eff027fbdc57f363597c3c3b-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_ansible__ansible-622a493ae03bd5e5cf517d336fc426e9d12208c7-v906c969b551b346ef54a2c0b41e04f632b7b73c2", + "instance_gravitational__teleport-82185f232ae8974258397e121b3bc2ed0c3729ed-v626ec2a48416b10a88641359a169d99e935ff037", + "instance_ansible__ansible-9a21e247786ebd294dafafca1105fcd770ff46c6-v67cdaa49f89b34e42b69d5b7830b3c3ad3d8803f", + "instance_protonmail__webclients-8afd9ce04c8dde9e150e1c2b50d32e7ee2efa3e7", + "instance_flipt-io__flipt-b22f5f02e40b225b6b93fff472914973422e97c6", + "instance_gravitational__teleport-2bb3bbbd8aff1164a2353381cb79e1dc93b90d28-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_ansible__ansible-de5858f48dc9e1ce9117034e0d7e76806f420ca8-v1055803c3a812189a1133297f7f5468579283f86", + "instance_gravitational__teleport-46aa81b1ce96ebb4ebed2ae53fd78cd44a05da6c-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_tutao__tutanota-8513a9e8114a8b42e64f4348335e0f23efa054c4-vee878bb72091875e912c52fc32bc60ec3760227b", + "instance_internetarchive__openlibrary-630221ab686c64e75a2ce253c893c033e4814b2e-v93c53c13d5f9b383ebb411ee7750b49dcd1a34c6", + "instance_internetarchive__openlibrary-7f6b722a10f822171501d027cad60afe53337732-ve8c8d62a2b60610a3c4631f5f23ed866bada9818", + "instance_nodebb__nodebb-a917210c5b2c20637094545401f85783905c074c-vf2cf3cbd463b7ad942381f1c6d077626485a1e9e", + "instance_internetarchive__openlibrary-1351c59fd43689753de1fca32c78d539a116ffc1-v29f82c9cf21d57b242f8d8b0e541525d259e2d63", + "instance_qutebrowser__qutebrowser-a25e8a09873838ca9efefd36ea8a45170bbeb95c-vc2f56a753b62a190ddb23cd330c257b9cf560d12", + "instance_protonmail__webclients-281a6b3f190f323ec2c0630999354fafb84b2880", + "instance_future-architect__vuls-999529a05b202b0fd29c6fca5039a4c47a3766bb", + "instance_element-hq__element-web-1077729a19c0ce902e713cf6fab42c91fb7907f1-vnan", + "instance_future-architect__vuls-be7b9114cc9545e68fb0ee7bc63d7ec53d1a00ad", + "instance_qutebrowser__qutebrowser-c580ebf0801e5a3ecabc54f327498bb753c6d5f2-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_gravitational__teleport-3ff19cf7c41f396ae468797d3aeb61515517edc9-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_future-architect__vuls-aaea15e516ece43978cf98e09e52080478b1d39f", + "instance_protonmail__webclients-f161c10cf7d31abf82e8d64d7a99c9fac5acfa18", + "instance_qutebrowser__qutebrowser-cf06f4e3708f886032d4d2a30108c2fddb042d81-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_internetarchive__openlibrary-3f580a5f244c299d936d73d9e327ba873b6401d9-v0f5aece3601a5b4419f7ccec1dbda2071be28ee4", + "instance_qutebrowser__qutebrowser-0aa57e4f7243024fa4bba8853306691b5dbd77b3-v5149fcda2a9a6fe1d35dfed1bade1444a11ef271", + "instance_future-architect__vuls-50580f6e98eeb36f53f27222f7f4fdfea0b21e8d" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1778280802.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.725, + "uncertainty": { + "standard_error": { + "value": 0.04445233515812304 + }, + "num_samples": 416 + }, + "details": { + "total_matched_trajectories": "416", + "stopping_reason_count_repetition_guard": "416", + "avg_total_tokens_target_model": "2075921.56", + "total_total_tokens_target_model": "863583369.0", + "avg_total_tokens_other_models": "1098.06", + "total_total_tokens_other_models": "456791.0", + "avg_total_tokens_all_models": "2077019.62", + "total_total_tokens_all_models": "864040160.0", + "avg_turn_count": "58.50", + "total_turn_count": "24337.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 30000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/anthropic/claude-opus-4-6/203094b7-060e-41dd-9650-3eee5da88ed1_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "c3642bbace0b100ddcfa6b1e7063d87c02fbde554192c25e053b04ca1247682f", + "total_rows": 416 + } +} \ No newline at end of file diff --git a/data/swebenchpro/anthropic/claude-opus-4-6/203094b7-060e-41dd-9650-3eee5da88ed1_samples.jsonl b/data/swebenchpro/anthropic/claude-opus-4-6/203094b7-060e-41dd-9650-3eee5da88ed1_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..33071497dc8e78377bc79712bf41f860fc9d2f34 --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-6/203094b7-060e-41dd-9650-3eee5da88ed1_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c3642bbace0b100ddcfa6b1e7063d87c02fbde554192c25e053b04ca1247682f +size 106933404 diff --git a/data/swebenchpro/anthropic/claude-opus-4-6/2043d614-8332-43f4-bada-9d4bed9a33f5.json b/data/swebenchpro/anthropic/claude-opus-4-6/2043d614-8332-43f4-bada-9d4bed9a33f5.json new file mode 100644 index 0000000000000000000000000000000000000000..68aef03050a6165f8a64bd2e4d96fe39a8ef44df --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-6/2043d614-8332-43f4-bada-9d4bed9a33f5.json @@ -0,0 +1,131 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/anthropic_claude-opus-4-6/1786163189.660763", + "evaluation_timestamp": "1778593563.0", + "retrieved_timestamp": "1786163189.660763", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-6", + "id": "anthropic/claude-opus-4-6", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on swebenchpro/S-adaptive+C/6f5659d0 for scorer harbor_scorer", + "source_data": { + "dataset_name": "swebenchpro", + "source_type": "hf_dataset", + "samples_number": 18, + "sample_ids": [ + "instance_protonmail__webclients-8afd9ce04c8dde9e150e1c2b50d32e7ee2efa3e7", + "instance_flipt-io__flipt-b22f5f02e40b225b6b93fff472914973422e97c6", + "instance_ansible__ansible-de5858f48dc9e1ce9117034e0d7e76806f420ca8-v1055803c3a812189a1133297f7f5468579283f86", + "instance_tutao__tutanota-8513a9e8114a8b42e64f4348335e0f23efa054c4-vee878bb72091875e912c52fc32bc60ec3760227b", + "instance_internetarchive__openlibrary-7f6b722a10f822171501d027cad60afe53337732-ve8c8d62a2b60610a3c4631f5f23ed866bada9818", + "instance_nodebb__nodebb-a917210c5b2c20637094545401f85783905c074c-vf2cf3cbd463b7ad942381f1c6d077626485a1e9e", + "instance_protonmail__webclients-281a6b3f190f323ec2c0630999354fafb84b2880", + "instance_future-architect__vuls-999529a05b202b0fd29c6fca5039a4c47a3766bb", + "instance_element-hq__element-web-1077729a19c0ce902e713cf6fab42c91fb7907f1-vnan", + "instance_future-architect__vuls-be7b9114cc9545e68fb0ee7bc63d7ec53d1a00ad", + "instance_qutebrowser__qutebrowser-c580ebf0801e5a3ecabc54f327498bb753c6d5f2-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_gravitational__teleport-3ff19cf7c41f396ae468797d3aeb61515517edc9-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_future-architect__vuls-aaea15e516ece43978cf98e09e52080478b1d39f", + "instance_protonmail__webclients-f161c10cf7d31abf82e8d64d7a99c9fac5acfa18", + "instance_qutebrowser__qutebrowser-cf06f4e3708f886032d4d2a30108c2fddb042d81-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_internetarchive__openlibrary-3f580a5f244c299d936d73d9e327ba873b6401d9-v0f5aece3601a5b4419f7ccec1dbda2071be28ee4", + "instance_qutebrowser__qutebrowser-0aa57e4f7243024fa4bba8853306691b5dbd77b3-v5149fcda2a9a6fe1d35dfed1bade1444a11ef271", + "instance_future-architect__vuls-50580f6e98eeb36f53f27222f7f4fdfea0b21e8d" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1778593563.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.6111111111111112, + "uncertainty": { + "standard_error": { + "value": 0.1182356373537617 + }, + "num_samples": 39 + }, + "details": { + "total_matched_trajectories": "39", + "stopping_reason_count_completed_on_successful_submit": "25", + "stopping_reason_count_repetition_guard": "14", + "avg_total_tokens_target_model": "1936353.13", + "total_total_tokens_target_model": "75517772.0", + "avg_total_tokens_other_models": "330.59", + "total_total_tokens_other_models": "12893.0", + "avg_total_tokens_all_models": "1936683.72", + "total_total_tokens_all_models": "75530665.0", + "avg_turn_count": "63.41", + "total_turn_count": "2473.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 30000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/anthropic/claude-opus-4-6/2043d614-8332-43f4-bada-9d4bed9a33f5_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "9d7ce2c9e7f8484ab45c6e67efd93425264f26bc0d0e683d2ba29bdea0d9ceda", + "total_rows": 39 + } +} \ No newline at end of file diff --git a/data/swebenchpro/anthropic/claude-opus-4-6/2043d614-8332-43f4-bada-9d4bed9a33f5_samples.jsonl b/data/swebenchpro/anthropic/claude-opus-4-6/2043d614-8332-43f4-bada-9d4bed9a33f5_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..7cf1cc9a5263e9237d4b74ee05d28beeb28ed0b7 --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-6/2043d614-8332-43f4-bada-9d4bed9a33f5_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9d7ce2c9e7f8484ab45c6e67efd93425264f26bc0d0e683d2ba29bdea0d9ceda +size 9906301 diff --git a/data/swebenchpro/anthropic/claude-opus-4-6/3d551c10-1417-43c2-9271-d6f6f9923c0e.json b/data/swebenchpro/anthropic/claude-opus-4-6/3d551c10-1417-43c2-9271-d6f6f9923c0e.json new file mode 100644 index 0000000000000000000000000000000000000000..57673d3153f3d6a2fd807d2053d983f1df239e27 --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-6/3d551c10-1417-43c2-9271-d6f6f9923c0e.json @@ -0,0 +1,138 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/anthropic_claude-opus-4-6/1786163182.335955", + "evaluation_timestamp": "1778711997.0", + "retrieved_timestamp": "1786163182.335955", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-6", + "id": "anthropic/claude-opus-4-6", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on swebenchpro/S-adaptive+C/+4ep/fcf25459 for scorer harbor_scorer", + "source_data": { + "dataset_name": "swebenchpro", + "source_type": "hf_dataset", + "samples_number": 24, + "sample_ids": [ + "instance_tutao__tutanota-d1aa0ecec288bfc800cfb9133b087c4f81ad8b38-vbc0d9ba8f0071fbe982809910959a6ff8884dbbf", + "instance_qutebrowser__qutebrowser-e64622cd2df5b521342cf4a62e0d4cb8f8c9ae5a-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_flipt-io__flipt-0b119520afca1cf25c470ff4288c464d4510b944", + "instance_future-architect__vuls-73f0adad95c4d227e2ccfa876c85cc95dd065e13", + "instance_protonmail__webclients-da91f084c0f532d9cc8ca385a701274d598057b8", + "instance_ansible__ansible-42355d181a11b51ebfc56f6f4b3d9c74e01cb13b-v1055803c3a812189a1133297f7f5468579283f86", + "instance_ansible__ansible-a02e22e902a69aeb465f16bf03f7f5a91b2cb828-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_protonmail__webclients-5f0745dd6993bb1430a951c62a49807c6635cd77", + "instance_tutao__tutanota-1ff82aa365763cee2d609c9d19360ad87fdf2ec7-vc4e41fd0029957297843cb9dec4a25c7c756f029", + "instance_nodebb__nodebb-8ca65b0c78c67c1653487c02d1135e1b702185e1-vf2cf3cbd463b7ad942381f1c6d077626485a1e9e", + "instance_qutebrowser__qutebrowser-2dd8966fdcf11972062c540b7a787e4d0de8d372-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_future-architect__vuls-030b2e03525d68d74cb749959aac2d7f3fc0effa", + "instance_nodebb__nodebb-b1f9ad5534bb3a44dab5364f659876a4b7fe34c1-vnan", + "instance_flipt-io__flipt-292fdaca9be39e6a921aaa8874c011d0fdd3e874", + "instance_gravitational__teleport-82185f232ae8974258397e121b3bc2ed0c3729ed-v626ec2a48416b10a88641359a169d99e935ff037", + "instance_gravitational__teleport-2bb3bbbd8aff1164a2353381cb79e1dc93b90d28-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_ansible__ansible-de5858f48dc9e1ce9117034e0d7e76806f420ca8-v1055803c3a812189a1133297f7f5468579283f86", + "instance_gravitational__teleport-46aa81b1ce96ebb4ebed2ae53fd78cd44a05da6c-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_internetarchive__openlibrary-630221ab686c64e75a2ce253c893c033e4814b2e-v93c53c13d5f9b383ebb411ee7750b49dcd1a34c6", + "instance_nodebb__nodebb-a917210c5b2c20637094545401f85783905c074c-vf2cf3cbd463b7ad942381f1c6d077626485a1e9e", + "instance_qutebrowser__qutebrowser-a25e8a09873838ca9efefd36ea8a45170bbeb95c-vc2f56a753b62a190ddb23cd330c257b9cf560d12", + "instance_protonmail__webclients-281a6b3f190f323ec2c0630999354fafb84b2880", + "instance_future-architect__vuls-999529a05b202b0fd29c6fca5039a4c47a3766bb", + "instance_qutebrowser__qutebrowser-cf06f4e3708f886032d4d2a30108c2fddb042d81-v2ef375ac784985212b1805e1d0431dc8f1b3c171" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1778711997.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.6041666666666666, + "uncertainty": { + "standard_error": { + "value": 0.09858218940372977 + }, + "num_samples": 96 + }, + "details": { + "total_matched_trajectories": "96", + "stopping_reason_count_completed_on_successful_submit": "58", + "stopping_reason_count_repetition_guard": "36", + "stopping_reason_count_completed_with_submit": "2", + "avg_total_tokens_target_model": "3803918.74", + "total_total_tokens_target_model": "365176199.0", + "avg_total_tokens_other_models": "454.41", + "total_total_tokens_other_models": "43623.0", + "avg_total_tokens_all_models": "3804373.15", + "total_total_tokens_all_models": "365219822.0", + "avg_turn_count": "84.11", + "total_turn_count": "8075.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "8", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 30000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "8", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/anthropic/claude-opus-4-6/3d551c10-1417-43c2-9271-d6f6f9923c0e_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "4a7ddf9e495dbd133670d11d9659363909a87632c50aeaac01e89ad45321d1fa", + "total_rows": 96 + } +} \ No newline at end of file diff --git a/data/swebenchpro/anthropic/claude-opus-4-6/3d551c10-1417-43c2-9271-d6f6f9923c0e_samples.jsonl b/data/swebenchpro/anthropic/claude-opus-4-6/3d551c10-1417-43c2-9271-d6f6f9923c0e_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..cda7384bdc7b79b9a5b4965d59d29ff10fefa343 --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-6/3d551c10-1417-43c2-9271-d6f6f9923c0e_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4a7ddf9e495dbd133670d11d9659363909a87632c50aeaac01e89ad45321d1fa +size 33504384 diff --git a/data/swebenchpro/anthropic/claude-opus-4-6/40761edf-7cb4-4e90-b50d-8a5ee7ef4ebb.json b/data/swebenchpro/anthropic/claude-opus-4-6/40761edf-7cb4-4e90-b50d-8a5ee7ef4ebb.json new file mode 100644 index 0000000000000000000000000000000000000000..5aab3b1954b7304750820615469a805eccda8876 --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-6/40761edf-7cb4-4e90-b50d-8a5ee7ef4ebb.json @@ -0,0 +1,166 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/anthropic_claude-opus-4-6/1786165626.842711", + "evaluation_timestamp": "1778516334.0", + "retrieved_timestamp": "1786165626.842711", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-6", + "id": "anthropic/claude-opus-4-6", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on swebenchpro/S-adaptive/+1ep/008245d5 for scorer harbor_scorer", + "source_data": { + "dataset_name": "swebenchpro", + "source_type": "hf_dataset", + "samples_number": 54, + "sample_ids": [ + "instance_flipt-io__flipt-0b119520afca1cf25c470ff4288c464d4510b944", + "instance_element-hq__element-web-f63160f38459fb552d00fcc60d4064977a9095a6-vnan", + "instance_ansible__ansible-f327e65d11bb905ed9f15996024f857a95592629-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_gravitational__teleport-73cc189b0e9636d418c4470ecce0d9af5dae2f02-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_internetarchive__openlibrary-3c48b4bb782189e0858e6c3fc7956046cf3e1cfb-v2d9a6c849c60ed19fd0858ce9e40b7cc8e097e59", + "instance_element-hq__element-web-9bf77963ee5e036d54b2a3ca202fbf6378464a5e-vnan", + "instance_qutebrowser__qutebrowser-70248f256f93ed9b1984494d0a1a919ddd774892-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_ansible__ansible-3db08adbb1cc6aa9941be5e0fc810132c6e1fa4b-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_internetarchive__openlibrary-25858f9f0c165df25742acf8309ce909773f0cdd-v13642507b4fc1f8d234172bf8129942da2c2ca26", + "instance_protonmail__webclients-01ea5214d11e0df8b7170d91bafd34f23cb0f2b1", + "instance_tutao__tutanota-4b4e45949096bb288f2b522f657610e480efa3e8-vee878bb72091875e912c52fc32bc60ec3760227b", + "instance_flipt-io__flipt-05d7234fa582df632f70a7cd10194d61bd7043b9", + "instance_ansible__ansible-a02e22e902a69aeb465f16bf03f7f5a91b2cb828-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_protonmail__webclients-5f0745dd6993bb1430a951c62a49807c6635cd77", + "instance_tutao__tutanota-1ff82aa365763cee2d609c9d19360ad87fdf2ec7-vc4e41fd0029957297843cb9dec4a25c7c756f029", + "instance_internetarchive__openlibrary-30bc73a1395fba2300087c7f307e54bb5372b60a-v76304ecdb3a5954fcf13feb710e8c40fcf24b73c", + "instance_nodebb__nodebb-8ca65b0c78c67c1653487c02d1135e1b702185e1-vf2cf3cbd463b7ad942381f1c6d077626485a1e9e", + "instance_nodebb__nodebb-22368b996ee0e5f11a5189b400b33af3cc8d925a-v4fbcfae8b15e4ce5d132c408bca69ebb9cf146ed", + "instance_protonmail__webclients-2c3559cad02d1090985dba7e8eb5a129144d9811", + "instance_protonmail__webclients-01b519cd49e6a24d9a05d2eb97f54e420740072e", + "instance_ansible__ansible-77658704217d5f166404fc67997203c25381cb6e-v390e508d27db7a51eece36bb6d9698b63a5b638a", + "instance_qutebrowser__qutebrowser-2dd8966fdcf11972062c540b7a787e4d0de8d372-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_ansible__ansible-cd9c4eb5a6b2bfaf4a6709f001ce3d0c92c1eed2-v0f01c69f1e2528b935359cfe578530722bca2c59", + "instance_protonmail__webclients-ac23d1efa1a6ab7e62724779317ba44c28d78cfd", + "instance_future-architect__vuls-030b2e03525d68d74cb749959aac2d7f3fc0effa", + "instance_nodebb__nodebb-b1f9ad5534bb3a44dab5364f659876a4b7fe34c1-vnan", + "instance_flipt-io__flipt-292fdaca9be39e6a921aaa8874c011d0fdd3e874", + "instance_qutebrowser__qutebrowser-ef5ba1a0360b39f9eff027fbdc57f363597c3c3b-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_ansible__ansible-622a493ae03bd5e5cf517d336fc426e9d12208c7-v906c969b551b346ef54a2c0b41e04f632b7b73c2", + "instance_gravitational__teleport-82185f232ae8974258397e121b3bc2ed0c3729ed-v626ec2a48416b10a88641359a169d99e935ff037", + "instance_ansible__ansible-9a21e247786ebd294dafafca1105fcd770ff46c6-v67cdaa49f89b34e42b69d5b7830b3c3ad3d8803f", + "instance_protonmail__webclients-8afd9ce04c8dde9e150e1c2b50d32e7ee2efa3e7", + "instance_flipt-io__flipt-b22f5f02e40b225b6b93fff472914973422e97c6", + "instance_gravitational__teleport-2bb3bbbd8aff1164a2353381cb79e1dc93b90d28-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_ansible__ansible-de5858f48dc9e1ce9117034e0d7e76806f420ca8-v1055803c3a812189a1133297f7f5468579283f86", + "instance_gravitational__teleport-46aa81b1ce96ebb4ebed2ae53fd78cd44a05da6c-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_tutao__tutanota-8513a9e8114a8b42e64f4348335e0f23efa054c4-vee878bb72091875e912c52fc32bc60ec3760227b", + "instance_internetarchive__openlibrary-630221ab686c64e75a2ce253c893c033e4814b2e-v93c53c13d5f9b383ebb411ee7750b49dcd1a34c6", + "instance_internetarchive__openlibrary-7f6b722a10f822171501d027cad60afe53337732-ve8c8d62a2b60610a3c4631f5f23ed866bada9818", + "instance_nodebb__nodebb-a917210c5b2c20637094545401f85783905c074c-vf2cf3cbd463b7ad942381f1c6d077626485a1e9e", + "instance_internetarchive__openlibrary-1351c59fd43689753de1fca32c78d539a116ffc1-v29f82c9cf21d57b242f8d8b0e541525d259e2d63", + "instance_qutebrowser__qutebrowser-a25e8a09873838ca9efefd36ea8a45170bbeb95c-vc2f56a753b62a190ddb23cd330c257b9cf560d12", + "instance_protonmail__webclients-281a6b3f190f323ec2c0630999354fafb84b2880", + "instance_future-architect__vuls-999529a05b202b0fd29c6fca5039a4c47a3766bb", + "instance_element-hq__element-web-1077729a19c0ce902e713cf6fab42c91fb7907f1-vnan", + "instance_future-architect__vuls-be7b9114cc9545e68fb0ee7bc63d7ec53d1a00ad", + "instance_qutebrowser__qutebrowser-c580ebf0801e5a3ecabc54f327498bb753c6d5f2-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_gravitational__teleport-3ff19cf7c41f396ae468797d3aeb61515517edc9-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_future-architect__vuls-aaea15e516ece43978cf98e09e52080478b1d39f", + "instance_protonmail__webclients-f161c10cf7d31abf82e8d64d7a99c9fac5acfa18", + "instance_qutebrowser__qutebrowser-cf06f4e3708f886032d4d2a30108c2fddb042d81-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_internetarchive__openlibrary-3f580a5f244c299d936d73d9e327ba873b6401d9-v0f5aece3601a5b4419f7ccec1dbda2071be28ee4", + "instance_qutebrowser__qutebrowser-0aa57e4f7243024fa4bba8853306691b5dbd77b3-v5149fcda2a9a6fe1d35dfed1bade1444a11ef271", + "instance_future-architect__vuls-50580f6e98eeb36f53f27222f7f4fdfea0b21e8d" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1778516334.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.7222222222222222, + "uncertainty": { + "standard_error": { + "value": 0.061524237278109795 + }, + "num_samples": 54 + }, + "details": { + "total_matched_trajectories": "54", + "stopping_reason_count_repetition_guard": "54", + "avg_total_tokens_target_model": "2112863.80", + "total_total_tokens_target_model": "114094645.0", + "avg_total_tokens_other_models": "1153.15", + "total_total_tokens_other_models": "62270.0", + "avg_total_tokens_all_models": "2114016.94", + "total_total_tokens_all_models": "114156915.0", + "avg_turn_count": "58.78", + "total_turn_count": "3174.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 30000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/anthropic/claude-opus-4-6/40761edf-7cb4-4e90-b50d-8a5ee7ef4ebb_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "14ac8fb845eeb6b96bfdb188506d961691cd24f3c1e5632be5f9b9d6ab72821c", + "total_rows": 54 + } +} \ No newline at end of file diff --git a/data/swebenchpro/anthropic/claude-opus-4-6/40761edf-7cb4-4e90-b50d-8a5ee7ef4ebb_samples.jsonl b/data/swebenchpro/anthropic/claude-opus-4-6/40761edf-7cb4-4e90-b50d-8a5ee7ef4ebb_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..a5129d0cdb9dc934ab48cb7b4c1ce35ac7320c92 --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-6/40761edf-7cb4-4e90-b50d-8a5ee7ef4ebb_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:14ac8fb845eeb6b96bfdb188506d961691cd24f3c1e5632be5f9b9d6ab72821c +size 14128799 diff --git a/data/swebenchpro/anthropic/claude-opus-4-6/76e9535b-4105-4411-9115-843e0cc78b33.json b/data/swebenchpro/anthropic/claude-opus-4-6/76e9535b-4105-4411-9115-843e0cc78b33.json new file mode 100644 index 0000000000000000000000000000000000000000..9329fb310a2f5d2c4481899729c20bcdf858b2d8 --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-6/76e9535b-4105-4411-9115-843e0cc78b33.json @@ -0,0 +1,291 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/anthropic_claude-opus-4-6/1786182954.597854", + "evaluation_timestamp": "1775602861.0", + "retrieved_timestamp": "1786182954.597854", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.196.dev25+gb973b316,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-6", + "id": "anthropic/claude-opus-4-6", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on eval_task for scorer harbor_scorer", + "source_data": { + "dataset_name": "swebenchpro", + "source_type": "hf_dataset", + "samples_number": 731, + "sample_ids": [ + "instance_flipt-io__flipt-e91615cf07966da41756017a7d571f9fc0fdbe80", + "instance_ansible__ansible-8127abbc298cabf04aaa89a478fc5e5e3432a6fc-v30a923fb5c164d6cd18280c02422f75e611e8fb2", + "instance_protonmail__webclients-da91f084c0f532d9cc8ca385a701274d598057b8", + "instance_qutebrowser__qutebrowser-deeb15d6f009b3ca0c3bd503a7cef07462bd16b4-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_internetarchive__openlibrary-5069b09e5f64428dce59b33455c8bb17fe577070-v8717e18970bcdc4e0d2cea3b1527752b21e74866", + "instance_protonmail__webclients-c5a2089ca2bfe9aa1d85a664b8ad87ef843a1c9c", + "instance_qutebrowser__qutebrowser-fd6790fe8c02b144ab2464f1fc8ab3d02ce3c476-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_navidrome__navidrome-eebfbc5381a1e506ff17b5f1371d1ad83d5fd642", + "instance_navidrome__navidrome-812dc2090f20ac4f8ac271b6ed95be5889d1a3ca", + "instance_nodebb__nodebb-0c81642997ea1d827dbd02c311db9d4976112cd4-vf2cf3cbd463b7ad942381f1c6d077626485a1e9e", + "instance_internetarchive__openlibrary-c506c1b0b678892af5cb22c1c1dbc35d96787a0a-v0f5aece3601a5b4419f7ccec1dbda2071be28ee4", + "instance_flipt-io__flipt-5c7037ececb0bead0a8eb56054e224bcd7ac5922", + "instance_qutebrowser__qutebrowser-01d1d1494411380d97cac14614a829d3a69cecaf-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_internetarchive__openlibrary-7f7e53aa4cf74a4f8549a5bcd4810c527e2f6d7e-v13642507b4fc1f8d234172bf8129942da2c2ca26", + "instance_internetarchive__openlibrary-53e02a22972e9253aeded0e1981e6845e1e521fe-vfa6ff903cb27f336e17654595dd900fa943dcd91", + "instance_ansible__ansible-34db57a47f875d11c4068567b9ec7ace174ec4cf-v1055803c3a812189a1133297f7f5468579283f86", + "instance_qutebrowser__qutebrowser-996487c43e4fcc265b541f9eca1e7930e3c5cf05-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_gravitational__teleport-96019ce0be7a2c8e36363f359eb7c943b41dde70", + "instance_protonmail__webclients-df60460f163fd5c34e844ab9015e3176f1ab1ac0", + "instance_future-architect__vuls-ad2edbb8448e2c41a097f1c0b52696c0f6c5924d", + "instance_ansible__ansible-9142be2f6cabbe6597c9254c5bb9186d17036d55-v0f01c69f1e2528b935359cfe578530722bca2c59", + "instance_element-hq__element-web-33299af5c9b7a7ec5a9c31d578d4ec5b18088fb7-vnan", + "instance_element-hq__element-web-772df3021201d9c73835a626df8dcb6334ad9a3e-vnan", + "instance_internetarchive__openlibrary-1be7de788a444f6255e89c10ef6aa608550604a8-v29f82c9cf21d57b242f8d8b0e541525d259e2d63", + "instance_internetarchive__openlibrary-630221ab686c64e75a2ce253c893c033e4814b2e-v93c53c13d5f9b383ebb411ee7750b49dcd1a34c6", + "instance_gravitational__teleport-f432a71a13e698b6e1c4672a2e9e9c1f32d35c12", + "instance_protonmail__webclients-0d0267c4438cf378bda90bc85eed3a3615871ac4", + "instance_internetarchive__openlibrary-dbbd9d539c6d4fd45d5be9662aa19b6d664b5137-v08d8e8889ec945ab821fb156c04c7d2e2810debb", + "instance_flipt-io__flipt-b2170346dc37cf42fda1386cd630f24821ad2ac5", + "instance_element-hq__element-web-aec454dd6feeb93000380523cbb0b3681c0275fd-vnan", + "instance_internetarchive__openlibrary-4a5d2a7d24c9e4c11d3069220c0685b736d5ecde-v13642507b4fc1f8d234172bf8129942da2c2ca26", + "instance_qutebrowser__qutebrowser-ec2dcfce9eee9f808efc17a1b99e227fc4421dea-v5149fcda2a9a6fe1d35dfed1bade1444a11ef271", + "instance_gravitational__teleport-1330415d33a27594c948a36d9d7701f496229e9f", + "instance_protonmail__webclients-3a6790f480309130b5d6332dce6c9d5ccca13ee3", + "instance_flipt-io__flipt-492cc0b158200089dceede3b1aba0ed28df3fb1d", + "instance_future-architect__vuls-edb324c3d9ec3b107bf947f00e38af99d05b3e16", + "instance_ansible__ansible-bec27fb4c0a40c5f8bbcf26a475704227d65ee73-v30a923fb5c164d6cd18280c02422f75e611e8fb2", + "instance_gravitational__teleport-c1b1c6a1541c478d7777a48fca993cc8206c73b9", + "instance_tutao__tutanota-1e516e989b3c0221f4af6b297d9c0e4c43e4adc3-vbc0d9ba8f0071fbe982809910959a6ff8884dbbf", + "instance_flipt-io__flipt-406f9396ad65696d58865b3a6283109cd4eaf40e", + "instance_future-architect__vuls-139f3a81b66c47e6d8f70ce6c4afe7a9196a6ea8", + "instance_flipt-io__flipt-3b2c25ee8a3ac247c3fad13ad8d64ace34ec8ee7", + "instance_gravitational__teleport-0ac7334939981cf85b9591ac295c3816954e287e", + "instance_qutebrowser__qutebrowser-e15d26630934d0b6415ed2295ac42fd570a57620-va0fd88aac89cde702ec1ba84877234da33adce8a", + "instance_ansible__ansible-5e88cd9972f10b66dd97e1ee684c910c6a2dd25e-v906c969b551b346ef54a2c0b41e04f632b7b73c2", + "instance_element-hq__element-web-923ad4323b2006b2b180544429455ffe7d4a6cc3-vnan", + "instance_protonmail__webclients-4feccbc9990980aee26ea29035f8f931d6089895", + "instance_flipt-io__flipt-a0cbc0cb65ae601270bdbe3f5313e2dfd49c80e4", + "instance_gravitational__teleport-b4e7cd3a5e246736d3fe8d6886af55030b232277", + "instance_ansible__ansible-6cc97447aac5816745278f3735af128afb255c81-v0f01c69f1e2528b935359cfe578530722bca2c59", + "instance_ansible__ansible-f327e65d11bb905ed9f15996024f857a95592629-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_qutebrowser__qutebrowser-5fdc83e5da6222fe61163395baaad7ae57fa2cb4-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_navidrome__navidrome-3982ba725883e71d4e3e618c61d5140eeb8d850a", + "instance_qutebrowser__qutebrowser-ebfe9b7aa0c4ba9d451f993e08955004aaec4345-v059c6fdc75567943479b23ebca7c07b5e9a7f34c", + "instance_ansible__ansible-709484969c8a4ffd74b839a673431a8c5caa6457-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_gravitational__teleport-ad41b3c15414b28a6cec8c25424a19bfa7abd0e9-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_protonmail__webclients-2c3559cad02d1090985dba7e8eb5a129144d9811", + "instance_flipt-io__flipt-5af0757e96dec4962a076376d1bedc79de0d4249", + "instance_element-hq__element-web-56c7fc1948923b4b3f3507799e725ac16bcf8018-vnan", + "instance_flipt-io__flipt-381b90f718435c4694380b5fcd0d5cf8e3b5a25a", + "instance_flipt-io__flipt-6fd0f9e2587f14ac1fdd1c229f0bcae0468c8daa", + "instance_ansible__ansible-4c5ce5a1a9e79a845aff4978cfeb72a0d4ecf7d6-v1055803c3a812189a1133297f7f5468579283f86", + "instance_navidrome__navidrome-7073d18b54da7e53274d11c9e2baef1242e8769e", + "instance_internetarchive__openlibrary-0a90f9f0256e4f933523e9842799e39f95ae29ce-v76304ecdb3a5954fcf13feb710e8c40fcf24b73c", + "instance_nodebb__nodebb-be43cd25974681c9743d424238b7536c357dc8d3-vf2cf3cbd463b7ad942381f1c6d077626485a1e9e", + "instance_flipt-io__flipt-05d7234fa582df632f70a7cd10194d61bd7043b9", + "instance_nodebb__nodebb-76c6e30282906ac664f2c9278fc90999b27b1f48-vd59a5728dfc977f44533186ace531248c2917516", + "instance_qutebrowser__qutebrowser-6b320dc18662580e1313d2548fdd6231d2a97e6d-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_future-architect__vuls-01441351c3407abfc21c48a38e28828e1b504e0c", + "instance_ansible__ansible-1a4644ff15355fd696ac5b9d074a566a80fe7ca3-v30a923fb5c164d6cd18280c02422f75e611e8fb2", + "instance_navidrome__navidrome-09ae41a2da66264c60ef307882362d2e2d8d8b89", + "instance_internetarchive__openlibrary-5c6c22f3d2edf2f1b10f5dc335e32cb6a5f40341-v76304ecdb3a5954fcf13feb710e8c40fcf24b73c", + "instance_navidrome__navidrome-bf2bcb12799b21069f137749e0c331f761d1f693", + "instance_flipt-io__flipt-2ce8a0331e8a8f63f2c1b555db8277ffe5aa2e63", + "instance_ansible__ansible-a1569ea4ca6af5480cf0b7b3135f5e12add28a44-v0f01c69f1e2528b935359cfe578530722bca2c59", + "instance_gravitational__teleport-82185f232ae8974258397e121b3bc2ed0c3729ed-v626ec2a48416b10a88641359a169d99e935ff037", + "instance_flipt-io__flipt-72d06db14d58692bfb4d07b1aa745a37b35956f3", + "instance_ansible__ansible-b6290e1d156af608bd79118d209a64a051c55001-v390e508d27db7a51eece36bb6d9698b63a5b638a", + "instance_gravitational__teleport-32bcd71591c234f0d8b091ec01f1f5cbfdc0f13c-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_ansible__ansible-a02e22e902a69aeb465f16bf03f7f5a91b2cb828-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_gravitational__teleport-eda668c30d9d3b56d9c69197b120b01013611186", + "instance_nodebb__nodebb-0f788b8eaa4bba3c142d171fd941d015c53b65fc-v0ec6d6c2baf3cb4797482ce4829bc25cd5716649", + "instance_gravitational__teleport-47530e1fd8bfb84ec096ebcbbc29990f30829655-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_qutebrowser__qutebrowser-e64622cd2df5b521342cf4a62e0d4cb8f8c9ae5a-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_element-hq__element-web-494d9de6f0a94ffb491e74744d2735bce02dc0ab-vnan", + "instance_element-hq__element-web-9bf77963ee5e036d54b2a3ca202fbf6378464a5e-vnan", + "instance_nodebb__nodebb-9c576a0758690f45a6ca03b5884c601e473bf2c1-vd59a5728dfc977f44533186ace531248c2917516", + "instance_internetarchive__openlibrary-25858f9f0c165df25742acf8309ce909773f0cdd-v13642507b4fc1f8d234172bf8129942da2c2ca26", + "instance_protonmail__webclients-f161c10cf7d31abf82e8d64d7a99c9fac5acfa18", + "instance_gravitational__teleport-cb712e3f0b06dadc679f895daef8072cae400c26-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_gravitational__teleport-46a13210519461c7cec8d643bfbe750265775b41", + "instance_ansible__ansible-e40889e7112ae00a21a2c74312b330e67a766cc0-v1055803c3a812189a1133297f7f5468579283f86", + "instance_gravitational__teleport-6eaaf3a27e64f4ef4ef855bd35d7ec338cf17460-v626ec2a48416b10a88641359a169d99e935ff037", + "instance_internetarchive__openlibrary-0d13e6b4bf80bced6c0946b969b9a1b6963f6bce-v76304ecdb3a5954fcf13feb710e8c40fcf24b73c", + "instance_protonmail__webclients-2f2f6c311c6128fe86976950d3c0c2db07b03921", + "instance_ansible__ansible-5260527c4a71bfed99d803e687dd19619423b134-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_qutebrowser__qutebrowser-8d05f0282a271bfd45e614238bd1b555c58b3fc1-v35616345bb8052ea303186706cec663146f0f184", + "instance_qutebrowser__qutebrowser-50efac08f623644a85441bbe02ab9347d2b71a9d-v9f8e9d96c85c85a605e382f1510bd08563afc566", + "instance_flipt-io__flipt-dae029cba7cdb98dfb1a6b416c00d324241e6063", + "instance_nodebb__nodebb-18c45b44613aecd53e9f60457b9812049ab2998d-v0495b863a912fbff5749c67e860612b91825407c", + "instance_future-architect__vuls-e4728e388120b311c4ed469e4f942e0347a2689b-v264a82e2f4818e30f5a25e4da53b27ba119f62b5", + "instance_qutebrowser__qutebrowser-5e0d6dc1483cb3336ea0e3dcbd4fe4aa00fc1742-v5149fcda2a9a6fe1d35dfed1bade1444a11ef271", + "instance_flipt-io__flipt-c12967bc73fdf02054cf3ef8498c05e25f0a18c0", + "instance_qutebrowser__qutebrowser-ff1c025ad3210506fc76e1f604d8c8c27637d88e-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_ansible__ansible-83909bfa22573777e3db5688773bda59721962ad-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_qutebrowser__qutebrowser-1a9e74bfaf9a9db2a510dc14572d33ded6040a57-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_gravitational__teleport-24cafecd8721891092210afc55f6413ab46ca211-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_internetarchive__openlibrary-4b7ea2977be2747496ba792a678940baa985f7ea-v0f5aece3601a5b4419f7ccec1dbda2071be28ee4", + "instance_internetarchive__openlibrary-fad4a40acf5ff5f06cd7441a5c7baf41a7d81fe4-vfa6ff903cb27f336e17654595dd900fa943dcd91", + "instance_qutebrowser__qutebrowser-3fd8e12949b8feda401930574facf09dd4180bba", + "instance_ansible__ansible-164881d871964aa64e0f911d03ae270acbad253c-v390e508d27db7a51eece36bb6d9698b63a5b638a", + "instance_ansible__ansible-fb144c44144f8bd3542e71f5db62b6d322c7bd85-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_element-hq__element-web-582a1b093fc0b77538052f45cbb9c7295f991b51-vnan", + "instance_element-hq__element-web-ec0f940ef0e8e3b61078f145f34dc40d1938e6c5-vnan", + "instance_internetarchive__openlibrary-6fdbbeee4c0a7e976ff3e46fb1d36f4eb110c428-v08d8e8889ec945ab821fb156c04c7d2e2810debb", + "instance_flipt-io__flipt-7161f7b876773a911afdd804b281e52681cb7321", + "instance_navidrome__navidrome-28389fb05e1523564dfc61fa43ed8eb8a10f938c", + "instance_ansible__ansible-5f4e332e3762999d94af27746db29ff1729252c1-v0f01c69f1e2528b935359cfe578530722bca2c59", + "instance_nodebb__nodebb-0e07f3c9bace416cbab078a30eae972868c0a8a3-vf2cf3cbd463b7ad942381f1c6d077626485a1e9e", + "instance_nodebb__nodebb-da0211b1a001d45d73b4c84c6417a4f1b0312575-vf2cf3cbd463b7ad942381f1c6d077626485a1e9e", + "instance_navidrome__navidrome-31799662706fedddf5bcc1a76b50409d1f91d327", + "instance_future-architect__vuls-50580f6e98eeb36f53f27222f7f4fdfea0b21e8d", + "instance_internetarchive__openlibrary-60725705782832a2cb22e17c49697948a42a9d03-v298a7a812ceed28c4c18355a091f1b268fe56d86", + "instance_ansible__ansible-3db08adbb1cc6aa9941be5e0fc810132c6e1fa4b-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_internetarchive__openlibrary-8a5a63af6e0be406aa6c8c9b6d5f28b2f1b6af5a-v0f5aece3601a5b4419f7ccec1dbda2071be28ee4", + "instance_future-architect__vuls-e1df74cbc1a1d1889428b3333a3b2405c4651993", + "instance_internetarchive__openlibrary-123e6e5e1c85b9c07d1e98f70bfc480bc8016890-v2733ff199fb72f0d033a30dc62cb0a4742e3a7f4", + "instance_protonmail__webclients-ae36cb23a1682dcfd69587c1b311ae0227e28f39", + "instance_protonmail__webclients-8be4f6cb9380fcd2e67bcb18cef931ae0d4b869c", + "instance_qutebrowser__qutebrowser-a25e8a09873838ca9efefd36ea8a45170bbeb95c-vc2f56a753b62a190ddb23cd330c257b9cf560d12", + "instance_qutebrowser__qutebrowser-44e64199ed38003253f0296badd4a447645067b6-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_gravitational__teleport-7744f72c6eb631791434b648ba41083b5f6d2278-vce94f93ad1030e3136852817f2423c1b3ac37bc4", + "instance_tutao__tutanota-40e94dee2bcec2b63f362da283123e9df1874cc1-vc4e41fd0029957297843cb9dec4a25c7c756f029", + "instance_future-architect__vuls-c11ba27509f733d7d280bdf661cbbe2e7a99df4c", + "instance_flipt-io__flipt-b2cd6a6dd73ca91b519015fd5924fde8d17f3f06", + "instance_gravitational__teleport-6a14edcf1ff010172fdbac622d0a474ed6af46de", + "instance_flipt-io__flipt-36e62baffae2132f78f9d34dc300a9baa2d7ae0e", + "instance_navidrome__navidrome-c90468b895f6171e33e937ff20dc915c995274f0", + "instance_ansible__ansible-ecea15c508f0e081525be036cf76bbb56dbcdd9d-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_ansible__ansible-e22e103cdf8edc56ff7d9b848a58f94f1471a263-v1055803c3a812189a1133297f7f5468579283f86", + "instance_future-architect__vuls-dc496468b9e9fb73371f9606cdcdb0f8e12e70ca", + "instance_gravitational__teleport-c782838c3a174fdff80cafd8cd3b1aa4dae8beb2", + "instance_internetarchive__openlibrary-91efee627df01e32007abf2d6ebf73f9d9053076-vbee42ad1b72fb23c6a1c874868a720b370983ed2", + "instance_qutebrowser__qutebrowser-52708364b5f91e198defb022d1a5b4b3ebd9b563-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_ansible__ansible-eea46a0d1b99a6dadedbb6a3502d599235fa7ec3-v390e508d27db7a51eece36bb6d9698b63a5b638a", + "instance_protonmail__webclients-c6f65d205c401350a226bb005f42fac1754b0b5b", + "instance_ansible__ansible-cb94c0cc550df9e98f1247bc71d8c2b861c75049-v1055803c3a812189a1133297f7f5468579283f86", + "instance_nodebb__nodebb-05f2236193f407cf8e2072757fbd6bb170bc13f0-vf2cf3cbd463b7ad942381f1c6d077626485a1e9e", + "instance_navidrome__navidrome-55730514ea59d5f1d0b8e3f8745569c29bdbf7b4", + "instance_navidrome__navidrome-0130c6dc13438b48cf0fdfab08a89e357b5517c9", + "instance_flipt-io__flipt-507170da0f7f4da330f6732bffdf11c4df7fc192", + "instance_tutao__tutanota-fe240cbf7f0fdd6744ef7bef8cb61676bcdbb621-vc4e41fd0029957297843cb9dec4a25c7c756f029", + "instance_qutebrowser__qutebrowser-233cb1cc48635130e5602549856a6fa4ab4c087f-v35616345bb8052ea303186706cec663146f0f184", + "instance_internetarchive__openlibrary-f8cc11d9c1575fdba5ac66aee0befca970da8d64-v13642507b4fc1f8d234172bf8129942da2c2ca26", + "instance_qutebrowser__qutebrowser-fec187c2cb53d769c2682b35ca77858a811414a8-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_tutao__tutanota-da4edb7375c10f47f4ed3860a591c5e6557f7b5c-vbc0d9ba8f0071fbe982809910959a6ff8884dbbf", + "instance_ansible__ansible-be2c376ab87e3e872ca21697508f12c6909cf85a-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_future-architect__vuls-457a3a9627fb9a0800d0aecf1d4713fb634a9011", + "instance_internetarchive__openlibrary-d40ec88713dc95ea791b252f92d2f7b75e107440-v13642507b4fc1f8d234172bf8129942da2c2ca26", + "instance_protonmail__webclients-863d524b5717b9d33ce08a0f0535e3fd8e8d1ed8", + "instance_internetarchive__openlibrary-308a35d6999427c02b1dbf5211c033ad3b352556-ve8c8d62a2b60610a3c4631f5f23ed866bada9818", + "instance_flipt-io__flipt-15b76cada1ef29cfa56b0fba36754be36243dded", + "instance_future-architect__vuls-61c39637f2f3809e1b5dad05f0c57c799dce1587", + "instance_ansible__ansible-5d253a13807e884b7ce0b6b57a963a45e2f0322c-v1055803c3a812189a1133297f7f5468579283f86", + "instance_element-hq__element-web-44b98896a79ede48f5ad7ff22619a39d5f6ff03c-vnan", + "instance_qutebrowser__qutebrowser-f91ace96223cac8161c16dd061907e138fe85111-v059c6fdc75567943479b23ebca7c07b5e9a7f34c", + "instance_gravitational__teleport-bb562408da4adeae16e025be65e170959d1ec492-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_gravitational__teleport-78b0d8c72637df1129fb6ff84fc49ef4b5ab1288", + "instance_internetarchive__openlibrary-3f7db6bbbcc7c418b3db72d157c6aed1d45b2ccf-v430f20c722405e462d9ef44dee7d34c41e76fe7a", + "instance_future-architect__vuls-e52fa8d6ed1d23e36f2a86e5d3efe9aa057a1b0d", + "instance_navidrome__navidrome-3bc9e75b2843f91f6a1e9b604e321c2bd4fd442a", + "instance_tutao__tutanota-fb32e5f9d9fc152a00144d56dd0af01760a2d4dc-vc4e41fd0029957297843cb9dec4a25c7c756f029", + "instance_gravitational__teleport-af5e2517de7d18406b614e413aca61c319312171-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_ansible__ansible-a20a52701402a12f91396549df04ac55809f68e9-v1055803c3a812189a1133297f7f5468579283f86", + "instance_protonmail__webclients-1501eb765873b2884b6f1944fd242ecfc9d6b103", + "instance_gravitational__teleport-db89206db6c2969266e664c7c0fb51b70e958b64", + "instance_ansible__ansible-be59caa59bf47ca78a4760eb7ff38568372a8260-v1055803c3a812189a1133297f7f5468579283f86", + "instance_flipt-io__flipt-65581fef4aa807540cb933753d085feb0d7e736f", + "instance_gravitational__teleport-4e1c39639edf1ab494dd7562844c8b277b5cfa18-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_future-architect__vuls-f0b3a8b1db98eb1bd32685f1c36c41a99c3452ed", + "instance_navidrome__navidrome-b65e76293a917ee2dfc5d4b373b1c62e054d0dca", + "instance_element-hq__element-web-33e8edb3d508d6eefb354819ca693b7accc695e7", + "instance_flipt-io__flipt-e50808c03e4b9d25a6a78af9c61a3b1616ea356b", + "instance_future-architect__vuls-e3c27e1817d68248043bd09d63cc31f3344a6f2c", + "instance_element-hq__element-web-ad26925bb6628260cfe0fcf90ec0a8cba381f4a4-vnan", + "instance_nodebb__nodebb-1ea9481af6125ffd6da0592ed439aa62af0bca11-vd59a5728dfc977f44533186ace531248c2917516", + "instance_ansible__ansible-489156378c8e97374a75a544c7c9c2c0dd8146d1-v390e508d27db7a51eece36bb6d9698b63a5b638a", + "instance_internetarchive__openlibrary-bb152d23c004f3d68986877143bb0f83531fe401-ve8c8d62a2b60610a3c4631f5f23ed866bada9818", + "instance_qutebrowser__qutebrowser-0b621cb0ce2b54d3f93d8d41d8ff4257888a87e5-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_qutebrowser__qutebrowser-1af602b258b97aaba69d2585ed499d95e2303ac2-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_future-architect__vuls-e1fab805afcfc92a2a615371d0ec1e667503c254-v264a82e2f4818e30f5a25e4da53b27ba119f62b5", + "instance_internetarchive__openlibrary-7edd1ef09d91fe0b435707633c5cc9af41dedddf-v76304ecdb3a5954fcf13feb710e8c40fcf24b73c", + "instance_qutebrowser__qutebrowser-99029144b5109bb1b2a53964a7c129e009980cd9-va0fd88aac89cde702ec1ba84877234da33adce8a", + "instance_ansible__ansible-12734fa21c08a0ce8c84e533abdc560db2eb1955-v7eee2454f617569fd6889f2211f75bc02a35f9f8", + "instance_protonmail__webclients-7e54526774e577c0ebb58ced7ba8bef349a69fec", + "instance_gravitational__teleport-1a77b7945a022ab86858029d30ac7ad0d5239d00-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_nodebb__nodebb-8ca65b0c78c67c1653487c02d1135e1b702185e1-vf2cf3cbd463b7ad942381f1c6d077626485a1e9e", + "instance_flipt-io__flipt-9f8127f225a86245fa35dca4885c2daef824ee55", + "instance_gravitational__teleport-ac2fb2f9b4fd1896b554d3011df23d3d71295779", + "instance_ansible__ansible-f02a62db509dc7463fab642c9c3458b9bc3476cc-v390e508d27db7a51eece36bb6d9698b63a5b638a" + ], + "additional_details": { + "shuffled": "True" + } + }, + "evaluation_timestamp": "1775602861.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.56, + "uncertainty": { + "standard_error": { + "value": 0.03518793763172075 + }, + "num_samples": 200 + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"If you do not submit an answer, you will receive a score of zero.\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [10, \"Your answer has been saved. Consider whether your previous approach is likely correct or whether a different approach would be more promising. Then either refine your previous answer or try a substantially different approach, and submit when ready.\", \"\"], \"submit\": null, \"on_continue\": null, \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\"}, \"params_passed\": {\"prompt\": [\"If you do not submit an answer, you will receive a score of zero.\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [10, \"Your answer has been saved. Consider whether your previous approach is likely correct or whether a different approach would be more promising. Then either refine your previous answer or try a substantially different approach, and submit when ready.\", \"\"], \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 20000000 + }, + "sandbox": { + "type": "k8s" + } + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/anthropic/claude-opus-4-6/76e9535b-4105-4411-9115-843e0cc78b33_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "d30b9454bf4294b36c27be7a85193c4950b212c8383f529300bedf7daa91aaa0", + "total_rows": 200 + } +} \ No newline at end of file diff --git a/data/swebenchpro/anthropic/claude-opus-4-6/76e9535b-4105-4411-9115-843e0cc78b33_samples.jsonl b/data/swebenchpro/anthropic/claude-opus-4-6/76e9535b-4105-4411-9115-843e0cc78b33_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..95057276ebccd8761ae965b3e46c2fff4f313b09 --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-6/76e9535b-4105-4411-9115-843e0cc78b33_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d30b9454bf4294b36c27be7a85193c4950b212c8383f529300bedf7daa91aaa0 +size 107676552 diff --git a/data/swebenchpro/anthropic/claude-opus-4-6/76fc88bd-e7b8-4ce4-8416-de8553a855e3.json b/data/swebenchpro/anthropic/claude-opus-4-6/76fc88bd-e7b8-4ce4-8416-de8553a855e3.json new file mode 100644 index 0000000000000000000000000000000000000000..0fca94ff78195d8cadb43d81def83e0cbc5cb79b --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-6/76fc88bd-e7b8-4ce4-8416-de8553a855e3.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/anthropic_claude-opus-4-6/1786167435.70252", + "evaluation_timestamp": "1778559237.0", + "retrieved_timestamp": "1786167435.70252", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-6", + "id": "anthropic/claude-opus-4-6", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/anthropic/claude-opus-4-6/76fc88bd-e7b8-4ce4-8416-de8553a855e3_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "dd5b593f66587bb35ebad15b85e09b7783a48dd2b876fcc303ff5e6836266f4c", + "total_rows": 82 + } +} \ No newline at end of file diff --git a/data/swebenchpro/anthropic/claude-opus-4-6/76fc88bd-e7b8-4ce4-8416-de8553a855e3_samples.jsonl b/data/swebenchpro/anthropic/claude-opus-4-6/76fc88bd-e7b8-4ce4-8416-de8553a855e3_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..0b8fb215ff3d03abc72d754da73c327ebf82e8ab --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-6/76fc88bd-e7b8-4ce4-8416-de8553a855e3_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dd5b593f66587bb35ebad15b85e09b7783a48dd2b876fcc303ff5e6836266f4c +size 22051381 diff --git a/data/swebenchpro/anthropic/claude-opus-4-6/780e0862-7b0a-4fb9-8da2-f4e77b98af28.json b/data/swebenchpro/anthropic/claude-opus-4-6/780e0862-7b0a-4fb9-8da2-f4e77b98af28.json new file mode 100644 index 0000000000000000000000000000000000000000..b49af9db6ae263f06b48fc0f48671b9d798e61c3 --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-6/780e0862-7b0a-4fb9-8da2-f4e77b98af28.json @@ -0,0 +1,116 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/anthropic_claude-opus-4-6/1786163145.250454", + "evaluation_timestamp": "1778703021.0", + "retrieved_timestamp": "1786163145.250454", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-6", + "id": "anthropic/claude-opus-4-6", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on swebenchpro/S-adaptive+C/+1ep/e281512c for scorer harbor_scorer", + "source_data": { + "dataset_name": "swebenchpro", + "source_type": "hf_dataset", + "samples_number": 7, + "sample_ids": [ + "instance_qutebrowser__qutebrowser-2e961080a85d660148937ee8f0f6b3445a8f2c01-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_qutebrowser__qutebrowser-01d1d1494411380d97cac14614a829d3a69cecaf-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_element-hq__element-web-f3534b42df3dcfe36dc48bddbf14034085af6d30-vnan", + "instance_internetarchive__openlibrary-fdbc0d8f418333c7e575c40b661b582c301ef7ac-v13642507b4fc1f8d234172bf8129942da2c2ca26", + "instance_element-hq__element-web-b007ea81b2ccd001b00f332bee65070aa7fc00f9-vnan", + "instance_nodebb__nodebb-0f788b8eaa4bba3c142d171fd941d015c53b65fc-v0ec6d6c2baf3cb4797482ce4829bc25cd5716649", + "instance_protonmail__webclients-f161c10cf7d31abf82e8d64d7a99c9fac5acfa18" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1778703021.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 1.0, + "uncertainty": { + "num_samples": 7 + }, + "details": { + "total_matched_trajectories": "7", + "stopping_reason_count_completed_on_successful_submit": "7", + "avg_total_tokens_target_model": "731446.57", + "total_total_tokens_target_model": "5120126.0", + "avg_total_tokens_other_models": "0.00", + "total_total_tokens_other_models": "0.0", + "avg_total_tokens_all_models": "731446.57", + "total_total_tokens_all_models": "5120126.0", + "avg_turn_count": "39.29", + "total_turn_count": "275.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "8", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 30000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "8", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/anthropic/claude-opus-4-6/780e0862-7b0a-4fb9-8da2-f4e77b98af28_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "b13ee3a09ec0486c00a3639cacdd5ce26d65c025129e5b0a0663b05a35474f79", + "total_rows": 7 + } +} \ No newline at end of file diff --git a/data/swebenchpro/anthropic/claude-opus-4-6/780e0862-7b0a-4fb9-8da2-f4e77b98af28_samples.jsonl b/data/swebenchpro/anthropic/claude-opus-4-6/780e0862-7b0a-4fb9-8da2-f4e77b98af28_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..b9728a9826cb5863a9a19de39b7546031773dbee --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-6/780e0862-7b0a-4fb9-8da2-f4e77b98af28_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b13ee3a09ec0486c00a3639cacdd5ce26d65c025129e5b0a0663b05a35474f79 +size 1072605 diff --git a/data/swebenchpro/anthropic/claude-opus-4-6/ac1a8d68-ad20-497a-99a0-4ff237e2c045.json b/data/swebenchpro/anthropic/claude-opus-4-6/ac1a8d68-ad20-497a-99a0-4ff237e2c045.json new file mode 100644 index 0000000000000000000000000000000000000000..1119e65a60d63d751d4fe95484096ca378f33043 --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-6/ac1a8d68-ad20-497a-99a0-4ff237e2c045.json @@ -0,0 +1,154 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/anthropic_claude-opus-4-6/1786163165.262504", + "evaluation_timestamp": "1778704859.0", + "retrieved_timestamp": "1786163165.262504", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-6", + "id": "anthropic/claude-opus-4-6", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on swebenchpro/S-adaptive+C/+3ep/9bb10350 for scorer harbor_scorer", + "source_data": { + "dataset_name": "swebenchpro", + "source_type": "hf_dataset", + "samples_number": 41, + "sample_ids": [ + "instance_flipt-io__flipt-b68b8960b8a08540d5198d78c665a7eb0bea4008", + "instance_future-architect__vuls-78b52d6a7f480bd610b692de9bf0c86f57332f23", + "instance_flipt-io__flipt-e42da21a07a5ae35835ec54f74004ebd58713874", + "instance_navidrome__navidrome-d8e794317f788198227e10fb667e10496b3eb99a", + "instance_future-architect__vuls-f0b3a8b1db98eb1bd32685f1c36c41a99c3452ed", + "instance_future-architect__vuls-6eff6a9329a65cc412e79b8f82444dfa3d0f0b5a", + "instance_ansible__ansible-b2a289dcbb702003377221e25f62c8a3608f0e89-v173091e2e36d38c978002990795f66cfc0af30ad", + "instance_internetarchive__openlibrary-f8cc11d9c1575fdba5ac66aee0befca970da8d64-v13642507b4fc1f8d234172bf8129942da2c2ca26", + "instance_qutebrowser__qutebrowser-e70f5b03187bdd40e8bf70f5f3ead840f52d1f42-v02ad04386d5238fe2d1a1be450df257370de4b6a", + "instance_gravitational__teleport-ac2fb2f9b4fd1896b554d3011df23d3d71295779", + "instance_nodebb__nodebb-70b4a0e2aebebe8f2f559de6680093d96a697b2f-vnan", + "instance_ansible__ansible-e22e103cdf8edc56ff7d9b848a58f94f1471a263-v1055803c3a812189a1133297f7f5468579283f86", + "instance_ansible__ansible-164881d871964aa64e0f911d03ae270acbad253c-v390e508d27db7a51eece36bb6d9698b63a5b638a", + "instance_internetarchive__openlibrary-9c392b60e2c6fa1d68cb68084b4b4ff04d0cb35c-v2d9a6c849c60ed19fd0858ce9e40b7cc8e097e59", + "instance_ansible__ansible-811093f0225caa4dd33890933150a81c6a6d5226-v1055803c3a812189a1133297f7f5468579283f86", + "instance_flipt-io__flipt-cd2f3b0a9d4d8b8a6d3d56afab65851ecdc408e8", + "instance_element-hq__element-web-f63160f38459fb552d00fcc60d4064977a9095a6-vnan", + "instance_ansible__ansible-f327e65d11bb905ed9f15996024f857a95592629-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_gravitational__teleport-73cc189b0e9636d418c4470ecce0d9af5dae2f02-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_future-architect__vuls-878c25bf5a9c9fd88ac32eb843f5636834d5712d", + "instance_element-hq__element-web-9bf77963ee5e036d54b2a3ca202fbf6378464a5e-vnan", + "instance_qutebrowser__qutebrowser-70248f256f93ed9b1984494d0a1a919ddd774892-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_flipt-io__flipt-b2cd6a6dd73ca91b519015fd5924fde8d17f3f06", + "instance_flipt-io__flipt-e2bd19dafa7166c96b082fb2a59eb54b4be0d778", + "instance_element-hq__element-web-75c2c1a572fa45d1ea1d1a96e9e36e303332ecaa-vnan", + "instance_element-hq__element-web-2760bfc8369f1bee640d6d7a7e910783143d4c5f-vnan", + "instance_internetarchive__openlibrary-25858f9f0c165df25742acf8309ce909773f0cdd-v13642507b4fc1f8d234172bf8129942da2c2ca26", + "instance_protonmail__webclients-01ea5214d11e0df8b7170d91bafd34f23cb0f2b1", + "instance_flipt-io__flipt-05d7234fa582df632f70a7cd10194d61bd7043b9", + "instance_protonmail__webclients-2c3559cad02d1090985dba7e8eb5a129144d9811", + "instance_protonmail__webclients-01b519cd49e6a24d9a05d2eb97f54e420740072e", + "instance_protonmail__webclients-ac23d1efa1a6ab7e62724779317ba44c28d78cfd", + "instance_tutao__tutanota-8513a9e8114a8b42e64f4348335e0f23efa054c4-vee878bb72091875e912c52fc32bc60ec3760227b", + "instance_internetarchive__openlibrary-7f6b722a10f822171501d027cad60afe53337732-ve8c8d62a2b60610a3c4631f5f23ed866bada9818", + "instance_internetarchive__openlibrary-1351c59fd43689753de1fca32c78d539a116ffc1-v29f82c9cf21d57b242f8d8b0e541525d259e2d63", + "instance_element-hq__element-web-1077729a19c0ce902e713cf6fab42c91fb7907f1-vnan", + "instance_future-architect__vuls-be7b9114cc9545e68fb0ee7bc63d7ec53d1a00ad", + "instance_gravitational__teleport-3ff19cf7c41f396ae468797d3aeb61515517edc9-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_future-architect__vuls-aaea15e516ece43978cf98e09e52080478b1d39f", + "instance_internetarchive__openlibrary-3f580a5f244c299d936d73d9e327ba873b6401d9-v0f5aece3601a5b4419f7ccec1dbda2071be28ee4", + "instance_qutebrowser__qutebrowser-0aa57e4f7243024fa4bba8853306691b5dbd77b3-v5149fcda2a9a6fe1d35dfed1bade1444a11ef271" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1778704859.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.6097560975609756, + "uncertainty": { + "standard_error": { + "value": 0.07712872341874095 + }, + "num_samples": 123 + }, + "details": { + "total_matched_trajectories": "123", + "stopping_reason_count_completed_on_successful_submit": "75", + "stopping_reason_count_repetition_guard": "48", + "avg_total_tokens_target_model": "3941133.56", + "total_total_tokens_target_model": "484759428.0", + "avg_total_tokens_other_models": "455.85", + "total_total_tokens_other_models": "56070.0", + "avg_total_tokens_all_models": "3941589.41", + "total_total_tokens_all_models": "484815498.0", + "avg_turn_count": "83.28", + "total_turn_count": "10243.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "8", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 30000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "8", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/anthropic/claude-opus-4-6/ac1a8d68-ad20-497a-99a0-4ff237e2c045_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "a28e10af8c713ae2dc48ab50e5a0d0ed6c1bb97a5463de1f0f661b0d7ae165ff", + "total_rows": 123 + } +} \ No newline at end of file diff --git a/data/swebenchpro/anthropic/claude-opus-4-6/ac1a8d68-ad20-497a-99a0-4ff237e2c045_samples.jsonl b/data/swebenchpro/anthropic/claude-opus-4-6/ac1a8d68-ad20-497a-99a0-4ff237e2c045_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..13d46dd42d0d2d10c83531c7d1a8911de125b966 --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-6/ac1a8d68-ad20-497a-99a0-4ff237e2c045_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a28e10af8c713ae2dc48ab50e5a0d0ed6c1bb97a5463de1f0f661b0d7ae165ff +size 42498403 diff --git a/data/swebenchpro/anthropic/claude-opus-4-6/b22b8bc7-baa1-4791-9246-e27683cb39f2.json b/data/swebenchpro/anthropic/claude-opus-4-6/b22b8bc7-baa1-4791-9246-e27683cb39f2.json new file mode 100644 index 0000000000000000000000000000000000000000..f3d2e8b3162fd5d78b851623dfc1a41174f3a462 --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-6/b22b8bc7-baa1-4791-9246-e27683cb39f2.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/anthropic_claude-opus-4-6/1786167422.67993", + "evaluation_timestamp": "1778665168.0", + "retrieved_timestamp": "1786167422.67993", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-6", + "id": "anthropic/claude-opus-4-6", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/anthropic/claude-opus-4-6/b22b8bc7-baa1-4791-9246-e27683cb39f2_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "9dd9a912cd66254dce64b71b0b1f7db521e3f83f67d84cd04131c44e2f0ced1f", + "total_rows": 1 + } +} \ No newline at end of file diff --git a/data/swebenchpro/anthropic/claude-opus-4-6/b22b8bc7-baa1-4791-9246-e27683cb39f2_samples.jsonl b/data/swebenchpro/anthropic/claude-opus-4-6/b22b8bc7-baa1-4791-9246-e27683cb39f2_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..483cde89e3825242378c36635de3b442d780e9e8 --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-6/b22b8bc7-baa1-4791-9246-e27683cb39f2_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9dd9a912cd66254dce64b71b0b1f7db521e3f83f67d84cd04131c44e2f0ced1f +size 93271 diff --git a/data/swebenchpro/anthropic/claude-opus-4-6/b4525dd4-443a-4c95-9a54-ae34198baa07.json b/data/swebenchpro/anthropic/claude-opus-4-6/b4525dd4-443a-4c95-9a54-ae34198baa07.json new file mode 100644 index 0000000000000000000000000000000000000000..7b7762bfeff9dcbe50cfb222dab4a847ba5c07b5 --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-6/b4525dd4-443a-4c95-9a54-ae34198baa07.json @@ -0,0 +1,136 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/anthropic_claude-opus-4-6/1786163149.197734", + "evaluation_timestamp": "1778703021.0", + "retrieved_timestamp": "1786163149.197734", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-6", + "id": "anthropic/claude-opus-4-6", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on swebenchpro/S-adaptive+C/+2ep/7943e937 for scorer harbor_scorer", + "source_data": { + "dataset_name": "swebenchpro", + "source_type": "hf_dataset", + "samples_number": 23, + "sample_ids": [ + "instance_nodebb__nodebb-84e065752f6d7fbe5c08cbf50cb173ffb866b8fa-vf2cf3cbd463b7ad942381f1c6d077626485a1e9e", + "instance_ansible__ansible-29aea9ff3466e4cd2ed00524b9e56738d568ce8b-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_qutebrowser__qutebrowser-e15d26630934d0b6415ed2295ac42fd570a57620-va0fd88aac89cde702ec1ba84877234da33adce8a", + "instance_ansible__ansible-502270c804c33d3bc963930dc85e0f4ca359674d-v7eee2454f617569fd6889f2211f75bc02a35f9f8", + "instance_internetarchive__openlibrary-3c48b4bb782189e0858e6c3fc7956046cf3e1cfb-v2d9a6c849c60ed19fd0858ce9e40b7cc8e097e59", + "instance_qutebrowser__qutebrowser-cc360cd4a34a126274c7b51f3b63afbaf3e05a02-v5fc38aaf22415ab0b70567368332beee7955b367", + "instance_qutebrowser__qutebrowser-9ed748effa8f3bcd804612d9291da017b514e12f-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_ansible__ansible-1ee70fc272aff6bf3415357c6e13c5de5b928d9b-v1055803c3a812189a1133297f7f5468579283f86", + "instance_qutebrowser__qutebrowser-f631cd4422744160d9dcf7a0455da532ce973315-v35616345bb8052ea303186706cec663146f0f184", + "instance_ansible__ansible-3db08adbb1cc6aa9941be5e0fc810132c6e1fa4b-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_qutebrowser__qutebrowser-e5340c449f23608803c286da0563b62f58ba25b0-v059c6fdc75567943479b23ebca7c07b5e9a7f34c", + "instance_internetarchive__openlibrary-09865f5fb549694d969f0a8e49b9d204ef1853ca-ve8c8d62a2b60610a3c4631f5f23ed866bada9818", + "instance_internetarchive__openlibrary-c506c1b0b678892af5cb22c1c1dbc35d96787a0a-v0f5aece3601a5b4419f7ccec1dbda2071be28ee4", + "instance_tutao__tutanota-4b4e45949096bb288f2b522f657610e480efa3e8-vee878bb72091875e912c52fc32bc60ec3760227b", + "instance_internetarchive__openlibrary-30bc73a1395fba2300087c7f307e54bb5372b60a-v76304ecdb3a5954fcf13feb710e8c40fcf24b73c", + "instance_nodebb__nodebb-22368b996ee0e5f11a5189b400b33af3cc8d925a-v4fbcfae8b15e4ce5d132c408bca69ebb9cf146ed", + "instance_ansible__ansible-77658704217d5f166404fc67997203c25381cb6e-v390e508d27db7a51eece36bb6d9698b63a5b638a", + "instance_ansible__ansible-cd9c4eb5a6b2bfaf4a6709f001ce3d0c92c1eed2-v0f01c69f1e2528b935359cfe578530722bca2c59", + "instance_qutebrowser__qutebrowser-ef5ba1a0360b39f9eff027fbdc57f363597c3c3b-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_ansible__ansible-622a493ae03bd5e5cf517d336fc426e9d12208c7-v906c969b551b346ef54a2c0b41e04f632b7b73c2", + "instance_ansible__ansible-9a21e247786ebd294dafafca1105fcd770ff46c6-v67cdaa49f89b34e42b69d5b7830b3c3ad3d8803f", + "instance_flipt-io__flipt-b22f5f02e40b225b6b93fff472914973422e97c6", + "instance_future-architect__vuls-50580f6e98eeb36f53f27222f7f4fdfea0b21e8d" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1778703021.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.9130434782608695, + "uncertainty": { + "standard_error": { + "value": 0.060073850409370216 + }, + "num_samples": 46 + }, + "details": { + "total_matched_trajectories": "46", + "stopping_reason_count_completed_on_successful_submit": "42", + "stopping_reason_count_repetition_guard": "4", + "avg_total_tokens_target_model": "695829.41", + "total_total_tokens_target_model": "32008153.0", + "avg_total_tokens_other_models": "81.41", + "total_total_tokens_other_models": "3745.0", + "avg_total_tokens_all_models": "695910.83", + "total_total_tokens_all_models": "32011898.0", + "avg_turn_count": "30.76", + "total_turn_count": "1415.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "8", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 30000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "8", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/anthropic/claude-opus-4-6/b4525dd4-443a-4c95-9a54-ae34198baa07_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "97cfbda258e72c3c60b751d985e8eadb8b91baeed8f609533cd83d909aa4965b", + "total_rows": 46 + } +} \ No newline at end of file diff --git a/data/swebenchpro/anthropic/claude-opus-4-6/b4525dd4-443a-4c95-9a54-ae34198baa07_samples.jsonl b/data/swebenchpro/anthropic/claude-opus-4-6/b4525dd4-443a-4c95-9a54-ae34198baa07_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..7a70d2bf4aa1d6b3937a0a7da49e182cb8c7fbe0 --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-6/b4525dd4-443a-4c95-9a54-ae34198baa07_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:97cfbda258e72c3c60b751d985e8eadb8b91baeed8f609533cd83d909aa4965b +size 6725284 diff --git a/data/swebenchpro/anthropic/claude-opus-4-6/bc2f04f0-bbeb-4160-8746-da5a7d97681f.json b/data/swebenchpro/anthropic/claude-opus-4-6/bc2f04f0-bbeb-4160-8746-da5a7d97681f.json new file mode 100644 index 0000000000000000000000000000000000000000..9216734c558ffcfd34ceb35cbd8b68ec490bf6aa --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-6/bc2f04f0-bbeb-4160-8746-da5a7d97681f.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/anthropic_claude-opus-4-6/1786167458.50664", + "evaluation_timestamp": "1778590202.0", + "retrieved_timestamp": "1786167458.50664", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-6", + "id": "anthropic/claude-opus-4-6", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/anthropic/claude-opus-4-6/bc2f04f0-bbeb-4160-8746-da5a7d97681f_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "0b40b8f9bf90c5d29cb23c63599fc1892ee36774ae7b0257226313e176586abf", + "total_rows": 100 + } +} \ No newline at end of file diff --git a/data/swebenchpro/anthropic/claude-opus-4-6/bc2f04f0-bbeb-4160-8746-da5a7d97681f_samples.jsonl b/data/swebenchpro/anthropic/claude-opus-4-6/bc2f04f0-bbeb-4160-8746-da5a7d97681f_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..a0164f74e38563692d4139ec426f241671eff5c8 --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-6/bc2f04f0-bbeb-4160-8746-da5a7d97681f_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0b40b8f9bf90c5d29cb23c63599fc1892ee36774ae7b0257226313e176586abf +size 20396454 diff --git a/data/swebenchpro/anthropic/claude-opus-4-6/f26f1ce1-2163-4648-b868-b1044208d07b.json b/data/swebenchpro/anthropic/claude-opus-4-6/f26f1ce1-2163-4648-b868-b1044208d07b.json new file mode 100644 index 0000000000000000000000000000000000000000..b56224457e543e0d099b2f40dee0dbe4492b26c7 --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-6/f26f1ce1-2163-4648-b868-b1044208d07b.json @@ -0,0 +1,291 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/anthropic_claude-opus-4-6/1786182503.699074", + "evaluation_timestamp": "1775482110.0", + "retrieved_timestamp": "1786182503.699074", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev7+g741aa0d2,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-6", + "id": "anthropic/claude-opus-4-6", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on eval_task for scorer harbor_scorer", + "source_data": { + "dataset_name": "swebenchpro", + "source_type": "hf_dataset", + "samples_number": 731, + "sample_ids": [ + "instance_flipt-io__flipt-e91615cf07966da41756017a7d571f9fc0fdbe80", + "instance_ansible__ansible-8127abbc298cabf04aaa89a478fc5e5e3432a6fc-v30a923fb5c164d6cd18280c02422f75e611e8fb2", + "instance_protonmail__webclients-da91f084c0f532d9cc8ca385a701274d598057b8", + "instance_qutebrowser__qutebrowser-deeb15d6f009b3ca0c3bd503a7cef07462bd16b4-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_internetarchive__openlibrary-5069b09e5f64428dce59b33455c8bb17fe577070-v8717e18970bcdc4e0d2cea3b1527752b21e74866", + "instance_protonmail__webclients-c5a2089ca2bfe9aa1d85a664b8ad87ef843a1c9c", + "instance_qutebrowser__qutebrowser-fd6790fe8c02b144ab2464f1fc8ab3d02ce3c476-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_navidrome__navidrome-eebfbc5381a1e506ff17b5f1371d1ad83d5fd642", + "instance_navidrome__navidrome-812dc2090f20ac4f8ac271b6ed95be5889d1a3ca", + "instance_nodebb__nodebb-0c81642997ea1d827dbd02c311db9d4976112cd4-vf2cf3cbd463b7ad942381f1c6d077626485a1e9e", + "instance_internetarchive__openlibrary-c506c1b0b678892af5cb22c1c1dbc35d96787a0a-v0f5aece3601a5b4419f7ccec1dbda2071be28ee4", + "instance_flipt-io__flipt-5c7037ececb0bead0a8eb56054e224bcd7ac5922", + "instance_qutebrowser__qutebrowser-01d1d1494411380d97cac14614a829d3a69cecaf-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_internetarchive__openlibrary-7f7e53aa4cf74a4f8549a5bcd4810c527e2f6d7e-v13642507b4fc1f8d234172bf8129942da2c2ca26", + "instance_internetarchive__openlibrary-53e02a22972e9253aeded0e1981e6845e1e521fe-vfa6ff903cb27f336e17654595dd900fa943dcd91", + "instance_ansible__ansible-34db57a47f875d11c4068567b9ec7ace174ec4cf-v1055803c3a812189a1133297f7f5468579283f86", + "instance_qutebrowser__qutebrowser-996487c43e4fcc265b541f9eca1e7930e3c5cf05-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_gravitational__teleport-96019ce0be7a2c8e36363f359eb7c943b41dde70", + "instance_protonmail__webclients-df60460f163fd5c34e844ab9015e3176f1ab1ac0", + "instance_future-architect__vuls-ad2edbb8448e2c41a097f1c0b52696c0f6c5924d", + "instance_ansible__ansible-9142be2f6cabbe6597c9254c5bb9186d17036d55-v0f01c69f1e2528b935359cfe578530722bca2c59", + "instance_element-hq__element-web-33299af5c9b7a7ec5a9c31d578d4ec5b18088fb7-vnan", + "instance_element-hq__element-web-772df3021201d9c73835a626df8dcb6334ad9a3e-vnan", + "instance_internetarchive__openlibrary-1be7de788a444f6255e89c10ef6aa608550604a8-v29f82c9cf21d57b242f8d8b0e541525d259e2d63", + "instance_internetarchive__openlibrary-630221ab686c64e75a2ce253c893c033e4814b2e-v93c53c13d5f9b383ebb411ee7750b49dcd1a34c6", + "instance_gravitational__teleport-f432a71a13e698b6e1c4672a2e9e9c1f32d35c12", + "instance_protonmail__webclients-0d0267c4438cf378bda90bc85eed3a3615871ac4", + "instance_internetarchive__openlibrary-dbbd9d539c6d4fd45d5be9662aa19b6d664b5137-v08d8e8889ec945ab821fb156c04c7d2e2810debb", + "instance_flipt-io__flipt-b2170346dc37cf42fda1386cd630f24821ad2ac5", + "instance_element-hq__element-web-aec454dd6feeb93000380523cbb0b3681c0275fd-vnan", + "instance_internetarchive__openlibrary-4a5d2a7d24c9e4c11d3069220c0685b736d5ecde-v13642507b4fc1f8d234172bf8129942da2c2ca26", + "instance_qutebrowser__qutebrowser-ec2dcfce9eee9f808efc17a1b99e227fc4421dea-v5149fcda2a9a6fe1d35dfed1bade1444a11ef271", + "instance_gravitational__teleport-1330415d33a27594c948a36d9d7701f496229e9f", + "instance_protonmail__webclients-3a6790f480309130b5d6332dce6c9d5ccca13ee3", + "instance_flipt-io__flipt-492cc0b158200089dceede3b1aba0ed28df3fb1d", + "instance_future-architect__vuls-edb324c3d9ec3b107bf947f00e38af99d05b3e16", + "instance_ansible__ansible-bec27fb4c0a40c5f8bbcf26a475704227d65ee73-v30a923fb5c164d6cd18280c02422f75e611e8fb2", + "instance_gravitational__teleport-c1b1c6a1541c478d7777a48fca993cc8206c73b9", + "instance_tutao__tutanota-1e516e989b3c0221f4af6b297d9c0e4c43e4adc3-vbc0d9ba8f0071fbe982809910959a6ff8884dbbf", + "instance_flipt-io__flipt-406f9396ad65696d58865b3a6283109cd4eaf40e", + "instance_future-architect__vuls-139f3a81b66c47e6d8f70ce6c4afe7a9196a6ea8", + "instance_flipt-io__flipt-3b2c25ee8a3ac247c3fad13ad8d64ace34ec8ee7", + "instance_gravitational__teleport-0ac7334939981cf85b9591ac295c3816954e287e", + "instance_qutebrowser__qutebrowser-e15d26630934d0b6415ed2295ac42fd570a57620-va0fd88aac89cde702ec1ba84877234da33adce8a", + "instance_ansible__ansible-5e88cd9972f10b66dd97e1ee684c910c6a2dd25e-v906c969b551b346ef54a2c0b41e04f632b7b73c2", + "instance_element-hq__element-web-923ad4323b2006b2b180544429455ffe7d4a6cc3-vnan", + "instance_protonmail__webclients-4feccbc9990980aee26ea29035f8f931d6089895", + "instance_flipt-io__flipt-a0cbc0cb65ae601270bdbe3f5313e2dfd49c80e4", + "instance_gravitational__teleport-b4e7cd3a5e246736d3fe8d6886af55030b232277", + "instance_ansible__ansible-6cc97447aac5816745278f3735af128afb255c81-v0f01c69f1e2528b935359cfe578530722bca2c59", + "instance_ansible__ansible-f327e65d11bb905ed9f15996024f857a95592629-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_qutebrowser__qutebrowser-5fdc83e5da6222fe61163395baaad7ae57fa2cb4-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_navidrome__navidrome-3982ba725883e71d4e3e618c61d5140eeb8d850a", + "instance_qutebrowser__qutebrowser-ebfe9b7aa0c4ba9d451f993e08955004aaec4345-v059c6fdc75567943479b23ebca7c07b5e9a7f34c", + "instance_ansible__ansible-709484969c8a4ffd74b839a673431a8c5caa6457-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_gravitational__teleport-ad41b3c15414b28a6cec8c25424a19bfa7abd0e9-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_protonmail__webclients-2c3559cad02d1090985dba7e8eb5a129144d9811", + "instance_flipt-io__flipt-5af0757e96dec4962a076376d1bedc79de0d4249", + "instance_element-hq__element-web-56c7fc1948923b4b3f3507799e725ac16bcf8018-vnan", + "instance_flipt-io__flipt-381b90f718435c4694380b5fcd0d5cf8e3b5a25a", + "instance_flipt-io__flipt-6fd0f9e2587f14ac1fdd1c229f0bcae0468c8daa", + "instance_ansible__ansible-4c5ce5a1a9e79a845aff4978cfeb72a0d4ecf7d6-v1055803c3a812189a1133297f7f5468579283f86", + "instance_navidrome__navidrome-7073d18b54da7e53274d11c9e2baef1242e8769e", + "instance_internetarchive__openlibrary-0a90f9f0256e4f933523e9842799e39f95ae29ce-v76304ecdb3a5954fcf13feb710e8c40fcf24b73c", + "instance_nodebb__nodebb-be43cd25974681c9743d424238b7536c357dc8d3-vf2cf3cbd463b7ad942381f1c6d077626485a1e9e", + "instance_flipt-io__flipt-05d7234fa582df632f70a7cd10194d61bd7043b9", + "instance_nodebb__nodebb-76c6e30282906ac664f2c9278fc90999b27b1f48-vd59a5728dfc977f44533186ace531248c2917516", + "instance_qutebrowser__qutebrowser-6b320dc18662580e1313d2548fdd6231d2a97e6d-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_future-architect__vuls-01441351c3407abfc21c48a38e28828e1b504e0c", + "instance_ansible__ansible-1a4644ff15355fd696ac5b9d074a566a80fe7ca3-v30a923fb5c164d6cd18280c02422f75e611e8fb2", + "instance_navidrome__navidrome-09ae41a2da66264c60ef307882362d2e2d8d8b89", + "instance_internetarchive__openlibrary-5c6c22f3d2edf2f1b10f5dc335e32cb6a5f40341-v76304ecdb3a5954fcf13feb710e8c40fcf24b73c", + "instance_navidrome__navidrome-bf2bcb12799b21069f137749e0c331f761d1f693", + "instance_flipt-io__flipt-2ce8a0331e8a8f63f2c1b555db8277ffe5aa2e63", + "instance_ansible__ansible-a1569ea4ca6af5480cf0b7b3135f5e12add28a44-v0f01c69f1e2528b935359cfe578530722bca2c59", + "instance_gravitational__teleport-82185f232ae8974258397e121b3bc2ed0c3729ed-v626ec2a48416b10a88641359a169d99e935ff037", + "instance_flipt-io__flipt-72d06db14d58692bfb4d07b1aa745a37b35956f3", + "instance_ansible__ansible-b6290e1d156af608bd79118d209a64a051c55001-v390e508d27db7a51eece36bb6d9698b63a5b638a", + "instance_gravitational__teleport-32bcd71591c234f0d8b091ec01f1f5cbfdc0f13c-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_ansible__ansible-a02e22e902a69aeb465f16bf03f7f5a91b2cb828-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_gravitational__teleport-eda668c30d9d3b56d9c69197b120b01013611186", + "instance_nodebb__nodebb-0f788b8eaa4bba3c142d171fd941d015c53b65fc-v0ec6d6c2baf3cb4797482ce4829bc25cd5716649", + "instance_gravitational__teleport-47530e1fd8bfb84ec096ebcbbc29990f30829655-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_qutebrowser__qutebrowser-e64622cd2df5b521342cf4a62e0d4cb8f8c9ae5a-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_element-hq__element-web-494d9de6f0a94ffb491e74744d2735bce02dc0ab-vnan", + "instance_element-hq__element-web-9bf77963ee5e036d54b2a3ca202fbf6378464a5e-vnan", + "instance_nodebb__nodebb-9c576a0758690f45a6ca03b5884c601e473bf2c1-vd59a5728dfc977f44533186ace531248c2917516", + "instance_internetarchive__openlibrary-25858f9f0c165df25742acf8309ce909773f0cdd-v13642507b4fc1f8d234172bf8129942da2c2ca26", + "instance_protonmail__webclients-f161c10cf7d31abf82e8d64d7a99c9fac5acfa18", + "instance_gravitational__teleport-cb712e3f0b06dadc679f895daef8072cae400c26-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_gravitational__teleport-46a13210519461c7cec8d643bfbe750265775b41", + "instance_ansible__ansible-e40889e7112ae00a21a2c74312b330e67a766cc0-v1055803c3a812189a1133297f7f5468579283f86", + "instance_gravitational__teleport-6eaaf3a27e64f4ef4ef855bd35d7ec338cf17460-v626ec2a48416b10a88641359a169d99e935ff037", + "instance_internetarchive__openlibrary-0d13e6b4bf80bced6c0946b969b9a1b6963f6bce-v76304ecdb3a5954fcf13feb710e8c40fcf24b73c", + "instance_protonmail__webclients-2f2f6c311c6128fe86976950d3c0c2db07b03921", + "instance_ansible__ansible-5260527c4a71bfed99d803e687dd19619423b134-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_qutebrowser__qutebrowser-8d05f0282a271bfd45e614238bd1b555c58b3fc1-v35616345bb8052ea303186706cec663146f0f184", + "instance_qutebrowser__qutebrowser-50efac08f623644a85441bbe02ab9347d2b71a9d-v9f8e9d96c85c85a605e382f1510bd08563afc566", + "instance_flipt-io__flipt-dae029cba7cdb98dfb1a6b416c00d324241e6063", + "instance_nodebb__nodebb-18c45b44613aecd53e9f60457b9812049ab2998d-v0495b863a912fbff5749c67e860612b91825407c", + "instance_future-architect__vuls-e4728e388120b311c4ed469e4f942e0347a2689b-v264a82e2f4818e30f5a25e4da53b27ba119f62b5", + "instance_qutebrowser__qutebrowser-5e0d6dc1483cb3336ea0e3dcbd4fe4aa00fc1742-v5149fcda2a9a6fe1d35dfed1bade1444a11ef271", + "instance_flipt-io__flipt-c12967bc73fdf02054cf3ef8498c05e25f0a18c0", + "instance_qutebrowser__qutebrowser-ff1c025ad3210506fc76e1f604d8c8c27637d88e-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_ansible__ansible-83909bfa22573777e3db5688773bda59721962ad-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_qutebrowser__qutebrowser-1a9e74bfaf9a9db2a510dc14572d33ded6040a57-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_gravitational__teleport-24cafecd8721891092210afc55f6413ab46ca211-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_internetarchive__openlibrary-4b7ea2977be2747496ba792a678940baa985f7ea-v0f5aece3601a5b4419f7ccec1dbda2071be28ee4", + "instance_internetarchive__openlibrary-fad4a40acf5ff5f06cd7441a5c7baf41a7d81fe4-vfa6ff903cb27f336e17654595dd900fa943dcd91", + "instance_qutebrowser__qutebrowser-3fd8e12949b8feda401930574facf09dd4180bba", + "instance_ansible__ansible-164881d871964aa64e0f911d03ae270acbad253c-v390e508d27db7a51eece36bb6d9698b63a5b638a", + "instance_ansible__ansible-fb144c44144f8bd3542e71f5db62b6d322c7bd85-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_element-hq__element-web-582a1b093fc0b77538052f45cbb9c7295f991b51-vnan", + "instance_element-hq__element-web-ec0f940ef0e8e3b61078f145f34dc40d1938e6c5-vnan", + "instance_internetarchive__openlibrary-6fdbbeee4c0a7e976ff3e46fb1d36f4eb110c428-v08d8e8889ec945ab821fb156c04c7d2e2810debb", + "instance_flipt-io__flipt-7161f7b876773a911afdd804b281e52681cb7321", + "instance_navidrome__navidrome-28389fb05e1523564dfc61fa43ed8eb8a10f938c", + "instance_ansible__ansible-5f4e332e3762999d94af27746db29ff1729252c1-v0f01c69f1e2528b935359cfe578530722bca2c59", + "instance_nodebb__nodebb-0e07f3c9bace416cbab078a30eae972868c0a8a3-vf2cf3cbd463b7ad942381f1c6d077626485a1e9e", + "instance_nodebb__nodebb-da0211b1a001d45d73b4c84c6417a4f1b0312575-vf2cf3cbd463b7ad942381f1c6d077626485a1e9e", + "instance_navidrome__navidrome-31799662706fedddf5bcc1a76b50409d1f91d327", + "instance_future-architect__vuls-50580f6e98eeb36f53f27222f7f4fdfea0b21e8d", + "instance_internetarchive__openlibrary-60725705782832a2cb22e17c49697948a42a9d03-v298a7a812ceed28c4c18355a091f1b268fe56d86", + "instance_ansible__ansible-3db08adbb1cc6aa9941be5e0fc810132c6e1fa4b-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_internetarchive__openlibrary-8a5a63af6e0be406aa6c8c9b6d5f28b2f1b6af5a-v0f5aece3601a5b4419f7ccec1dbda2071be28ee4", + "instance_future-architect__vuls-e1df74cbc1a1d1889428b3333a3b2405c4651993", + "instance_internetarchive__openlibrary-123e6e5e1c85b9c07d1e98f70bfc480bc8016890-v2733ff199fb72f0d033a30dc62cb0a4742e3a7f4", + "instance_protonmail__webclients-ae36cb23a1682dcfd69587c1b311ae0227e28f39", + "instance_protonmail__webclients-8be4f6cb9380fcd2e67bcb18cef931ae0d4b869c", + "instance_qutebrowser__qutebrowser-a25e8a09873838ca9efefd36ea8a45170bbeb95c-vc2f56a753b62a190ddb23cd330c257b9cf560d12", + "instance_qutebrowser__qutebrowser-44e64199ed38003253f0296badd4a447645067b6-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_gravitational__teleport-7744f72c6eb631791434b648ba41083b5f6d2278-vce94f93ad1030e3136852817f2423c1b3ac37bc4", + "instance_tutao__tutanota-40e94dee2bcec2b63f362da283123e9df1874cc1-vc4e41fd0029957297843cb9dec4a25c7c756f029", + "instance_future-architect__vuls-c11ba27509f733d7d280bdf661cbbe2e7a99df4c", + "instance_flipt-io__flipt-b2cd6a6dd73ca91b519015fd5924fde8d17f3f06", + "instance_gravitational__teleport-6a14edcf1ff010172fdbac622d0a474ed6af46de", + "instance_flipt-io__flipt-36e62baffae2132f78f9d34dc300a9baa2d7ae0e", + "instance_navidrome__navidrome-c90468b895f6171e33e937ff20dc915c995274f0", + "instance_ansible__ansible-ecea15c508f0e081525be036cf76bbb56dbcdd9d-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_ansible__ansible-e22e103cdf8edc56ff7d9b848a58f94f1471a263-v1055803c3a812189a1133297f7f5468579283f86", + "instance_future-architect__vuls-dc496468b9e9fb73371f9606cdcdb0f8e12e70ca", + "instance_gravitational__teleport-c782838c3a174fdff80cafd8cd3b1aa4dae8beb2", + "instance_internetarchive__openlibrary-91efee627df01e32007abf2d6ebf73f9d9053076-vbee42ad1b72fb23c6a1c874868a720b370983ed2", + "instance_qutebrowser__qutebrowser-52708364b5f91e198defb022d1a5b4b3ebd9b563-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_ansible__ansible-eea46a0d1b99a6dadedbb6a3502d599235fa7ec3-v390e508d27db7a51eece36bb6d9698b63a5b638a", + "instance_protonmail__webclients-c6f65d205c401350a226bb005f42fac1754b0b5b", + "instance_ansible__ansible-cb94c0cc550df9e98f1247bc71d8c2b861c75049-v1055803c3a812189a1133297f7f5468579283f86", + "instance_nodebb__nodebb-05f2236193f407cf8e2072757fbd6bb170bc13f0-vf2cf3cbd463b7ad942381f1c6d077626485a1e9e", + "instance_navidrome__navidrome-55730514ea59d5f1d0b8e3f8745569c29bdbf7b4", + "instance_navidrome__navidrome-0130c6dc13438b48cf0fdfab08a89e357b5517c9", + "instance_flipt-io__flipt-507170da0f7f4da330f6732bffdf11c4df7fc192", + "instance_tutao__tutanota-fe240cbf7f0fdd6744ef7bef8cb61676bcdbb621-vc4e41fd0029957297843cb9dec4a25c7c756f029", + "instance_qutebrowser__qutebrowser-233cb1cc48635130e5602549856a6fa4ab4c087f-v35616345bb8052ea303186706cec663146f0f184", + "instance_internetarchive__openlibrary-f8cc11d9c1575fdba5ac66aee0befca970da8d64-v13642507b4fc1f8d234172bf8129942da2c2ca26", + "instance_qutebrowser__qutebrowser-fec187c2cb53d769c2682b35ca77858a811414a8-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_tutao__tutanota-da4edb7375c10f47f4ed3860a591c5e6557f7b5c-vbc0d9ba8f0071fbe982809910959a6ff8884dbbf", + "instance_ansible__ansible-be2c376ab87e3e872ca21697508f12c6909cf85a-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_future-architect__vuls-457a3a9627fb9a0800d0aecf1d4713fb634a9011", + "instance_internetarchive__openlibrary-d40ec88713dc95ea791b252f92d2f7b75e107440-v13642507b4fc1f8d234172bf8129942da2c2ca26", + "instance_protonmail__webclients-863d524b5717b9d33ce08a0f0535e3fd8e8d1ed8", + "instance_internetarchive__openlibrary-308a35d6999427c02b1dbf5211c033ad3b352556-ve8c8d62a2b60610a3c4631f5f23ed866bada9818", + "instance_flipt-io__flipt-15b76cada1ef29cfa56b0fba36754be36243dded", + "instance_future-architect__vuls-61c39637f2f3809e1b5dad05f0c57c799dce1587", + "instance_ansible__ansible-5d253a13807e884b7ce0b6b57a963a45e2f0322c-v1055803c3a812189a1133297f7f5468579283f86", + "instance_element-hq__element-web-44b98896a79ede48f5ad7ff22619a39d5f6ff03c-vnan", + "instance_qutebrowser__qutebrowser-f91ace96223cac8161c16dd061907e138fe85111-v059c6fdc75567943479b23ebca7c07b5e9a7f34c", + "instance_gravitational__teleport-bb562408da4adeae16e025be65e170959d1ec492-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_gravitational__teleport-78b0d8c72637df1129fb6ff84fc49ef4b5ab1288", + "instance_internetarchive__openlibrary-3f7db6bbbcc7c418b3db72d157c6aed1d45b2ccf-v430f20c722405e462d9ef44dee7d34c41e76fe7a", + "instance_future-architect__vuls-e52fa8d6ed1d23e36f2a86e5d3efe9aa057a1b0d", + "instance_navidrome__navidrome-3bc9e75b2843f91f6a1e9b604e321c2bd4fd442a", + "instance_tutao__tutanota-fb32e5f9d9fc152a00144d56dd0af01760a2d4dc-vc4e41fd0029957297843cb9dec4a25c7c756f029", + "instance_gravitational__teleport-af5e2517de7d18406b614e413aca61c319312171-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_ansible__ansible-a20a52701402a12f91396549df04ac55809f68e9-v1055803c3a812189a1133297f7f5468579283f86", + "instance_protonmail__webclients-1501eb765873b2884b6f1944fd242ecfc9d6b103", + "instance_gravitational__teleport-db89206db6c2969266e664c7c0fb51b70e958b64", + "instance_ansible__ansible-be59caa59bf47ca78a4760eb7ff38568372a8260-v1055803c3a812189a1133297f7f5468579283f86", + "instance_flipt-io__flipt-65581fef4aa807540cb933753d085feb0d7e736f", + "instance_gravitational__teleport-4e1c39639edf1ab494dd7562844c8b277b5cfa18-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_future-architect__vuls-f0b3a8b1db98eb1bd32685f1c36c41a99c3452ed", + "instance_navidrome__navidrome-b65e76293a917ee2dfc5d4b373b1c62e054d0dca", + "instance_element-hq__element-web-33e8edb3d508d6eefb354819ca693b7accc695e7", + "instance_flipt-io__flipt-e50808c03e4b9d25a6a78af9c61a3b1616ea356b", + "instance_future-architect__vuls-e3c27e1817d68248043bd09d63cc31f3344a6f2c", + "instance_element-hq__element-web-ad26925bb6628260cfe0fcf90ec0a8cba381f4a4-vnan", + "instance_nodebb__nodebb-1ea9481af6125ffd6da0592ed439aa62af0bca11-vd59a5728dfc977f44533186ace531248c2917516", + "instance_ansible__ansible-489156378c8e97374a75a544c7c9c2c0dd8146d1-v390e508d27db7a51eece36bb6d9698b63a5b638a", + "instance_internetarchive__openlibrary-bb152d23c004f3d68986877143bb0f83531fe401-ve8c8d62a2b60610a3c4631f5f23ed866bada9818", + "instance_qutebrowser__qutebrowser-0b621cb0ce2b54d3f93d8d41d8ff4257888a87e5-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_qutebrowser__qutebrowser-1af602b258b97aaba69d2585ed499d95e2303ac2-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_future-architect__vuls-e1fab805afcfc92a2a615371d0ec1e667503c254-v264a82e2f4818e30f5a25e4da53b27ba119f62b5", + "instance_internetarchive__openlibrary-7edd1ef09d91fe0b435707633c5cc9af41dedddf-v76304ecdb3a5954fcf13feb710e8c40fcf24b73c", + "instance_qutebrowser__qutebrowser-99029144b5109bb1b2a53964a7c129e009980cd9-va0fd88aac89cde702ec1ba84877234da33adce8a", + "instance_ansible__ansible-12734fa21c08a0ce8c84e533abdc560db2eb1955-v7eee2454f617569fd6889f2211f75bc02a35f9f8", + "instance_protonmail__webclients-7e54526774e577c0ebb58ced7ba8bef349a69fec", + "instance_gravitational__teleport-1a77b7945a022ab86858029d30ac7ad0d5239d00-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_nodebb__nodebb-8ca65b0c78c67c1653487c02d1135e1b702185e1-vf2cf3cbd463b7ad942381f1c6d077626485a1e9e", + "instance_flipt-io__flipt-9f8127f225a86245fa35dca4885c2daef824ee55", + "instance_gravitational__teleport-ac2fb2f9b4fd1896b554d3011df23d3d71295779", + "instance_ansible__ansible-f02a62db509dc7463fab642c9c3458b9bc3476cc-v390e508d27db7a51eece36bb6d9698b63a5b638a" + ], + "additional_details": { + "shuffled": "True" + } + }, + "evaluation_timestamp": "1775482110.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.6, + "uncertainty": { + "standard_error": { + "value": 0.03472794481039396 + }, + "num_samples": 200 + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"If you do not submit an answer, you will receive a score of zero.\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"Your submission was incorrect. Please review your work, identify what went wrong, and try again.\", \"to_float\"], \"submit\": null, \"on_continue\": null, \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"If you do not submit an answer, you will receive a score of zero.\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"Your submission was incorrect. Please review your work, identify what went wrong, and try again.\", \"to_float\"], \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 15000000 + }, + "sandbox": { + "type": "k8s" + } + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/anthropic/claude-opus-4-6/f26f1ce1-2163-4648-b868-b1044208d07b_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "3bd21d74c271f8e3a4dc2798b384e786127424e496f6672cf169fb9ebd365697", + "total_rows": 200 + } +} \ No newline at end of file diff --git a/data/swebenchpro/anthropic/claude-opus-4-6/f26f1ce1-2163-4648-b868-b1044208d07b_samples.jsonl b/data/swebenchpro/anthropic/claude-opus-4-6/f26f1ce1-2163-4648-b868-b1044208d07b_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..c17788c50fd6a080bf6835619fb5e43845e3fab8 --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-6/f26f1ce1-2163-4648-b868-b1044208d07b_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3bd21d74c271f8e3a4dc2798b384e786127424e496f6672cf169fb9ebd365697 +size 113926361 diff --git a/data/swebenchpro/anthropic/claude-opus-4-6/f3754b8b-5029-4cff-ae21-cbc60c5b0385.json b/data/swebenchpro/anthropic/claude-opus-4-6/f3754b8b-5029-4cff-ae21-cbc60c5b0385.json new file mode 100644 index 0000000000000000000000000000000000000000..fa11a3a50065bb45e8be1c2c0642cb3a8e5397a0 --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-6/f3754b8b-5029-4cff-ae21-cbc60c5b0385.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/anthropic_claude-opus-4-6/1786167443.984703", + "evaluation_timestamp": "1778665168.0", + "retrieved_timestamp": "1786167443.984703", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-6", + "id": "anthropic/claude-opus-4-6", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/anthropic/claude-opus-4-6/f3754b8b-5029-4cff-ae21-cbc60c5b0385_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "d66190c9a93a33829e9350571e20d401f4c358d34e86328ad4e446cfc1dc0903", + "total_rows": 5 + } +} \ No newline at end of file diff --git a/data/swebenchpro/anthropic/claude-opus-4-6/f3754b8b-5029-4cff-ae21-cbc60c5b0385_samples.jsonl b/data/swebenchpro/anthropic/claude-opus-4-6/f3754b8b-5029-4cff-ae21-cbc60c5b0385_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..e98d37b55e549b1d4d9393abbccca98a2fcbe14c --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-6/f3754b8b-5029-4cff-ae21-cbc60c5b0385_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d66190c9a93a33829e9350571e20d401f4c358d34e86328ad4e446cfc1dc0903 +size 430240 diff --git a/data/swebenchpro/anthropic/claude-opus-4-6/fda59d66-de33-4604-9b30-942b276e3abf.json b/data/swebenchpro/anthropic/claude-opus-4-6/fda59d66-de33-4604-9b30-942b276e3abf.json new file mode 100644 index 0000000000000000000000000000000000000000..dbd38bc4b9f2f57e4806a4f1125e14db925c67f0 --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-6/fda59d66-de33-4604-9b30-942b276e3abf.json @@ -0,0 +1,127 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/anthropic_claude-opus-4-6/1786165634.463937", + "evaluation_timestamp": "1778540511.0", + "retrieved_timestamp": "1786165634.463937", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-6", + "id": "anthropic/claude-opus-4-6", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on swebenchpro/S-adaptive/+2ep/5f6bbef4 for scorer harbor_scorer", + "source_data": { + "dataset_name": "swebenchpro", + "source_type": "hf_dataset", + "samples_number": 15, + "sample_ids": [ + "instance_future-architect__vuls-878c25bf5a9c9fd88ac32eb843f5636834d5712d", + "instance_future-architect__vuls-73f0adad95c4d227e2ccfa876c85cc95dd065e13", + "instance_protonmail__webclients-da91f084c0f532d9cc8ca385a701274d598057b8", + "instance_flipt-io__flipt-b2cd6a6dd73ca91b519015fd5924fde8d17f3f06", + "instance_qutebrowser__qutebrowser-cc360cd4a34a126274c7b51f3b63afbaf3e05a02-v5fc38aaf22415ab0b70567368332beee7955b367", + "instance_flipt-io__flipt-e2bd19dafa7166c96b082fb2a59eb54b4be0d778", + "instance_qutebrowser__qutebrowser-9ed748effa8f3bcd804612d9291da017b514e12f-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_ansible__ansible-1ee70fc272aff6bf3415357c6e13c5de5b928d9b-v1055803c3a812189a1133297f7f5468579283f86", + "instance_element-hq__element-web-75c2c1a572fa45d1ea1d1a96e9e36e303332ecaa-vnan", + "instance_qutebrowser__qutebrowser-f631cd4422744160d9dcf7a0455da532ce973315-v35616345bb8052ea303186706cec663146f0f184", + "instance_qutebrowser__qutebrowser-e5340c449f23608803c286da0563b62f58ba25b0-v059c6fdc75567943479b23ebca7c07b5e9a7f34c", + "instance_ansible__ansible-42355d181a11b51ebfc56f6f4b3d9c74e01cb13b-v1055803c3a812189a1133297f7f5468579283f86", + "instance_internetarchive__openlibrary-09865f5fb549694d969f0a8e49b9d204ef1853ca-ve8c8d62a2b60610a3c4631f5f23ed866bada9818", + "instance_internetarchive__openlibrary-c506c1b0b678892af5cb22c1c1dbc35d96787a0a-v0f5aece3601a5b4419f7ccec1dbda2071be28ee4", + "instance_element-hq__element-web-2760bfc8369f1bee640d6d7a7e910783143d4c5f-vnan" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1778540511.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.7333333333333333, + "uncertainty": { + "standard_error": { + "value": 0.11818736805705576 + }, + "num_samples": 30 + }, + "details": { + "total_matched_trajectories": "30", + "stopping_reason_count_repetition_guard": "30", + "avg_total_tokens_target_model": "2404326.00", + "total_total_tokens_target_model": "72129780.0", + "avg_total_tokens_other_models": "1063.23", + "total_total_tokens_other_models": "31897.0", + "avg_total_tokens_all_models": "2405389.23", + "total_total_tokens_all_models": "72161677.0", + "avg_turn_count": "58.20", + "total_turn_count": "1746.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 30000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/anthropic/claude-opus-4-6/fda59d66-de33-4604-9b30-942b276e3abf_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "d6ffb5fea3a7da1a9af3bf48394928a2292a50486fd00b22e1e098201b17d84b", + "total_rows": 30 + } +} \ No newline at end of file diff --git a/data/swebenchpro/anthropic/claude-opus-4-6/fda59d66-de33-4604-9b30-942b276e3abf_samples.jsonl b/data/swebenchpro/anthropic/claude-opus-4-6/fda59d66-de33-4604-9b30-942b276e3abf_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..1b936c16bf6f7f9e0fc24bae5c51b4acfc78a05d --- /dev/null +++ b/data/swebenchpro/anthropic/claude-opus-4-6/fda59d66-de33-4604-9b30-942b276e3abf_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d6ffb5fea3a7da1a9af3bf48394928a2292a50486fd00b22e1e098201b17d84b +size 8620789 diff --git a/data/swebenchpro/openai/gpt-5-2025-08-07/1772c3b3-ecf8-43d1-a666-61f404183c95.json b/data/swebenchpro/openai/gpt-5-2025-08-07/1772c3b3-ecf8-43d1-a666-61f404183c95.json new file mode 100644 index 0000000000000000000000000000000000000000..49a50a38a197a35a6f1137605b2781f252143dcc --- /dev/null +++ b/data/swebenchpro/openai/gpt-5-2025-08-07/1772c3b3-ecf8-43d1-a666-61f404183c95.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/openai_gpt-5-2025-08-07/1786169172.976431", + "evaluation_timestamp": "1777544308.0", + "retrieved_timestamp": "1786169172.976431", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5-2025-08-07", + "id": "openai/gpt-5-2025-08-07", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/openai/gpt-5-2025-08-07/1772c3b3-ecf8-43d1-a666-61f404183c95_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "0d823ab2de523796197bdbeb174925453917a058dcbae3c3d824121f0da26cec", + "total_rows": 68 + } +} \ No newline at end of file diff --git a/data/swebenchpro/openai/gpt-5-2025-08-07/1772c3b3-ecf8-43d1-a666-61f404183c95_samples.jsonl b/data/swebenchpro/openai/gpt-5-2025-08-07/1772c3b3-ecf8-43d1-a666-61f404183c95_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..97fa15dad2533ddf9b60387884a472223c27abf6 --- /dev/null +++ b/data/swebenchpro/openai/gpt-5-2025-08-07/1772c3b3-ecf8-43d1-a666-61f404183c95_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0d823ab2de523796197bdbeb174925453917a058dcbae3c3d824121f0da26cec +size 34173159 diff --git a/data/swebenchpro/openai/gpt-5-2025-08-07/26d10ecf-b41e-4a9b-8e83-29d5187b314c.json b/data/swebenchpro/openai/gpt-5-2025-08-07/26d10ecf-b41e-4a9b-8e83-29d5187b314c.json new file mode 100644 index 0000000000000000000000000000000000000000..0e3d2d8d28c6e96d1b69c010b39f9de48ad41321 --- /dev/null +++ b/data/swebenchpro/openai/gpt-5-2025-08-07/26d10ecf-b41e-4a9b-8e83-29d5187b314c.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/openai_gpt-5-2025-08-07/1786169103.446545", + "evaluation_timestamp": "1776601454.0", + "retrieved_timestamp": "1786169103.446545", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5-2025-08-07", + "id": "openai/gpt-5-2025-08-07", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/openai/gpt-5-2025-08-07/26d10ecf-b41e-4a9b-8e83-29d5187b314c_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "daa75f77de8d0a277b98c6385043123546f47848226412d58cce79841b3e7d83", + "total_rows": 10 + } +} \ No newline at end of file diff --git a/data/swebenchpro/openai/gpt-5-2025-08-07/26d10ecf-b41e-4a9b-8e83-29d5187b314c_samples.jsonl b/data/swebenchpro/openai/gpt-5-2025-08-07/26d10ecf-b41e-4a9b-8e83-29d5187b314c_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..efd1e59ba0b614b779083efb7007f8481e4e1264 --- /dev/null +++ b/data/swebenchpro/openai/gpt-5-2025-08-07/26d10ecf-b41e-4a9b-8e83-29d5187b314c_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:daa75f77de8d0a277b98c6385043123546f47848226412d58cce79841b3e7d83 +size 5890723 diff --git a/data/swebenchpro/openai/gpt-5-2025-08-07/297572f0-b841-4303-8887-75ac51a4e50b.json b/data/swebenchpro/openai/gpt-5-2025-08-07/297572f0-b841-4303-8887-75ac51a4e50b.json new file mode 100644 index 0000000000000000000000000000000000000000..dea5c47bf90747cbdd966f6c9d1f483f83a29ec1 --- /dev/null +++ b/data/swebenchpro/openai/gpt-5-2025-08-07/297572f0-b841-4303-8887-75ac51a4e50b.json @@ -0,0 +1,110 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/openai_gpt-5-2025-08-07/1786164488.667102", + "evaluation_timestamp": "1777914565.0", + "retrieved_timestamp": "1786164488.667102", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5-2025-08-07", + "id": "openai/gpt-5-2025-08-07", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on swebenchpro/S-adaptive/+2ep/e04ec096 for scorer harbor_scorer", + "source_data": { + "dataset_name": "swebenchpro", + "source_type": "hf_dataset", + "samples_number": 1, + "sample_ids": [ + "instance_ansible__ansible-de5858f48dc9e1ce9117034e0d7e76806f420ca8-v1055803c3a812189a1133297f7f5468579283f86" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1777914565.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.0, + "uncertainty": { + "num_samples": 2 + }, + "details": { + "total_matched_trajectories": "2", + "stopping_reason_count_repetition_guard": "2", + "avg_total_tokens_target_model": "14462263.50", + "total_total_tokens_target_model": "28924527.0", + "avg_total_tokens_other_models": "1767.50", + "total_total_tokens_other_models": "3535.0", + "avg_total_tokens_all_models": "14464031.00", + "total_total_tokens_all_models": "28928062.0", + "avg_turn_count": "188.00", + "total_turn_count": "376.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 30000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/openai/gpt-5-2025-08-07/297572f0-b841-4303-8887-75ac51a4e50b_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "2691ecbbe8f6199789b362567f7cdf948755457076660436d2a95334855d90a8", + "total_rows": 2 + } +} \ No newline at end of file diff --git a/data/swebenchpro/openai/gpt-5-2025-08-07/297572f0-b841-4303-8887-75ac51a4e50b_samples.jsonl b/data/swebenchpro/openai/gpt-5-2025-08-07/297572f0-b841-4303-8887-75ac51a4e50b_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..27d0f499c57154291fe0f56a1694616a46574652 --- /dev/null +++ b/data/swebenchpro/openai/gpt-5-2025-08-07/297572f0-b841-4303-8887-75ac51a4e50b_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2691ecbbe8f6199789b362567f7cdf948755457076660436d2a95334855d90a8 +size 2932384 diff --git a/data/swebenchpro/openai/gpt-5-2025-08-07/2e3b5a36-3ad8-4ce8-b2b4-05b5c6c41844.json b/data/swebenchpro/openai/gpt-5-2025-08-07/2e3b5a36-3ad8-4ce8-b2b4-05b5c6c41844.json new file mode 100644 index 0000000000000000000000000000000000000000..33a71063755092be2e1142e3c40fff5165e9d81a --- /dev/null +++ b/data/swebenchpro/openai/gpt-5-2025-08-07/2e3b5a36-3ad8-4ce8-b2b4-05b5c6c41844.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/openai_gpt-5-2025-08-07/1786169121.460512", + "evaluation_timestamp": "1776647316.0", + "retrieved_timestamp": "1786169121.460512", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5-2025-08-07", + "id": "openai/gpt-5-2025-08-07", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/openai/gpt-5-2025-08-07/2e3b5a36-3ad8-4ce8-b2b4-05b5c6c41844_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "68a8038fa6d7be994c193b90a12f91165628a2cfa58a5420db0d32e370640dbf", + "total_rows": 33 + } +} \ No newline at end of file diff --git a/data/swebenchpro/openai/gpt-5-2025-08-07/2e3b5a36-3ad8-4ce8-b2b4-05b5c6c41844_samples.jsonl b/data/swebenchpro/openai/gpt-5-2025-08-07/2e3b5a36-3ad8-4ce8-b2b4-05b5c6c41844_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..2d213b35cb0517de31fe01b3842669073e56699f --- /dev/null +++ b/data/swebenchpro/openai/gpt-5-2025-08-07/2e3b5a36-3ad8-4ce8-b2b4-05b5c6c41844_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68a8038fa6d7be994c193b90a12f91165628a2cfa58a5420db0d32e370640dbf +size 15155615 diff --git a/data/swebenchpro/openai/gpt-5-2025-08-07/2f3f4cf2-0f46-4ee3-84b9-9693b0da255c.json b/data/swebenchpro/openai/gpt-5-2025-08-07/2f3f4cf2-0f46-4ee3-84b9-9693b0da255c.json new file mode 100644 index 0000000000000000000000000000000000000000..9d1b062d7791bd0c39cb9e77f3d3dea79704776b --- /dev/null +++ b/data/swebenchpro/openai/gpt-5-2025-08-07/2f3f4cf2-0f46-4ee3-84b9-9693b0da255c.json @@ -0,0 +1,116 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/openai_gpt-5-2025-08-07/1786164484.555601", + "evaluation_timestamp": "1776820365.0", + "retrieved_timestamp": "1786164484.555601", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5-2025-08-07", + "id": "openai/gpt-5-2025-08-07", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on swebenchpro/S-adaptive/+1ep/a32dc7ea for scorer harbor_scorer", + "source_data": { + "dataset_name": "swebenchpro", + "source_type": "hf_dataset", + "samples_number": 7, + "sample_ids": [ + "instance_ansible__ansible-b2a289dcbb702003377221e25f62c8a3608f0e89-v173091e2e36d38c978002990795f66cfc0af30ad", + "instance_qutebrowser__qutebrowser-e64622cd2df5b521342cf4a62e0d4cb8f8c9ae5a-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_gravitational__teleport-ac2fb2f9b4fd1896b554d3011df23d3d71295779", + "instance_nodebb__nodebb-70b4a0e2aebebe8f2f559de6680093d96a697b2f-vnan", + "instance_future-architect__vuls-73f0adad95c4d227e2ccfa876c85cc95dd065e13", + "instance_flipt-io__flipt-b2cd6a6dd73ca91b519015fd5924fde8d17f3f06", + "instance_ansible__ansible-1ee70fc272aff6bf3415357c6e13c5de5b928d9b-v1055803c3a812189a1133297f7f5468579283f86" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1776820365.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.5714285714285714, + "uncertainty": { + "standard_error": { + "value": 0.20203050891044214 + }, + "num_samples": 7 + }, + "details": { + "total_matched_trajectories": "7", + "stopping_reason_count_repetition_guard": "7", + "avg_total_tokens_target_model": "5193580.57", + "total_total_tokens_target_model": "36355064.0", + "avg_total_tokens_other_models": "1751.14", + "total_total_tokens_other_models": "12258.0", + "avg_total_tokens_all_models": "5195331.71", + "total_total_tokens_all_models": "36367322.0", + "avg_turn_count": "91.71", + "total_turn_count": "642.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 30000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/openai/gpt-5-2025-08-07/2f3f4cf2-0f46-4ee3-84b9-9693b0da255c_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "786d6838e72efe9b1c5a0ac503057f12668557a0965a73933cad7915042d9ee0", + "total_rows": 7 + } +} \ No newline at end of file diff --git a/data/swebenchpro/openai/gpt-5-2025-08-07/2f3f4cf2-0f46-4ee3-84b9-9693b0da255c_samples.jsonl b/data/swebenchpro/openai/gpt-5-2025-08-07/2f3f4cf2-0f46-4ee3-84b9-9693b0da255c_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..8b42b47d05123a2e7b730ffe3bd8eed37d2db029 --- /dev/null +++ b/data/swebenchpro/openai/gpt-5-2025-08-07/2f3f4cf2-0f46-4ee3-84b9-9693b0da255c_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:786d6838e72efe9b1c5a0ac503057f12668557a0965a73933cad7915042d9ee0 +size 3846165 diff --git a/data/swebenchpro/openai/gpt-5-2025-08-07/35a45176-07a1-4976-84f6-b0bc6f356e4d.json b/data/swebenchpro/openai/gpt-5-2025-08-07/35a45176-07a1-4976-84f6-b0bc6f356e4d.json new file mode 100644 index 0000000000000000000000000000000000000000..eec54416225b6aebee2b0f41cd6a814a4e81249d --- /dev/null +++ b/data/swebenchpro/openai/gpt-5-2025-08-07/35a45176-07a1-4976-84f6-b0bc6f356e4d.json @@ -0,0 +1,105 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/openai_gpt-5-2025-08-07/1786165466.099898", + "evaluation_timestamp": "1776629382.0", + "retrieved_timestamp": "1786165466.099898", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5-2025-08-07", + "id": "openai/gpt-5-2025-08-07", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on swebenchpro/S-adaptive/+3ep/60d44e75 for scorer harbor_scorer", + "source_data": { + "dataset_name": "swebenchpro", + "source_type": "hf_dataset", + "samples_number": 1, + "sample_ids": [ + "instance_navidrome__navidrome-d8e794317f788198227e10fb667e10496b3eb99a" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1776629382.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.0, + "uncertainty": { + "num_samples": 3 + }, + "details": { + "total_matched_trajectories": "3", + "stopping_reason_count_repetition_guard": "3", + "avg_total_tokens_target_model": "7456488.33", + "total_total_tokens_target_model": "22369465.0", + "avg_total_tokens_other_models": "2290.00", + "total_total_tokens_other_models": "6870.0", + "avg_total_tokens_all_models": "7458778.33", + "total_total_tokens_all_models": "22376335.0", + "avg_turn_count": "138.33", + "total_turn_count": "415.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 30000000 + }, + "sandbox": {} + }, + "additional_details": { + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/openai/gpt-5-2025-08-07/35a45176-07a1-4976-84f6-b0bc6f356e4d_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "8fd0a87695a6e4796bca1584c40cecdb8f60b68be6d3b8d9b7545d844d19dd70", + "total_rows": 3 + } +} \ No newline at end of file diff --git a/data/swebenchpro/openai/gpt-5-2025-08-07/35a45176-07a1-4976-84f6-b0bc6f356e4d_samples.jsonl b/data/swebenchpro/openai/gpt-5-2025-08-07/35a45176-07a1-4976-84f6-b0bc6f356e4d_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..8067e1ab7dac7d56e58ed4815cdcbde239cee1d4 --- /dev/null +++ b/data/swebenchpro/openai/gpt-5-2025-08-07/35a45176-07a1-4976-84f6-b0bc6f356e4d_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8fd0a87695a6e4796bca1584c40cecdb8f60b68be6d3b8d9b7545d844d19dd70 +size 2377845 diff --git a/data/swebenchpro/openai/gpt-5-2025-08-07/3f135786-851d-43e8-bb84-ea5bec5b4c66.json b/data/swebenchpro/openai/gpt-5-2025-08-07/3f135786-851d-43e8-bb84-ea5bec5b4c66.json new file mode 100644 index 0000000000000000000000000000000000000000..200dcc07b4bec79be15418062b8c98c0d0133fd8 --- /dev/null +++ b/data/swebenchpro/openai/gpt-5-2025-08-07/3f135786-851d-43e8-bb84-ea5bec5b4c66.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/openai_gpt-5-2025-08-07/1786169162.03383", + "evaluation_timestamp": "1776720505.0", + "retrieved_timestamp": "1786169162.03383", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5-2025-08-07", + "id": "openai/gpt-5-2025-08-07", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/openai/gpt-5-2025-08-07/3f135786-851d-43e8-bb84-ea5bec5b4c66_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "8c60e58ca4b254383237f1893370ea23715049ecc99891b089e1d3a7f007863c", + "total_rows": 42 + } +} \ No newline at end of file diff --git a/data/swebenchpro/openai/gpt-5-2025-08-07/3f135786-851d-43e8-bb84-ea5bec5b4c66_samples.jsonl b/data/swebenchpro/openai/gpt-5-2025-08-07/3f135786-851d-43e8-bb84-ea5bec5b4c66_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..ff13b98a2f548d8f4a1d96c0f8d40aa182ed04c2 --- /dev/null +++ b/data/swebenchpro/openai/gpt-5-2025-08-07/3f135786-851d-43e8-bb84-ea5bec5b4c66_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8c60e58ca4b254383237f1893370ea23715049ecc99891b089e1d3a7f007863c +size 27297497 diff --git a/data/swebenchpro/openai/gpt-5-2025-08-07/46d7ce44-6eec-4696-a54d-cf8cb8028142.json b/data/swebenchpro/openai/gpt-5-2025-08-07/46d7ce44-6eec-4696-a54d-cf8cb8028142.json new file mode 100644 index 0000000000000000000000000000000000000000..167e5cfa84a0fb5cdca12e8f1ee2fbca243dfe57 --- /dev/null +++ b/data/swebenchpro/openai/gpt-5-2025-08-07/46d7ce44-6eec-4696-a54d-cf8cb8028142.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/openai_gpt-5-2025-08-07/1786169106.703725", + "evaluation_timestamp": "1776680886.0", + "retrieved_timestamp": "1786169106.703725", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5-2025-08-07", + "id": "openai/gpt-5-2025-08-07", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/openai/gpt-5-2025-08-07/46d7ce44-6eec-4696-a54d-cf8cb8028142_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "c170e46237e5fb7abe1bfa3dc56a97b3aec6dd6eb270a3c60be97a9479bf1805", + "total_rows": 18 + } +} \ No newline at end of file diff --git a/data/swebenchpro/openai/gpt-5-2025-08-07/46d7ce44-6eec-4696-a54d-cf8cb8028142_samples.jsonl b/data/swebenchpro/openai/gpt-5-2025-08-07/46d7ce44-6eec-4696-a54d-cf8cb8028142_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..b9ce54c4230511d50ab63967c4dfaaff8ee1399c --- /dev/null +++ b/data/swebenchpro/openai/gpt-5-2025-08-07/46d7ce44-6eec-4696-a54d-cf8cb8028142_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c170e46237e5fb7abe1bfa3dc56a97b3aec6dd6eb270a3c60be97a9479bf1805 +size 6804070 diff --git a/data/swebenchpro/openai/gpt-5-2025-08-07/53d73f0a-184e-45b4-8906-c96e7a2ee378.json b/data/swebenchpro/openai/gpt-5-2025-08-07/53d73f0a-184e-45b4-8906-c96e7a2ee378.json new file mode 100644 index 0000000000000000000000000000000000000000..03967b1fd97af56d0df7f95ee1aa94bafeeccaf4 --- /dev/null +++ b/data/swebenchpro/openai/gpt-5-2025-08-07/53d73f0a-184e-45b4-8906-c96e7a2ee378.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/openai_gpt-5-2025-08-07/1786167364.265289", + "evaluation_timestamp": "1778228784.0", + "retrieved_timestamp": "1786167364.265289", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5-2025-08-07", + "id": "openai/gpt-5-2025-08-07", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/openai/gpt-5-2025-08-07/53d73f0a-184e-45b4-8906-c96e7a2ee378_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "7dad0c05531f2be83e41dc12a76997d9f95d620dd64f0a55546b9d2879299788", + "total_rows": 57 + } +} \ No newline at end of file diff --git a/data/swebenchpro/openai/gpt-5-2025-08-07/53d73f0a-184e-45b4-8906-c96e7a2ee378_samples.jsonl b/data/swebenchpro/openai/gpt-5-2025-08-07/53d73f0a-184e-45b4-8906-c96e7a2ee378_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..96745308cea33dc7986ac4b0f8a2c77630a47fa2 --- /dev/null +++ b/data/swebenchpro/openai/gpt-5-2025-08-07/53d73f0a-184e-45b4-8906-c96e7a2ee378_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7dad0c05531f2be83e41dc12a76997d9f95d620dd64f0a55546b9d2879299788 +size 21379955 diff --git a/data/swebenchpro/openai/gpt-5-2025-08-07/540c16aa-dcec-432e-8d7d-ad35e610404d.json b/data/swebenchpro/openai/gpt-5-2025-08-07/540c16aa-dcec-432e-8d7d-ad35e610404d.json new file mode 100644 index 0000000000000000000000000000000000000000..28a8c5ded8a90cbf130df2aad74e4fdd304ad866 --- /dev/null +++ b/data/swebenchpro/openai/gpt-5-2025-08-07/540c16aa-dcec-432e-8d7d-ad35e610404d.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/openai_gpt-5-2025-08-07/1786167343.771896", + "evaluation_timestamp": "1778055257.0", + "retrieved_timestamp": "1786167343.771896", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5-2025-08-07", + "id": "openai/gpt-5-2025-08-07", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/openai/gpt-5-2025-08-07/540c16aa-dcec-432e-8d7d-ad35e610404d_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "069b67133e64101f810dd03c9fd5d0a3e2ab37faf3442c9f054e17c4433276d2", + "total_rows": 20 + } +} \ No newline at end of file diff --git a/data/swebenchpro/openai/gpt-5-2025-08-07/540c16aa-dcec-432e-8d7d-ad35e610404d_samples.jsonl b/data/swebenchpro/openai/gpt-5-2025-08-07/540c16aa-dcec-432e-8d7d-ad35e610404d_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..8284c5d5b065bb68439d71da1507a05a6f713227 --- /dev/null +++ b/data/swebenchpro/openai/gpt-5-2025-08-07/540c16aa-dcec-432e-8d7d-ad35e610404d_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:069b67133e64101f810dd03c9fd5d0a3e2ab37faf3442c9f054e17c4433276d2 +size 8072406 diff --git a/data/swebenchpro/openai/gpt-5-2025-08-07/58e44077-0585-4229-b4e4-19ce1ea51f4b.json b/data/swebenchpro/openai/gpt-5-2025-08-07/58e44077-0585-4229-b4e4-19ce1ea51f4b.json new file mode 100644 index 0000000000000000000000000000000000000000..084a70baaa07fd9e8850eeb452371a8ede6edb5b --- /dev/null +++ b/data/swebenchpro/openai/gpt-5-2025-08-07/58e44077-0585-4229-b4e4-19ce1ea51f4b.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/openai_gpt-5-2025-08-07/1786169141.441444", + "evaluation_timestamp": "1777126007.0", + "retrieved_timestamp": "1786169141.441444", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5-2025-08-07", + "id": "openai/gpt-5-2025-08-07", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/openai/gpt-5-2025-08-07/58e44077-0585-4229-b4e4-19ce1ea51f4b_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "da4c65507844500b8431b83fbd8f5f637483145c5f74cd4ffc1bdd374afa0c13", + "total_rows": 12 + } +} \ No newline at end of file diff --git a/data/swebenchpro/openai/gpt-5-2025-08-07/58e44077-0585-4229-b4e4-19ce1ea51f4b_samples.jsonl b/data/swebenchpro/openai/gpt-5-2025-08-07/58e44077-0585-4229-b4e4-19ce1ea51f4b_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..43b26566f69e54ac551914c2fe9af1850d310960 --- /dev/null +++ b/data/swebenchpro/openai/gpt-5-2025-08-07/58e44077-0585-4229-b4e4-19ce1ea51f4b_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:da4c65507844500b8431b83fbd8f5f637483145c5f74cd4ffc1bdd374afa0c13 +size 5463090 diff --git a/data/swebenchpro/openai/gpt-5-2025-08-07/5bc6c4d1-1f11-4c24-9b59-fa6afe7e4109.json b/data/swebenchpro/openai/gpt-5-2025-08-07/5bc6c4d1-1f11-4c24-9b59-fa6afe7e4109.json new file mode 100644 index 0000000000000000000000000000000000000000..20268aec42140d6260600a08e06437842b4e395d --- /dev/null +++ b/data/swebenchpro/openai/gpt-5-2025-08-07/5bc6c4d1-1f11-4c24-9b59-fa6afe7e4109.json @@ -0,0 +1,112 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/openai_gpt-5-2025-08-07/1786162967.823979", + "evaluation_timestamp": "1778590202.0", + "retrieved_timestamp": "1786162967.823979", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5-2025-08-07", + "id": "openai/gpt-5-2025-08-07", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on swebenchpro/S-adaptive+C/+1ep/d9dd46ac for scorer harbor_scorer", + "source_data": { + "dataset_name": "swebenchpro", + "source_type": "hf_dataset", + "samples_number": 3, + "sample_ids": [ + "instance_ansible__ansible-42355d181a11b51ebfc56f6f4b3d9c74e01cb13b-v1055803c3a812189a1133297f7f5468579283f86", + "instance_internetarchive__openlibrary-c506c1b0b678892af5cb22c1c1dbc35d96787a0a-v0f5aece3601a5b4419f7ccec1dbda2071be28ee4", + "instance_protonmail__webclients-5f0745dd6993bb1430a951c62a49807c6635cd77" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1778590202.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 1.0, + "uncertainty": { + "num_samples": 3 + }, + "details": { + "total_matched_trajectories": "3", + "stopping_reason_count_completed_on_successful_submit": "3", + "avg_total_tokens_target_model": "3650248.67", + "total_total_tokens_target_model": "10950746.0", + "avg_total_tokens_other_models": "0.00", + "total_total_tokens_other_models": "0.0", + "avg_total_tokens_all_models": "3650248.67", + "total_total_tokens_all_models": "10950746.0", + "avg_turn_count": "69.33", + "total_turn_count": "208.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 30000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/openai/gpt-5-2025-08-07/5bc6c4d1-1f11-4c24-9b59-fa6afe7e4109_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "c1a484084876aadcaf7c48b3e87d147d099e8fbdafc1fb853df47b51eda64099", + "total_rows": 3 + } +} \ No newline at end of file diff --git a/data/swebenchpro/openai/gpt-5-2025-08-07/5bc6c4d1-1f11-4c24-9b59-fa6afe7e4109_samples.jsonl b/data/swebenchpro/openai/gpt-5-2025-08-07/5bc6c4d1-1f11-4c24-9b59-fa6afe7e4109_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..bac907531ddf616a7785b3c7f336a68d1f25abbb --- /dev/null +++ b/data/swebenchpro/openai/gpt-5-2025-08-07/5bc6c4d1-1f11-4c24-9b59-fa6afe7e4109_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c1a484084876aadcaf7c48b3e87d147d099e8fbdafc1fb853df47b51eda64099 +size 1372054 diff --git a/data/swebenchpro/openai/gpt-5-2025-08-07/609d01c6-6374-41b4-afab-86109b14b42c.json b/data/swebenchpro/openai/gpt-5-2025-08-07/609d01c6-6374-41b4-afab-86109b14b42c.json new file mode 100644 index 0000000000000000000000000000000000000000..b2c9f466de51c6e1dc6b09aa22892deec3678cfc --- /dev/null +++ b/data/swebenchpro/openai/gpt-5-2025-08-07/609d01c6-6374-41b4-afab-86109b14b42c.json @@ -0,0 +1,164 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/openai_gpt-5-2025-08-07/1786162974.031185", + "evaluation_timestamp": "1778513006.0", + "retrieved_timestamp": "1786162974.031185", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5-2025-08-07", + "id": "openai/gpt-5-2025-08-07", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on swebenchpro/S-adaptive+C/+1ep/ddffc120 for scorer harbor_scorer", + "source_data": { + "dataset_name": "swebenchpro", + "source_type": "hf_dataset", + "samples_number": 51, + "sample_ids": [ + "instance_element-hq__element-web-f63160f38459fb552d00fcc60d4064977a9095a6-vnan", + "instance_gravitational__teleport-73cc189b0e9636d418c4470ecce0d9af5dae2f02-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_future-architect__vuls-73f0adad95c4d227e2ccfa876c85cc95dd065e13", + "instance_protonmail__webclients-da91f084c0f532d9cc8ca385a701274d598057b8", + "instance_flipt-io__flipt-b2cd6a6dd73ca91b519015fd5924fde8d17f3f06", + "instance_qutebrowser__qutebrowser-cc360cd4a34a126274c7b51f3b63afbaf3e05a02-v5fc38aaf22415ab0b70567368332beee7955b367", + "instance_qutebrowser__qutebrowser-9ed748effa8f3bcd804612d9291da017b514e12f-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_ansible__ansible-1ee70fc272aff6bf3415357c6e13c5de5b928d9b-v1055803c3a812189a1133297f7f5468579283f86", + "instance_qutebrowser__qutebrowser-f631cd4422744160d9dcf7a0455da532ce973315-v35616345bb8052ea303186706cec663146f0f184", + "instance_ansible__ansible-3db08adbb1cc6aa9941be5e0fc810132c6e1fa4b-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_qutebrowser__qutebrowser-e5340c449f23608803c286da0563b62f58ba25b0-v059c6fdc75567943479b23ebca7c07b5e9a7f34c", + "instance_ansible__ansible-42355d181a11b51ebfc56f6f4b3d9c74e01cb13b-v1055803c3a812189a1133297f7f5468579283f86", + "instance_internetarchive__openlibrary-c506c1b0b678892af5cb22c1c1dbc35d96787a0a-v0f5aece3601a5b4419f7ccec1dbda2071be28ee4", + "instance_protonmail__webclients-5f0745dd6993bb1430a951c62a49807c6635cd77", + "instance_nodebb__nodebb-22368b996ee0e5f11a5189b400b33af3cc8d925a-v4fbcfae8b15e4ce5d132c408bca69ebb9cf146ed", + "instance_protonmail__webclients-2c3559cad02d1090985dba7e8eb5a129144d9811", + "instance_protonmail__webclients-01b519cd49e6a24d9a05d2eb97f54e420740072e", + "instance_ansible__ansible-77658704217d5f166404fc67997203c25381cb6e-v390e508d27db7a51eece36bb6d9698b63a5b638a", + "instance_qutebrowser__qutebrowser-2dd8966fdcf11972062c540b7a787e4d0de8d372-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_ansible__ansible-cd9c4eb5a6b2bfaf4a6709f001ce3d0c92c1eed2-v0f01c69f1e2528b935359cfe578530722bca2c59", + "instance_protonmail__webclients-ac23d1efa1a6ab7e62724779317ba44c28d78cfd", + "instance_future-architect__vuls-030b2e03525d68d74cb749959aac2d7f3fc0effa", + "instance_nodebb__nodebb-b1f9ad5534bb3a44dab5364f659876a4b7fe34c1-vnan", + "instance_flipt-io__flipt-292fdaca9be39e6a921aaa8874c011d0fdd3e874", + "instance_qutebrowser__qutebrowser-ef5ba1a0360b39f9eff027fbdc57f363597c3c3b-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_ansible__ansible-622a493ae03bd5e5cf517d336fc426e9d12208c7-v906c969b551b346ef54a2c0b41e04f632b7b73c2", + "instance_gravitational__teleport-82185f232ae8974258397e121b3bc2ed0c3729ed-v626ec2a48416b10a88641359a169d99e935ff037", + "instance_ansible__ansible-9a21e247786ebd294dafafca1105fcd770ff46c6-v67cdaa49f89b34e42b69d5b7830b3c3ad3d8803f", + "instance_protonmail__webclients-8afd9ce04c8dde9e150e1c2b50d32e7ee2efa3e7", + "instance_flipt-io__flipt-b22f5f02e40b225b6b93fff472914973422e97c6", + "instance_gravitational__teleport-2bb3bbbd8aff1164a2353381cb79e1dc93b90d28-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_ansible__ansible-de5858f48dc9e1ce9117034e0d7e76806f420ca8-v1055803c3a812189a1133297f7f5468579283f86", + "instance_gravitational__teleport-46aa81b1ce96ebb4ebed2ae53fd78cd44a05da6c-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_tutao__tutanota-8513a9e8114a8b42e64f4348335e0f23efa054c4-vee878bb72091875e912c52fc32bc60ec3760227b", + "instance_internetarchive__openlibrary-630221ab686c64e75a2ce253c893c033e4814b2e-v93c53c13d5f9b383ebb411ee7750b49dcd1a34c6", + "instance_internetarchive__openlibrary-7f6b722a10f822171501d027cad60afe53337732-ve8c8d62a2b60610a3c4631f5f23ed866bada9818", + "instance_nodebb__nodebb-a917210c5b2c20637094545401f85783905c074c-vf2cf3cbd463b7ad942381f1c6d077626485a1e9e", + "instance_internetarchive__openlibrary-1351c59fd43689753de1fca32c78d539a116ffc1-v29f82c9cf21d57b242f8d8b0e541525d259e2d63", + "instance_qutebrowser__qutebrowser-a25e8a09873838ca9efefd36ea8a45170bbeb95c-vc2f56a753b62a190ddb23cd330c257b9cf560d12", + "instance_protonmail__webclients-281a6b3f190f323ec2c0630999354fafb84b2880", + "instance_future-architect__vuls-999529a05b202b0fd29c6fca5039a4c47a3766bb", + "instance_element-hq__element-web-1077729a19c0ce902e713cf6fab42c91fb7907f1-vnan", + "instance_future-architect__vuls-be7b9114cc9545e68fb0ee7bc63d7ec53d1a00ad", + "instance_qutebrowser__qutebrowser-c580ebf0801e5a3ecabc54f327498bb753c6d5f2-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_gravitational__teleport-3ff19cf7c41f396ae468797d3aeb61515517edc9-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_future-architect__vuls-aaea15e516ece43978cf98e09e52080478b1d39f", + "instance_protonmail__webclients-f161c10cf7d31abf82e8d64d7a99c9fac5acfa18", + "instance_qutebrowser__qutebrowser-cf06f4e3708f886032d4d2a30108c2fddb042d81-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_internetarchive__openlibrary-3f580a5f244c299d936d73d9e327ba873b6401d9-v0f5aece3601a5b4419f7ccec1dbda2071be28ee4", + "instance_qutebrowser__qutebrowser-0aa57e4f7243024fa4bba8853306691b5dbd77b3-v5149fcda2a9a6fe1d35dfed1bade1444a11ef271", + "instance_future-architect__vuls-50580f6e98eeb36f53f27222f7f4fdfea0b21e8d" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1778513006.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.625, + "uncertainty": { + "standard_error": { + "value": 0.07061658535825402 + }, + "num_samples": 48 + }, + "details": { + "total_matched_trajectories": "48", + "stopping_reason_count_completed_on_successful_submit": "30", + "stopping_reason_count_repetition_guard": "18", + "avg_total_tokens_target_model": "2193883.88", + "total_total_tokens_target_model": "105306426.0", + "avg_total_tokens_other_models": "1142.46", + "total_total_tokens_other_models": "54838.0", + "avg_total_tokens_all_models": "2195026.33", + "total_total_tokens_all_models": "105361264.0", + "avg_turn_count": "55.31", + "total_turn_count": "2655.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 30000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/openai/gpt-5-2025-08-07/609d01c6-6374-41b4-afab-86109b14b42c_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "1b9e71a05f395bd4fcae03751271c1306ff890a8c57a6a0f3d2186a556f16175", + "total_rows": 48 + } +} \ No newline at end of file diff --git a/data/swebenchpro/openai/gpt-5-2025-08-07/609d01c6-6374-41b4-afab-86109b14b42c_samples.jsonl b/data/swebenchpro/openai/gpt-5-2025-08-07/609d01c6-6374-41b4-afab-86109b14b42c_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..33b75afbeda6daae03eb75cc54976535fe386b96 --- /dev/null +++ b/data/swebenchpro/openai/gpt-5-2025-08-07/609d01c6-6374-41b4-afab-86109b14b42c_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1b9e71a05f395bd4fcae03751271c1306ff890a8c57a6a0f3d2186a556f16175 +size 19262781 diff --git a/data/swebenchpro/openai/gpt-5-2025-08-07/7b9a6d8c-ca3c-4e45-bc4b-d3be8f99c7c8.json b/data/swebenchpro/openai/gpt-5-2025-08-07/7b9a6d8c-ca3c-4e45-bc4b-d3be8f99c7c8.json new file mode 100644 index 0000000000000000000000000000000000000000..634410fef2d2779153ab7412696cd2c492163967 --- /dev/null +++ b/data/swebenchpro/openai/gpt-5-2025-08-07/7b9a6d8c-ca3c-4e45-bc4b-d3be8f99c7c8.json @@ -0,0 +1,127 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/openai_gpt-5-2025-08-07/1786165443.384571", + "evaluation_timestamp": "1776696546.0", + "retrieved_timestamp": "1786165443.384571", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5-2025-08-07", + "id": "openai/gpt-5-2025-08-07", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on swebenchpro/S-adaptive/+2ep/f34ebfa1 for scorer harbor_scorer", + "source_data": { + "dataset_name": "swebenchpro", + "source_type": "hf_dataset", + "samples_number": 18, + "sample_ids": [ + "instance_internetarchive__openlibrary-fdbc0d8f418333c7e575c40b661b582c301ef7ac-v13642507b4fc1f8d234172bf8129942da2c2ca26", + "instance_ansible__ansible-b2a289dcbb702003377221e25f62c8a3608f0e89-v173091e2e36d38c978002990795f66cfc0af30ad", + "instance_internetarchive__openlibrary-f8cc11d9c1575fdba5ac66aee0befca970da8d64-v13642507b4fc1f8d234172bf8129942da2c2ca26", + "instance_qutebrowser__qutebrowser-e70f5b03187bdd40e8bf70f5f3ead840f52d1f42-v02ad04386d5238fe2d1a1be450df257370de4b6a", + "instance_tutao__tutanota-d1aa0ecec288bfc800cfb9133b087c4f81ad8b38-vbc0d9ba8f0071fbe982809910959a6ff8884dbbf", + "instance_qutebrowser__qutebrowser-e64622cd2df5b521342cf4a62e0d4cb8f8c9ae5a-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_gravitational__teleport-ac2fb2f9b4fd1896b554d3011df23d3d71295779", + "instance_protonmail__webclients-cb8cc309c6968b0a2a5fe4288d0ae0a969ff31e1", + "instance_ansible__ansible-11c1777d56664b1acb56b387a1ad6aeadef1391d-v0f01c69f1e2528b935359cfe578530722bca2c59", + "instance_ansible__ansible-29aea9ff3466e4cd2ed00524b9e56738d568ce8b-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_nodebb__nodebb-70b4a0e2aebebe8f2f559de6680093d96a697b2f-vnan", + "instance_ansible__ansible-e22e103cdf8edc56ff7d9b848a58f94f1471a263-v1055803c3a812189a1133297f7f5468579283f86", + "instance_qutebrowser__qutebrowser-e15d26630934d0b6415ed2295ac42fd570a57620-va0fd88aac89cde702ec1ba84877234da33adce8a", + "instance_ansible__ansible-164881d871964aa64e0f911d03ae270acbad253c-v390e508d27db7a51eece36bb6d9698b63a5b638a", + "instance_ansible__ansible-502270c804c33d3bc963930dc85e0f4ca359674d-v7eee2454f617569fd6889f2211f75bc02a35f9f8", + "instance_internetarchive__openlibrary-9c392b60e2c6fa1d68cb68084b4b4ff04d0cb35c-v2d9a6c849c60ed19fd0858ce9e40b7cc8e097e59", + "instance_element-hq__element-web-b007ea81b2ccd001b00f332bee65070aa7fc00f9-vnan", + "instance_internetarchive__openlibrary-7cbfb812ef0e1f9716e2d6e85d538a96fcb79d13-vfa6ff903cb27f336e17654595dd900fa943dcd91" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1776696546.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.8125, + "uncertainty": { + "standard_error": { + "value": 0.07739239842086129 + }, + "num_samples": 28 + }, + "details": { + "total_matched_trajectories": "28", + "stopping_reason_count_repetition_guard": "28", + "avg_total_tokens_target_model": "2413882.36", + "total_total_tokens_target_model": "67588706.0", + "avg_total_tokens_other_models": "1421.18", + "total_total_tokens_other_models": "39793.0", + "avg_total_tokens_all_models": "2415303.54", + "total_total_tokens_all_models": "67628499.0", + "avg_turn_count": "55.29", + "total_turn_count": "1548.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "3600", + "attempt_timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 30000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/openai/gpt-5-2025-08-07/7b9a6d8c-ca3c-4e45-bc4b-d3be8f99c7c8_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "28ec3009adbcb273d9ac4a7c70fd35d8e9244c47f6c31bd3bb06cf592840f7d6", + "total_rows": 28 + } +} \ No newline at end of file diff --git a/data/swebenchpro/openai/gpt-5-2025-08-07/7b9a6d8c-ca3c-4e45-bc4b-d3be8f99c7c8_samples.jsonl b/data/swebenchpro/openai/gpt-5-2025-08-07/7b9a6d8c-ca3c-4e45-bc4b-d3be8f99c7c8_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..6deee998c44233229b75aa064abdf7f59762035a --- /dev/null +++ b/data/swebenchpro/openai/gpt-5-2025-08-07/7b9a6d8c-ca3c-4e45-bc4b-d3be8f99c7c8_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:28ec3009adbcb273d9ac4a7c70fd35d8e9244c47f6c31bd3bb06cf592840f7d6 +size 10873297 diff --git a/data/swebenchpro/openai/gpt-5-2025-08-07/7c3771b0-fa13-4612-96e7-467c8e63ed22.json b/data/swebenchpro/openai/gpt-5-2025-08-07/7c3771b0-fa13-4612-96e7-467c8e63ed22.json new file mode 100644 index 0000000000000000000000000000000000000000..fb2771545e51e28272a73a4365469f3e4f903ee9 --- /dev/null +++ b/data/swebenchpro/openai/gpt-5-2025-08-07/7c3771b0-fa13-4612-96e7-467c8e63ed22.json @@ -0,0 +1,116 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/openai_gpt-5-2025-08-07/1786165463.063482", + "evaluation_timestamp": "1776629382.0", + "retrieved_timestamp": "1786165463.063482", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5-2025-08-07", + "id": "openai/gpt-5-2025-08-07", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on swebenchpro/S-adaptive/+2ep/b8e02e64 for scorer harbor_scorer", + "source_data": { + "dataset_name": "swebenchpro", + "source_type": "hf_dataset", + "samples_number": 9, + "sample_ids": [ + "instance_qutebrowser__qutebrowser-2e961080a85d660148937ee8f0f6b3445a8f2c01-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_flipt-io__flipt-b68b8960b8a08540d5198d78c665a7eb0bea4008", + "instance_qutebrowser__qutebrowser-01d1d1494411380d97cac14614a829d3a69cecaf-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_future-architect__vuls-78b52d6a7f480bd610b692de9bf0c86f57332f23", + "instance_flipt-io__flipt-e42da21a07a5ae35835ec54f74004ebd58713874", + "instance_nodebb__nodebb-84e065752f6d7fbe5c08cbf50cb173ffb866b8fa-vf2cf3cbd463b7ad942381f1c6d077626485a1e9e", + "instance_element-hq__element-web-f3534b42df3dcfe36dc48bddbf14034085af6d30-vnan", + "instance_future-architect__vuls-f0b3a8b1db98eb1bd32685f1c36c41a99c3452ed", + "instance_future-architect__vuls-6eff6a9329a65cc412e79b8f82444dfa3d0f0b5a" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1776629382.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.4444444444444444, + "uncertainty": { + "standard_error": { + "value": 0.17568209223157663 + }, + "num_samples": 18 + }, + "details": { + "total_matched_trajectories": "18", + "stopping_reason_count_repetition_guard": "18", + "avg_total_tokens_target_model": "6571998.72", + "total_total_tokens_target_model": "118295977.0", + "avg_total_tokens_other_models": "1840.94", + "total_total_tokens_other_models": "33137.0", + "avg_total_tokens_all_models": "6573839.67", + "total_total_tokens_all_models": "118329114.0", + "avg_turn_count": "110.56", + "total_turn_count": "1990.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 30000000 + }, + "sandbox": {} + }, + "additional_details": { + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/openai/gpt-5-2025-08-07/7c3771b0-fa13-4612-96e7-467c8e63ed22_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "21a61992cbc0d40132366de405897ee4e112d1f0662c5ddbe5b897ceba3e7e2a", + "total_rows": 18 + } +} \ No newline at end of file diff --git a/data/swebenchpro/openai/gpt-5-2025-08-07/7c3771b0-fa13-4612-96e7-467c8e63ed22_samples.jsonl b/data/swebenchpro/openai/gpt-5-2025-08-07/7c3771b0-fa13-4612-96e7-467c8e63ed22_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..490774d132d181d1482aa2b8f7b074ef9384ffa0 --- /dev/null +++ b/data/swebenchpro/openai/gpt-5-2025-08-07/7c3771b0-fa13-4612-96e7-467c8e63ed22_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:21a61992cbc0d40132366de405897ee4e112d1f0662c5ddbe5b897ceba3e7e2a +size 11899255 diff --git a/data/swebenchpro/openai/gpt-5-2025-08-07/83c45e91-1a1e-42aa-88f9-74c3bad3fc35.json b/data/swebenchpro/openai/gpt-5-2025-08-07/83c45e91-1a1e-42aa-88f9-74c3bad3fc35.json new file mode 100644 index 0000000000000000000000000000000000000000..88692f3d065a5651dded651e65890011bd8fd6ea --- /dev/null +++ b/data/swebenchpro/openai/gpt-5-2025-08-07/83c45e91-1a1e-42aa-88f9-74c3bad3fc35.json @@ -0,0 +1,122 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/openai_gpt-5-2025-08-07/1786164478.758943", + "evaluation_timestamp": "1777909399.0", + "retrieved_timestamp": "1786164478.758943", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5-2025-08-07", + "id": "openai/gpt-5-2025-08-07", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on swebenchpro/S-adaptive/+1ep/fbb66e57 for scorer harbor_scorer", + "source_data": { + "dataset_name": "swebenchpro", + "source_type": "hf_dataset", + "samples_number": 10, + "sample_ids": [ + "instance_nodebb__nodebb-22368b996ee0e5f11a5189b400b33af3cc8d925a-v4fbcfae8b15e4ce5d132c408bca69ebb9cf146ed", + "instance_protonmail__webclients-ac23d1efa1a6ab7e62724779317ba44c28d78cfd", + "instance_nodebb__nodebb-b1f9ad5534bb3a44dab5364f659876a4b7fe34c1-vnan", + "instance_flipt-io__flipt-292fdaca9be39e6a921aaa8874c011d0fdd3e874", + "instance_qutebrowser__qutebrowser-ef5ba1a0360b39f9eff027fbdc57f363597c3c3b-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_ansible__ansible-622a493ae03bd5e5cf517d336fc426e9d12208c7-v906c969b551b346ef54a2c0b41e04f632b7b73c2", + "instance_gravitational__teleport-82185f232ae8974258397e121b3bc2ed0c3729ed-v626ec2a48416b10a88641359a169d99e935ff037", + "instance_ansible__ansible-9a21e247786ebd294dafafca1105fcd770ff46c6-v67cdaa49f89b34e42b69d5b7830b3c3ad3d8803f", + "instance_protonmail__webclients-8afd9ce04c8dde9e150e1c2b50d32e7ee2efa3e7", + "instance_flipt-io__flipt-b22f5f02e40b225b6b93fff472914973422e97c6" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1777909399.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.6, + "uncertainty": { + "standard_error": { + "value": 0.1632993161855452 + }, + "num_samples": 10 + }, + "details": { + "total_matched_trajectories": "10", + "stopping_reason_count_repetition_guard": "10", + "avg_total_tokens_target_model": "1950212.80", + "total_total_tokens_target_model": "19502128.0", + "avg_total_tokens_other_models": "1756.90", + "total_total_tokens_other_models": "17569.0", + "avg_total_tokens_all_models": "1951969.70", + "total_total_tokens_all_models": "19519697.0", + "avg_turn_count": "59.80", + "total_turn_count": "598.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 30000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/openai/gpt-5-2025-08-07/83c45e91-1a1e-42aa-88f9-74c3bad3fc35_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "b048eab12723f7a707bcdd2f849c57b6f818ea1b4c7e9f13f0aa25fe6bcd58e9", + "total_rows": 10 + } +} \ No newline at end of file diff --git a/data/swebenchpro/openai/gpt-5-2025-08-07/83c45e91-1a1e-42aa-88f9-74c3bad3fc35_samples.jsonl b/data/swebenchpro/openai/gpt-5-2025-08-07/83c45e91-1a1e-42aa-88f9-74c3bad3fc35_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..87719637eb17baf4bd57f1d20e50e1487fd3fcbf --- /dev/null +++ b/data/swebenchpro/openai/gpt-5-2025-08-07/83c45e91-1a1e-42aa-88f9-74c3bad3fc35_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b048eab12723f7a707bcdd2f849c57b6f818ea1b4c7e9f13f0aa25fe6bcd58e9 +size 4032634 diff --git a/data/swebenchpro/openai/gpt-5-2025-08-07/8847e6cc-dd41-486e-a35c-5d0c46bd70ce.json b/data/swebenchpro/openai/gpt-5-2025-08-07/8847e6cc-dd41-486e-a35c-5d0c46bd70ce.json new file mode 100644 index 0000000000000000000000000000000000000000..25dcda50f750ab38f2a8eff246ac9f492d5de83d --- /dev/null +++ b/data/swebenchpro/openai/gpt-5-2025-08-07/8847e6cc-dd41-486e-a35c-5d0c46bd70ce.json @@ -0,0 +1,110 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/openai_gpt-5-2025-08-07/1786162977.736362", + "evaluation_timestamp": "1778193699.0", + "retrieved_timestamp": "1786162977.736362", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5-2025-08-07", + "id": "openai/gpt-5-2025-08-07", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on swebenchpro/S-adaptive+C/+2ep/60d44e75 for scorer harbor_scorer", + "source_data": { + "dataset_name": "swebenchpro", + "source_type": "hf_dataset", + "samples_number": 1, + "sample_ids": [ + "instance_navidrome__navidrome-d8e794317f788198227e10fb667e10496b3eb99a" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1778193699.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.0, + "uncertainty": { + "num_samples": 2 + }, + "details": { + "total_matched_trajectories": "2", + "stopping_reason_count_repetition_guard": "2", + "avg_total_tokens_target_model": "7248959.50", + "total_total_tokens_target_model": "14497919.0", + "avg_total_tokens_other_models": "3519.00", + "total_total_tokens_other_models": "7038.0", + "avg_total_tokens_all_models": "7252478.50", + "total_total_tokens_all_models": "14504957.0", + "avg_turn_count": "141.50", + "total_turn_count": "283.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 30000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/openai/gpt-5-2025-08-07/8847e6cc-dd41-486e-a35c-5d0c46bd70ce_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "f74b8308cb8617431856c473b385f75883ea43a4a3fbe0a44daebba1aa201beb", + "total_rows": 2 + } +} \ No newline at end of file diff --git a/data/swebenchpro/openai/gpt-5-2025-08-07/8847e6cc-dd41-486e-a35c-5d0c46bd70ce_samples.jsonl b/data/swebenchpro/openai/gpt-5-2025-08-07/8847e6cc-dd41-486e-a35c-5d0c46bd70ce_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..21b135387eabfdf5700e9c69aeaecc47697d8a20 --- /dev/null +++ b/data/swebenchpro/openai/gpt-5-2025-08-07/8847e6cc-dd41-486e-a35c-5d0c46bd70ce_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f74b8308cb8617431856c473b385f75883ea43a4a3fbe0a44daebba1aa201beb +size 1665440 diff --git a/data/swebenchpro/openai/gpt-5-2025-08-07/8b6794a7-8f5d-481a-9330-90073549048b.json b/data/swebenchpro/openai/gpt-5-2025-08-07/8b6794a7-8f5d-481a-9330-90073549048b.json new file mode 100644 index 0000000000000000000000000000000000000000..67c9bbb3f813182327c495ef12c25840268f0f7f --- /dev/null +++ b/data/swebenchpro/openai/gpt-5-2025-08-07/8b6794a7-8f5d-481a-9330-90073549048b.json @@ -0,0 +1,112 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/openai_gpt-5-2025-08-07/1786164481.384418", + "evaluation_timestamp": "1777375719.0", + "retrieved_timestamp": "1786164481.384418", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5-2025-08-07", + "id": "openai/gpt-5-2025-08-07", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on swebenchpro/S-adaptive/+1ep/30d2601a for scorer harbor_scorer", + "source_data": { + "dataset_name": "swebenchpro", + "source_type": "hf_dataset", + "samples_number": 3, + "sample_ids": [ + "instance_ansible__ansible-42355d181a11b51ebfc56f6f4b3d9c74e01cb13b-v1055803c3a812189a1133297f7f5468579283f86", + "instance_internetarchive__openlibrary-09865f5fb549694d969f0a8e49b9d204ef1853ca-ve8c8d62a2b60610a3c4631f5f23ed866bada9818", + "instance_protonmail__webclients-01ea5214d11e0df8b7170d91bafd34f23cb0f2b1" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1777375719.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.6666666666666666, + "uncertainty": { + "standard_error": { + "value": 0.33333333333333337 + }, + "num_samples": 3 + }, + "details": { + "total_matched_trajectories": "3", + "stopping_reason_count_repetition_guard": "3", + "avg_total_tokens_target_model": "6829736.00", + "total_total_tokens_target_model": "20489208.0", + "avg_total_tokens_other_models": "1753.33", + "total_total_tokens_other_models": "5260.0", + "avg_total_tokens_all_models": "6831489.33", + "total_total_tokens_all_models": "20494468.0", + "avg_turn_count": "108.33", + "total_turn_count": "325.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 30000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/openai/gpt-5-2025-08-07/8b6794a7-8f5d-481a-9330-90073549048b_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "62d160cb2f3453f48e879d0720fc5a84ac7ce9838bd1f36ff40dbb91f076a8d9", + "total_rows": 3 + } +} \ No newline at end of file diff --git a/data/swebenchpro/openai/gpt-5-2025-08-07/8b6794a7-8f5d-481a-9330-90073549048b_samples.jsonl b/data/swebenchpro/openai/gpt-5-2025-08-07/8b6794a7-8f5d-481a-9330-90073549048b_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..3c522e7a0da8d4fe319ce59974b03055b85f29aa --- /dev/null +++ b/data/swebenchpro/openai/gpt-5-2025-08-07/8b6794a7-8f5d-481a-9330-90073549048b_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:62d160cb2f3453f48e879d0720fc5a84ac7ce9838bd1f36ff40dbb91f076a8d9 +size 1745886 diff --git a/data/swebenchpro/openai/gpt-5-2025-08-07/94269635-3b9e-4ce9-8dcc-c9c8503a40fb.json b/data/swebenchpro/openai/gpt-5-2025-08-07/94269635-3b9e-4ce9-8dcc-c9c8503a40fb.json new file mode 100644 index 0000000000000000000000000000000000000000..be5826deb9b1033ef313ab3dd6c23e3919c640dc --- /dev/null +++ b/data/swebenchpro/openai/gpt-5-2025-08-07/94269635-3b9e-4ce9-8dcc-c9c8503a40fb.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/openai_gpt-5-2025-08-07/1786169134.180185", + "evaluation_timestamp": "1776520455.0", + "retrieved_timestamp": "1786169134.180185", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5-2025-08-07", + "id": "openai/gpt-5-2025-08-07", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/openai/gpt-5-2025-08-07/94269635-3b9e-4ce9-8dcc-c9c8503a40fb_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "659e3eb42bf74e48f73301e5543231bbe2b962819a6fed65fd777006f95fd37c", + "total_rows": 50 + } +} \ No newline at end of file diff --git a/data/swebenchpro/openai/gpt-5-2025-08-07/94269635-3b9e-4ce9-8dcc-c9c8503a40fb_samples.jsonl b/data/swebenchpro/openai/gpt-5-2025-08-07/94269635-3b9e-4ce9-8dcc-c9c8503a40fb_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..d8c8f67f880ad1e8b678ac17f2638b8a26bc11c8 --- /dev/null +++ b/data/swebenchpro/openai/gpt-5-2025-08-07/94269635-3b9e-4ce9-8dcc-c9c8503a40fb_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:659e3eb42bf74e48f73301e5543231bbe2b962819a6fed65fd777006f95fd37c +size 26961860 diff --git a/data/swebenchpro/openai/gpt-5-2025-08-07/9529cbb8-dd6c-44ac-96ae-bba0cb1a60db.json b/data/swebenchpro/openai/gpt-5-2025-08-07/9529cbb8-dd6c-44ac-96ae-bba0cb1a60db.json new file mode 100644 index 0000000000000000000000000000000000000000..aa495849315d0dfa7405a4ba71074bd0cd05ade0 --- /dev/null +++ b/data/swebenchpro/openai/gpt-5-2025-08-07/9529cbb8-dd6c-44ac-96ae-bba0cb1a60db.json @@ -0,0 +1,108 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/openai_gpt-5-2025-08-07/1786165452.126038", + "evaluation_timestamp": "1777399517.0", + "retrieved_timestamp": "1786165452.126038", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5-2025-08-07", + "id": "openai/gpt-5-2025-08-07", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on swebenchpro/S-adaptive/+2ep/03ed0e33 for scorer harbor_scorer", + "source_data": { + "dataset_name": "swebenchpro", + "source_type": "hf_dataset", + "samples_number": 2, + "sample_ids": [ + "instance_flipt-io__flipt-05d7234fa582df632f70a7cd10194d61bd7043b9", + "instance_protonmail__webclients-5f0745dd6993bb1430a951c62a49807c6635cd77" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1777399517.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.0, + "uncertainty": { + "num_samples": 4 + }, + "details": { + "total_matched_trajectories": "4", + "stopping_reason_count_repetition_guard": "4", + "avg_total_tokens_target_model": "16301202.00", + "total_total_tokens_target_model": "65204808.0", + "avg_total_tokens_other_models": "1837.25", + "total_total_tokens_other_models": "7349.0", + "avg_total_tokens_all_models": "16303039.25", + "total_total_tokens_all_models": "65212157.0", + "avg_turn_count": "238.75", + "total_turn_count": "955.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 30000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/openai/gpt-5-2025-08-07/9529cbb8-dd6c-44ac-96ae-bba0cb1a60db_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "ddba1b5bb6efaa95d20ac78798d37ebb0021368a3db398b467e2075caa3b4c9f", + "total_rows": 4 + } +} \ No newline at end of file diff --git a/data/swebenchpro/openai/gpt-5-2025-08-07/9529cbb8-dd6c-44ac-96ae-bba0cb1a60db_samples.jsonl b/data/swebenchpro/openai/gpt-5-2025-08-07/9529cbb8-dd6c-44ac-96ae-bba0cb1a60db_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..eab17851a03b5dd7a4a87dd72880d54e476df84a --- /dev/null +++ b/data/swebenchpro/openai/gpt-5-2025-08-07/9529cbb8-dd6c-44ac-96ae-bba0cb1a60db_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ddba1b5bb6efaa95d20ac78798d37ebb0021368a3db398b467e2075caa3b4c9f +size 5328960 diff --git a/data/swebenchpro/openai/gpt-5-2025-08-07/96e19259-80de-40fc-8840-b09ac4b16b45.json b/data/swebenchpro/openai/gpt-5-2025-08-07/96e19259-80de-40fc-8840-b09ac4b16b45.json new file mode 100644 index 0000000000000000000000000000000000000000..424be35b97a9a5b74d2296cc9295652899e99ed1 --- /dev/null +++ b/data/swebenchpro/openai/gpt-5-2025-08-07/96e19259-80de-40fc-8840-b09ac4b16b45.json @@ -0,0 +1,130 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/openai_gpt-5-2025-08-07/1786162963.358636", + "evaluation_timestamp": "1778265533.0", + "retrieved_timestamp": "1786162963.358636", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5-2025-08-07", + "id": "openai/gpt-5-2025-08-07", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on swebenchpro/S-adaptive+C/+1ep/d035b118 for scorer harbor_scorer", + "source_data": { + "dataset_name": "swebenchpro", + "source_type": "hf_dataset", + "samples_number": 17, + "sample_ids": [ + "instance_future-architect__vuls-f0b3a8b1db98eb1bd32685f1c36c41a99c3452ed", + "instance_future-architect__vuls-6eff6a9329a65cc412e79b8f82444dfa3d0f0b5a", + "instance_internetarchive__openlibrary-fdbc0d8f418333c7e575c40b661b582c301ef7ac-v13642507b4fc1f8d234172bf8129942da2c2ca26", + "instance_ansible__ansible-b2a289dcbb702003377221e25f62c8a3608f0e89-v173091e2e36d38c978002990795f66cfc0af30ad", + "instance_internetarchive__openlibrary-f8cc11d9c1575fdba5ac66aee0befca970da8d64-v13642507b4fc1f8d234172bf8129942da2c2ca26", + "instance_qutebrowser__qutebrowser-e70f5b03187bdd40e8bf70f5f3ead840f52d1f42-v02ad04386d5238fe2d1a1be450df257370de4b6a", + "instance_qutebrowser__qutebrowser-e64622cd2df5b521342cf4a62e0d4cb8f8c9ae5a-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_protonmail__webclients-cb8cc309c6968b0a2a5fe4288d0ae0a969ff31e1", + "instance_ansible__ansible-11c1777d56664b1acb56b387a1ad6aeadef1391d-v0f01c69f1e2528b935359cfe578530722bca2c59", + "instance_ansible__ansible-29aea9ff3466e4cd2ed00524b9e56738d568ce8b-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_nodebb__nodebb-70b4a0e2aebebe8f2f559de6680093d96a697b2f-vnan", + "instance_ansible__ansible-e22e103cdf8edc56ff7d9b848a58f94f1471a263-v1055803c3a812189a1133297f7f5468579283f86", + "instance_qutebrowser__qutebrowser-e15d26630934d0b6415ed2295ac42fd570a57620-va0fd88aac89cde702ec1ba84877234da33adce8a", + "instance_ansible__ansible-164881d871964aa64e0f911d03ae270acbad253c-v390e508d27db7a51eece36bb6d9698b63a5b638a", + "instance_ansible__ansible-502270c804c33d3bc963930dc85e0f4ca359674d-v7eee2454f617569fd6889f2211f75bc02a35f9f8", + "instance_internetarchive__openlibrary-9c392b60e2c6fa1d68cb68084b4b4ff04d0cb35c-v2d9a6c849c60ed19fd0858ce9e40b7cc8e097e59", + "instance_element-hq__element-web-b007ea81b2ccd001b00f332bee65070aa7fc00f9-vnan" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1778265533.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.8235294117647058, + "uncertainty": { + "standard_error": { + "value": 0.09530501027070383 + }, + "num_samples": 17 + }, + "details": { + "total_matched_trajectories": "17", + "stopping_reason_count_completed_on_successful_submit": "14", + "stopping_reason_count_repetition_guard": "3", + "avg_total_tokens_target_model": "1899419.18", + "total_total_tokens_target_model": "32290126.0", + "avg_total_tokens_other_models": "491.29", + "total_total_tokens_other_models": "8352.0", + "avg_total_tokens_all_models": "1899910.47", + "total_total_tokens_all_models": "32298478.0", + "avg_turn_count": "46.53", + "total_turn_count": "791.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 30000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/openai/gpt-5-2025-08-07/96e19259-80de-40fc-8840-b09ac4b16b45_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "35869f148646395674ac8787a4a60af7d51fb5163110f4cb24ff46c00aa96a9f", + "total_rows": 17 + } +} \ No newline at end of file diff --git a/data/swebenchpro/openai/gpt-5-2025-08-07/96e19259-80de-40fc-8840-b09ac4b16b45_samples.jsonl b/data/swebenchpro/openai/gpt-5-2025-08-07/96e19259-80de-40fc-8840-b09ac4b16b45_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..bd7c45f8320070767f6356cf81dcc2bc053d3502 --- /dev/null +++ b/data/swebenchpro/openai/gpt-5-2025-08-07/96e19259-80de-40fc-8840-b09ac4b16b45_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:35869f148646395674ac8787a4a60af7d51fb5163110f4cb24ff46c00aa96a9f +size 5452625 diff --git a/data/swebenchpro/openai/gpt-5-2025-08-07/a10f3ba0-45b2-4ce2-8f38-20925cdeeff0.json b/data/swebenchpro/openai/gpt-5-2025-08-07/a10f3ba0-45b2-4ce2-8f38-20925cdeeff0.json new file mode 100644 index 0000000000000000000000000000000000000000..98ea8891daf4bd687893d85eb2136580892d2025 --- /dev/null +++ b/data/swebenchpro/openai/gpt-5-2025-08-07/a10f3ba0-45b2-4ce2-8f38-20925cdeeff0.json @@ -0,0 +1,186 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/openai_gpt-5-2025-08-07/1786163026.34849", + "evaluation_timestamp": "1778276135.0", + "retrieved_timestamp": "1786163026.34849", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5-2025-08-07", + "id": "openai/gpt-5-2025-08-07", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on swebenchpro/S-adaptive+C/4a748958 for scorer harbor_scorer", + "source_data": { + "dataset_name": "swebenchpro", + "source_type": "hf_dataset", + "samples_number": 72, + "sample_ids": [ + "instance_ansible__ansible-811093f0225caa4dd33890933150a81c6a6d5226-v1055803c3a812189a1133297f7f5468579283f86", + "instance_internetarchive__openlibrary-7cbfb812ef0e1f9716e2d6e85d538a96fcb79d13-vfa6ff903cb27f336e17654595dd900fa943dcd91", + "instance_nodebb__nodebb-0f788b8eaa4bba3c142d171fd941d015c53b65fc-v0ec6d6c2baf3cb4797482ce4829bc25cd5716649", + "instance_flipt-io__flipt-0b119520afca1cf25c470ff4288c464d4510b944", + "instance_element-hq__element-web-f63160f38459fb552d00fcc60d4064977a9095a6-vnan", + "instance_ansible__ansible-f327e65d11bb905ed9f15996024f857a95592629-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_gravitational__teleport-73cc189b0e9636d418c4470ecce0d9af5dae2f02-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_internetarchive__openlibrary-3c48b4bb782189e0858e6c3fc7956046cf3e1cfb-v2d9a6c849c60ed19fd0858ce9e40b7cc8e097e59", + "instance_future-architect__vuls-878c25bf5a9c9fd88ac32eb843f5636834d5712d", + "instance_element-hq__element-web-9bf77963ee5e036d54b2a3ca202fbf6378464a5e-vnan", + "instance_qutebrowser__qutebrowser-70248f256f93ed9b1984494d0a1a919ddd774892-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_future-architect__vuls-73f0adad95c4d227e2ccfa876c85cc95dd065e13", + "instance_protonmail__webclients-da91f084c0f532d9cc8ca385a701274d598057b8", + "instance_flipt-io__flipt-b2cd6a6dd73ca91b519015fd5924fde8d17f3f06", + "instance_qutebrowser__qutebrowser-cc360cd4a34a126274c7b51f3b63afbaf3e05a02-v5fc38aaf22415ab0b70567368332beee7955b367", + "instance_flipt-io__flipt-e2bd19dafa7166c96b082fb2a59eb54b4be0d778", + "instance_qutebrowser__qutebrowser-9ed748effa8f3bcd804612d9291da017b514e12f-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_ansible__ansible-1ee70fc272aff6bf3415357c6e13c5de5b928d9b-v1055803c3a812189a1133297f7f5468579283f86", + "instance_element-hq__element-web-75c2c1a572fa45d1ea1d1a96e9e36e303332ecaa-vnan", + "instance_qutebrowser__qutebrowser-f631cd4422744160d9dcf7a0455da532ce973315-v35616345bb8052ea303186706cec663146f0f184", + "instance_ansible__ansible-3db08adbb1cc6aa9941be5e0fc810132c6e1fa4b-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_qutebrowser__qutebrowser-e5340c449f23608803c286da0563b62f58ba25b0-v059c6fdc75567943479b23ebca7c07b5e9a7f34c", + "instance_ansible__ansible-42355d181a11b51ebfc56f6f4b3d9c74e01cb13b-v1055803c3a812189a1133297f7f5468579283f86", + "instance_internetarchive__openlibrary-09865f5fb549694d969f0a8e49b9d204ef1853ca-ve8c8d62a2b60610a3c4631f5f23ed866bada9818", + "instance_internetarchive__openlibrary-c506c1b0b678892af5cb22c1c1dbc35d96787a0a-v0f5aece3601a5b4419f7ccec1dbda2071be28ee4", + "instance_element-hq__element-web-2760bfc8369f1bee640d6d7a7e910783143d4c5f-vnan", + "instance_internetarchive__openlibrary-25858f9f0c165df25742acf8309ce909773f0cdd-v13642507b4fc1f8d234172bf8129942da2c2ca26", + "instance_protonmail__webclients-01ea5214d11e0df8b7170d91bafd34f23cb0f2b1", + "instance_tutao__tutanota-4b4e45949096bb288f2b522f657610e480efa3e8-vee878bb72091875e912c52fc32bc60ec3760227b", + "instance_flipt-io__flipt-05d7234fa582df632f70a7cd10194d61bd7043b9", + "instance_ansible__ansible-a02e22e902a69aeb465f16bf03f7f5a91b2cb828-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_protonmail__webclients-5f0745dd6993bb1430a951c62a49807c6635cd77", + "instance_tutao__tutanota-1ff82aa365763cee2d609c9d19360ad87fdf2ec7-vc4e41fd0029957297843cb9dec4a25c7c756f029", + "instance_internetarchive__openlibrary-30bc73a1395fba2300087c7f307e54bb5372b60a-v76304ecdb3a5954fcf13feb710e8c40fcf24b73c", + "instance_nodebb__nodebb-8ca65b0c78c67c1653487c02d1135e1b702185e1-vf2cf3cbd463b7ad942381f1c6d077626485a1e9e", + "instance_nodebb__nodebb-22368b996ee0e5f11a5189b400b33af3cc8d925a-v4fbcfae8b15e4ce5d132c408bca69ebb9cf146ed", + "instance_protonmail__webclients-2c3559cad02d1090985dba7e8eb5a129144d9811", + "instance_protonmail__webclients-01b519cd49e6a24d9a05d2eb97f54e420740072e", + "instance_ansible__ansible-77658704217d5f166404fc67997203c25381cb6e-v390e508d27db7a51eece36bb6d9698b63a5b638a", + "instance_qutebrowser__qutebrowser-2dd8966fdcf11972062c540b7a787e4d0de8d372-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_ansible__ansible-cd9c4eb5a6b2bfaf4a6709f001ce3d0c92c1eed2-v0f01c69f1e2528b935359cfe578530722bca2c59", + "instance_protonmail__webclients-ac23d1efa1a6ab7e62724779317ba44c28d78cfd", + "instance_future-architect__vuls-030b2e03525d68d74cb749959aac2d7f3fc0effa", + "instance_nodebb__nodebb-b1f9ad5534bb3a44dab5364f659876a4b7fe34c1-vnan", + "instance_flipt-io__flipt-292fdaca9be39e6a921aaa8874c011d0fdd3e874", + "instance_qutebrowser__qutebrowser-ef5ba1a0360b39f9eff027fbdc57f363597c3c3b-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_ansible__ansible-622a493ae03bd5e5cf517d336fc426e9d12208c7-v906c969b551b346ef54a2c0b41e04f632b7b73c2", + "instance_gravitational__teleport-82185f232ae8974258397e121b3bc2ed0c3729ed-v626ec2a48416b10a88641359a169d99e935ff037", + "instance_ansible__ansible-9a21e247786ebd294dafafca1105fcd770ff46c6-v67cdaa49f89b34e42b69d5b7830b3c3ad3d8803f", + "instance_protonmail__webclients-8afd9ce04c8dde9e150e1c2b50d32e7ee2efa3e7", + "instance_flipt-io__flipt-b22f5f02e40b225b6b93fff472914973422e97c6", + "instance_gravitational__teleport-2bb3bbbd8aff1164a2353381cb79e1dc93b90d28-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_ansible__ansible-de5858f48dc9e1ce9117034e0d7e76806f420ca8-v1055803c3a812189a1133297f7f5468579283f86", + "instance_gravitational__teleport-46aa81b1ce96ebb4ebed2ae53fd78cd44a05da6c-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_tutao__tutanota-8513a9e8114a8b42e64f4348335e0f23efa054c4-vee878bb72091875e912c52fc32bc60ec3760227b", + "instance_internetarchive__openlibrary-630221ab686c64e75a2ce253c893c033e4814b2e-v93c53c13d5f9b383ebb411ee7750b49dcd1a34c6", + "instance_internetarchive__openlibrary-7f6b722a10f822171501d027cad60afe53337732-ve8c8d62a2b60610a3c4631f5f23ed866bada9818", + "instance_nodebb__nodebb-a917210c5b2c20637094545401f85783905c074c-vf2cf3cbd463b7ad942381f1c6d077626485a1e9e", + "instance_internetarchive__openlibrary-1351c59fd43689753de1fca32c78d539a116ffc1-v29f82c9cf21d57b242f8d8b0e541525d259e2d63", + "instance_qutebrowser__qutebrowser-a25e8a09873838ca9efefd36ea8a45170bbeb95c-vc2f56a753b62a190ddb23cd330c257b9cf560d12", + "instance_protonmail__webclients-281a6b3f190f323ec2c0630999354fafb84b2880", + "instance_future-architect__vuls-999529a05b202b0fd29c6fca5039a4c47a3766bb", + "instance_element-hq__element-web-1077729a19c0ce902e713cf6fab42c91fb7907f1-vnan", + "instance_future-architect__vuls-be7b9114cc9545e68fb0ee7bc63d7ec53d1a00ad", + "instance_qutebrowser__qutebrowser-c580ebf0801e5a3ecabc54f327498bb753c6d5f2-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_gravitational__teleport-3ff19cf7c41f396ae468797d3aeb61515517edc9-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_future-architect__vuls-aaea15e516ece43978cf98e09e52080478b1d39f", + "instance_protonmail__webclients-f161c10cf7d31abf82e8d64d7a99c9fac5acfa18", + "instance_qutebrowser__qutebrowser-cf06f4e3708f886032d4d2a30108c2fddb042d81-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_internetarchive__openlibrary-3f580a5f244c299d936d73d9e327ba873b6401d9-v0f5aece3601a5b4419f7ccec1dbda2071be28ee4", + "instance_qutebrowser__qutebrowser-0aa57e4f7243024fa4bba8853306691b5dbd77b3-v5149fcda2a9a6fe1d35dfed1bade1444a11ef271", + "instance_future-architect__vuls-50580f6e98eeb36f53f27222f7f4fdfea0b21e8d" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1778276135.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.6726851851851852, + "uncertainty": { + "standard_error": { + "value": 0.05035768595658604 + }, + "num_samples": 284 + }, + "details": { + "total_matched_trajectories": "284", + "stopping_reason_count_completed_on_successful_submit": "190", + "stopping_reason_count_repetition_guard": "93", + "stopping_reason_count_completed_with_submit": "1", + "avg_total_tokens_target_model": "2853624.80", + "total_total_tokens_target_model": "810429443.0", + "avg_total_tokens_other_models": "937.96", + "total_total_tokens_other_models": "266380.0", + "avg_total_tokens_all_models": "2854562.76", + "total_total_tokens_all_models": "810695823.0", + "avg_turn_count": "62.45", + "total_turn_count": "17737.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 30000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/openai/gpt-5-2025-08-07/a10f3ba0-45b2-4ce2-8f38-20925cdeeff0_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "5dca14a6f71d7c0f86a8da53f2aa61c59ee6f9743cfb52086de07f80a2845d31", + "total_rows": 284 + } +} \ No newline at end of file diff --git a/data/swebenchpro/openai/gpt-5-2025-08-07/a10f3ba0-45b2-4ce2-8f38-20925cdeeff0_samples.jsonl b/data/swebenchpro/openai/gpt-5-2025-08-07/a10f3ba0-45b2-4ce2-8f38-20925cdeeff0_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..d15532cb1b54f65d244272d9f344ebd4555cbe48 --- /dev/null +++ b/data/swebenchpro/openai/gpt-5-2025-08-07/a10f3ba0-45b2-4ce2-8f38-20925cdeeff0_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5dca14a6f71d7c0f86a8da53f2aa61c59ee6f9743cfb52086de07f80a2845d31 +size 126784247 diff --git a/data/swebenchpro/openai/gpt-5-2025-08-07/c3d67ce3-338a-474f-a849-388dd0f5536b.json b/data/swebenchpro/openai/gpt-5-2025-08-07/c3d67ce3-338a-474f-a849-388dd0f5536b.json new file mode 100644 index 0000000000000000000000000000000000000000..8a28fff76bbefb896aa53f71c14e55c782ea636e --- /dev/null +++ b/data/swebenchpro/openai/gpt-5-2025-08-07/c3d67ce3-338a-474f-a849-388dd0f5536b.json @@ -0,0 +1,133 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/openai_gpt-5-2025-08-07/1786165478.745534", + "evaluation_timestamp": "1777924672.0", + "retrieved_timestamp": "1786165478.745534", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5-2025-08-07", + "id": "openai/gpt-5-2025-08-07", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on swebenchpro/S-adaptive/1d28034f for scorer harbor_scorer", + "source_data": { + "dataset_name": "swebenchpro", + "source_type": "hf_dataset", + "samples_number": 20, + "sample_ids": [ + "instance_gravitational__teleport-2bb3bbbd8aff1164a2353381cb79e1dc93b90d28-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_gravitational__teleport-46aa81b1ce96ebb4ebed2ae53fd78cd44a05da6c-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_tutao__tutanota-8513a9e8114a8b42e64f4348335e0f23efa054c4-vee878bb72091875e912c52fc32bc60ec3760227b", + "instance_internetarchive__openlibrary-630221ab686c64e75a2ce253c893c033e4814b2e-v93c53c13d5f9b383ebb411ee7750b49dcd1a34c6", + "instance_internetarchive__openlibrary-7f6b722a10f822171501d027cad60afe53337732-ve8c8d62a2b60610a3c4631f5f23ed866bada9818", + "instance_nodebb__nodebb-a917210c5b2c20637094545401f85783905c074c-vf2cf3cbd463b7ad942381f1c6d077626485a1e9e", + "instance_internetarchive__openlibrary-1351c59fd43689753de1fca32c78d539a116ffc1-v29f82c9cf21d57b242f8d8b0e541525d259e2d63", + "instance_qutebrowser__qutebrowser-a25e8a09873838ca9efefd36ea8a45170bbeb95c-vc2f56a753b62a190ddb23cd330c257b9cf560d12", + "instance_protonmail__webclients-281a6b3f190f323ec2c0630999354fafb84b2880", + "instance_future-architect__vuls-999529a05b202b0fd29c6fca5039a4c47a3766bb", + "instance_element-hq__element-web-1077729a19c0ce902e713cf6fab42c91fb7907f1-vnan", + "instance_future-architect__vuls-be7b9114cc9545e68fb0ee7bc63d7ec53d1a00ad", + "instance_qutebrowser__qutebrowser-c580ebf0801e5a3ecabc54f327498bb753c6d5f2-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_gravitational__teleport-3ff19cf7c41f396ae468797d3aeb61515517edc9-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_future-architect__vuls-aaea15e516ece43978cf98e09e52080478b1d39f", + "instance_protonmail__webclients-f161c10cf7d31abf82e8d64d7a99c9fac5acfa18", + "instance_qutebrowser__qutebrowser-cf06f4e3708f886032d4d2a30108c2fddb042d81-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_internetarchive__openlibrary-3f580a5f244c299d936d73d9e327ba873b6401d9-v0f5aece3601a5b4419f7ccec1dbda2071be28ee4", + "instance_qutebrowser__qutebrowser-0aa57e4f7243024fa4bba8853306691b5dbd77b3-v5149fcda2a9a6fe1d35dfed1bade1444a11ef271", + "instance_future-architect__vuls-50580f6e98eeb36f53f27222f7f4fdfea0b21e8d" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1777924672.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.55, + "uncertainty": { + "standard_error": { + "value": 0.10039396081648065 + }, + "num_samples": 100 + }, + "details": { + "total_matched_trajectories": "100", + "stopping_reason_count_repetition_guard": "99", + "stopping_reason_count_token_limit": "1", + "avg_total_tokens_target_model": "3032772.34", + "total_total_tokens_target_model": "303277234.0", + "avg_total_tokens_other_models": "1778.21", + "total_total_tokens_other_models": "177821.0", + "avg_total_tokens_all_models": "3034550.55", + "total_total_tokens_all_models": "303455055.0", + "avg_turn_count": "66.20", + "total_turn_count": "6620.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 30000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/openai/gpt-5-2025-08-07/c3d67ce3-338a-474f-a849-388dd0f5536b_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "258d0e1e00715d5bf5288f1ad7c2bd5309fbf2d68c65d5cbbea9627220f7bef0", + "total_rows": 100 + } +} \ No newline at end of file diff --git a/data/swebenchpro/openai/gpt-5-2025-08-07/c3d67ce3-338a-474f-a849-388dd0f5536b_samples.jsonl b/data/swebenchpro/openai/gpt-5-2025-08-07/c3d67ce3-338a-474f-a849-388dd0f5536b_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..061964f38d8f4c482d93b92659338a6925669f1b --- /dev/null +++ b/data/swebenchpro/openai/gpt-5-2025-08-07/c3d67ce3-338a-474f-a849-388dd0f5536b_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:258d0e1e00715d5bf5288f1ad7c2bd5309fbf2d68c65d5cbbea9627220f7bef0 +size 46110313 diff --git a/data/swebenchpro/openai/gpt-5-2025-08-07/caf650f6-9fc2-473d-9a9c-5191e8856e7f.json b/data/swebenchpro/openai/gpt-5-2025-08-07/caf650f6-9fc2-473d-9a9c-5191e8856e7f.json new file mode 100644 index 0000000000000000000000000000000000000000..06ad32c7a226da7e6da5ef579d499bde0589f1c3 --- /dev/null +++ b/data/swebenchpro/openai/gpt-5-2025-08-07/caf650f6-9fc2-473d-9a9c-5191e8856e7f.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/openai_gpt-5-2025-08-07/1786169138.334466", + "evaluation_timestamp": "1776625526.0", + "retrieved_timestamp": "1786169138.334466", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5-2025-08-07", + "id": "openai/gpt-5-2025-08-07", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/openai/gpt-5-2025-08-07/caf650f6-9fc2-473d-9a9c-5191e8856e7f_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "e3a26e6d991e002709d80328c6c1ebf13a1e56d19dfb8c2973f175ec5badf841", + "total_rows": 9 + } +} \ No newline at end of file diff --git a/data/swebenchpro/openai/gpt-5-2025-08-07/caf650f6-9fc2-473d-9a9c-5191e8856e7f_samples.jsonl b/data/swebenchpro/openai/gpt-5-2025-08-07/caf650f6-9fc2-473d-9a9c-5191e8856e7f_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..62f33516cdf3c78bc448aac277e06da557760942 --- /dev/null +++ b/data/swebenchpro/openai/gpt-5-2025-08-07/caf650f6-9fc2-473d-9a9c-5191e8856e7f_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e3a26e6d991e002709d80328c6c1ebf13a1e56d19dfb8c2973f175ec5badf841 +size 4263121 diff --git a/data/swebenchpro/openai/gpt-5-2025-08-07/d275ff7e-0028-44b4-99ba-7047e489e20a.json b/data/swebenchpro/openai/gpt-5-2025-08-07/d275ff7e-0028-44b4-99ba-7047e489e20a.json new file mode 100644 index 0000000000000000000000000000000000000000..34f5160e801a8f4fc6d1504a1c720664d919e74d --- /dev/null +++ b/data/swebenchpro/openai/gpt-5-2025-08-07/d275ff7e-0028-44b4-99ba-7047e489e20a.json @@ -0,0 +1,111 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/openai_gpt-5-2025-08-07/1786165447.142733", + "evaluation_timestamp": "1777042477.0", + "retrieved_timestamp": "1786165447.142733", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5-2025-08-07", + "id": "openai/gpt-5-2025-08-07", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on swebenchpro/S-adaptive/+2ep/b226c519 for scorer harbor_scorer", + "source_data": { + "dataset_name": "swebenchpro", + "source_type": "hf_dataset", + "samples_number": 2, + "sample_ids": [ + "instance_internetarchive__openlibrary-fdbc0d8f418333c7e575c40b661b582c301ef7ac-v13642507b4fc1f8d234172bf8129942da2c2ca26", + "instance_ansible__ansible-502270c804c33d3bc963930dc85e0f4ca359674d-v7eee2454f617569fd6889f2211f75bc02a35f9f8" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1777042477.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.75, + "uncertainty": { + "standard_error": { + "value": 0.25 + }, + "num_samples": 4 + }, + "details": { + "total_matched_trajectories": "4", + "stopping_reason_count_repetition_guard": "4", + "avg_total_tokens_target_model": "1172225.75", + "total_total_tokens_target_model": "4688903.0", + "avg_total_tokens_other_models": "1305.75", + "total_total_tokens_other_models": "5223.0", + "avg_total_tokens_all_models": "1173531.50", + "total_total_tokens_all_models": "4694126.0", + "avg_turn_count": "38.25", + "total_turn_count": "153.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 30000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/openai/gpt-5-2025-08-07/d275ff7e-0028-44b4-99ba-7047e489e20a_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "6fc622fccf06515cedda0a519140a6f19cb387ce58af2d8228fe8d3f1831df0f", + "total_rows": 4 + } +} \ No newline at end of file diff --git a/data/swebenchpro/openai/gpt-5-2025-08-07/d275ff7e-0028-44b4-99ba-7047e489e20a_samples.jsonl b/data/swebenchpro/openai/gpt-5-2025-08-07/d275ff7e-0028-44b4-99ba-7047e489e20a_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..ca047db4c3ac0ad1f1b39324af4aa54c0c8c2a19 --- /dev/null +++ b/data/swebenchpro/openai/gpt-5-2025-08-07/d275ff7e-0028-44b4-99ba-7047e489e20a_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6fc622fccf06515cedda0a519140a6f19cb387ce58af2d8228fe8d3f1831df0f +size 1276248 diff --git a/data/swebenchpro/openai/gpt-5-2025-08-07/d5a0b05f-1380-4ae6-835a-f4313cb2c190.json b/data/swebenchpro/openai/gpt-5-2025-08-07/d5a0b05f-1380-4ae6-835a-f4313cb2c190.json new file mode 100644 index 0000000000000000000000000000000000000000..5c8cbf8578bbb099970f8ec753a50514dfe14bde --- /dev/null +++ b/data/swebenchpro/openai/gpt-5-2025-08-07/d5a0b05f-1380-4ae6-835a-f4313cb2c190.json @@ -0,0 +1,110 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/openai_gpt-5-2025-08-07/1786162990.688909", + "evaluation_timestamp": "1778557458.0", + "retrieved_timestamp": "1786162990.688909", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5-2025-08-07", + "id": "openai/gpt-5-2025-08-07", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on swebenchpro/S-adaptive+C/+3ep/e778fe33 for scorer harbor_scorer", + "source_data": { + "dataset_name": "swebenchpro", + "source_type": "hf_dataset", + "samples_number": 1, + "sample_ids": [ + "instance_element-hq__element-web-2760bfc8369f1bee640d6d7a7e910783143d4c5f-vnan" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1778557458.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 1.0, + "uncertainty": { + "num_samples": 3 + }, + "details": { + "total_matched_trajectories": "3", + "stopping_reason_count_completed_on_successful_submit": "3", + "avg_total_tokens_target_model": "4704863.33", + "total_total_tokens_target_model": "14114590.0", + "avg_total_tokens_other_models": "0.00", + "total_total_tokens_other_models": "0.0", + "avg_total_tokens_all_models": "4704863.33", + "total_total_tokens_all_models": "14114590.0", + "avg_turn_count": "82.00", + "total_turn_count": "246.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 30000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/openai/gpt-5-2025-08-07/d5a0b05f-1380-4ae6-835a-f4313cb2c190_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "977bd7eb7e75c10019ad5036e40efc84e4604661ad34f20c91b82661774e3802", + "total_rows": 3 + } +} \ No newline at end of file diff --git a/data/swebenchpro/openai/gpt-5-2025-08-07/d5a0b05f-1380-4ae6-835a-f4313cb2c190_samples.jsonl b/data/swebenchpro/openai/gpt-5-2025-08-07/d5a0b05f-1380-4ae6-835a-f4313cb2c190_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..de9147275704bb2e478a0cb81f574ed46538c14c --- /dev/null +++ b/data/swebenchpro/openai/gpt-5-2025-08-07/d5a0b05f-1380-4ae6-835a-f4313cb2c190_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:977bd7eb7e75c10019ad5036e40efc84e4604661ad34f20c91b82661774e3802 +size 1948325 diff --git a/data/swebenchpro/openai/gpt-5-2025-08-07/d8b3320b-76f8-40bc-8095-cb3bc821006a.json b/data/swebenchpro/openai/gpt-5-2025-08-07/d8b3320b-76f8-40bc-8095-cb3bc821006a.json new file mode 100644 index 0000000000000000000000000000000000000000..eb324f18f64ae82df58cd31cdee6d49baa6bfa4c --- /dev/null +++ b/data/swebenchpro/openai/gpt-5-2025-08-07/d8b3320b-76f8-40bc-8095-cb3bc821006a.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/openai_gpt-5-2025-08-07/1786169099.884851", + "evaluation_timestamp": "1776601454.0", + "retrieved_timestamp": "1786169099.884851", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5-2025-08-07", + "id": "openai/gpt-5-2025-08-07", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/openai/gpt-5-2025-08-07/d8b3320b-76f8-40bc-8095-cb3bc821006a_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "5eb030e7ef24d09f14cbf070318aaf391d47b1f2723c161c615d30dab7703c5b", + "total_rows": 10 + } +} \ No newline at end of file diff --git a/data/swebenchpro/openai/gpt-5-2025-08-07/d8b3320b-76f8-40bc-8095-cb3bc821006a_samples.jsonl b/data/swebenchpro/openai/gpt-5-2025-08-07/d8b3320b-76f8-40bc-8095-cb3bc821006a_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..55ac58c4c751428308ea5b679425e5a71b3849bc --- /dev/null +++ b/data/swebenchpro/openai/gpt-5-2025-08-07/d8b3320b-76f8-40bc-8095-cb3bc821006a_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5eb030e7ef24d09f14cbf070318aaf391d47b1f2723c161c615d30dab7703c5b +size 6859091 diff --git a/data/swebenchpro/openai/gpt-5-2025-08-07/dbc6d392-55e8-4437-ad54-5a8d12b7f3d4.json b/data/swebenchpro/openai/gpt-5-2025-08-07/dbc6d392-55e8-4437-ad54-5a8d12b7f3d4.json new file mode 100644 index 0000000000000000000000000000000000000000..6c3ea9cb9e1eb311e932fb379addbb9520a048fa --- /dev/null +++ b/data/swebenchpro/openai/gpt-5-2025-08-07/dbc6d392-55e8-4437-ad54-5a8d12b7f3d4.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/openai_gpt-5-2025-08-07/1786169112.857201", + "evaluation_timestamp": "1777478369.0", + "retrieved_timestamp": "1786169112.857201", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5-2025-08-07", + "id": "openai/gpt-5-2025-08-07", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/openai/gpt-5-2025-08-07/dbc6d392-55e8-4437-ad54-5a8d12b7f3d4_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "811dce63d581907e04ee3d94f50894674333a39f40f3aa62b49b517fcf72800e", + "total_rows": 20 + } +} \ No newline at end of file diff --git a/data/swebenchpro/openai/gpt-5-2025-08-07/dbc6d392-55e8-4437-ad54-5a8d12b7f3d4_samples.jsonl b/data/swebenchpro/openai/gpt-5-2025-08-07/dbc6d392-55e8-4437-ad54-5a8d12b7f3d4_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..c68ad74349725ddacd1826806e7fb33cf022b96f --- /dev/null +++ b/data/swebenchpro/openai/gpt-5-2025-08-07/dbc6d392-55e8-4437-ad54-5a8d12b7f3d4_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:811dce63d581907e04ee3d94f50894674333a39f40f3aa62b49b517fcf72800e +size 10034126 diff --git a/data/swebenchpro/openai/gpt-5-2025-08-07/e0aa1911-0bdf-4da5-b34f-6fd7ae708a4b.json b/data/swebenchpro/openai/gpt-5-2025-08-07/e0aa1911-0bdf-4da5-b34f-6fd7ae708a4b.json new file mode 100644 index 0000000000000000000000000000000000000000..2818111eca0e8edadcb3e95a8ce7fb88823e1b4b --- /dev/null +++ b/data/swebenchpro/openai/gpt-5-2025-08-07/e0aa1911-0bdf-4da5-b34f-6fd7ae708a4b.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/openai_gpt-5-2025-08-07/1786169179.716416", + "evaluation_timestamp": "1777337600.0", + "retrieved_timestamp": "1786169179.716416", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5-2025-08-07", + "id": "openai/gpt-5-2025-08-07", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/openai/gpt-5-2025-08-07/e0aa1911-0bdf-4da5-b34f-6fd7ae708a4b_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "47dca7ff5838cb3cdb6d44c7c8dbe16c76e47f2b6a032d2380666570f84c818b", + "total_rows": 14 + } +} \ No newline at end of file diff --git a/data/swebenchpro/openai/gpt-5-2025-08-07/e0aa1911-0bdf-4da5-b34f-6fd7ae708a4b_samples.jsonl b/data/swebenchpro/openai/gpt-5-2025-08-07/e0aa1911-0bdf-4da5-b34f-6fd7ae708a4b_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..72f7a0af235768f523e005fbf7f5ab74ec1a042d --- /dev/null +++ b/data/swebenchpro/openai/gpt-5-2025-08-07/e0aa1911-0bdf-4da5-b34f-6fd7ae708a4b_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:47dca7ff5838cb3cdb6d44c7c8dbe16c76e47f2b6a032d2380666570f84c818b +size 12838867 diff --git a/data/swebenchpro/openai/gpt-5-2025-08-07/e3d1794b-100b-4426-9ad9-d6aef5798754.json b/data/swebenchpro/openai/gpt-5-2025-08-07/e3d1794b-100b-4426-9ad9-d6aef5798754.json new file mode 100644 index 0000000000000000000000000000000000000000..8a6f159eb95aad1aa15468868da0dc28d5d3b171 --- /dev/null +++ b/data/swebenchpro/openai/gpt-5-2025-08-07/e3d1794b-100b-4426-9ad9-d6aef5798754.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/openai_gpt-5-2025-08-07/1786169152.56011", + "evaluation_timestamp": "1777317713.0", + "retrieved_timestamp": "1786169152.56011", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5-2025-08-07", + "id": "openai/gpt-5-2025-08-07", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/openai/gpt-5-2025-08-07/e3d1794b-100b-4426-9ad9-d6aef5798754_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "bbddd2ecc7985efbc4c70ba6fe77954f5c0ad82f1a10a819bfb8444b23cbd2a0", + "total_rows": 35 + } +} \ No newline at end of file diff --git a/data/swebenchpro/openai/gpt-5-2025-08-07/e3d1794b-100b-4426-9ad9-d6aef5798754_samples.jsonl b/data/swebenchpro/openai/gpt-5-2025-08-07/e3d1794b-100b-4426-9ad9-d6aef5798754_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..78d3dd29d40d26e2e9b3cadd425459ff30ae8a7b --- /dev/null +++ b/data/swebenchpro/openai/gpt-5-2025-08-07/e3d1794b-100b-4426-9ad9-d6aef5798754_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bbddd2ecc7985efbc4c70ba6fe77954f5c0ad82f1a10a819bfb8444b23cbd2a0 +size 18147342 diff --git a/data/swebenchpro/openai/gpt-5-2025-08-07/e3f1fbbc-2dee-4f91-9196-87dec243beab.json b/data/swebenchpro/openai/gpt-5-2025-08-07/e3f1fbbc-2dee-4f91-9196-87dec243beab.json new file mode 100644 index 0000000000000000000000000000000000000000..a21bccfef2dff91a1b5d26ac2700ddb1c6430b12 --- /dev/null +++ b/data/swebenchpro/openai/gpt-5-2025-08-07/e3f1fbbc-2dee-4f91-9196-87dec243beab.json @@ -0,0 +1,116 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/openai_gpt-5-2025-08-07/1786162985.375415", + "evaluation_timestamp": "1778272107.0", + "retrieved_timestamp": "1786162985.375415", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5-2025-08-07", + "id": "openai/gpt-5-2025-08-07", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on swebenchpro/S-adaptive+C/+2ep/5d93d0a6 for scorer harbor_scorer", + "source_data": { + "dataset_name": "swebenchpro", + "source_type": "hf_dataset", + "samples_number": 3, + "sample_ids": [ + "instance_tutao__tutanota-d1aa0ecec288bfc800cfb9133b087c4f81ad8b38-vbc0d9ba8f0071fbe982809910959a6ff8884dbbf", + "instance_gravitational__teleport-ac2fb2f9b4fd1896b554d3011df23d3d71295779", + "instance_flipt-io__flipt-cd2f3b0a9d4d8b8a6d3d56afab65851ecdc408e8" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1778272107.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.3333333333333333, + "uncertainty": { + "standard_error": { + "value": 0.33333333333333337 + }, + "num_samples": 6 + }, + "details": { + "total_matched_trajectories": "6", + "stopping_reason_count_repetition_guard": "4", + "stopping_reason_count_completed_on_successful_submit": "2", + "avg_total_tokens_target_model": "4957639.50", + "total_total_tokens_target_model": "29745837.0", + "avg_total_tokens_other_models": "1712.00", + "total_total_tokens_other_models": "10272.0", + "avg_total_tokens_all_models": "4959351.50", + "total_total_tokens_all_models": "29756109.0", + "avg_turn_count": "93.83", + "total_turn_count": "563.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 30000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/openai/gpt-5-2025-08-07/e3f1fbbc-2dee-4f91-9196-87dec243beab_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "0708ec260e4c19bb3e44e9c876f184a7881ec776bcb1838bf44d1a660ace0158", + "total_rows": 6 + } +} \ No newline at end of file diff --git a/data/swebenchpro/openai/gpt-5-2025-08-07/e3f1fbbc-2dee-4f91-9196-87dec243beab_samples.jsonl b/data/swebenchpro/openai/gpt-5-2025-08-07/e3f1fbbc-2dee-4f91-9196-87dec243beab_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..293198ddce8623a7fc744ff9d0811b4019f334fd --- /dev/null +++ b/data/swebenchpro/openai/gpt-5-2025-08-07/e3f1fbbc-2dee-4f91-9196-87dec243beab_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0708ec260e4c19bb3e44e9c876f184a7881ec776bcb1838bf44d1a660ace0158 +size 3716239 diff --git a/data/swebenchpro/openai/gpt-5-2025-08-07/e7523c02-f02d-415b-a9d6-e424947785a3.json b/data/swebenchpro/openai/gpt-5-2025-08-07/e7523c02-f02d-415b-a9d6-e424947785a3.json new file mode 100644 index 0000000000000000000000000000000000000000..a538b306a1e3b7f5da57f519acff4fc285ac9bd5 --- /dev/null +++ b/data/swebenchpro/openai/gpt-5-2025-08-07/e7523c02-f02d-415b-a9d6-e424947785a3.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/openai_gpt-5-2025-08-07/1786167383.986484", + "evaluation_timestamp": "1778005529.0", + "retrieved_timestamp": "1786167383.986484", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5-2025-08-07", + "id": "openai/gpt-5-2025-08-07", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/openai/gpt-5-2025-08-07/e7523c02-f02d-415b-a9d6-e424947785a3_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "004931eb665f6e079be23514a6c394d8d6e9246b04d70686c0d25e12a42db209", + "total_rows": 24 + } +} \ No newline at end of file diff --git a/data/swebenchpro/openai/gpt-5-2025-08-07/e7523c02-f02d-415b-a9d6-e424947785a3_samples.jsonl b/data/swebenchpro/openai/gpt-5-2025-08-07/e7523c02-f02d-415b-a9d6-e424947785a3_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..8ab9c874eae4364b54f446c938c01d49dfe4633a --- /dev/null +++ b/data/swebenchpro/openai/gpt-5-2025-08-07/e7523c02-f02d-415b-a9d6-e424947785a3_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:004931eb665f6e079be23514a6c394d8d6e9246b04d70686c0d25e12a42db209 +size 30404427 diff --git a/data/swebenchpro/openai/gpt-5-2025-08-07/eb9400de-8efa-4bc7-b85d-7a3600288465.json b/data/swebenchpro/openai/gpt-5-2025-08-07/eb9400de-8efa-4bc7-b85d-7a3600288465.json new file mode 100644 index 0000000000000000000000000000000000000000..fb5d03dcfa19a7682690a7a08b264463adde6977 --- /dev/null +++ b/data/swebenchpro/openai/gpt-5-2025-08-07/eb9400de-8efa-4bc7-b85d-7a3600288465.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/openai_gpt-5-2025-08-07/1786167353.527707", + "evaluation_timestamp": "1777981659.0", + "retrieved_timestamp": "1786167353.527707", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5-2025-08-07", + "id": "openai/gpt-5-2025-08-07", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/openai/gpt-5-2025-08-07/eb9400de-8efa-4bc7-b85d-7a3600288465_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "e56e36206efaa15f9e6662ab613a2f2b21fcc80731b56922817b331c02687f8b", + "total_rows": 8 + } +} \ No newline at end of file diff --git a/data/swebenchpro/openai/gpt-5-2025-08-07/eb9400de-8efa-4bc7-b85d-7a3600288465_samples.jsonl b/data/swebenchpro/openai/gpt-5-2025-08-07/eb9400de-8efa-4bc7-b85d-7a3600288465_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..480786cc5f386e4ce76f0cd90035b9cb068d8a6f --- /dev/null +++ b/data/swebenchpro/openai/gpt-5-2025-08-07/eb9400de-8efa-4bc7-b85d-7a3600288465_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e56e36206efaa15f9e6662ab613a2f2b21fcc80731b56922817b331c02687f8b +size 14715179 diff --git a/data/swebenchpro/openai/gpt-5-2025-08-07/f33875f4-a454-4bf8-977a-2f54363bbc4b.json b/data/swebenchpro/openai/gpt-5-2025-08-07/f33875f4-a454-4bf8-977a-2f54363bbc4b.json new file mode 100644 index 0000000000000000000000000000000000000000..c0ad9fc4ab8a515ce69545167dcaf8d102ed4753 --- /dev/null +++ b/data/swebenchpro/openai/gpt-5-2025-08-07/f33875f4-a454-4bf8-977a-2f54363bbc4b.json @@ -0,0 +1,110 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/openai_gpt-5-2025-08-07/1786162988.13798", + "evaluation_timestamp": "1778196561.0", + "retrieved_timestamp": "1786162988.13798", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5-2025-08-07", + "id": "openai/gpt-5-2025-08-07", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on swebenchpro/S-adaptive+C/+3ep/52f921f6 for scorer harbor_scorer", + "source_data": { + "dataset_name": "swebenchpro", + "source_type": "hf_dataset", + "samples_number": 1, + "sample_ids": [ + "instance_future-architect__vuls-78b52d6a7f480bd610b692de9bf0c86f57332f23" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1778196561.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.0, + "uncertainty": { + "num_samples": 3 + }, + "details": { + "total_matched_trajectories": "3", + "stopping_reason_count_repetition_guard": "3", + "avg_total_tokens_target_model": "9383695.67", + "total_total_tokens_target_model": "28151087.0", + "avg_total_tokens_other_models": "4214.00", + "total_total_tokens_other_models": "12642.0", + "avg_total_tokens_all_models": "9387909.67", + "total_total_tokens_all_models": "28163729.0", + "avg_turn_count": "143.33", + "total_turn_count": "430.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 30000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/openai/gpt-5-2025-08-07/f33875f4-a454-4bf8-977a-2f54363bbc4b_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "3aff05325d5d4634d8b12f90a5056831f03f6f7b386bf17ca0bc31e17a59a1ed", + "total_rows": 3 + } +} \ No newline at end of file diff --git a/data/swebenchpro/openai/gpt-5-2025-08-07/f33875f4-a454-4bf8-977a-2f54363bbc4b_samples.jsonl b/data/swebenchpro/openai/gpt-5-2025-08-07/f33875f4-a454-4bf8-977a-2f54363bbc4b_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..5e6b1e232967684f6e3e9b6c45a757a8e0748474 --- /dev/null +++ b/data/swebenchpro/openai/gpt-5-2025-08-07/f33875f4-a454-4bf8-977a-2f54363bbc4b_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3aff05325d5d4634d8b12f90a5056831f03f6f7b386bf17ca0bc31e17a59a1ed +size 2742039 diff --git a/data/swebenchpro/openai/gpt-5-2025-08-07/f71d6f8b-7d22-409c-882f-f498268b5d49.json b/data/swebenchpro/openai/gpt-5-2025-08-07/f71d6f8b-7d22-409c-882f-f498268b5d49.json new file mode 100644 index 0000000000000000000000000000000000000000..ca5167c17030208d15130f4ccafec62d980953c3 --- /dev/null +++ b/data/swebenchpro/openai/gpt-5-2025-08-07/f71d6f8b-7d22-409c-882f-f498268b5d49.json @@ -0,0 +1,116 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/openai_gpt-5-2025-08-07/1786162965.857748", + "evaluation_timestamp": "1778192086.0", + "retrieved_timestamp": "1786162965.857748", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5-2025-08-07", + "id": "openai/gpt-5-2025-08-07", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on swebenchpro/S-adaptive+C/+1ep/75a8a5be for scorer harbor_scorer", + "source_data": { + "dataset_name": "swebenchpro", + "source_type": "hf_dataset", + "samples_number": 3, + "sample_ids": [ + "instance_flipt-io__flipt-e42da21a07a5ae35835ec54f74004ebd58713874", + "instance_nodebb__nodebb-84e065752f6d7fbe5c08cbf50cb173ffb866b8fa-vf2cf3cbd463b7ad942381f1c6d077626485a1e9e", + "instance_element-hq__element-web-f3534b42df3dcfe36dc48bddbf14034085af6d30-vnan" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1778192086.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.6666666666666666, + "uncertainty": { + "standard_error": { + "value": 0.33333333333333337 + }, + "num_samples": 3 + }, + "details": { + "total_matched_trajectories": "3", + "stopping_reason_count_completed_on_successful_submit": "2", + "stopping_reason_count_repetition_guard": "1", + "avg_total_tokens_target_model": "5927368.67", + "total_total_tokens_target_model": "17782106.0", + "avg_total_tokens_other_models": "1138.33", + "total_total_tokens_other_models": "3415.0", + "avg_total_tokens_all_models": "5928507.00", + "total_total_tokens_all_models": "17785521.0", + "avg_turn_count": "103.67", + "total_turn_count": "311.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 30000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/openai/gpt-5-2025-08-07/f71d6f8b-7d22-409c-882f-f498268b5d49_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "d6933702d1b04d18eb13a011dd321e442125b5865305f266236b46089c8f57ba", + "total_rows": 3 + } +} \ No newline at end of file diff --git a/data/swebenchpro/openai/gpt-5-2025-08-07/f71d6f8b-7d22-409c-882f-f498268b5d49_samples.jsonl b/data/swebenchpro/openai/gpt-5-2025-08-07/f71d6f8b-7d22-409c-882f-f498268b5d49_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..6bba4df2bd0132b445ada82b6ca077675f97bb7f --- /dev/null +++ b/data/swebenchpro/openai/gpt-5-2025-08-07/f71d6f8b-7d22-409c-882f-f498268b5d49_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d6933702d1b04d18eb13a011dd321e442125b5865305f266236b46089c8f57ba +size 1873626 diff --git a/data/swebenchpro/openai/gpt-5-2025-08-07/fd7c6402-96c7-4692-a952-38941761e599.json b/data/swebenchpro/openai/gpt-5-2025-08-07/fd7c6402-96c7-4692-a952-38941761e599.json new file mode 100644 index 0000000000000000000000000000000000000000..0676e4ee89b8c95c3b571ac87c02c0f3b6154b0f --- /dev/null +++ b/data/swebenchpro/openai/gpt-5-2025-08-07/fd7c6402-96c7-4692-a952-38941761e599.json @@ -0,0 +1,124 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/openai_gpt-5-2025-08-07/1786162981.79786", + "evaluation_timestamp": "1778547641.0", + "retrieved_timestamp": "1786162981.79786", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5-2025-08-07", + "id": "openai/gpt-5-2025-08-07", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on swebenchpro/S-adaptive+C/+2ep/1aa97ae6 for scorer harbor_scorer", + "source_data": { + "dataset_name": "swebenchpro", + "source_type": "hf_dataset", + "samples_number": 11, + "sample_ids": [ + "instance_flipt-io__flipt-e2bd19dafa7166c96b082fb2a59eb54b4be0d778", + "instance_element-hq__element-web-75c2c1a572fa45d1ea1d1a96e9e36e303332ecaa-vnan", + "instance_internetarchive__openlibrary-09865f5fb549694d969f0a8e49b9d204ef1853ca-ve8c8d62a2b60610a3c4631f5f23ed866bada9818", + "instance_internetarchive__openlibrary-25858f9f0c165df25742acf8309ce909773f0cdd-v13642507b4fc1f8d234172bf8129942da2c2ca26", + "instance_protonmail__webclients-01ea5214d11e0df8b7170d91bafd34f23cb0f2b1", + "instance_tutao__tutanota-4b4e45949096bb288f2b522f657610e480efa3e8-vee878bb72091875e912c52fc32bc60ec3760227b", + "instance_flipt-io__flipt-05d7234fa582df632f70a7cd10194d61bd7043b9", + "instance_ansible__ansible-a02e22e902a69aeb465f16bf03f7f5a91b2cb828-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_tutao__tutanota-1ff82aa365763cee2d609c9d19360ad87fdf2ec7-vc4e41fd0029957297843cb9dec4a25c7c756f029", + "instance_internetarchive__openlibrary-30bc73a1395fba2300087c7f307e54bb5372b60a-v76304ecdb3a5954fcf13feb710e8c40fcf24b73c", + "instance_nodebb__nodebb-8ca65b0c78c67c1653487c02d1135e1b702185e1-vf2cf3cbd463b7ad942381f1c6d077626485a1e9e" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1778547641.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.7272727272727273, + "uncertainty": { + "standard_error": { + "value": 0.14083575804390605 + }, + "num_samples": 22 + }, + "details": { + "total_matched_trajectories": "22", + "stopping_reason_count_completed_on_successful_submit": "16", + "stopping_reason_count_repetition_guard": "6", + "avg_total_tokens_target_model": "3278050.95", + "total_total_tokens_target_model": "72117121.0", + "avg_total_tokens_other_models": "946.86", + "total_total_tokens_other_models": "20831.0", + "avg_total_tokens_all_models": "3278997.82", + "total_total_tokens_all_models": "72137952.0", + "avg_turn_count": "68.50", + "total_turn_count": "1507.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 30000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/openai/gpt-5-2025-08-07/fd7c6402-96c7-4692-a952-38941761e599_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "681391b0fc8a9cc52144920ddff926010e0184034a24c9f72d768a089eed2068", + "total_rows": 22 + } +} \ No newline at end of file diff --git a/data/swebenchpro/openai/gpt-5-2025-08-07/fd7c6402-96c7-4692-a952-38941761e599_samples.jsonl b/data/swebenchpro/openai/gpt-5-2025-08-07/fd7c6402-96c7-4692-a952-38941761e599_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..aa1b84495b86a6186a603b045c048e9f30aa6ecc --- /dev/null +++ b/data/swebenchpro/openai/gpt-5-2025-08-07/fd7c6402-96c7-4692-a952-38941761e599_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:681391b0fc8a9cc52144920ddff926010e0184034a24c9f72d768a089eed2068 +size 10420498 diff --git a/data/swebenchpro/openai/gpt-5.2-2025-12-11/12e426bd-2a11-43a5-bd74-26827ba2f590.json b/data/swebenchpro/openai/gpt-5.2-2025-12-11/12e426bd-2a11-43a5-bd74-26827ba2f590.json new file mode 100644 index 0000000000000000000000000000000000000000..5dca6b8425f526be7703fecd274d24d2482b5cf7 --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.2-2025-12-11/12e426bd-2a11-43a5-bd74-26827ba2f590.json @@ -0,0 +1,151 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/openai_gpt-5.2-2025-12-11/1786162901.449261", + "evaluation_timestamp": "1778703817.0", + "retrieved_timestamp": "1786162901.449261", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on swebenchpro/S-adaptive+C/+4ep/e0e0be23 for scorer harbor_scorer", + "source_data": { + "dataset_name": "swebenchpro", + "source_type": "hf_dataset", + "samples_number": 38, + "sample_ids": [ + "instance_flipt-io__flipt-e42da21a07a5ae35835ec54f74004ebd58713874", + "instance_nodebb__nodebb-84e065752f6d7fbe5c08cbf50cb173ffb866b8fa-vf2cf3cbd463b7ad942381f1c6d077626485a1e9e", + "instance_element-hq__element-web-f3534b42df3dcfe36dc48bddbf14034085af6d30-vnan", + "instance_navidrome__navidrome-d8e794317f788198227e10fb667e10496b3eb99a", + "instance_internetarchive__openlibrary-fdbc0d8f418333c7e575c40b661b582c301ef7ac-v13642507b4fc1f8d234172bf8129942da2c2ca26", + "instance_ansible__ansible-b2a289dcbb702003377221e25f62c8a3608f0e89-v173091e2e36d38c978002990795f66cfc0af30ad", + "instance_qutebrowser__qutebrowser-e70f5b03187bdd40e8bf70f5f3ead840f52d1f42-v02ad04386d5238fe2d1a1be450df257370de4b6a", + "instance_tutao__tutanota-d1aa0ecec288bfc800cfb9133b087c4f81ad8b38-vbc0d9ba8f0071fbe982809910959a6ff8884dbbf", + "instance_ansible__ansible-11c1777d56664b1acb56b387a1ad6aeadef1391d-v0f01c69f1e2528b935359cfe578530722bca2c59", + "instance_ansible__ansible-29aea9ff3466e4cd2ed00524b9e56738d568ce8b-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_nodebb__nodebb-70b4a0e2aebebe8f2f559de6680093d96a697b2f-vnan", + "instance_ansible__ansible-502270c804c33d3bc963930dc85e0f4ca359674d-v7eee2454f617569fd6889f2211f75bc02a35f9f8", + "instance_internetarchive__openlibrary-9c392b60e2c6fa1d68cb68084b4b4ff04d0cb35c-v2d9a6c849c60ed19fd0858ce9e40b7cc8e097e59", + "instance_element-hq__element-web-b007ea81b2ccd001b00f332bee65070aa7fc00f9-vnan", + "instance_flipt-io__flipt-cd2f3b0a9d4d8b8a6d3d56afab65851ecdc408e8", + "instance_internetarchive__openlibrary-7cbfb812ef0e1f9716e2d6e85d538a96fcb79d13-vfa6ff903cb27f336e17654595dd900fa943dcd91", + "instance_flipt-io__flipt-0b119520afca1cf25c470ff4288c464d4510b944", + "instance_element-hq__element-web-f63160f38459fb552d00fcc60d4064977a9095a6-vnan", + "instance_gravitational__teleport-73cc189b0e9636d418c4470ecce0d9af5dae2f02-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_internetarchive__openlibrary-3c48b4bb782189e0858e6c3fc7956046cf3e1cfb-v2d9a6c849c60ed19fd0858ce9e40b7cc8e097e59", + "instance_future-architect__vuls-878c25bf5a9c9fd88ac32eb843f5636834d5712d", + "instance_qutebrowser__qutebrowser-70248f256f93ed9b1984494d0a1a919ddd774892-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_future-architect__vuls-73f0adad95c4d227e2ccfa876c85cc95dd065e13", + "instance_qutebrowser__qutebrowser-cc360cd4a34a126274c7b51f3b63afbaf3e05a02-v5fc38aaf22415ab0b70567368332beee7955b367", + "instance_flipt-io__flipt-e2bd19dafa7166c96b082fb2a59eb54b4be0d778", + "instance_qutebrowser__qutebrowser-9ed748effa8f3bcd804612d9291da017b514e12f-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_ansible__ansible-1ee70fc272aff6bf3415357c6e13c5de5b928d9b-v1055803c3a812189a1133297f7f5468579283f86", + "instance_element-hq__element-web-75c2c1a572fa45d1ea1d1a96e9e36e303332ecaa-vnan", + "instance_qutebrowser__qutebrowser-f631cd4422744160d9dcf7a0455da532ce973315-v35616345bb8052ea303186706cec663146f0f184", + "instance_qutebrowser__qutebrowser-e5340c449f23608803c286da0563b62f58ba25b0-v059c6fdc75567943479b23ebca7c07b5e9a7f34c", + "instance_ansible__ansible-42355d181a11b51ebfc56f6f4b3d9c74e01cb13b-v1055803c3a812189a1133297f7f5468579283f86", + "instance_internetarchive__openlibrary-09865f5fb549694d969f0a8e49b9d204ef1853ca-ve8c8d62a2b60610a3c4631f5f23ed866bada9818", + "instance_element-hq__element-web-2760bfc8369f1bee640d6d7a7e910783143d4c5f-vnan", + "instance_protonmail__webclients-01ea5214d11e0df8b7170d91bafd34f23cb0f2b1", + "instance_tutao__tutanota-4b4e45949096bb288f2b522f657610e480efa3e8-vee878bb72091875e912c52fc32bc60ec3760227b", + "instance_protonmail__webclients-5f0745dd6993bb1430a951c62a49807c6635cd77", + "instance_tutao__tutanota-1ff82aa365763cee2d609c9d19360ad87fdf2ec7-vc4e41fd0029957297843cb9dec4a25c7c756f029", + "instance_internetarchive__openlibrary-30bc73a1395fba2300087c7f307e54bb5372b60a-v76304ecdb3a5954fcf13feb710e8c40fcf24b73c" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1778703817.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.7719298245614035, + "uncertainty": { + "standard_error": { + "value": 0.06461734413403099 + }, + "num_samples": 141 + }, + "details": { + "total_matched_trajectories": "141", + "stopping_reason_count_completed_on_successful_submit": "110", + "stopping_reason_count_repetition_guard": "31", + "avg_total_tokens_target_model": "3357797.60", + "total_total_tokens_target_model": "473449462.0", + "avg_total_tokens_other_models": "279.80", + "total_total_tokens_other_models": "39452.0", + "avg_total_tokens_all_models": "3358077.40", + "total_total_tokens_all_models": "473488914.0", + "avg_turn_count": "80.96", + "total_turn_count": "11416.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "8", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 30000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "8", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/openai/gpt-5.2-2025-12-11/12e426bd-2a11-43a5-bd74-26827ba2f590_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "87e4bffba8555eb536da3e66c2d8aa47d6cfb2658a7d0abcf3b2d9b8425c8a5b", + "total_rows": 141 + } +} \ No newline at end of file diff --git a/data/swebenchpro/openai/gpt-5.2-2025-12-11/12e426bd-2a11-43a5-bd74-26827ba2f590_samples.jsonl b/data/swebenchpro/openai/gpt-5.2-2025-12-11/12e426bd-2a11-43a5-bd74-26827ba2f590_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..064a4acb54e42ba036a2cb70a34f7f9ed6532b2f --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.2-2025-12-11/12e426bd-2a11-43a5-bd74-26827ba2f590_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:87e4bffba8555eb536da3e66c2d8aa47d6cfb2658a7d0abcf3b2d9b8425c8a5b +size 75984912 diff --git a/data/swebenchpro/openai/gpt-5.2-2025-12-11/15d21311-3f01-460e-8505-622d3e97b363.json b/data/swebenchpro/openai/gpt-5.2-2025-12-11/15d21311-3f01-460e-8505-622d3e97b363.json new file mode 100644 index 0000000000000000000000000000000000000000..10f5fcc23a50f01c82e08ac37e6cfbd148abbb02 --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.2-2025-12-11/15d21311-3f01-460e-8505-622d3e97b363.json @@ -0,0 +1,304 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/openai_gpt-5.2-2025-12-11/1786181825.423334", + "evaluation_timestamp": "1774867708.0", + "retrieved_timestamp": "1786181825.423334", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.196.dev25+gb973b316,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on eval_task for scorer harbor_scorer", + "source_data": { + "dataset_name": "swebenchpro", + "source_type": "hf_dataset", + "samples_number": 731, + "sample_ids": [ + "instance_flipt-io__flipt-e91615cf07966da41756017a7d571f9fc0fdbe80", + "instance_ansible__ansible-8127abbc298cabf04aaa89a478fc5e5e3432a6fc-v30a923fb5c164d6cd18280c02422f75e611e8fb2", + "instance_protonmail__webclients-da91f084c0f532d9cc8ca385a701274d598057b8", + "instance_qutebrowser__qutebrowser-deeb15d6f009b3ca0c3bd503a7cef07462bd16b4-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_internetarchive__openlibrary-5069b09e5f64428dce59b33455c8bb17fe577070-v8717e18970bcdc4e0d2cea3b1527752b21e74866", + "instance_protonmail__webclients-c5a2089ca2bfe9aa1d85a664b8ad87ef843a1c9c", + "instance_qutebrowser__qutebrowser-fd6790fe8c02b144ab2464f1fc8ab3d02ce3c476-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_navidrome__navidrome-eebfbc5381a1e506ff17b5f1371d1ad83d5fd642", + "instance_navidrome__navidrome-812dc2090f20ac4f8ac271b6ed95be5889d1a3ca", + "instance_nodebb__nodebb-0c81642997ea1d827dbd02c311db9d4976112cd4-vf2cf3cbd463b7ad942381f1c6d077626485a1e9e", + "instance_internetarchive__openlibrary-c506c1b0b678892af5cb22c1c1dbc35d96787a0a-v0f5aece3601a5b4419f7ccec1dbda2071be28ee4", + "instance_flipt-io__flipt-5c7037ececb0bead0a8eb56054e224bcd7ac5922", + "instance_qutebrowser__qutebrowser-01d1d1494411380d97cac14614a829d3a69cecaf-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_internetarchive__openlibrary-7f7e53aa4cf74a4f8549a5bcd4810c527e2f6d7e-v13642507b4fc1f8d234172bf8129942da2c2ca26", + "instance_internetarchive__openlibrary-53e02a22972e9253aeded0e1981e6845e1e521fe-vfa6ff903cb27f336e17654595dd900fa943dcd91", + "instance_ansible__ansible-34db57a47f875d11c4068567b9ec7ace174ec4cf-v1055803c3a812189a1133297f7f5468579283f86", + "instance_qutebrowser__qutebrowser-996487c43e4fcc265b541f9eca1e7930e3c5cf05-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_gravitational__teleport-96019ce0be7a2c8e36363f359eb7c943b41dde70", + "instance_protonmail__webclients-df60460f163fd5c34e844ab9015e3176f1ab1ac0", + "instance_future-architect__vuls-ad2edbb8448e2c41a097f1c0b52696c0f6c5924d", + "instance_ansible__ansible-9142be2f6cabbe6597c9254c5bb9186d17036d55-v0f01c69f1e2528b935359cfe578530722bca2c59", + "instance_element-hq__element-web-33299af5c9b7a7ec5a9c31d578d4ec5b18088fb7-vnan", + "instance_element-hq__element-web-772df3021201d9c73835a626df8dcb6334ad9a3e-vnan", + "instance_internetarchive__openlibrary-1be7de788a444f6255e89c10ef6aa608550604a8-v29f82c9cf21d57b242f8d8b0e541525d259e2d63", + "instance_internetarchive__openlibrary-630221ab686c64e75a2ce253c893c033e4814b2e-v93c53c13d5f9b383ebb411ee7750b49dcd1a34c6", + "instance_gravitational__teleport-f432a71a13e698b6e1c4672a2e9e9c1f32d35c12", + "instance_protonmail__webclients-0d0267c4438cf378bda90bc85eed3a3615871ac4", + "instance_internetarchive__openlibrary-dbbd9d539c6d4fd45d5be9662aa19b6d664b5137-v08d8e8889ec945ab821fb156c04c7d2e2810debb", + "instance_flipt-io__flipt-b2170346dc37cf42fda1386cd630f24821ad2ac5", + "instance_element-hq__element-web-aec454dd6feeb93000380523cbb0b3681c0275fd-vnan", + "instance_internetarchive__openlibrary-4a5d2a7d24c9e4c11d3069220c0685b736d5ecde-v13642507b4fc1f8d234172bf8129942da2c2ca26", + "instance_qutebrowser__qutebrowser-ec2dcfce9eee9f808efc17a1b99e227fc4421dea-v5149fcda2a9a6fe1d35dfed1bade1444a11ef271", + "instance_gravitational__teleport-1330415d33a27594c948a36d9d7701f496229e9f", + "instance_protonmail__webclients-3a6790f480309130b5d6332dce6c9d5ccca13ee3", + "instance_flipt-io__flipt-492cc0b158200089dceede3b1aba0ed28df3fb1d", + "instance_future-architect__vuls-edb324c3d9ec3b107bf947f00e38af99d05b3e16", + "instance_ansible__ansible-bec27fb4c0a40c5f8bbcf26a475704227d65ee73-v30a923fb5c164d6cd18280c02422f75e611e8fb2", + "instance_gravitational__teleport-c1b1c6a1541c478d7777a48fca993cc8206c73b9", + "instance_tutao__tutanota-1e516e989b3c0221f4af6b297d9c0e4c43e4adc3-vbc0d9ba8f0071fbe982809910959a6ff8884dbbf", + "instance_flipt-io__flipt-406f9396ad65696d58865b3a6283109cd4eaf40e", + "instance_future-architect__vuls-139f3a81b66c47e6d8f70ce6c4afe7a9196a6ea8", + "instance_flipt-io__flipt-3b2c25ee8a3ac247c3fad13ad8d64ace34ec8ee7", + "instance_gravitational__teleport-0ac7334939981cf85b9591ac295c3816954e287e", + "instance_qutebrowser__qutebrowser-e15d26630934d0b6415ed2295ac42fd570a57620-va0fd88aac89cde702ec1ba84877234da33adce8a", + "instance_ansible__ansible-5e88cd9972f10b66dd97e1ee684c910c6a2dd25e-v906c969b551b346ef54a2c0b41e04f632b7b73c2", + "instance_element-hq__element-web-923ad4323b2006b2b180544429455ffe7d4a6cc3-vnan", + "instance_protonmail__webclients-4feccbc9990980aee26ea29035f8f931d6089895", + "instance_flipt-io__flipt-a0cbc0cb65ae601270bdbe3f5313e2dfd49c80e4", + "instance_gravitational__teleport-b4e7cd3a5e246736d3fe8d6886af55030b232277", + "instance_ansible__ansible-6cc97447aac5816745278f3735af128afb255c81-v0f01c69f1e2528b935359cfe578530722bca2c59", + "instance_ansible__ansible-f327e65d11bb905ed9f15996024f857a95592629-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_qutebrowser__qutebrowser-5fdc83e5da6222fe61163395baaad7ae57fa2cb4-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_navidrome__navidrome-3982ba725883e71d4e3e618c61d5140eeb8d850a", + "instance_qutebrowser__qutebrowser-ebfe9b7aa0c4ba9d451f993e08955004aaec4345-v059c6fdc75567943479b23ebca7c07b5e9a7f34c", + "instance_ansible__ansible-709484969c8a4ffd74b839a673431a8c5caa6457-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_gravitational__teleport-ad41b3c15414b28a6cec8c25424a19bfa7abd0e9-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_protonmail__webclients-2c3559cad02d1090985dba7e8eb5a129144d9811", + "instance_flipt-io__flipt-5af0757e96dec4962a076376d1bedc79de0d4249", + "instance_element-hq__element-web-56c7fc1948923b4b3f3507799e725ac16bcf8018-vnan", + "instance_flipt-io__flipt-381b90f718435c4694380b5fcd0d5cf8e3b5a25a", + "instance_flipt-io__flipt-6fd0f9e2587f14ac1fdd1c229f0bcae0468c8daa", + "instance_ansible__ansible-4c5ce5a1a9e79a845aff4978cfeb72a0d4ecf7d6-v1055803c3a812189a1133297f7f5468579283f86", + "instance_navidrome__navidrome-7073d18b54da7e53274d11c9e2baef1242e8769e", + "instance_internetarchive__openlibrary-0a90f9f0256e4f933523e9842799e39f95ae29ce-v76304ecdb3a5954fcf13feb710e8c40fcf24b73c", + "instance_nodebb__nodebb-be43cd25974681c9743d424238b7536c357dc8d3-vf2cf3cbd463b7ad942381f1c6d077626485a1e9e", + "instance_flipt-io__flipt-05d7234fa582df632f70a7cd10194d61bd7043b9", + "instance_nodebb__nodebb-76c6e30282906ac664f2c9278fc90999b27b1f48-vd59a5728dfc977f44533186ace531248c2917516", + "instance_qutebrowser__qutebrowser-6b320dc18662580e1313d2548fdd6231d2a97e6d-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_future-architect__vuls-01441351c3407abfc21c48a38e28828e1b504e0c", + "instance_ansible__ansible-1a4644ff15355fd696ac5b9d074a566a80fe7ca3-v30a923fb5c164d6cd18280c02422f75e611e8fb2", + "instance_navidrome__navidrome-09ae41a2da66264c60ef307882362d2e2d8d8b89", + "instance_internetarchive__openlibrary-5c6c22f3d2edf2f1b10f5dc335e32cb6a5f40341-v76304ecdb3a5954fcf13feb710e8c40fcf24b73c", + "instance_navidrome__navidrome-bf2bcb12799b21069f137749e0c331f761d1f693", + "instance_flipt-io__flipt-2ce8a0331e8a8f63f2c1b555db8277ffe5aa2e63", + "instance_ansible__ansible-a1569ea4ca6af5480cf0b7b3135f5e12add28a44-v0f01c69f1e2528b935359cfe578530722bca2c59", + "instance_gravitational__teleport-82185f232ae8974258397e121b3bc2ed0c3729ed-v626ec2a48416b10a88641359a169d99e935ff037", + "instance_flipt-io__flipt-72d06db14d58692bfb4d07b1aa745a37b35956f3", + "instance_ansible__ansible-b6290e1d156af608bd79118d209a64a051c55001-v390e508d27db7a51eece36bb6d9698b63a5b638a", + "instance_gravitational__teleport-32bcd71591c234f0d8b091ec01f1f5cbfdc0f13c-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_ansible__ansible-a02e22e902a69aeb465f16bf03f7f5a91b2cb828-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_gravitational__teleport-eda668c30d9d3b56d9c69197b120b01013611186", + "instance_nodebb__nodebb-0f788b8eaa4bba3c142d171fd941d015c53b65fc-v0ec6d6c2baf3cb4797482ce4829bc25cd5716649", + "instance_gravitational__teleport-47530e1fd8bfb84ec096ebcbbc29990f30829655-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_qutebrowser__qutebrowser-e64622cd2df5b521342cf4a62e0d4cb8f8c9ae5a-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_element-hq__element-web-494d9de6f0a94ffb491e74744d2735bce02dc0ab-vnan", + "instance_element-hq__element-web-9bf77963ee5e036d54b2a3ca202fbf6378464a5e-vnan", + "instance_nodebb__nodebb-9c576a0758690f45a6ca03b5884c601e473bf2c1-vd59a5728dfc977f44533186ace531248c2917516", + "instance_internetarchive__openlibrary-25858f9f0c165df25742acf8309ce909773f0cdd-v13642507b4fc1f8d234172bf8129942da2c2ca26", + "instance_protonmail__webclients-f161c10cf7d31abf82e8d64d7a99c9fac5acfa18", + "instance_gravitational__teleport-cb712e3f0b06dadc679f895daef8072cae400c26-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_gravitational__teleport-46a13210519461c7cec8d643bfbe750265775b41", + "instance_ansible__ansible-e40889e7112ae00a21a2c74312b330e67a766cc0-v1055803c3a812189a1133297f7f5468579283f86", + "instance_gravitational__teleport-6eaaf3a27e64f4ef4ef855bd35d7ec338cf17460-v626ec2a48416b10a88641359a169d99e935ff037", + "instance_internetarchive__openlibrary-0d13e6b4bf80bced6c0946b969b9a1b6963f6bce-v76304ecdb3a5954fcf13feb710e8c40fcf24b73c", + "instance_protonmail__webclients-2f2f6c311c6128fe86976950d3c0c2db07b03921", + "instance_ansible__ansible-5260527c4a71bfed99d803e687dd19619423b134-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_qutebrowser__qutebrowser-8d05f0282a271bfd45e614238bd1b555c58b3fc1-v35616345bb8052ea303186706cec663146f0f184", + "instance_qutebrowser__qutebrowser-50efac08f623644a85441bbe02ab9347d2b71a9d-v9f8e9d96c85c85a605e382f1510bd08563afc566", + "instance_flipt-io__flipt-dae029cba7cdb98dfb1a6b416c00d324241e6063", + "instance_nodebb__nodebb-18c45b44613aecd53e9f60457b9812049ab2998d-v0495b863a912fbff5749c67e860612b91825407c", + "instance_future-architect__vuls-e4728e388120b311c4ed469e4f942e0347a2689b-v264a82e2f4818e30f5a25e4da53b27ba119f62b5", + "instance_qutebrowser__qutebrowser-5e0d6dc1483cb3336ea0e3dcbd4fe4aa00fc1742-v5149fcda2a9a6fe1d35dfed1bade1444a11ef271", + "instance_flipt-io__flipt-c12967bc73fdf02054cf3ef8498c05e25f0a18c0", + "instance_qutebrowser__qutebrowser-ff1c025ad3210506fc76e1f604d8c8c27637d88e-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_ansible__ansible-83909bfa22573777e3db5688773bda59721962ad-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_qutebrowser__qutebrowser-1a9e74bfaf9a9db2a510dc14572d33ded6040a57-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_gravitational__teleport-24cafecd8721891092210afc55f6413ab46ca211-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_internetarchive__openlibrary-4b7ea2977be2747496ba792a678940baa985f7ea-v0f5aece3601a5b4419f7ccec1dbda2071be28ee4", + "instance_internetarchive__openlibrary-fad4a40acf5ff5f06cd7441a5c7baf41a7d81fe4-vfa6ff903cb27f336e17654595dd900fa943dcd91", + "instance_qutebrowser__qutebrowser-3fd8e12949b8feda401930574facf09dd4180bba", + "instance_ansible__ansible-164881d871964aa64e0f911d03ae270acbad253c-v390e508d27db7a51eece36bb6d9698b63a5b638a", + "instance_ansible__ansible-fb144c44144f8bd3542e71f5db62b6d322c7bd85-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_element-hq__element-web-582a1b093fc0b77538052f45cbb9c7295f991b51-vnan", + "instance_element-hq__element-web-ec0f940ef0e8e3b61078f145f34dc40d1938e6c5-vnan", + "instance_internetarchive__openlibrary-6fdbbeee4c0a7e976ff3e46fb1d36f4eb110c428-v08d8e8889ec945ab821fb156c04c7d2e2810debb", + "instance_flipt-io__flipt-7161f7b876773a911afdd804b281e52681cb7321", + "instance_navidrome__navidrome-28389fb05e1523564dfc61fa43ed8eb8a10f938c", + "instance_ansible__ansible-5f4e332e3762999d94af27746db29ff1729252c1-v0f01c69f1e2528b935359cfe578530722bca2c59", + "instance_nodebb__nodebb-0e07f3c9bace416cbab078a30eae972868c0a8a3-vf2cf3cbd463b7ad942381f1c6d077626485a1e9e", + "instance_nodebb__nodebb-da0211b1a001d45d73b4c84c6417a4f1b0312575-vf2cf3cbd463b7ad942381f1c6d077626485a1e9e", + "instance_navidrome__navidrome-31799662706fedddf5bcc1a76b50409d1f91d327", + "instance_future-architect__vuls-50580f6e98eeb36f53f27222f7f4fdfea0b21e8d", + "instance_internetarchive__openlibrary-60725705782832a2cb22e17c49697948a42a9d03-v298a7a812ceed28c4c18355a091f1b268fe56d86", + "instance_ansible__ansible-3db08adbb1cc6aa9941be5e0fc810132c6e1fa4b-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_internetarchive__openlibrary-8a5a63af6e0be406aa6c8c9b6d5f28b2f1b6af5a-v0f5aece3601a5b4419f7ccec1dbda2071be28ee4", + "instance_future-architect__vuls-e1df74cbc1a1d1889428b3333a3b2405c4651993", + "instance_internetarchive__openlibrary-123e6e5e1c85b9c07d1e98f70bfc480bc8016890-v2733ff199fb72f0d033a30dc62cb0a4742e3a7f4", + "instance_protonmail__webclients-ae36cb23a1682dcfd69587c1b311ae0227e28f39", + "instance_protonmail__webclients-8be4f6cb9380fcd2e67bcb18cef931ae0d4b869c", + "instance_qutebrowser__qutebrowser-a25e8a09873838ca9efefd36ea8a45170bbeb95c-vc2f56a753b62a190ddb23cd330c257b9cf560d12", + "instance_qutebrowser__qutebrowser-44e64199ed38003253f0296badd4a447645067b6-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_gravitational__teleport-7744f72c6eb631791434b648ba41083b5f6d2278-vce94f93ad1030e3136852817f2423c1b3ac37bc4", + "instance_tutao__tutanota-40e94dee2bcec2b63f362da283123e9df1874cc1-vc4e41fd0029957297843cb9dec4a25c7c756f029", + "instance_future-architect__vuls-c11ba27509f733d7d280bdf661cbbe2e7a99df4c", + "instance_flipt-io__flipt-b2cd6a6dd73ca91b519015fd5924fde8d17f3f06", + "instance_gravitational__teleport-6a14edcf1ff010172fdbac622d0a474ed6af46de", + "instance_flipt-io__flipt-36e62baffae2132f78f9d34dc300a9baa2d7ae0e", + "instance_navidrome__navidrome-c90468b895f6171e33e937ff20dc915c995274f0", + "instance_ansible__ansible-ecea15c508f0e081525be036cf76bbb56dbcdd9d-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_ansible__ansible-e22e103cdf8edc56ff7d9b848a58f94f1471a263-v1055803c3a812189a1133297f7f5468579283f86", + "instance_future-architect__vuls-dc496468b9e9fb73371f9606cdcdb0f8e12e70ca", + "instance_gravitational__teleport-c782838c3a174fdff80cafd8cd3b1aa4dae8beb2", + "instance_internetarchive__openlibrary-91efee627df01e32007abf2d6ebf73f9d9053076-vbee42ad1b72fb23c6a1c874868a720b370983ed2", + "instance_qutebrowser__qutebrowser-52708364b5f91e198defb022d1a5b4b3ebd9b563-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_ansible__ansible-eea46a0d1b99a6dadedbb6a3502d599235fa7ec3-v390e508d27db7a51eece36bb6d9698b63a5b638a", + "instance_protonmail__webclients-c6f65d205c401350a226bb005f42fac1754b0b5b", + "instance_ansible__ansible-cb94c0cc550df9e98f1247bc71d8c2b861c75049-v1055803c3a812189a1133297f7f5468579283f86", + "instance_nodebb__nodebb-05f2236193f407cf8e2072757fbd6bb170bc13f0-vf2cf3cbd463b7ad942381f1c6d077626485a1e9e", + "instance_navidrome__navidrome-55730514ea59d5f1d0b8e3f8745569c29bdbf7b4", + "instance_navidrome__navidrome-0130c6dc13438b48cf0fdfab08a89e357b5517c9", + "instance_flipt-io__flipt-507170da0f7f4da330f6732bffdf11c4df7fc192", + "instance_tutao__tutanota-fe240cbf7f0fdd6744ef7bef8cb61676bcdbb621-vc4e41fd0029957297843cb9dec4a25c7c756f029", + "instance_qutebrowser__qutebrowser-233cb1cc48635130e5602549856a6fa4ab4c087f-v35616345bb8052ea303186706cec663146f0f184", + "instance_internetarchive__openlibrary-f8cc11d9c1575fdba5ac66aee0befca970da8d64-v13642507b4fc1f8d234172bf8129942da2c2ca26", + "instance_qutebrowser__qutebrowser-fec187c2cb53d769c2682b35ca77858a811414a8-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_tutao__tutanota-da4edb7375c10f47f4ed3860a591c5e6557f7b5c-vbc0d9ba8f0071fbe982809910959a6ff8884dbbf", + "instance_ansible__ansible-be2c376ab87e3e872ca21697508f12c6909cf85a-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_future-architect__vuls-457a3a9627fb9a0800d0aecf1d4713fb634a9011", + "instance_internetarchive__openlibrary-d40ec88713dc95ea791b252f92d2f7b75e107440-v13642507b4fc1f8d234172bf8129942da2c2ca26", + "instance_protonmail__webclients-863d524b5717b9d33ce08a0f0535e3fd8e8d1ed8", + "instance_internetarchive__openlibrary-308a35d6999427c02b1dbf5211c033ad3b352556-ve8c8d62a2b60610a3c4631f5f23ed866bada9818", + "instance_flipt-io__flipt-15b76cada1ef29cfa56b0fba36754be36243dded", + "instance_future-architect__vuls-61c39637f2f3809e1b5dad05f0c57c799dce1587", + "instance_ansible__ansible-5d253a13807e884b7ce0b6b57a963a45e2f0322c-v1055803c3a812189a1133297f7f5468579283f86", + "instance_element-hq__element-web-44b98896a79ede48f5ad7ff22619a39d5f6ff03c-vnan", + "instance_qutebrowser__qutebrowser-f91ace96223cac8161c16dd061907e138fe85111-v059c6fdc75567943479b23ebca7c07b5e9a7f34c", + "instance_gravitational__teleport-bb562408da4adeae16e025be65e170959d1ec492-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_gravitational__teleport-78b0d8c72637df1129fb6ff84fc49ef4b5ab1288", + "instance_internetarchive__openlibrary-3f7db6bbbcc7c418b3db72d157c6aed1d45b2ccf-v430f20c722405e462d9ef44dee7d34c41e76fe7a", + "instance_future-architect__vuls-e52fa8d6ed1d23e36f2a86e5d3efe9aa057a1b0d", + "instance_navidrome__navidrome-3bc9e75b2843f91f6a1e9b604e321c2bd4fd442a", + "instance_tutao__tutanota-fb32e5f9d9fc152a00144d56dd0af01760a2d4dc-vc4e41fd0029957297843cb9dec4a25c7c756f029", + "instance_gravitational__teleport-af5e2517de7d18406b614e413aca61c319312171-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_ansible__ansible-a20a52701402a12f91396549df04ac55809f68e9-v1055803c3a812189a1133297f7f5468579283f86", + "instance_protonmail__webclients-1501eb765873b2884b6f1944fd242ecfc9d6b103", + "instance_gravitational__teleport-db89206db6c2969266e664c7c0fb51b70e958b64", + "instance_ansible__ansible-be59caa59bf47ca78a4760eb7ff38568372a8260-v1055803c3a812189a1133297f7f5468579283f86", + "instance_flipt-io__flipt-65581fef4aa807540cb933753d085feb0d7e736f", + "instance_gravitational__teleport-4e1c39639edf1ab494dd7562844c8b277b5cfa18-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_future-architect__vuls-f0b3a8b1db98eb1bd32685f1c36c41a99c3452ed", + "instance_navidrome__navidrome-b65e76293a917ee2dfc5d4b373b1c62e054d0dca", + "instance_element-hq__element-web-33e8edb3d508d6eefb354819ca693b7accc695e7", + "instance_flipt-io__flipt-e50808c03e4b9d25a6a78af9c61a3b1616ea356b", + "instance_future-architect__vuls-e3c27e1817d68248043bd09d63cc31f3344a6f2c", + "instance_element-hq__element-web-ad26925bb6628260cfe0fcf90ec0a8cba381f4a4-vnan", + "instance_nodebb__nodebb-1ea9481af6125ffd6da0592ed439aa62af0bca11-vd59a5728dfc977f44533186ace531248c2917516", + "instance_ansible__ansible-489156378c8e97374a75a544c7c9c2c0dd8146d1-v390e508d27db7a51eece36bb6d9698b63a5b638a", + "instance_internetarchive__openlibrary-bb152d23c004f3d68986877143bb0f83531fe401-ve8c8d62a2b60610a3c4631f5f23ed866bada9818", + "instance_qutebrowser__qutebrowser-0b621cb0ce2b54d3f93d8d41d8ff4257888a87e5-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_qutebrowser__qutebrowser-1af602b258b97aaba69d2585ed499d95e2303ac2-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_future-architect__vuls-e1fab805afcfc92a2a615371d0ec1e667503c254-v264a82e2f4818e30f5a25e4da53b27ba119f62b5", + "instance_internetarchive__openlibrary-7edd1ef09d91fe0b435707633c5cc9af41dedddf-v76304ecdb3a5954fcf13feb710e8c40fcf24b73c", + "instance_qutebrowser__qutebrowser-99029144b5109bb1b2a53964a7c129e009980cd9-va0fd88aac89cde702ec1ba84877234da33adce8a", + "instance_ansible__ansible-12734fa21c08a0ce8c84e533abdc560db2eb1955-v7eee2454f617569fd6889f2211f75bc02a35f9f8", + "instance_protonmail__webclients-7e54526774e577c0ebb58ced7ba8bef349a69fec", + "instance_gravitational__teleport-1a77b7945a022ab86858029d30ac7ad0d5239d00-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_nodebb__nodebb-8ca65b0c78c67c1653487c02d1135e1b702185e1-vf2cf3cbd463b7ad942381f1c6d077626485a1e9e", + "instance_flipt-io__flipt-9f8127f225a86245fa35dca4885c2daef824ee55", + "instance_gravitational__teleport-ac2fb2f9b4fd1896b554d3011df23d3d71295779", + "instance_ansible__ansible-f02a62db509dc7463fab642c9c3458b9bc3476cc-v390e508d27db7a51eece36bb6d9698b63a5b638a" + ], + "additional_details": { + "shuffled": "True" + } + }, + "evaluation_timestamp": "1774867708.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.525, + "uncertainty": { + "standard_error": { + "value": 0.03539972744976418 + }, + "num_samples": 200 + }, + "details": { + "total_matched_trajectories": "25", + "stopping_reason_count_completed_on_successful_submit": "18", + "stopping_reason_count_repetition_guard": "7", + "avg_total_tokens_target_model": "2621540.52", + "total_total_tokens_target_model": "65538513.0", + "avg_total_tokens_other_models": "0.00", + "total_total_tokens_other_models": "0.0", + "avg_total_tokens_all_models": "2621540.52", + "total_total_tokens_all_models": "65538513.0", + "avg_turn_count": "68.32", + "total_turn_count": "1708.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"If you do not submit an answer, you will receive a score of zero.\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"Your submission was incorrect. Please review your work, identify what went wrong, and try again.\", \"to_float\"], \"submit\": null, \"on_continue\": null, \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\"}, \"params_passed\": {\"prompt\": [\"If you do not submit an answer, you will receive a score of zero.\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"Your submission was incorrect. Please review your work, identify what went wrong, and try again.\", \"to_float\"], \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 30000000 + }, + "sandbox": { + "type": "k8s" + } + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/openai/gpt-5.2-2025-12-11/15d21311-3f01-460e-8505-622d3e97b363_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "866e22cf2b29acbd3fc50784e60c1372ff98f5c687088b9ad5b8968c76a7b12f", + "total_rows": 200 + } +} \ No newline at end of file diff --git a/data/swebenchpro/openai/gpt-5.2-2025-12-11/15d21311-3f01-460e-8505-622d3e97b363_samples.jsonl b/data/swebenchpro/openai/gpt-5.2-2025-12-11/15d21311-3f01-460e-8505-622d3e97b363_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..c0bc62f95f1cde594b5fbaaa71942fb494e45d78 --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.2-2025-12-11/15d21311-3f01-460e-8505-622d3e97b363_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:866e22cf2b29acbd3fc50784e60c1372ff98f5c687088b9ad5b8968c76a7b12f +size 365725052 diff --git a/data/swebenchpro/openai/gpt-5.2-2025-12-11/1b361fb6-ae6a-456d-bf90-51f9a400c7cf.json b/data/swebenchpro/openai/gpt-5.2-2025-12-11/1b361fb6-ae6a-456d-bf90-51f9a400c7cf.json new file mode 100644 index 0000000000000000000000000000000000000000..e6a36a7262dde5288ecdbed149fa5a99f30c2657 --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.2-2025-12-11/1b361fb6-ae6a-456d-bf90-51f9a400c7cf.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/openai_gpt-5.2-2025-12-11/1786169005.641518", + "evaluation_timestamp": "1776520455.0", + "retrieved_timestamp": "1786169005.641518", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/openai/gpt-5.2-2025-12-11/1b361fb6-ae6a-456d-bf90-51f9a400c7cf_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "a718b389ce9bace69f2ec03bf07ed66f268f7a23e86867c3db0e06e87608a16b", + "total_rows": 20 + } +} \ No newline at end of file diff --git a/data/swebenchpro/openai/gpt-5.2-2025-12-11/1b361fb6-ae6a-456d-bf90-51f9a400c7cf_samples.jsonl b/data/swebenchpro/openai/gpt-5.2-2025-12-11/1b361fb6-ae6a-456d-bf90-51f9a400c7cf_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..540fcfca7afcc3a4c12bd52e89ce05fd7ac01de6 --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.2-2025-12-11/1b361fb6-ae6a-456d-bf90-51f9a400c7cf_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a718b389ce9bace69f2ec03bf07ed66f268f7a23e86867c3db0e06e87608a16b +size 17249337 diff --git a/data/swebenchpro/openai/gpt-5.2-2025-12-11/1e8b37bb-0a85-42c2-bfd7-f34ae917b917.json b/data/swebenchpro/openai/gpt-5.2-2025-12-11/1e8b37bb-0a85-42c2-bfd7-f34ae917b917.json new file mode 100644 index 0000000000000000000000000000000000000000..008f552aef5390ba8335cc8aec1544811372d649 --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.2-2025-12-11/1e8b37bb-0a85-42c2-bfd7-f34ae917b917.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/openai_gpt-5.2-2025-12-11/1786167104.985348", + "evaluation_timestamp": "1778627169.0", + "retrieved_timestamp": "1786167104.985348", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/openai/gpt-5.2-2025-12-11/1e8b37bb-0a85-42c2-bfd7-f34ae917b917_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "499b5daabd3337236fa336210080bd3c20f2a0066b1acfe5cfb62f861f8a8e41", + "total_rows": 38 + } +} \ No newline at end of file diff --git a/data/swebenchpro/openai/gpt-5.2-2025-12-11/1e8b37bb-0a85-42c2-bfd7-f34ae917b917_samples.jsonl b/data/swebenchpro/openai/gpt-5.2-2025-12-11/1e8b37bb-0a85-42c2-bfd7-f34ae917b917_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..0d4aea79409cd555a7cabbfdfc1460b063244920 --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.2-2025-12-11/1e8b37bb-0a85-42c2-bfd7-f34ae917b917_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:499b5daabd3337236fa336210080bd3c20f2a0066b1acfe5cfb62f861f8a8e41 +size 21017935 diff --git a/data/swebenchpro/openai/gpt-5.2-2025-12-11/27ca0295-2e85-4b10-8d94-ae6fb833c4ff.json b/data/swebenchpro/openai/gpt-5.2-2025-12-11/27ca0295-2e85-4b10-8d94-ae6fb833c4ff.json new file mode 100644 index 0000000000000000000000000000000000000000..8b0b3be259bba0b7fe3a0f029b12ba32e8953242 --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.2-2025-12-11/27ca0295-2e85-4b10-8d94-ae6fb833c4ff.json @@ -0,0 +1,111 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/openai_gpt-5.2-2025-12-11/1786164377.299649", + "evaluation_timestamp": "1776696546.0", + "retrieved_timestamp": "1786164377.299649", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on swebenchpro/S-adaptive/+3ep/a4a47e32 for scorer harbor_scorer", + "source_data": { + "dataset_name": "swebenchpro", + "source_type": "hf_dataset", + "samples_number": 2, + "sample_ids": [ + "instance_future-architect__vuls-78b52d6a7f480bd610b692de9bf0c86f57332f23", + "instance_nodebb__nodebb-84e065752f6d7fbe5c08cbf50cb173ffb866b8fa-vf2cf3cbd463b7ad942381f1c6d077626485a1e9e" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1776696546.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.5, + "uncertainty": { + "standard_error": { + "value": 0.5 + }, + "num_samples": 2 + }, + "details": { + "total_matched_trajectories": "2", + "stopping_reason_count_repetition_guard": "2", + "avg_total_tokens_target_model": "14063622.50", + "total_total_tokens_target_model": "28127245.0", + "avg_total_tokens_other_models": "1698.00", + "total_total_tokens_other_models": "3396.0", + "avg_total_tokens_all_models": "14065320.50", + "total_total_tokens_all_models": "28130641.0", + "avg_turn_count": "212.00", + "total_turn_count": "424.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "3600", + "attempt_timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 30000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/openai/gpt-5.2-2025-12-11/27ca0295-2e85-4b10-8d94-ae6fb833c4ff_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "ce15866a29f5bdfec428c5a076d262dda25fb6c9a39a214ae7b830ad303f2410", + "total_rows": 2 + } +} \ No newline at end of file diff --git a/data/swebenchpro/openai/gpt-5.2-2025-12-11/27ca0295-2e85-4b10-8d94-ae6fb833c4ff_samples.jsonl b/data/swebenchpro/openai/gpt-5.2-2025-12-11/27ca0295-2e85-4b10-8d94-ae6fb833c4ff_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..1c5ec00d0c25f1a15c55178787a217cba15016bf --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.2-2025-12-11/27ca0295-2e85-4b10-8d94-ae6fb833c4ff_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ce15866a29f5bdfec428c5a076d262dda25fb6c9a39a214ae7b830ad303f2410 +size 2282201 diff --git a/data/swebenchpro/openai/gpt-5.2-2025-12-11/296cb7f0-2087-44a1-be46-b3638f2b1190.json b/data/swebenchpro/openai/gpt-5.2-2025-12-11/296cb7f0-2087-44a1-be46-b3638f2b1190.json new file mode 100644 index 0000000000000000000000000000000000000000..204821dea72311e4a3b537ba13d63d6e40b4bc51 --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.2-2025-12-11/296cb7f0-2087-44a1-be46-b3638f2b1190.json @@ -0,0 +1,120 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/openai_gpt-5.2-2025-12-11/1786164396.292615", + "evaluation_timestamp": "1776714136.0", + "retrieved_timestamp": "1786164396.292615", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on swebenchpro/S-adaptive/+4ep/563e805c for scorer harbor_scorer", + "source_data": { + "dataset_name": "swebenchpro", + "source_type": "hf_dataset", + "samples_number": 10, + "sample_ids": [ + "instance_internetarchive__openlibrary-fdbc0d8f418333c7e575c40b661b582c301ef7ac-v13642507b4fc1f8d234172bf8129942da2c2ca26", + "instance_ansible__ansible-b2a289dcbb702003377221e25f62c8a3608f0e89-v173091e2e36d38c978002990795f66cfc0af30ad", + "instance_internetarchive__openlibrary-f8cc11d9c1575fdba5ac66aee0befca970da8d64-v13642507b4fc1f8d234172bf8129942da2c2ca26", + "instance_qutebrowser__qutebrowser-e70f5b03187bdd40e8bf70f5f3ead840f52d1f42-v02ad04386d5238fe2d1a1be450df257370de4b6a", + "instance_tutao__tutanota-d1aa0ecec288bfc800cfb9133b087c4f81ad8b38-vbc0d9ba8f0071fbe982809910959a6ff8884dbbf", + "instance_qutebrowser__qutebrowser-e64622cd2df5b521342cf4a62e0d4cb8f8c9ae5a-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_gravitational__teleport-ac2fb2f9b4fd1896b554d3011df23d3d71295779", + "instance_protonmail__webclients-cb8cc309c6968b0a2a5fe4288d0ae0a969ff31e1", + "instance_ansible__ansible-11c1777d56664b1acb56b387a1ad6aeadef1391d-v0f01c69f1e2528b935359cfe578530722bca2c59", + "instance_ansible__ansible-29aea9ff3466e4cd2ed00524b9e56738d568ce8b-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1776714136.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.875, + "uncertainty": { + "standard_error": { + "value": 0.10034662148993578 + }, + "num_samples": 35 + }, + "details": { + "total_matched_trajectories": "35", + "stopping_reason_count_repetition_guard": "34", + "stopping_reason_count_token_limit": "1", + "avg_total_tokens_target_model": "4501421.31", + "total_total_tokens_target_model": "157549746.0", + "avg_total_tokens_other_models": "1143.17", + "total_total_tokens_other_models": "40011.0", + "avg_total_tokens_all_models": "4502564.49", + "total_total_tokens_all_models": "157589757.0", + "avg_turn_count": "114.03", + "total_turn_count": "3991.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "3600", + "attempt_timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 30000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/openai/gpt-5.2-2025-12-11/296cb7f0-2087-44a1-be46-b3638f2b1190_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "1a0f89f86fe93ac85ce7f358c47a869189191220ee555cb3ca2f1a7826536429", + "total_rows": 35 + } +} \ No newline at end of file diff --git a/data/swebenchpro/openai/gpt-5.2-2025-12-11/296cb7f0-2087-44a1-be46-b3638f2b1190_samples.jsonl b/data/swebenchpro/openai/gpt-5.2-2025-12-11/296cb7f0-2087-44a1-be46-b3638f2b1190_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..c06448c03d4bba6a6fea8461578e53cf051bf8de --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.2-2025-12-11/296cb7f0-2087-44a1-be46-b3638f2b1190_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1a0f89f86fe93ac85ce7f358c47a869189191220ee555cb3ca2f1a7826536429 +size 25197624 diff --git a/data/swebenchpro/openai/gpt-5.2-2025-12-11/2b2b63ce-bd56-402e-bbac-b440924d2140.json b/data/swebenchpro/openai/gpt-5.2-2025-12-11/2b2b63ce-bd56-402e-bbac-b440924d2140.json new file mode 100644 index 0000000000000000000000000000000000000000..b4cdbc4b63af2134fc25ef34a5a31ecabe53a1d8 --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.2-2025-12-11/2b2b63ce-bd56-402e-bbac-b440924d2140.json @@ -0,0 +1,134 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/openai_gpt-5.2-2025-12-11/1786164350.896046", + "evaluation_timestamp": "1778590202.0", + "retrieved_timestamp": "1786164350.896046", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on swebenchpro/S-adaptive/+1ep/b6f51969 for scorer harbor_scorer", + "source_data": { + "dataset_name": "swebenchpro", + "source_type": "hf_dataset", + "samples_number": 22, + "sample_ids": [ + "instance_qutebrowser__qutebrowser-e5340c449f23608803c286da0563b62f58ba25b0-v059c6fdc75567943479b23ebca7c07b5e9a7f34c", + "instance_tutao__tutanota-4b4e45949096bb288f2b522f657610e480efa3e8-vee878bb72091875e912c52fc32bc60ec3760227b", + "instance_protonmail__webclients-01b519cd49e6a24d9a05d2eb97f54e420740072e", + "instance_future-architect__vuls-030b2e03525d68d74cb749959aac2d7f3fc0effa", + "instance_qutebrowser__qutebrowser-ef5ba1a0360b39f9eff027fbdc57f363597c3c3b-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_gravitational__teleport-82185f232ae8974258397e121b3bc2ed0c3729ed-v626ec2a48416b10a88641359a169d99e935ff037", + "instance_ansible__ansible-de5858f48dc9e1ce9117034e0d7e76806f420ca8-v1055803c3a812189a1133297f7f5468579283f86", + "instance_gravitational__teleport-46aa81b1ce96ebb4ebed2ae53fd78cd44a05da6c-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_nodebb__nodebb-a917210c5b2c20637094545401f85783905c074c-vf2cf3cbd463b7ad942381f1c6d077626485a1e9e", + "instance_internetarchive__openlibrary-1351c59fd43689753de1fca32c78d539a116ffc1-v29f82c9cf21d57b242f8d8b0e541525d259e2d63", + "instance_qutebrowser__qutebrowser-a25e8a09873838ca9efefd36ea8a45170bbeb95c-vc2f56a753b62a190ddb23cd330c257b9cf560d12", + "instance_protonmail__webclients-281a6b3f190f323ec2c0630999354fafb84b2880", + "instance_element-hq__element-web-1077729a19c0ce902e713cf6fab42c91fb7907f1-vnan", + "instance_future-architect__vuls-be7b9114cc9545e68fb0ee7bc63d7ec53d1a00ad", + "instance_qutebrowser__qutebrowser-c580ebf0801e5a3ecabc54f327498bb753c6d5f2-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_gravitational__teleport-3ff19cf7c41f396ae468797d3aeb61515517edc9-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_future-architect__vuls-aaea15e516ece43978cf98e09e52080478b1d39f", + "instance_protonmail__webclients-f161c10cf7d31abf82e8d64d7a99c9fac5acfa18", + "instance_qutebrowser__qutebrowser-cf06f4e3708f886032d4d2a30108c2fddb042d81-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_internetarchive__openlibrary-3f580a5f244c299d936d73d9e327ba873b6401d9-v0f5aece3601a5b4419f7ccec1dbda2071be28ee4", + "instance_qutebrowser__qutebrowser-0aa57e4f7243024fa4bba8853306691b5dbd77b3-v5149fcda2a9a6fe1d35dfed1bade1444a11ef271", + "instance_future-architect__vuls-50580f6e98eeb36f53f27222f7f4fdfea0b21e8d" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1778590202.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.55, + "uncertainty": { + "standard_error": { + "value": 0.11413288653790231 + }, + "num_samples": 20 + }, + "details": { + "total_matched_trajectories": "20", + "stopping_reason_count_repetition_guard": "20", + "avg_total_tokens_target_model": "4591510.55", + "total_total_tokens_target_model": "91830211.0", + "avg_total_tokens_other_models": "1143.35", + "total_total_tokens_other_models": "22867.0", + "avg_total_tokens_all_models": "4592653.90", + "total_total_tokens_all_models": "91853078.0", + "avg_turn_count": "105.05", + "total_turn_count": "2101.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 30000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/openai/gpt-5.2-2025-12-11/2b2b63ce-bd56-402e-bbac-b440924d2140_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "9a1c423a51d4350e9a727a336ed619397be2fb70d29ccbe81f64472181ee2b00", + "total_rows": 20 + } +} \ No newline at end of file diff --git a/data/swebenchpro/openai/gpt-5.2-2025-12-11/2b2b63ce-bd56-402e-bbac-b440924d2140_samples.jsonl b/data/swebenchpro/openai/gpt-5.2-2025-12-11/2b2b63ce-bd56-402e-bbac-b440924d2140_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..6437eed3b128d8fe6a21942835b87bf618f51b74 --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.2-2025-12-11/2b2b63ce-bd56-402e-bbac-b440924d2140_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9a1c423a51d4350e9a727a336ed619397be2fb70d29ccbe81f64472181ee2b00 +size 14132277 diff --git a/data/swebenchpro/openai/gpt-5.2-2025-12-11/3dd56e91-b511-482f-8b92-14527c8229c3.json b/data/swebenchpro/openai/gpt-5.2-2025-12-11/3dd56e91-b511-482f-8b92-14527c8229c3.json new file mode 100644 index 0000000000000000000000000000000000000000..c742a1ed55dce1d11ae58e95b5c54979e5e29a17 --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.2-2025-12-11/3dd56e91-b511-482f-8b92-14527c8229c3.json @@ -0,0 +1,110 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/openai_gpt-5.2-2025-12-11/1786164381.501596", + "evaluation_timestamp": "1777672101.0", + "retrieved_timestamp": "1786164381.501596", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on swebenchpro/S-adaptive/+3ep/92a36c4d for scorer harbor_scorer", + "source_data": { + "dataset_name": "swebenchpro", + "source_type": "hf_dataset", + "samples_number": 4, + "sample_ids": [ + "instance_nodebb__nodebb-70b4a0e2aebebe8f2f559de6680093d96a697b2f-vnan", + "instance_ansible__ansible-e22e103cdf8edc56ff7d9b848a58f94f1471a263-v1055803c3a812189a1133297f7f5468579283f86", + "instance_qutebrowser__qutebrowser-e15d26630934d0b6415ed2295ac42fd570a57620-va0fd88aac89cde702ec1ba84877234da33adce8a", + "instance_ansible__ansible-164881d871964aa64e0f911d03ae270acbad253c-v390e508d27db7a51eece36bb6d9698b63a5b638a" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1777672101.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 1.0, + "uncertainty": { + "num_samples": 12 + }, + "details": { + "total_matched_trajectories": "12", + "stopping_reason_count_repetition_guard": "12", + "avg_total_tokens_target_model": "3269158.42", + "total_total_tokens_target_model": "39229901.0", + "avg_total_tokens_other_models": "1067.17", + "total_total_tokens_other_models": "12806.0", + "avg_total_tokens_all_models": "3270225.58", + "total_total_tokens_all_models": "39242707.0", + "avg_turn_count": "94.92", + "total_turn_count": "1139.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 30000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/openai/gpt-5.2-2025-12-11/3dd56e91-b511-482f-8b92-14527c8229c3_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "0775c7578e86652a4651c475eb5f74e121bfde6800529bf47090b097aa77ed26", + "total_rows": 12 + } +} \ No newline at end of file diff --git a/data/swebenchpro/openai/gpt-5.2-2025-12-11/3dd56e91-b511-482f-8b92-14527c8229c3_samples.jsonl b/data/swebenchpro/openai/gpt-5.2-2025-12-11/3dd56e91-b511-482f-8b92-14527c8229c3_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..8c585904b92793549486beeba535f88b6867b5fc --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.2-2025-12-11/3dd56e91-b511-482f-8b92-14527c8229c3_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0775c7578e86652a4651c475eb5f74e121bfde6800529bf47090b097aa77ed26 +size 7250749 diff --git a/data/swebenchpro/openai/gpt-5.2-2025-12-11/43859143-c545-4be8-a695-f47a6c75483d.json b/data/swebenchpro/openai/gpt-5.2-2025-12-11/43859143-c545-4be8-a695-f47a6c75483d.json new file mode 100644 index 0000000000000000000000000000000000000000..8bb636dbfbf62200454515428d4c8f1b43346b2b --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.2-2025-12-11/43859143-c545-4be8-a695-f47a6c75483d.json @@ -0,0 +1,119 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/openai_gpt-5.2-2025-12-11/1786164322.835142", + "evaluation_timestamp": "1776804206.0", + "retrieved_timestamp": "1786164322.835142", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on swebenchpro/S-adaptive/+1ep/65735fdf for scorer harbor_scorer", + "source_data": { + "dataset_name": "swebenchpro", + "source_type": "hf_dataset", + "samples_number": 10, + "sample_ids": [ + "instance_flipt-io__flipt-b68b8960b8a08540d5198d78c665a7eb0bea4008", + "instance_flipt-io__flipt-e42da21a07a5ae35835ec54f74004ebd58713874", + "instance_element-hq__element-web-f3534b42df3dcfe36dc48bddbf14034085af6d30-vnan", + "instance_future-architect__vuls-f0b3a8b1db98eb1bd32685f1c36c41a99c3452ed", + "instance_future-architect__vuls-6eff6a9329a65cc412e79b8f82444dfa3d0f0b5a", + "instance_ansible__ansible-b2a289dcbb702003377221e25f62c8a3608f0e89-v173091e2e36d38c978002990795f66cfc0af30ad", + "instance_qutebrowser__qutebrowser-e70f5b03187bdd40e8bf70f5f3ead840f52d1f42-v02ad04386d5238fe2d1a1be450df257370de4b6a", + "instance_protonmail__webclients-cb8cc309c6968b0a2a5fe4288d0ae0a969ff31e1", + "instance_ansible__ansible-11c1777d56664b1acb56b387a1ad6aeadef1391d-v0f01c69f1e2528b935359cfe578530722bca2c59", + "instance_ansible__ansible-29aea9ff3466e4cd2ed00524b9e56738d568ce8b-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1776804206.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.7142857142857143, + "uncertainty": { + "standard_error": { + "value": 0.18442777839082938 + }, + "num_samples": 7 + }, + "details": { + "total_matched_trajectories": "7", + "stopping_reason_count_repetition_guard": "7", + "avg_total_tokens_target_model": "4413343.86", + "total_total_tokens_target_model": "30893407.0", + "avg_total_tokens_other_models": "1035.57", + "total_total_tokens_other_models": "7249.0", + "avg_total_tokens_all_models": "4414379.43", + "total_total_tokens_all_models": "30900656.0", + "avg_turn_count": "117.00", + "total_turn_count": "819.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 30000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/openai/gpt-5.2-2025-12-11/43859143-c545-4be8-a695-f47a6c75483d_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "78291fd69c48f7a1aaea455717249982b214c5b029670b3786761f767bc0ed05", + "total_rows": 7 + } +} \ No newline at end of file diff --git a/data/swebenchpro/openai/gpt-5.2-2025-12-11/43859143-c545-4be8-a695-f47a6c75483d_samples.jsonl b/data/swebenchpro/openai/gpt-5.2-2025-12-11/43859143-c545-4be8-a695-f47a6c75483d_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..b4eb0abae601f505990068d2419878ea65aa41b3 --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.2-2025-12-11/43859143-c545-4be8-a695-f47a6c75483d_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:78291fd69c48f7a1aaea455717249982b214c5b029670b3786761f767bc0ed05 +size 5173114 diff --git a/data/swebenchpro/openai/gpt-5.2-2025-12-11/4c4fc1e9-f7c4-4ff8-a615-e35afdfb96ce.json b/data/swebenchpro/openai/gpt-5.2-2025-12-11/4c4fc1e9-f7c4-4ff8-a615-e35afdfb96ce.json new file mode 100644 index 0000000000000000000000000000000000000000..df920248d027f5d44ac6c73fb7b7fa3e7378825c --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.2-2025-12-11/4c4fc1e9-f7c4-4ff8-a615-e35afdfb96ce.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/openai_gpt-5.2-2025-12-11/1786169025.378544", + "evaluation_timestamp": "1778541330.0", + "retrieved_timestamp": "1786169025.378544", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/openai/gpt-5.2-2025-12-11/4c4fc1e9-f7c4-4ff8-a615-e35afdfb96ce_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "847692f9b0894138e8653fe7c8731c0fbdfb32857ff8689fe957e4c8808238c8", + "total_rows": 51 + } +} \ No newline at end of file diff --git a/data/swebenchpro/openai/gpt-5.2-2025-12-11/4c4fc1e9-f7c4-4ff8-a615-e35afdfb96ce_samples.jsonl b/data/swebenchpro/openai/gpt-5.2-2025-12-11/4c4fc1e9-f7c4-4ff8-a615-e35afdfb96ce_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..268671b6dfe23d91a68ba88dd7aa7e004d98793c --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.2-2025-12-11/4c4fc1e9-f7c4-4ff8-a615-e35afdfb96ce_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:847692f9b0894138e8653fe7c8731c0fbdfb32857ff8689fe957e4c8808238c8 +size 27752697 diff --git a/data/swebenchpro/openai/gpt-5.2-2025-12-11/55a0d421-a10d-4aaa-945e-e8e9b0ba2531.json b/data/swebenchpro/openai/gpt-5.2-2025-12-11/55a0d421-a10d-4aaa-945e-e8e9b0ba2531.json new file mode 100644 index 0000000000000000000000000000000000000000..d13fff8b4151e9a8585ac640a2c54076cfc93932 --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.2-2025-12-11/55a0d421-a10d-4aaa-945e-e8e9b0ba2531.json @@ -0,0 +1,119 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/openai_gpt-5.2-2025-12-11/1786164365.569232", + "evaluation_timestamp": "1777901270.0", + "retrieved_timestamp": "1786164365.569232", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on swebenchpro/S-adaptive/+1ep/f580e732 for scorer harbor_scorer", + "source_data": { + "dataset_name": "swebenchpro", + "source_type": "hf_dataset", + "samples_number": 7, + "sample_ids": [ + "instance_ansible__ansible-502270c804c33d3bc963930dc85e0f4ca359674d-v7eee2454f617569fd6889f2211f75bc02a35f9f8", + "instance_ansible__ansible-811093f0225caa4dd33890933150a81c6a6d5226-v1055803c3a812189a1133297f7f5468579283f86", + "instance_nodebb__nodebb-0f788b8eaa4bba3c142d171fd941d015c53b65fc-v0ec6d6c2baf3cb4797482ce4829bc25cd5716649", + "instance_gravitational__teleport-73cc189b0e9636d418c4470ecce0d9af5dae2f02-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_internetarchive__openlibrary-3c48b4bb782189e0858e6c3fc7956046cf3e1cfb-v2d9a6c849c60ed19fd0858ce9e40b7cc8e097e59", + "instance_element-hq__element-web-9bf77963ee5e036d54b2a3ca202fbf6378464a5e-vnan", + "instance_future-architect__vuls-73f0adad95c4d227e2ccfa876c85cc95dd065e13" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1777901270.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.7142857142857143, + "uncertainty": { + "standard_error": { + "value": 0.18442777839082936 + }, + "num_samples": 7 + }, + "details": { + "total_matched_trajectories": "7", + "stopping_reason_count_repetition_guard": "7", + "avg_total_tokens_target_model": "6476429.86", + "total_total_tokens_target_model": "45335009.0", + "avg_total_tokens_other_models": "1256.00", + "total_total_tokens_other_models": "8792.0", + "avg_total_tokens_all_models": "6477685.86", + "total_total_tokens_all_models": "45343801.0", + "avg_turn_count": "134.86", + "total_turn_count": "944.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 30000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/openai/gpt-5.2-2025-12-11/55a0d421-a10d-4aaa-945e-e8e9b0ba2531_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "f27208969e301f7d6f9763d1a8cf626a6a3e63fb603597fdffe38d1b0f4889e9", + "total_rows": 7 + } +} \ No newline at end of file diff --git a/data/swebenchpro/openai/gpt-5.2-2025-12-11/55a0d421-a10d-4aaa-945e-e8e9b0ba2531_samples.jsonl b/data/swebenchpro/openai/gpt-5.2-2025-12-11/55a0d421-a10d-4aaa-945e-e8e9b0ba2531_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..90186365944531b8f0a551f0bc9664c5e6bdb1cd --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.2-2025-12-11/55a0d421-a10d-4aaa-945e-e8e9b0ba2531_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f27208969e301f7d6f9763d1a8cf626a6a3e63fb603597fdffe38d1b0f4889e9 +size 6803357 diff --git a/data/swebenchpro/openai/gpt-5.2-2025-12-11/6704a49f-3556-4385-b24a-e3c60651e56c.json b/data/swebenchpro/openai/gpt-5.2-2025-12-11/6704a49f-3556-4385-b24a-e3c60651e56c.json new file mode 100644 index 0000000000000000000000000000000000000000..bf06da5175711950d32f0e166f5672683d9362b2 --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.2-2025-12-11/6704a49f-3556-4385-b24a-e3c60651e56c.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/openai_gpt-5.2-2025-12-11/1786167132.998", + "evaluation_timestamp": "1776433475.0", + "retrieved_timestamp": "1786167132.998", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/openai/gpt-5.2-2025-12-11/6704a49f-3556-4385-b24a-e3c60651e56c_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "667c6926e6d89f873aae45cd5bf41877ad7918c8911f8ca427512e5ff3545559", + "total_rows": 50 + } +} \ No newline at end of file diff --git a/data/swebenchpro/openai/gpt-5.2-2025-12-11/6704a49f-3556-4385-b24a-e3c60651e56c_samples.jsonl b/data/swebenchpro/openai/gpt-5.2-2025-12-11/6704a49f-3556-4385-b24a-e3c60651e56c_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..9f6ebf750be0ec08c77138d8edcf81cacd851a41 --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.2-2025-12-11/6704a49f-3556-4385-b24a-e3c60651e56c_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:667c6926e6d89f873aae45cd5bf41877ad7918c8911f8ca427512e5ff3545559 +size 40279123 diff --git a/data/swebenchpro/openai/gpt-5.2-2025-12-11/6d3c43ae-bc18-4b9a-9ccc-44fd5419d27e.json b/data/swebenchpro/openai/gpt-5.2-2025-12-11/6d3c43ae-bc18-4b9a-9ccc-44fd5419d27e.json new file mode 100644 index 0000000000000000000000000000000000000000..71bb7d3c6c0a751e0726300aa69beef95f306c04 --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.2-2025-12-11/6d3c43ae-bc18-4b9a-9ccc-44fd5419d27e.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/openai_gpt-5.2-2025-12-11/1786168996.526732", + "evaluation_timestamp": "1776601454.0", + "retrieved_timestamp": "1786168996.526732", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/openai/gpt-5.2-2025-12-11/6d3c43ae-bc18-4b9a-9ccc-44fd5419d27e_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "6fea99a3c6cab6ed5ad23a36fd02a5f2d9274618d572d9866e323123c053c3d1", + "total_rows": 10 + } +} \ No newline at end of file diff --git a/data/swebenchpro/openai/gpt-5.2-2025-12-11/6d3c43ae-bc18-4b9a-9ccc-44fd5419d27e_samples.jsonl b/data/swebenchpro/openai/gpt-5.2-2025-12-11/6d3c43ae-bc18-4b9a-9ccc-44fd5419d27e_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..9792868367efa4b8dff1e59a543fafe57527f42e --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.2-2025-12-11/6d3c43ae-bc18-4b9a-9ccc-44fd5419d27e_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6fea99a3c6cab6ed5ad23a36fd02a5f2d9274618d572d9866e323123c053c3d1 +size 8035020 diff --git a/data/swebenchpro/openai/gpt-5.2-2025-12-11/6e11997f-0058-4ac2-bcfa-ae38a3f50231.json b/data/swebenchpro/openai/gpt-5.2-2025-12-11/6e11997f-0058-4ac2-bcfa-ae38a3f50231.json new file mode 100644 index 0000000000000000000000000000000000000000..9c899f713ad08f8d4d90f4ad030cb1c2b21613e9 --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.2-2025-12-11/6e11997f-0058-4ac2-bcfa-ae38a3f50231.json @@ -0,0 +1,125 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/openai_gpt-5.2-2025-12-11/1786164409.985019", + "evaluation_timestamp": "1777692145.0", + "retrieved_timestamp": "1786164409.985019", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on swebenchpro/S-adaptive/+4ep/458a6744 for scorer harbor_scorer", + "source_data": { + "dataset_name": "swebenchpro", + "source_type": "hf_dataset", + "samples_number": 15, + "sample_ids": [ + "instance_ansible__ansible-502270c804c33d3bc963930dc85e0f4ca359674d-v7eee2454f617569fd6889f2211f75bc02a35f9f8", + "instance_internetarchive__openlibrary-9c392b60e2c6fa1d68cb68084b4b4ff04d0cb35c-v2d9a6c849c60ed19fd0858ce9e40b7cc8e097e59", + "instance_element-hq__element-web-b007ea81b2ccd001b00f332bee65070aa7fc00f9-vnan", + "instance_ansible__ansible-811093f0225caa4dd33890933150a81c6a6d5226-v1055803c3a812189a1133297f7f5468579283f86", + "instance_flipt-io__flipt-cd2f3b0a9d4d8b8a6d3d56afab65851ecdc408e8", + "instance_nodebb__nodebb-0f788b8eaa4bba3c142d171fd941d015c53b65fc-v0ec6d6c2baf3cb4797482ce4829bc25cd5716649", + "instance_flipt-io__flipt-0b119520afca1cf25c470ff4288c464d4510b944", + "instance_element-hq__element-web-f63160f38459fb552d00fcc60d4064977a9095a6-vnan", + "instance_ansible__ansible-f327e65d11bb905ed9f15996024f857a95592629-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_gravitational__teleport-73cc189b0e9636d418c4470ecce0d9af5dae2f02-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_internetarchive__openlibrary-3c48b4bb782189e0858e6c3fc7956046cf3e1cfb-v2d9a6c849c60ed19fd0858ce9e40b7cc8e097e59", + "instance_future-architect__vuls-878c25bf5a9c9fd88ac32eb843f5636834d5712d", + "instance_element-hq__element-web-9bf77963ee5e036d54b2a3ca202fbf6378464a5e-vnan", + "instance_qutebrowser__qutebrowser-70248f256f93ed9b1984494d0a1a919ddd774892-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_future-architect__vuls-73f0adad95c4d227e2ccfa876c85cc95dd065e13" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1777692145.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.6444444444444444, + "uncertainty": { + "standard_error": { + "value": 0.12372809695177825 + }, + "num_samples": 53 + }, + "details": { + "total_matched_trajectories": "53", + "stopping_reason_count_repetition_guard": "50", + "stopping_reason_count_token_limit": "3", + "avg_total_tokens_target_model": "6646585.68", + "total_total_tokens_target_model": "352269041.0", + "avg_total_tokens_other_models": "1077.91", + "total_total_tokens_other_models": "57129.0", + "avg_total_tokens_all_models": "6647663.58", + "total_total_tokens_all_models": "352326170.0", + "avg_turn_count": "132.89", + "total_turn_count": "7043.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 30000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/openai/gpt-5.2-2025-12-11/6e11997f-0058-4ac2-bcfa-ae38a3f50231_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "4b57613962063cb1235b060d4825b2f36311fd539ea10e7377502d5988740881", + "total_rows": 53 + } +} \ No newline at end of file diff --git a/data/swebenchpro/openai/gpt-5.2-2025-12-11/6e11997f-0058-4ac2-bcfa-ae38a3f50231_samples.jsonl b/data/swebenchpro/openai/gpt-5.2-2025-12-11/6e11997f-0058-4ac2-bcfa-ae38a3f50231_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..e22c321089d2ba1d73ce5fa5cc18c479fb411bbb --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.2-2025-12-11/6e11997f-0058-4ac2-bcfa-ae38a3f50231_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4b57613962063cb1235b060d4825b2f36311fd539ea10e7377502d5988740881 +size 49377218 diff --git a/data/swebenchpro/openai/gpt-5.2-2025-12-11/71c2e24e-a41d-4d34-855b-fa9639df2ea4.json b/data/swebenchpro/openai/gpt-5.2-2025-12-11/71c2e24e-a41d-4d34-855b-fa9639df2ea4.json new file mode 100644 index 0000000000000000000000000000000000000000..9b348203450ef21016510791bfde8e4f992878c8 --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.2-2025-12-11/71c2e24e-a41d-4d34-855b-fa9639df2ea4.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/openai_gpt-5.2-2025-12-11/1786168999.586868", + "evaluation_timestamp": "1778665168.0", + "retrieved_timestamp": "1786168999.586868", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/openai/gpt-5.2-2025-12-11/71c2e24e-a41d-4d34-855b-fa9639df2ea4_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "9d18f04bb972e2f765648b9d722e2e22d52063912b10da3104a9fa6963245e40", + "total_rows": 1 + } +} \ No newline at end of file diff --git a/data/swebenchpro/openai/gpt-5.2-2025-12-11/71c2e24e-a41d-4d34-855b-fa9639df2ea4_samples.jsonl b/data/swebenchpro/openai/gpt-5.2-2025-12-11/71c2e24e-a41d-4d34-855b-fa9639df2ea4_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..5943e8bcd0e8eeef055054a463ace25b3cade0ac --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.2-2025-12-11/71c2e24e-a41d-4d34-855b-fa9639df2ea4_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9d18f04bb972e2f765648b9d722e2e22d52063912b10da3104a9fa6963245e40 +size 307953 diff --git a/data/swebenchpro/openai/gpt-5.2-2025-12-11/779ca895-d116-4c0c-a3bb-e590a6af4647.json b/data/swebenchpro/openai/gpt-5.2-2025-12-11/779ca895-d116-4c0c-a3bb-e590a6af4647.json new file mode 100644 index 0000000000000000000000000000000000000000..e0fd1e179e22ca23fd871ec8f9291ac681b13d0d --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.2-2025-12-11/779ca895-d116-4c0c-a3bb-e590a6af4647.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/openai_gpt-5.2-2025-12-11/1786167093.692096", + "evaluation_timestamp": "1780423668.0", + "retrieved_timestamp": "1786167093.692096", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/openai/gpt-5.2-2025-12-11/779ca895-d116-4c0c-a3bb-e590a6af4647_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "564590d7170aa62b29339f8f5123748ba107c7e8bcfc9eb1a9cba477b9a2c142", + "total_rows": 3 + } +} \ No newline at end of file diff --git a/data/swebenchpro/openai/gpt-5.2-2025-12-11/779ca895-d116-4c0c-a3bb-e590a6af4647_samples.jsonl b/data/swebenchpro/openai/gpt-5.2-2025-12-11/779ca895-d116-4c0c-a3bb-e590a6af4647_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..0bad8458be9b1afa4cb42dccaa67a5666c49cd84 --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.2-2025-12-11/779ca895-d116-4c0c-a3bb-e590a6af4647_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:564590d7170aa62b29339f8f5123748ba107c7e8bcfc9eb1a9cba477b9a2c142 +size 1021791 diff --git a/data/swebenchpro/openai/gpt-5.2-2025-12-11/7e593408-c6cb-477d-9172-601d2c0cd210.json b/data/swebenchpro/openai/gpt-5.2-2025-12-11/7e593408-c6cb-477d-9172-601d2c0cd210.json new file mode 100644 index 0000000000000000000000000000000000000000..640a8cf1cd72c061f495a501a397f28645e1775b --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.2-2025-12-11/7e593408-c6cb-477d-9172-601d2c0cd210.json @@ -0,0 +1,133 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/openai_gpt-5.2-2025-12-11/1786162874.303256", + "evaluation_timestamp": "1778854176.0", + "retrieved_timestamp": "1786162874.303256", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on swebenchpro/S-adaptive+C/+2ep/e6274898 for scorer harbor_scorer", + "source_data": { + "dataset_name": "swebenchpro", + "source_type": "hf_dataset", + "samples_number": 20, + "sample_ids": [ + "instance_ansible__ansible-811093f0225caa4dd33890933150a81c6a6d5226-v1055803c3a812189a1133297f7f5468579283f86", + "instance_future-architect__vuls-73f0adad95c4d227e2ccfa876c85cc95dd065e13", + "instance_tutao__tutanota-4b4e45949096bb288f2b522f657610e480efa3e8-vee878bb72091875e912c52fc32bc60ec3760227b", + "instance_qutebrowser__qutebrowser-2dd8966fdcf11972062c540b7a787e4d0de8d372-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_qutebrowser__qutebrowser-ef5ba1a0360b39f9eff027fbdc57f363597c3c3b-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_protonmail__webclients-8afd9ce04c8dde9e150e1c2b50d32e7ee2efa3e7", + "instance_gravitational__teleport-2bb3bbbd8aff1164a2353381cb79e1dc93b90d28-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_gravitational__teleport-46aa81b1ce96ebb4ebed2ae53fd78cd44a05da6c-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_tutao__tutanota-8513a9e8114a8b42e64f4348335e0f23efa054c4-vee878bb72091875e912c52fc32bc60ec3760227b", + "instance_internetarchive__openlibrary-7f6b722a10f822171501d027cad60afe53337732-ve8c8d62a2b60610a3c4631f5f23ed866bada9818", + "instance_nodebb__nodebb-a917210c5b2c20637094545401f85783905c074c-vf2cf3cbd463b7ad942381f1c6d077626485a1e9e", + "instance_internetarchive__openlibrary-1351c59fd43689753de1fca32c78d539a116ffc1-v29f82c9cf21d57b242f8d8b0e541525d259e2d63", + "instance_future-architect__vuls-999529a05b202b0fd29c6fca5039a4c47a3766bb", + "instance_element-hq__element-web-1077729a19c0ce902e713cf6fab42c91fb7907f1-vnan", + "instance_future-architect__vuls-be7b9114cc9545e68fb0ee7bc63d7ec53d1a00ad", + "instance_qutebrowser__qutebrowser-c580ebf0801e5a3ecabc54f327498bb753c6d5f2-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_future-architect__vuls-aaea15e516ece43978cf98e09e52080478b1d39f", + "instance_qutebrowser__qutebrowser-cf06f4e3708f886032d4d2a30108c2fddb042d81-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_internetarchive__openlibrary-3f580a5f244c299d936d73d9e327ba873b6401d9-v0f5aece3601a5b4419f7ccec1dbda2071be28ee4", + "instance_qutebrowser__qutebrowser-0aa57e4f7243024fa4bba8853306691b5dbd77b3-v5149fcda2a9a6fe1d35dfed1bade1444a11ef271" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1778854176.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.7105263157894737, + "uncertainty": { + "standard_error": { + "value": 0.09609167675529229 + }, + "num_samples": 35 + }, + "details": { + "total_matched_trajectories": "35", + "stopping_reason_count_completed_on_successful_submit": "24", + "stopping_reason_count_repetition_guard": "11", + "avg_total_tokens_target_model": "2568417.31", + "total_total_tokens_target_model": "89894606.0", + "avg_total_tokens_other_models": "325.91", + "total_total_tokens_other_models": "11407.0", + "avg_total_tokens_all_models": "2568743.23", + "total_total_tokens_all_models": "89906013.0", + "avg_turn_count": "74.54", + "total_turn_count": "2609.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "8", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 30000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "8", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/openai/gpt-5.2-2025-12-11/7e593408-c6cb-477d-9172-601d2c0cd210_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "fea8fa38bac16553b2152ed0b0291999734b96d931469480fdec1758eb965556", + "total_rows": 35 + } +} \ No newline at end of file diff --git a/data/swebenchpro/openai/gpt-5.2-2025-12-11/7e593408-c6cb-477d-9172-601d2c0cd210_samples.jsonl b/data/swebenchpro/openai/gpt-5.2-2025-12-11/7e593408-c6cb-477d-9172-601d2c0cd210_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..9be5573750b5495a9679e00980a76bd0eabbd39d --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.2-2025-12-11/7e593408-c6cb-477d-9172-601d2c0cd210_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fea8fa38bac16553b2152ed0b0291999734b96d931469480fdec1758eb965556 +size 17604032 diff --git a/data/swebenchpro/openai/gpt-5.2-2025-12-11/87de816a-ce96-4985-b6e1-67a24d33c773.json b/data/swebenchpro/openai/gpt-5.2-2025-12-11/87de816a-ce96-4985-b6e1-67a24d33c773.json new file mode 100644 index 0000000000000000000000000000000000000000..434935b8322f2ac545fe6f740f6b7286991344a5 --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.2-2025-12-11/87de816a-ce96-4985-b6e1-67a24d33c773.json @@ -0,0 +1,110 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/openai_gpt-5.2-2025-12-11/1786164337.544945", + "evaluation_timestamp": "1778678411.0", + "retrieved_timestamp": "1786164337.544945", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on swebenchpro/S-adaptive/+1ep/bb130bed for scorer harbor_scorer", + "source_data": { + "dataset_name": "swebenchpro", + "source_type": "hf_dataset", + "samples_number": 1, + "sample_ids": [ + "instance_element-hq__element-web-1077729a19c0ce902e713cf6fab42c91fb7907f1-vnan" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1778678411.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 1.0, + "uncertainty": { + "num_samples": 1 + }, + "details": { + "total_matched_trajectories": "1", + "stopping_reason_count_repetition_guard": "1", + "avg_total_tokens_target_model": "2849542.00", + "total_total_tokens_target_model": "2849542.0", + "avg_total_tokens_other_models": "1382.00", + "total_total_tokens_other_models": "1382.0", + "avg_total_tokens_all_models": "2850924.00", + "total_total_tokens_all_models": "2850924.0", + "avg_turn_count": "69.00", + "total_turn_count": "69.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 30000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/openai/gpt-5.2-2025-12-11/87de816a-ce96-4985-b6e1-67a24d33c773_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "be8dfa88513d99aabd1571d24efdaddc88bd2ae691c5a841e7976dfac30ad8c9", + "total_rows": 1 + } +} \ No newline at end of file diff --git a/data/swebenchpro/openai/gpt-5.2-2025-12-11/87de816a-ce96-4985-b6e1-67a24d33c773_samples.jsonl b/data/swebenchpro/openai/gpt-5.2-2025-12-11/87de816a-ce96-4985-b6e1-67a24d33c773_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..e517010ba5c693a8525aafc051d802dcbd2d946a --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.2-2025-12-11/87de816a-ce96-4985-b6e1-67a24d33c773_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:be8dfa88513d99aabd1571d24efdaddc88bd2ae691c5a841e7976dfac30ad8c9 +size 763502 diff --git a/data/swebenchpro/openai/gpt-5.2-2025-12-11/8f958c67-48d4-478c-a3f4-8ca2ec5fa9af.json b/data/swebenchpro/openai/gpt-5.2-2025-12-11/8f958c67-48d4-478c-a3f4-8ca2ec5fa9af.json new file mode 100644 index 0000000000000000000000000000000000000000..4e41a7fd07046af8c3186caef956d1874e38e2a2 --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.2-2025-12-11/8f958c67-48d4-478c-a3f4-8ca2ec5fa9af.json @@ -0,0 +1,110 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/openai_gpt-5.2-2025-12-11/1786162868.686594", + "evaluation_timestamp": "1778703021.0", + "retrieved_timestamp": "1786162868.686594", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on swebenchpro/S-adaptive+C/+2ep/aacb763d for scorer harbor_scorer", + "source_data": { + "dataset_name": "swebenchpro", + "source_type": "hf_dataset", + "samples_number": 1, + "sample_ids": [ + "instance_qutebrowser__qutebrowser-2e961080a85d660148937ee8f0f6b3445a8f2c01-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1778703021.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 1.0, + "uncertainty": { + "num_samples": 2 + }, + "details": { + "total_matched_trajectories": "2", + "stopping_reason_count_completed_on_successful_submit": "2", + "avg_total_tokens_target_model": "1302730.50", + "total_total_tokens_target_model": "2605461.0", + "avg_total_tokens_other_models": "0.00", + "total_total_tokens_other_models": "0.0", + "avg_total_tokens_all_models": "1302730.50", + "total_total_tokens_all_models": "2605461.0", + "avg_turn_count": "59.00", + "total_turn_count": "118.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "8", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 30000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "8", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/openai/gpt-5.2-2025-12-11/8f958c67-48d4-478c-a3f4-8ca2ec5fa9af_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "4e4283bc7fc10769109534654eb9ff492eafd0fe5e633de03172988a930a1973", + "total_rows": 2 + } +} \ No newline at end of file diff --git a/data/swebenchpro/openai/gpt-5.2-2025-12-11/8f958c67-48d4-478c-a3f4-8ca2ec5fa9af_samples.jsonl b/data/swebenchpro/openai/gpt-5.2-2025-12-11/8f958c67-48d4-478c-a3f4-8ca2ec5fa9af_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..7ae3c768500421a74b5191f48e3c86db1ecd3828 --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.2-2025-12-11/8f958c67-48d4-478c-a3f4-8ca2ec5fa9af_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4e4283bc7fc10769109534654eb9ff492eafd0fe5e633de03172988a930a1973 +size 652805 diff --git a/data/swebenchpro/openai/gpt-5.2-2025-12-11/90d98ab8-0d32-48ed-a088-2b9d0743917c.json b/data/swebenchpro/openai/gpt-5.2-2025-12-11/90d98ab8-0d32-48ed-a088-2b9d0743917c.json new file mode 100644 index 0000000000000000000000000000000000000000..f3301480a09ccbb0bf3fc37ce71a77aaa69b2cd8 --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.2-2025-12-11/90d98ab8-0d32-48ed-a088-2b9d0743917c.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/openai_gpt-5.2-2025-12-11/1786169015.22285", + "evaluation_timestamp": "1777620583.0", + "retrieved_timestamp": "1786169015.22285", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/openai/gpt-5.2-2025-12-11/90d98ab8-0d32-48ed-a088-2b9d0743917c_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "4e98903f21635917e5effd4b62538af92f14ead89425bb0d4e4d9abb71ae7ff6", + "total_rows": 4 + } +} \ No newline at end of file diff --git a/data/swebenchpro/openai/gpt-5.2-2025-12-11/90d98ab8-0d32-48ed-a088-2b9d0743917c_samples.jsonl b/data/swebenchpro/openai/gpt-5.2-2025-12-11/90d98ab8-0d32-48ed-a088-2b9d0743917c_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..c81314c9c6e6e94dbe1666457d7aa01e82a68a29 --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.2-2025-12-11/90d98ab8-0d32-48ed-a088-2b9d0743917c_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4e98903f21635917e5effd4b62538af92f14ead89425bb0d4e4d9abb71ae7ff6 +size 2486242 diff --git a/data/swebenchpro/openai/gpt-5.2-2025-12-11/924021a5-c2cb-46ac-9f07-7028122eefce.json b/data/swebenchpro/openai/gpt-5.2-2025-12-11/924021a5-c2cb-46ac-9f07-7028122eefce.json new file mode 100644 index 0000000000000000000000000000000000000000..2a0141536a62d807b95c64d7fe7b5499981d2e3d --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.2-2025-12-11/924021a5-c2cb-46ac-9f07-7028122eefce.json @@ -0,0 +1,112 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/openai_gpt-5.2-2025-12-11/1786164361.364863", + "evaluation_timestamp": "1776883764.0", + "retrieved_timestamp": "1786164361.364863", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on swebenchpro/S-adaptive/+1ep/c0771904 for scorer harbor_scorer", + "source_data": { + "dataset_name": "swebenchpro", + "source_type": "hf_dataset", + "samples_number": 3, + "sample_ids": [ + "instance_future-architect__vuls-f0b3a8b1db98eb1bd32685f1c36c41a99c3452ed", + "instance_future-architect__vuls-6eff6a9329a65cc412e79b8f82444dfa3d0f0b5a", + "instance_ansible__ansible-11c1777d56664b1acb56b387a1ad6aeadef1391d-v0f01c69f1e2528b935359cfe578530722bca2c59" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1776883764.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.6666666666666666, + "uncertainty": { + "standard_error": { + "value": 0.33333333333333337 + }, + "num_samples": 3 + }, + "details": { + "total_matched_trajectories": "3", + "stopping_reason_count_repetition_guard": "3", + "avg_total_tokens_target_model": "7436034.33", + "total_total_tokens_target_model": "22308103.0", + "avg_total_tokens_other_models": "1138.00", + "total_total_tokens_other_models": "3414.0", + "avg_total_tokens_all_models": "7437172.33", + "total_total_tokens_all_models": "22311517.0", + "avg_turn_count": "147.00", + "total_turn_count": "441.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 30000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/openai/gpt-5.2-2025-12-11/924021a5-c2cb-46ac-9f07-7028122eefce_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "620b00ed46ab30921df69df231cc970cac614b7a22bb7efe4961a4e7e3d84df3", + "total_rows": 3 + } +} \ No newline at end of file diff --git a/data/swebenchpro/openai/gpt-5.2-2025-12-11/924021a5-c2cb-46ac-9f07-7028122eefce_samples.jsonl b/data/swebenchpro/openai/gpt-5.2-2025-12-11/924021a5-c2cb-46ac-9f07-7028122eefce_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..f5ee3e00cba69091f82c5e0a0c091a3af80466c6 --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.2-2025-12-11/924021a5-c2cb-46ac-9f07-7028122eefce_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:620b00ed46ab30921df69df231cc970cac614b7a22bb7efe4961a4e7e3d84df3 +size 2420108 diff --git a/data/swebenchpro/openai/gpt-5.2-2025-12-11/932ce350-7cc1-46b8-b1d9-e2aefdabaf0e.json b/data/swebenchpro/openai/gpt-5.2-2025-12-11/932ce350-7cc1-46b8-b1d9-e2aefdabaf0e.json new file mode 100644 index 0000000000000000000000000000000000000000..435595f67d8fb3394caeeaecc5b1494f36354d9d --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.2-2025-12-11/932ce350-7cc1-46b8-b1d9-e2aefdabaf0e.json @@ -0,0 +1,116 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/openai_gpt-5.2-2025-12-11/1786164374.298276", + "evaluation_timestamp": "1776696546.0", + "retrieved_timestamp": "1786164374.298276", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on swebenchpro/S-adaptive/+2ep/6528bea6 for scorer harbor_scorer", + "source_data": { + "dataset_name": "swebenchpro", + "source_type": "hf_dataset", + "samples_number": 7, + "sample_ids": [ + "instance_flipt-io__flipt-b68b8960b8a08540d5198d78c665a7eb0bea4008", + "instance_qutebrowser__qutebrowser-01d1d1494411380d97cac14614a829d3a69cecaf-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_flipt-io__flipt-e42da21a07a5ae35835ec54f74004ebd58713874", + "instance_element-hq__element-web-f3534b42df3dcfe36dc48bddbf14034085af6d30-vnan", + "instance_navidrome__navidrome-d8e794317f788198227e10fb667e10496b3eb99a", + "instance_future-architect__vuls-f0b3a8b1db98eb1bd32685f1c36c41a99c3452ed", + "instance_future-architect__vuls-6eff6a9329a65cc412e79b8f82444dfa3d0f0b5a" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1776696546.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.2857142857142857, + "uncertainty": { + "standard_error": { + "value": 0.18442777839082938 + }, + "num_samples": 9 + }, + "details": { + "total_matched_trajectories": "9", + "stopping_reason_count_repetition_guard": "9", + "avg_total_tokens_target_model": "5306307.78", + "total_total_tokens_target_model": "47756770.0", + "avg_total_tokens_other_models": "1219.22", + "total_total_tokens_other_models": "10973.0", + "avg_total_tokens_all_models": "5307527.00", + "total_total_tokens_all_models": "47767743.0", + "avg_turn_count": "123.89", + "total_turn_count": "1115.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "3600", + "attempt_timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 30000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/openai/gpt-5.2-2025-12-11/932ce350-7cc1-46b8-b1d9-e2aefdabaf0e_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "48d4461320f1e431a1e91bf6fcd869ce16707e08aa99f930aa525f5cbcfd652e", + "total_rows": 9 + } +} \ No newline at end of file diff --git a/data/swebenchpro/openai/gpt-5.2-2025-12-11/932ce350-7cc1-46b8-b1d9-e2aefdabaf0e_samples.jsonl b/data/swebenchpro/openai/gpt-5.2-2025-12-11/932ce350-7cc1-46b8-b1d9-e2aefdabaf0e_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..c405a92f99c3c4a944aadf613a151980eab6089e --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.2-2025-12-11/932ce350-7cc1-46b8-b1d9-e2aefdabaf0e_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:48d4461320f1e431a1e91bf6fcd869ce16707e08aa99f930aa525f5cbcfd652e +size 7130576 diff --git a/data/swebenchpro/openai/gpt-5.2-2025-12-11/a12bcdc1-d777-469f-8a2a-fb4658644728.json b/data/swebenchpro/openai/gpt-5.2-2025-12-11/a12bcdc1-d777-469f-8a2a-fb4658644728.json new file mode 100644 index 0000000000000000000000000000000000000000..c3eaf1881ba385aca3e8b522ec6642c4f7310198 --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.2-2025-12-11/a12bcdc1-d777-469f-8a2a-fb4658644728.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/openai_gpt-5.2-2025-12-11/1786167090.961701", + "evaluation_timestamp": "1778584694.0", + "retrieved_timestamp": "1786167090.961701", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/openai/gpt-5.2-2025-12-11/a12bcdc1-d777-469f-8a2a-fb4658644728_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "03c22d9ee938661a37e81011a2496aaa64896437c8d313356740c8dd4af715b3", + "total_rows": 3 + } +} \ No newline at end of file diff --git a/data/swebenchpro/openai/gpt-5.2-2025-12-11/a12bcdc1-d777-469f-8a2a-fb4658644728_samples.jsonl b/data/swebenchpro/openai/gpt-5.2-2025-12-11/a12bcdc1-d777-469f-8a2a-fb4658644728_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..631acc3480e47c8b2adf38b7ea2338c352d7f7cd --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.2-2025-12-11/a12bcdc1-d777-469f-8a2a-fb4658644728_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:03c22d9ee938661a37e81011a2496aaa64896437c8d313356740c8dd4af715b3 +size 2303943 diff --git a/data/swebenchpro/openai/gpt-5.2-2025-12-11/a9957010-d5fc-4855-81ac-94b90ef0e832.json b/data/swebenchpro/openai/gpt-5.2-2025-12-11/a9957010-d5fc-4855-81ac-94b90ef0e832.json new file mode 100644 index 0000000000000000000000000000000000000000..72ddf7315d81066212fb29db460b6fd5fad3e928 --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.2-2025-12-11/a9957010-d5fc-4855-81ac-94b90ef0e832.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/openai_gpt-5.2-2025-12-11/1786169030.812218", + "evaluation_timestamp": "1776651987.0", + "retrieved_timestamp": "1786169030.812218", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/openai/gpt-5.2-2025-12-11/a9957010-d5fc-4855-81ac-94b90ef0e832_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "bc87addd2e1c73cb94965326b0a33c682c9eb8c76e7a616d505dbef0550ba852", + "total_rows": 9 + } +} \ No newline at end of file diff --git a/data/swebenchpro/openai/gpt-5.2-2025-12-11/a9957010-d5fc-4855-81ac-94b90ef0e832_samples.jsonl b/data/swebenchpro/openai/gpt-5.2-2025-12-11/a9957010-d5fc-4855-81ac-94b90ef0e832_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..e27725b52b360b6cfaade24da08e586aa56e6a47 --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.2-2025-12-11/a9957010-d5fc-4855-81ac-94b90ef0e832_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bc87addd2e1c73cb94965326b0a33c682c9eb8c76e7a616d505dbef0550ba852 +size 8820929 diff --git a/data/swebenchpro/openai/gpt-5.2-2025-12-11/b13227be-5ac9-4052-a5a1-804bf5166b96.json b/data/swebenchpro/openai/gpt-5.2-2025-12-11/b13227be-5ac9-4052-a5a1-804bf5166b96.json new file mode 100644 index 0000000000000000000000000000000000000000..d8407bbb01c972f13f77e43f770553dafd7094d1 --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.2-2025-12-11/b13227be-5ac9-4052-a5a1-804bf5166b96.json @@ -0,0 +1,291 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/openai_gpt-5.2-2025-12-11/1786182652.494316", + "evaluation_timestamp": "1775602861.0", + "retrieved_timestamp": "1786182652.494316", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.196.dev25+gb973b316,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on eval_task for scorer harbor_scorer", + "source_data": { + "dataset_name": "swebenchpro", + "source_type": "hf_dataset", + "samples_number": 731, + "sample_ids": [ + "instance_flipt-io__flipt-e91615cf07966da41756017a7d571f9fc0fdbe80", + "instance_ansible__ansible-8127abbc298cabf04aaa89a478fc5e5e3432a6fc-v30a923fb5c164d6cd18280c02422f75e611e8fb2", + "instance_protonmail__webclients-da91f084c0f532d9cc8ca385a701274d598057b8", + "instance_qutebrowser__qutebrowser-deeb15d6f009b3ca0c3bd503a7cef07462bd16b4-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_internetarchive__openlibrary-5069b09e5f64428dce59b33455c8bb17fe577070-v8717e18970bcdc4e0d2cea3b1527752b21e74866", + "instance_protonmail__webclients-c5a2089ca2bfe9aa1d85a664b8ad87ef843a1c9c", + "instance_qutebrowser__qutebrowser-fd6790fe8c02b144ab2464f1fc8ab3d02ce3c476-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_navidrome__navidrome-eebfbc5381a1e506ff17b5f1371d1ad83d5fd642", + "instance_navidrome__navidrome-812dc2090f20ac4f8ac271b6ed95be5889d1a3ca", + "instance_nodebb__nodebb-0c81642997ea1d827dbd02c311db9d4976112cd4-vf2cf3cbd463b7ad942381f1c6d077626485a1e9e", + "instance_internetarchive__openlibrary-c506c1b0b678892af5cb22c1c1dbc35d96787a0a-v0f5aece3601a5b4419f7ccec1dbda2071be28ee4", + "instance_flipt-io__flipt-5c7037ececb0bead0a8eb56054e224bcd7ac5922", + "instance_qutebrowser__qutebrowser-01d1d1494411380d97cac14614a829d3a69cecaf-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_internetarchive__openlibrary-7f7e53aa4cf74a4f8549a5bcd4810c527e2f6d7e-v13642507b4fc1f8d234172bf8129942da2c2ca26", + "instance_internetarchive__openlibrary-53e02a22972e9253aeded0e1981e6845e1e521fe-vfa6ff903cb27f336e17654595dd900fa943dcd91", + "instance_ansible__ansible-34db57a47f875d11c4068567b9ec7ace174ec4cf-v1055803c3a812189a1133297f7f5468579283f86", + "instance_qutebrowser__qutebrowser-996487c43e4fcc265b541f9eca1e7930e3c5cf05-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_gravitational__teleport-96019ce0be7a2c8e36363f359eb7c943b41dde70", + "instance_protonmail__webclients-df60460f163fd5c34e844ab9015e3176f1ab1ac0", + "instance_future-architect__vuls-ad2edbb8448e2c41a097f1c0b52696c0f6c5924d", + "instance_ansible__ansible-9142be2f6cabbe6597c9254c5bb9186d17036d55-v0f01c69f1e2528b935359cfe578530722bca2c59", + "instance_element-hq__element-web-33299af5c9b7a7ec5a9c31d578d4ec5b18088fb7-vnan", + "instance_element-hq__element-web-772df3021201d9c73835a626df8dcb6334ad9a3e-vnan", + "instance_internetarchive__openlibrary-1be7de788a444f6255e89c10ef6aa608550604a8-v29f82c9cf21d57b242f8d8b0e541525d259e2d63", + "instance_internetarchive__openlibrary-630221ab686c64e75a2ce253c893c033e4814b2e-v93c53c13d5f9b383ebb411ee7750b49dcd1a34c6", + "instance_gravitational__teleport-f432a71a13e698b6e1c4672a2e9e9c1f32d35c12", + "instance_protonmail__webclients-0d0267c4438cf378bda90bc85eed3a3615871ac4", + "instance_internetarchive__openlibrary-dbbd9d539c6d4fd45d5be9662aa19b6d664b5137-v08d8e8889ec945ab821fb156c04c7d2e2810debb", + "instance_flipt-io__flipt-b2170346dc37cf42fda1386cd630f24821ad2ac5", + "instance_element-hq__element-web-aec454dd6feeb93000380523cbb0b3681c0275fd-vnan", + "instance_internetarchive__openlibrary-4a5d2a7d24c9e4c11d3069220c0685b736d5ecde-v13642507b4fc1f8d234172bf8129942da2c2ca26", + "instance_qutebrowser__qutebrowser-ec2dcfce9eee9f808efc17a1b99e227fc4421dea-v5149fcda2a9a6fe1d35dfed1bade1444a11ef271", + "instance_gravitational__teleport-1330415d33a27594c948a36d9d7701f496229e9f", + "instance_protonmail__webclients-3a6790f480309130b5d6332dce6c9d5ccca13ee3", + "instance_flipt-io__flipt-492cc0b158200089dceede3b1aba0ed28df3fb1d", + "instance_future-architect__vuls-edb324c3d9ec3b107bf947f00e38af99d05b3e16", + "instance_ansible__ansible-bec27fb4c0a40c5f8bbcf26a475704227d65ee73-v30a923fb5c164d6cd18280c02422f75e611e8fb2", + "instance_gravitational__teleport-c1b1c6a1541c478d7777a48fca993cc8206c73b9", + "instance_tutao__tutanota-1e516e989b3c0221f4af6b297d9c0e4c43e4adc3-vbc0d9ba8f0071fbe982809910959a6ff8884dbbf", + "instance_flipt-io__flipt-406f9396ad65696d58865b3a6283109cd4eaf40e", + "instance_future-architect__vuls-139f3a81b66c47e6d8f70ce6c4afe7a9196a6ea8", + "instance_flipt-io__flipt-3b2c25ee8a3ac247c3fad13ad8d64ace34ec8ee7", + "instance_gravitational__teleport-0ac7334939981cf85b9591ac295c3816954e287e", + "instance_qutebrowser__qutebrowser-e15d26630934d0b6415ed2295ac42fd570a57620-va0fd88aac89cde702ec1ba84877234da33adce8a", + "instance_ansible__ansible-5e88cd9972f10b66dd97e1ee684c910c6a2dd25e-v906c969b551b346ef54a2c0b41e04f632b7b73c2", + "instance_element-hq__element-web-923ad4323b2006b2b180544429455ffe7d4a6cc3-vnan", + "instance_protonmail__webclients-4feccbc9990980aee26ea29035f8f931d6089895", + "instance_flipt-io__flipt-a0cbc0cb65ae601270bdbe3f5313e2dfd49c80e4", + "instance_gravitational__teleport-b4e7cd3a5e246736d3fe8d6886af55030b232277", + "instance_ansible__ansible-6cc97447aac5816745278f3735af128afb255c81-v0f01c69f1e2528b935359cfe578530722bca2c59", + "instance_ansible__ansible-f327e65d11bb905ed9f15996024f857a95592629-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_qutebrowser__qutebrowser-5fdc83e5da6222fe61163395baaad7ae57fa2cb4-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_navidrome__navidrome-3982ba725883e71d4e3e618c61d5140eeb8d850a", + "instance_qutebrowser__qutebrowser-ebfe9b7aa0c4ba9d451f993e08955004aaec4345-v059c6fdc75567943479b23ebca7c07b5e9a7f34c", + "instance_ansible__ansible-709484969c8a4ffd74b839a673431a8c5caa6457-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_gravitational__teleport-ad41b3c15414b28a6cec8c25424a19bfa7abd0e9-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_protonmail__webclients-2c3559cad02d1090985dba7e8eb5a129144d9811", + "instance_flipt-io__flipt-5af0757e96dec4962a076376d1bedc79de0d4249", + "instance_element-hq__element-web-56c7fc1948923b4b3f3507799e725ac16bcf8018-vnan", + "instance_flipt-io__flipt-381b90f718435c4694380b5fcd0d5cf8e3b5a25a", + "instance_flipt-io__flipt-6fd0f9e2587f14ac1fdd1c229f0bcae0468c8daa", + "instance_ansible__ansible-4c5ce5a1a9e79a845aff4978cfeb72a0d4ecf7d6-v1055803c3a812189a1133297f7f5468579283f86", + "instance_navidrome__navidrome-7073d18b54da7e53274d11c9e2baef1242e8769e", + "instance_internetarchive__openlibrary-0a90f9f0256e4f933523e9842799e39f95ae29ce-v76304ecdb3a5954fcf13feb710e8c40fcf24b73c", + "instance_nodebb__nodebb-be43cd25974681c9743d424238b7536c357dc8d3-vf2cf3cbd463b7ad942381f1c6d077626485a1e9e", + "instance_flipt-io__flipt-05d7234fa582df632f70a7cd10194d61bd7043b9", + "instance_nodebb__nodebb-76c6e30282906ac664f2c9278fc90999b27b1f48-vd59a5728dfc977f44533186ace531248c2917516", + "instance_qutebrowser__qutebrowser-6b320dc18662580e1313d2548fdd6231d2a97e6d-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_future-architect__vuls-01441351c3407abfc21c48a38e28828e1b504e0c", + "instance_ansible__ansible-1a4644ff15355fd696ac5b9d074a566a80fe7ca3-v30a923fb5c164d6cd18280c02422f75e611e8fb2", + "instance_navidrome__navidrome-09ae41a2da66264c60ef307882362d2e2d8d8b89", + "instance_internetarchive__openlibrary-5c6c22f3d2edf2f1b10f5dc335e32cb6a5f40341-v76304ecdb3a5954fcf13feb710e8c40fcf24b73c", + "instance_navidrome__navidrome-bf2bcb12799b21069f137749e0c331f761d1f693", + "instance_flipt-io__flipt-2ce8a0331e8a8f63f2c1b555db8277ffe5aa2e63", + "instance_ansible__ansible-a1569ea4ca6af5480cf0b7b3135f5e12add28a44-v0f01c69f1e2528b935359cfe578530722bca2c59", + "instance_gravitational__teleport-82185f232ae8974258397e121b3bc2ed0c3729ed-v626ec2a48416b10a88641359a169d99e935ff037", + "instance_flipt-io__flipt-72d06db14d58692bfb4d07b1aa745a37b35956f3", + "instance_ansible__ansible-b6290e1d156af608bd79118d209a64a051c55001-v390e508d27db7a51eece36bb6d9698b63a5b638a", + "instance_gravitational__teleport-32bcd71591c234f0d8b091ec01f1f5cbfdc0f13c-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_ansible__ansible-a02e22e902a69aeb465f16bf03f7f5a91b2cb828-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_gravitational__teleport-eda668c30d9d3b56d9c69197b120b01013611186", + "instance_nodebb__nodebb-0f788b8eaa4bba3c142d171fd941d015c53b65fc-v0ec6d6c2baf3cb4797482ce4829bc25cd5716649", + "instance_gravitational__teleport-47530e1fd8bfb84ec096ebcbbc29990f30829655-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_qutebrowser__qutebrowser-e64622cd2df5b521342cf4a62e0d4cb8f8c9ae5a-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_element-hq__element-web-494d9de6f0a94ffb491e74744d2735bce02dc0ab-vnan", + "instance_element-hq__element-web-9bf77963ee5e036d54b2a3ca202fbf6378464a5e-vnan", + "instance_nodebb__nodebb-9c576a0758690f45a6ca03b5884c601e473bf2c1-vd59a5728dfc977f44533186ace531248c2917516", + "instance_internetarchive__openlibrary-25858f9f0c165df25742acf8309ce909773f0cdd-v13642507b4fc1f8d234172bf8129942da2c2ca26", + "instance_protonmail__webclients-f161c10cf7d31abf82e8d64d7a99c9fac5acfa18", + "instance_gravitational__teleport-cb712e3f0b06dadc679f895daef8072cae400c26-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_gravitational__teleport-46a13210519461c7cec8d643bfbe750265775b41", + "instance_ansible__ansible-e40889e7112ae00a21a2c74312b330e67a766cc0-v1055803c3a812189a1133297f7f5468579283f86", + "instance_gravitational__teleport-6eaaf3a27e64f4ef4ef855bd35d7ec338cf17460-v626ec2a48416b10a88641359a169d99e935ff037", + "instance_internetarchive__openlibrary-0d13e6b4bf80bced6c0946b969b9a1b6963f6bce-v76304ecdb3a5954fcf13feb710e8c40fcf24b73c", + "instance_protonmail__webclients-2f2f6c311c6128fe86976950d3c0c2db07b03921", + "instance_ansible__ansible-5260527c4a71bfed99d803e687dd19619423b134-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_qutebrowser__qutebrowser-8d05f0282a271bfd45e614238bd1b555c58b3fc1-v35616345bb8052ea303186706cec663146f0f184", + "instance_qutebrowser__qutebrowser-50efac08f623644a85441bbe02ab9347d2b71a9d-v9f8e9d96c85c85a605e382f1510bd08563afc566", + "instance_flipt-io__flipt-dae029cba7cdb98dfb1a6b416c00d324241e6063", + "instance_nodebb__nodebb-18c45b44613aecd53e9f60457b9812049ab2998d-v0495b863a912fbff5749c67e860612b91825407c", + "instance_future-architect__vuls-e4728e388120b311c4ed469e4f942e0347a2689b-v264a82e2f4818e30f5a25e4da53b27ba119f62b5", + "instance_qutebrowser__qutebrowser-5e0d6dc1483cb3336ea0e3dcbd4fe4aa00fc1742-v5149fcda2a9a6fe1d35dfed1bade1444a11ef271", + "instance_flipt-io__flipt-c12967bc73fdf02054cf3ef8498c05e25f0a18c0", + "instance_qutebrowser__qutebrowser-ff1c025ad3210506fc76e1f604d8c8c27637d88e-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_ansible__ansible-83909bfa22573777e3db5688773bda59721962ad-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_qutebrowser__qutebrowser-1a9e74bfaf9a9db2a510dc14572d33ded6040a57-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_gravitational__teleport-24cafecd8721891092210afc55f6413ab46ca211-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_internetarchive__openlibrary-4b7ea2977be2747496ba792a678940baa985f7ea-v0f5aece3601a5b4419f7ccec1dbda2071be28ee4", + "instance_internetarchive__openlibrary-fad4a40acf5ff5f06cd7441a5c7baf41a7d81fe4-vfa6ff903cb27f336e17654595dd900fa943dcd91", + "instance_qutebrowser__qutebrowser-3fd8e12949b8feda401930574facf09dd4180bba", + "instance_ansible__ansible-164881d871964aa64e0f911d03ae270acbad253c-v390e508d27db7a51eece36bb6d9698b63a5b638a", + "instance_ansible__ansible-fb144c44144f8bd3542e71f5db62b6d322c7bd85-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_element-hq__element-web-582a1b093fc0b77538052f45cbb9c7295f991b51-vnan", + "instance_element-hq__element-web-ec0f940ef0e8e3b61078f145f34dc40d1938e6c5-vnan", + "instance_internetarchive__openlibrary-6fdbbeee4c0a7e976ff3e46fb1d36f4eb110c428-v08d8e8889ec945ab821fb156c04c7d2e2810debb", + "instance_flipt-io__flipt-7161f7b876773a911afdd804b281e52681cb7321", + "instance_navidrome__navidrome-28389fb05e1523564dfc61fa43ed8eb8a10f938c", + "instance_ansible__ansible-5f4e332e3762999d94af27746db29ff1729252c1-v0f01c69f1e2528b935359cfe578530722bca2c59", + "instance_nodebb__nodebb-0e07f3c9bace416cbab078a30eae972868c0a8a3-vf2cf3cbd463b7ad942381f1c6d077626485a1e9e", + "instance_nodebb__nodebb-da0211b1a001d45d73b4c84c6417a4f1b0312575-vf2cf3cbd463b7ad942381f1c6d077626485a1e9e", + "instance_navidrome__navidrome-31799662706fedddf5bcc1a76b50409d1f91d327", + "instance_future-architect__vuls-50580f6e98eeb36f53f27222f7f4fdfea0b21e8d", + "instance_internetarchive__openlibrary-60725705782832a2cb22e17c49697948a42a9d03-v298a7a812ceed28c4c18355a091f1b268fe56d86", + "instance_ansible__ansible-3db08adbb1cc6aa9941be5e0fc810132c6e1fa4b-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_internetarchive__openlibrary-8a5a63af6e0be406aa6c8c9b6d5f28b2f1b6af5a-v0f5aece3601a5b4419f7ccec1dbda2071be28ee4", + "instance_future-architect__vuls-e1df74cbc1a1d1889428b3333a3b2405c4651993", + "instance_internetarchive__openlibrary-123e6e5e1c85b9c07d1e98f70bfc480bc8016890-v2733ff199fb72f0d033a30dc62cb0a4742e3a7f4", + "instance_protonmail__webclients-ae36cb23a1682dcfd69587c1b311ae0227e28f39", + "instance_protonmail__webclients-8be4f6cb9380fcd2e67bcb18cef931ae0d4b869c", + "instance_qutebrowser__qutebrowser-a25e8a09873838ca9efefd36ea8a45170bbeb95c-vc2f56a753b62a190ddb23cd330c257b9cf560d12", + "instance_qutebrowser__qutebrowser-44e64199ed38003253f0296badd4a447645067b6-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_gravitational__teleport-7744f72c6eb631791434b648ba41083b5f6d2278-vce94f93ad1030e3136852817f2423c1b3ac37bc4", + "instance_tutao__tutanota-40e94dee2bcec2b63f362da283123e9df1874cc1-vc4e41fd0029957297843cb9dec4a25c7c756f029", + "instance_future-architect__vuls-c11ba27509f733d7d280bdf661cbbe2e7a99df4c", + "instance_flipt-io__flipt-b2cd6a6dd73ca91b519015fd5924fde8d17f3f06", + "instance_gravitational__teleport-6a14edcf1ff010172fdbac622d0a474ed6af46de", + "instance_flipt-io__flipt-36e62baffae2132f78f9d34dc300a9baa2d7ae0e", + "instance_navidrome__navidrome-c90468b895f6171e33e937ff20dc915c995274f0", + "instance_ansible__ansible-ecea15c508f0e081525be036cf76bbb56dbcdd9d-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_ansible__ansible-e22e103cdf8edc56ff7d9b848a58f94f1471a263-v1055803c3a812189a1133297f7f5468579283f86", + "instance_future-architect__vuls-dc496468b9e9fb73371f9606cdcdb0f8e12e70ca", + "instance_gravitational__teleport-c782838c3a174fdff80cafd8cd3b1aa4dae8beb2", + "instance_internetarchive__openlibrary-91efee627df01e32007abf2d6ebf73f9d9053076-vbee42ad1b72fb23c6a1c874868a720b370983ed2", + "instance_qutebrowser__qutebrowser-52708364b5f91e198defb022d1a5b4b3ebd9b563-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_ansible__ansible-eea46a0d1b99a6dadedbb6a3502d599235fa7ec3-v390e508d27db7a51eece36bb6d9698b63a5b638a", + "instance_protonmail__webclients-c6f65d205c401350a226bb005f42fac1754b0b5b", + "instance_ansible__ansible-cb94c0cc550df9e98f1247bc71d8c2b861c75049-v1055803c3a812189a1133297f7f5468579283f86", + "instance_nodebb__nodebb-05f2236193f407cf8e2072757fbd6bb170bc13f0-vf2cf3cbd463b7ad942381f1c6d077626485a1e9e", + "instance_navidrome__navidrome-55730514ea59d5f1d0b8e3f8745569c29bdbf7b4", + "instance_navidrome__navidrome-0130c6dc13438b48cf0fdfab08a89e357b5517c9", + "instance_flipt-io__flipt-507170da0f7f4da330f6732bffdf11c4df7fc192", + "instance_tutao__tutanota-fe240cbf7f0fdd6744ef7bef8cb61676bcdbb621-vc4e41fd0029957297843cb9dec4a25c7c756f029", + "instance_qutebrowser__qutebrowser-233cb1cc48635130e5602549856a6fa4ab4c087f-v35616345bb8052ea303186706cec663146f0f184", + "instance_internetarchive__openlibrary-f8cc11d9c1575fdba5ac66aee0befca970da8d64-v13642507b4fc1f8d234172bf8129942da2c2ca26", + "instance_qutebrowser__qutebrowser-fec187c2cb53d769c2682b35ca77858a811414a8-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_tutao__tutanota-da4edb7375c10f47f4ed3860a591c5e6557f7b5c-vbc0d9ba8f0071fbe982809910959a6ff8884dbbf", + "instance_ansible__ansible-be2c376ab87e3e872ca21697508f12c6909cf85a-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_future-architect__vuls-457a3a9627fb9a0800d0aecf1d4713fb634a9011", + "instance_internetarchive__openlibrary-d40ec88713dc95ea791b252f92d2f7b75e107440-v13642507b4fc1f8d234172bf8129942da2c2ca26", + "instance_protonmail__webclients-863d524b5717b9d33ce08a0f0535e3fd8e8d1ed8", + "instance_internetarchive__openlibrary-308a35d6999427c02b1dbf5211c033ad3b352556-ve8c8d62a2b60610a3c4631f5f23ed866bada9818", + "instance_flipt-io__flipt-15b76cada1ef29cfa56b0fba36754be36243dded", + "instance_future-architect__vuls-61c39637f2f3809e1b5dad05f0c57c799dce1587", + "instance_ansible__ansible-5d253a13807e884b7ce0b6b57a963a45e2f0322c-v1055803c3a812189a1133297f7f5468579283f86", + "instance_element-hq__element-web-44b98896a79ede48f5ad7ff22619a39d5f6ff03c-vnan", + "instance_qutebrowser__qutebrowser-f91ace96223cac8161c16dd061907e138fe85111-v059c6fdc75567943479b23ebca7c07b5e9a7f34c", + "instance_gravitational__teleport-bb562408da4adeae16e025be65e170959d1ec492-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_gravitational__teleport-78b0d8c72637df1129fb6ff84fc49ef4b5ab1288", + "instance_internetarchive__openlibrary-3f7db6bbbcc7c418b3db72d157c6aed1d45b2ccf-v430f20c722405e462d9ef44dee7d34c41e76fe7a", + "instance_future-architect__vuls-e52fa8d6ed1d23e36f2a86e5d3efe9aa057a1b0d", + "instance_navidrome__navidrome-3bc9e75b2843f91f6a1e9b604e321c2bd4fd442a", + "instance_tutao__tutanota-fb32e5f9d9fc152a00144d56dd0af01760a2d4dc-vc4e41fd0029957297843cb9dec4a25c7c756f029", + "instance_gravitational__teleport-af5e2517de7d18406b614e413aca61c319312171-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_ansible__ansible-a20a52701402a12f91396549df04ac55809f68e9-v1055803c3a812189a1133297f7f5468579283f86", + "instance_protonmail__webclients-1501eb765873b2884b6f1944fd242ecfc9d6b103", + "instance_gravitational__teleport-db89206db6c2969266e664c7c0fb51b70e958b64", + "instance_ansible__ansible-be59caa59bf47ca78a4760eb7ff38568372a8260-v1055803c3a812189a1133297f7f5468579283f86", + "instance_flipt-io__flipt-65581fef4aa807540cb933753d085feb0d7e736f", + "instance_gravitational__teleport-4e1c39639edf1ab494dd7562844c8b277b5cfa18-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_future-architect__vuls-f0b3a8b1db98eb1bd32685f1c36c41a99c3452ed", + "instance_navidrome__navidrome-b65e76293a917ee2dfc5d4b373b1c62e054d0dca", + "instance_element-hq__element-web-33e8edb3d508d6eefb354819ca693b7accc695e7", + "instance_flipt-io__flipt-e50808c03e4b9d25a6a78af9c61a3b1616ea356b", + "instance_future-architect__vuls-e3c27e1817d68248043bd09d63cc31f3344a6f2c", + "instance_element-hq__element-web-ad26925bb6628260cfe0fcf90ec0a8cba381f4a4-vnan", + "instance_nodebb__nodebb-1ea9481af6125ffd6da0592ed439aa62af0bca11-vd59a5728dfc977f44533186ace531248c2917516", + "instance_ansible__ansible-489156378c8e97374a75a544c7c9c2c0dd8146d1-v390e508d27db7a51eece36bb6d9698b63a5b638a", + "instance_internetarchive__openlibrary-bb152d23c004f3d68986877143bb0f83531fe401-ve8c8d62a2b60610a3c4631f5f23ed866bada9818", + "instance_qutebrowser__qutebrowser-0b621cb0ce2b54d3f93d8d41d8ff4257888a87e5-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_qutebrowser__qutebrowser-1af602b258b97aaba69d2585ed499d95e2303ac2-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_future-architect__vuls-e1fab805afcfc92a2a615371d0ec1e667503c254-v264a82e2f4818e30f5a25e4da53b27ba119f62b5", + "instance_internetarchive__openlibrary-7edd1ef09d91fe0b435707633c5cc9af41dedddf-v76304ecdb3a5954fcf13feb710e8c40fcf24b73c", + "instance_qutebrowser__qutebrowser-99029144b5109bb1b2a53964a7c129e009980cd9-va0fd88aac89cde702ec1ba84877234da33adce8a", + "instance_ansible__ansible-12734fa21c08a0ce8c84e533abdc560db2eb1955-v7eee2454f617569fd6889f2211f75bc02a35f9f8", + "instance_protonmail__webclients-7e54526774e577c0ebb58ced7ba8bef349a69fec", + "instance_gravitational__teleport-1a77b7945a022ab86858029d30ac7ad0d5239d00-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_nodebb__nodebb-8ca65b0c78c67c1653487c02d1135e1b702185e1-vf2cf3cbd463b7ad942381f1c6d077626485a1e9e", + "instance_flipt-io__flipt-9f8127f225a86245fa35dca4885c2daef824ee55", + "instance_gravitational__teleport-ac2fb2f9b4fd1896b554d3011df23d3d71295779", + "instance_ansible__ansible-f02a62db509dc7463fab642c9c3458b9bc3476cc-v390e508d27db7a51eece36bb6d9698b63a5b638a" + ], + "additional_details": { + "shuffled": "True" + } + }, + "evaluation_timestamp": "1775602861.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.425, + "uncertainty": { + "standard_error": { + "value": 0.03504304603451139 + }, + "num_samples": 200 + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"If you do not submit an answer, you will receive a score of zero.\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [10, \"Your answer has been saved. Consider whether your previous approach is likely correct or whether a different approach would be more promising. Then either refine your previous answer or try a substantially different approach, and submit when ready.\", \"\"], \"submit\": null, \"on_continue\": null, \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\"}, \"params_passed\": {\"prompt\": [\"If you do not submit an answer, you will receive a score of zero.\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [10, \"Your answer has been saved. Consider whether your previous approach is likely correct or whether a different approach would be more promising. Then either refine your previous answer or try a substantially different approach, and submit when ready.\", \"\"], \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "reasoning_effort": "\"high\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 20000000 + }, + "sandbox": { + "type": "k8s" + } + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/openai/gpt-5.2-2025-12-11/b13227be-5ac9-4052-a5a1-804bf5166b96_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "e222a6fa2dc22b8fa6e865aad220f4c35e9bc52e243d95a50e278f2d4835c9bc", + "total_rows": 200 + } +} \ No newline at end of file diff --git a/data/swebenchpro/openai/gpt-5.2-2025-12-11/b13227be-5ac9-4052-a5a1-804bf5166b96_samples.jsonl b/data/swebenchpro/openai/gpt-5.2-2025-12-11/b13227be-5ac9-4052-a5a1-804bf5166b96_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..c208ea7e2be64453ff296161533b0219f043e54e --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.2-2025-12-11/b13227be-5ac9-4052-a5a1-804bf5166b96_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e222a6fa2dc22b8fa6e865aad220f4c35e9bc52e243d95a50e278f2d4835c9bc +size 113607175 diff --git a/data/swebenchpro/openai/gpt-5.2-2025-12-11/c1b55282-8d8f-4aa0-a12f-9053458e99c9.json b/data/swebenchpro/openai/gpt-5.2-2025-12-11/c1b55282-8d8f-4aa0-a12f-9053458e99c9.json new file mode 100644 index 0000000000000000000000000000000000000000..3dec51dcc43fd7eef3da8b2bcad4ee133f3ae543 --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.2-2025-12-11/c1b55282-8d8f-4aa0-a12f-9053458e99c9.json @@ -0,0 +1,118 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/openai_gpt-5.2-2025-12-11/1786162862.054013", + "evaluation_timestamp": "1776520455.0", + "retrieved_timestamp": "1786162862.054013", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on swebenchpro/S-adaptive+C/+1ep/552bcbea for scorer harbor_scorer", + "source_data": { + "dataset_name": "swebenchpro", + "source_type": "hf_dataset", + "samples_number": 10, + "sample_ids": [ + "instance_ansible__ansible-f327e65d11bb905ed9f15996024f857a95592629-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_protonmail__webclients-da91f084c0f532d9cc8ca385a701274d598057b8", + "instance_flipt-io__flipt-b2cd6a6dd73ca91b519015fd5924fde8d17f3f06", + "instance_ansible__ansible-3db08adbb1cc6aa9941be5e0fc810132c6e1fa4b-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_internetarchive__openlibrary-c506c1b0b678892af5cb22c1c1dbc35d96787a0a-v0f5aece3601a5b4419f7ccec1dbda2071be28ee4", + "instance_internetarchive__openlibrary-25858f9f0c165df25742acf8309ce909773f0cdd-v13642507b4fc1f8d234172bf8129942da2c2ca26", + "instance_flipt-io__flipt-05d7234fa582df632f70a7cd10194d61bd7043b9", + "instance_ansible__ansible-a02e22e902a69aeb465f16bf03f7f5a91b2cb828-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_nodebb__nodebb-8ca65b0c78c67c1653487c02d1135e1b702185e1-vf2cf3cbd463b7ad942381f1c6d077626485a1e9e", + "instance_protonmail__webclients-2c3559cad02d1090985dba7e8eb5a129144d9811" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1776520455.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.8, + "uncertainty": { + "standard_error": { + "value": 0.13333333333333333 + }, + "num_samples": 10 + }, + "details": { + "total_matched_trajectories": "10", + "stopping_reason_count_completed_on_successful_submit": "8", + "stopping_reason_count_repetition_guard": "2", + "avg_total_tokens_target_model": "3946782.90", + "total_total_tokens_target_model": "39467829.0", + "avg_total_tokens_other_models": "0.00", + "total_total_tokens_other_models": "0.0", + "avg_total_tokens_all_models": "3946782.90", + "total_total_tokens_all_models": "39467829.0", + "avg_turn_count": "89.20", + "total_turn_count": "892.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 30000000 + }, + "sandbox": {} + }, + "additional_details": { + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/openai/gpt-5.2-2025-12-11/c1b55282-8d8f-4aa0-a12f-9053458e99c9_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "3eb171ba84f3c673db8b10eb985f307af3d99dad919682b31da7b34bab7319f3", + "total_rows": 10 + } +} \ No newline at end of file diff --git a/data/swebenchpro/openai/gpt-5.2-2025-12-11/c1b55282-8d8f-4aa0-a12f-9053458e99c9_samples.jsonl b/data/swebenchpro/openai/gpt-5.2-2025-12-11/c1b55282-8d8f-4aa0-a12f-9053458e99c9_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..13f2314b0f5601f12d518ad5aa9d9ef400f9dd9e --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.2-2025-12-11/c1b55282-8d8f-4aa0-a12f-9053458e99c9_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3eb171ba84f3c673db8b10eb985f307af3d99dad919682b31da7b34bab7319f3 +size 9735171 diff --git a/data/swebenchpro/openai/gpt-5.2-2025-12-11/c46cd570-b481-4991-8207-d53df20ee4d4.json b/data/swebenchpro/openai/gpt-5.2-2025-12-11/c46cd570-b481-4991-8207-d53df20ee4d4.json new file mode 100644 index 0000000000000000000000000000000000000000..2a03a2a46969e7babb6e4547938ebed09117da3f --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.2-2025-12-11/c46cd570-b481-4991-8207-d53df20ee4d4.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/openai_gpt-5.2-2025-12-11/1786167083.091274", + "evaluation_timestamp": "1776359446.0", + "retrieved_timestamp": "1786167083.091274", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/openai/gpt-5.2-2025-12-11/c46cd570-b481-4991-8207-d53df20ee4d4_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "65a9c99172b5cc75091c4f4ad5c0da415bea5f3defbcbb14d404d29890fd0133", + "total_rows": 20 + } +} \ No newline at end of file diff --git a/data/swebenchpro/openai/gpt-5.2-2025-12-11/c46cd570-b481-4991-8207-d53df20ee4d4_samples.jsonl b/data/swebenchpro/openai/gpt-5.2-2025-12-11/c46cd570-b481-4991-8207-d53df20ee4d4_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..a1f828ffc200f6cdadbe169bd6a4c2a8893e52b1 --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.2-2025-12-11/c46cd570-b481-4991-8207-d53df20ee4d4_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:65a9c99172b5cc75091c4f4ad5c0da415bea5f3defbcbb14d404d29890fd0133 +size 18454149 diff --git a/data/swebenchpro/openai/gpt-5.2-2025-12-11/c6178af5-b8a0-4b75-baa8-6a969a3b6828.json b/data/swebenchpro/openai/gpt-5.2-2025-12-11/c6178af5-b8a0-4b75-baa8-6a969a3b6828.json new file mode 100644 index 0000000000000000000000000000000000000000..0c8df88198287cc20b6781aa5aa3e498dc91299d --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.2-2025-12-11/c6178af5-b8a0-4b75-baa8-6a969a3b6828.json @@ -0,0 +1,113 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/openai_gpt-5.2-2025-12-11/1786162915.633871", + "evaluation_timestamp": "1776443008.0", + "retrieved_timestamp": "1786162915.633871", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on swebenchpro/S-adaptive+C/+4ep/497a67ea for scorer harbor_scorer", + "source_data": { + "dataset_name": "swebenchpro", + "source_type": "hf_dataset", + "samples_number": 5, + "sample_ids": [ + "instance_gravitational__teleport-82185f232ae8974258397e121b3bc2ed0c3729ed-v626ec2a48416b10a88641359a169d99e935ff037", + "instance_internetarchive__openlibrary-630221ab686c64e75a2ce253c893c033e4814b2e-v93c53c13d5f9b383ebb411ee7750b49dcd1a34c6", + "instance_qutebrowser__qutebrowser-a25e8a09873838ca9efefd36ea8a45170bbeb95c-vc2f56a753b62a190ddb23cd330c257b9cf560d12", + "instance_protonmail__webclients-f161c10cf7d31abf82e8d64d7a99c9fac5acfa18", + "instance_future-architect__vuls-50580f6e98eeb36f53f27222f7f4fdfea0b21e8d" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1776443008.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.6, + "uncertainty": { + "standard_error": { + "value": 0.24494897427831783 + }, + "num_samples": 20 + }, + "details": { + "total_matched_trajectories": "20", + "stopping_reason_count_repetition_guard": "13", + "stopping_reason_count_completed_on_successful_submit": "7", + "avg_total_tokens_target_model": "4026583.55", + "total_total_tokens_target_model": "80531671.0", + "avg_total_tokens_other_models": "0.00", + "total_total_tokens_other_models": "0.0", + "avg_total_tokens_all_models": "4026583.55", + "total_total_tokens_all_models": "80531671.0", + "avg_turn_count": "98.35", + "total_turn_count": "1967.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 30000000 + }, + "sandbox": {} + }, + "additional_details": { + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/openai/gpt-5.2-2025-12-11/c6178af5-b8a0-4b75-baa8-6a969a3b6828_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "140839527b461f09900b798c56d8f145ba98d71ad00bde1eb75798f8e09a240c", + "total_rows": 20 + } +} \ No newline at end of file diff --git a/data/swebenchpro/openai/gpt-5.2-2025-12-11/c6178af5-b8a0-4b75-baa8-6a969a3b6828_samples.jsonl b/data/swebenchpro/openai/gpt-5.2-2025-12-11/c6178af5-b8a0-4b75-baa8-6a969a3b6828_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..b12928f30da891b23147f6235c167bf01edca31b --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.2-2025-12-11/c6178af5-b8a0-4b75-baa8-6a969a3b6828_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:140839527b461f09900b798c56d8f145ba98d71ad00bde1eb75798f8e09a240c +size 40356294 diff --git a/data/swebenchpro/openai/gpt-5.2-2025-12-11/cec418ea-9a25-4f6c-9f97-f112e70b9043.json b/data/swebenchpro/openai/gpt-5.2-2025-12-11/cec418ea-9a25-4f6c-9f97-f112e70b9043.json new file mode 100644 index 0000000000000000000000000000000000000000..5bb6fcca2bfcad3e9036dcd34130348fe69f4b63 --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.2-2025-12-11/cec418ea-9a25-4f6c-9f97-f112e70b9043.json @@ -0,0 +1,119 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/openai_gpt-5.2-2025-12-11/1786164388.211554", + "evaluation_timestamp": "1778564151.0", + "retrieved_timestamp": "1786164388.211554", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on swebenchpro/S-adaptive/+3ep/71cb16d6 for scorer harbor_scorer", + "source_data": { + "dataset_name": "swebenchpro", + "source_type": "hf_dataset", + "samples_number": 7, + "sample_ids": [ + "instance_protonmail__webclients-01ea5214d11e0df8b7170d91bafd34f23cb0f2b1", + "instance_nodebb__nodebb-22368b996ee0e5f11a5189b400b33af3cc8d925a-v4fbcfae8b15e4ce5d132c408bca69ebb9cf146ed", + "instance_ansible__ansible-cd9c4eb5a6b2bfaf4a6709f001ce3d0c92c1eed2-v0f01c69f1e2528b935359cfe578530722bca2c59", + "instance_gravitational__teleport-82185f232ae8974258397e121b3bc2ed0c3729ed-v626ec2a48416b10a88641359a169d99e935ff037", + "instance_gravitational__teleport-2bb3bbbd8aff1164a2353381cb79e1dc93b90d28-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_ansible__ansible-de5858f48dc9e1ce9117034e0d7e76806f420ca8-v1055803c3a812189a1133297f7f5468579283f86", + "instance_future-architect__vuls-999529a05b202b0fd29c6fca5039a4c47a3766bb" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1778564151.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.42857142857142855, + "uncertainty": { + "standard_error": { + "value": 0.20203050891044214 + }, + "num_samples": 19 + }, + "details": { + "total_matched_trajectories": "19", + "stopping_reason_count_repetition_guard": "19", + "avg_total_tokens_target_model": "3608752.53", + "total_total_tokens_target_model": "68566298.0", + "avg_total_tokens_other_models": "1068.11", + "total_total_tokens_other_models": "20294.0", + "avg_total_tokens_all_models": "3609820.63", + "total_total_tokens_all_models": "68586592.0", + "avg_turn_count": "82.32", + "total_turn_count": "1564.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 30000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/openai/gpt-5.2-2025-12-11/cec418ea-9a25-4f6c-9f97-f112e70b9043_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "16f0720f36ecc7dddd05354bbd95a5e90759f567b9a92e8c17b85479cd7ef8b0", + "total_rows": 19 + } +} \ No newline at end of file diff --git a/data/swebenchpro/openai/gpt-5.2-2025-12-11/cec418ea-9a25-4f6c-9f97-f112e70b9043_samples.jsonl b/data/swebenchpro/openai/gpt-5.2-2025-12-11/cec418ea-9a25-4f6c-9f97-f112e70b9043_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..f70bab27f1db8edc5804a8fb58686f937813cfb2 --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.2-2025-12-11/cec418ea-9a25-4f6c-9f97-f112e70b9043_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:16f0720f36ecc7dddd05354bbd95a5e90759f567b9a92e8c17b85479cd7ef8b0 +size 11569679 diff --git a/data/swebenchpro/openai/gpt-5.2-2025-12-11/d2776738-9142-412f-bae3-09be0a9c1f49.json b/data/swebenchpro/openai/gpt-5.2-2025-12-11/d2776738-9142-412f-bae3-09be0a9c1f49.json new file mode 100644 index 0000000000000000000000000000000000000000..4b47eda7afbc96edad5937b0b64b41925966bbd9 --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.2-2025-12-11/d2776738-9142-412f-bae3-09be0a9c1f49.json @@ -0,0 +1,116 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/openai_gpt-5.2-2025-12-11/1786162878.508313", + "evaluation_timestamp": "1778875207.0", + "retrieved_timestamp": "1786162878.508313", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on swebenchpro/S-adaptive+C/+3ep/65c91019 for scorer harbor_scorer", + "source_data": { + "dataset_name": "swebenchpro", + "source_type": "hf_dataset", + "samples_number": 3, + "sample_ids": [ + "instance_ansible__ansible-de5858f48dc9e1ce9117034e0d7e76806f420ca8-v1055803c3a812189a1133297f7f5468579283f86", + "instance_protonmail__webclients-281a6b3f190f323ec2c0630999354fafb84b2880", + "instance_gravitational__teleport-3ff19cf7c41f396ae468797d3aeb61515517edc9-vee9b09fb20c43af7e520f57e9239bbcf46b7113d" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1778875207.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.3333333333333333, + "uncertainty": { + "standard_error": { + "value": 0.33333333333333337 + }, + "num_samples": 7 + }, + "details": { + "total_matched_trajectories": "7", + "stopping_reason_count_repetition_guard": "4", + "stopping_reason_count_completed_on_successful_submit": "3", + "avg_total_tokens_target_model": "4719705.86", + "total_total_tokens_target_model": "33037941.0", + "avg_total_tokens_other_models": "594.86", + "total_total_tokens_other_models": "4164.0", + "avg_total_tokens_all_models": "4720300.71", + "total_total_tokens_all_models": "33042105.0", + "avg_turn_count": "107.86", + "total_turn_count": "755.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "8", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 30000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "8", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/openai/gpt-5.2-2025-12-11/d2776738-9142-412f-bae3-09be0a9c1f49_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "bb18c14d390bc21dd34b9e0f024ae1a6bbd351d9084b2da79a23126013e952eb", + "total_rows": 7 + } +} \ No newline at end of file diff --git a/data/swebenchpro/openai/gpt-5.2-2025-12-11/d2776738-9142-412f-bae3-09be0a9c1f49_samples.jsonl b/data/swebenchpro/openai/gpt-5.2-2025-12-11/d2776738-9142-412f-bae3-09be0a9c1f49_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..0e94e7fbd80939465a0a8751b994416f8843d349 --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.2-2025-12-11/d2776738-9142-412f-bae3-09be0a9c1f49_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bb18c14d390bc21dd34b9e0f024ae1a6bbd351d9084b2da79a23126013e952eb +size 5423919 diff --git a/data/swebenchpro/openai/gpt-5.2-2025-12-11/d550b10d-7cb6-4c91-8d24-b43c4e3ab7fb.json b/data/swebenchpro/openai/gpt-5.2-2025-12-11/d550b10d-7cb6-4c91-8d24-b43c4e3ab7fb.json new file mode 100644 index 0000000000000000000000000000000000000000..831647bad0bc24a7a0b15c049a293121cf130df5 --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.2-2025-12-11/d550b10d-7cb6-4c91-8d24-b43c4e3ab7fb.json @@ -0,0 +1,105 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/openai_gpt-5.2-2025-12-11/1786164319.176323", + "evaluation_timestamp": "1776680886.0", + "retrieved_timestamp": "1786164319.176323", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on swebenchpro/S-adaptive/+1ep/aacb763d for scorer harbor_scorer", + "source_data": { + "dataset_name": "swebenchpro", + "source_type": "hf_dataset", + "samples_number": 1, + "sample_ids": [ + "instance_qutebrowser__qutebrowser-2e961080a85d660148937ee8f0f6b3445a8f2c01-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1776680886.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 1.0, + "uncertainty": { + "num_samples": 1 + }, + "details": { + "total_matched_trajectories": "1", + "stopping_reason_count_repetition_guard": "1", + "avg_total_tokens_target_model": "3764525.00", + "total_total_tokens_target_model": "3764525.0", + "avg_total_tokens_other_models": "1429.00", + "total_total_tokens_other_models": "1429.0", + "avg_total_tokens_all_models": "3765954.00", + "total_total_tokens_all_models": "3765954.0", + "avg_turn_count": "100.00", + "total_turn_count": "100.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 30000000 + }, + "sandbox": {} + }, + "additional_details": { + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/openai/gpt-5.2-2025-12-11/d550b10d-7cb6-4c91-8d24-b43c4e3ab7fb_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "1d27abb47840960d2e8a8e72d6b6cd3ee03b36bf32e4275ef36b05e8728798d7", + "total_rows": 1 + } +} \ No newline at end of file diff --git a/data/swebenchpro/openai/gpt-5.2-2025-12-11/d550b10d-7cb6-4c91-8d24-b43c4e3ab7fb_samples.jsonl b/data/swebenchpro/openai/gpt-5.2-2025-12-11/d550b10d-7cb6-4c91-8d24-b43c4e3ab7fb_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..9dd33970d08a8ff074e9a53154e94a674a2f4fa2 --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.2-2025-12-11/d550b10d-7cb6-4c91-8d24-b43c4e3ab7fb_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1d27abb47840960d2e8a8e72d6b6cd3ee03b36bf32e4275ef36b05e8728798d7 +size 643983 diff --git a/data/swebenchpro/openai/gpt-5.2-2025-12-11/d8fbabc2-0ebd-4c40-b252-1c4faa2227d5.json b/data/swebenchpro/openai/gpt-5.2-2025-12-11/d8fbabc2-0ebd-4c40-b252-1c4faa2227d5.json new file mode 100644 index 0000000000000000000000000000000000000000..2dbb788b2ffb2567e303be5817a867d9c64fbb8b --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.2-2025-12-11/d8fbabc2-0ebd-4c40-b252-1c4faa2227d5.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/openai_gpt-5.2-2025-12-11/1786169011.766374", + "evaluation_timestamp": "1777544308.0", + "retrieved_timestamp": "1786169011.766374", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/openai/gpt-5.2-2025-12-11/d8fbabc2-0ebd-4c40-b252-1c4faa2227d5_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "aded9dc6a6f91fe7c83667bf5bdd018df119d3fa17057d384cd44627231d3b2b", + "total_rows": 19 + } +} \ No newline at end of file diff --git a/data/swebenchpro/openai/gpt-5.2-2025-12-11/d8fbabc2-0ebd-4c40-b252-1c4faa2227d5_samples.jsonl b/data/swebenchpro/openai/gpt-5.2-2025-12-11/d8fbabc2-0ebd-4c40-b252-1c4faa2227d5_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..5306034f4c81c56336a32f840cc34f20fab1359e --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.2-2025-12-11/d8fbabc2-0ebd-4c40-b252-1c4faa2227d5_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:aded9dc6a6f91fe7c83667bf5bdd018df119d3fa17057d384cd44627231d3b2b +size 13882052 diff --git a/data/swebenchpro/openai/gpt-5.2-2025-12-11/dc897f0c-b7be-4156-ac3f-b06c8aac87d3.json b/data/swebenchpro/openai/gpt-5.2-2025-12-11/dc897f0c-b7be-4156-ac3f-b06c8aac87d3.json new file mode 100644 index 0000000000000000000000000000000000000000..d8a607099c0022c9a0e29597bdaac05fb7999ed1 --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.2-2025-12-11/dc897f0c-b7be-4156-ac3f-b06c8aac87d3.json @@ -0,0 +1,133 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/openai_gpt-5.2-2025-12-11/1786162866.450672", + "evaluation_timestamp": "1778854176.0", + "retrieved_timestamp": "1786162866.450672", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on swebenchpro/S-adaptive+C/+1ep/1a8aef33 for scorer harbor_scorer", + "source_data": { + "dataset_name": "swebenchpro", + "source_type": "hf_dataset", + "samples_number": 20, + "sample_ids": [ + "instance_nodebb__nodebb-84e065752f6d7fbe5c08cbf50cb173ffb866b8fa-vf2cf3cbd463b7ad942381f1c6d077626485a1e9e", + "instance_future-architect__vuls-6eff6a9329a65cc412e79b8f82444dfa3d0f0b5a", + "instance_protonmail__webclients-cb8cc309c6968b0a2a5fe4288d0ae0a969ff31e1", + "instance_element-hq__element-web-b007ea81b2ccd001b00f332bee65070aa7fc00f9-vnan", + "instance_flipt-io__flipt-0b119520afca1cf25c470ff4288c464d4510b944", + "instance_element-hq__element-web-f63160f38459fb552d00fcc60d4064977a9095a6-vnan", + "instance_qutebrowser__qutebrowser-9ed748effa8f3bcd804612d9291da017b514e12f-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_qutebrowser__qutebrowser-e5340c449f23608803c286da0563b62f58ba25b0-v059c6fdc75567943479b23ebca7c07b5e9a7f34c", + "instance_protonmail__webclients-5f0745dd6993bb1430a951c62a49807c6635cd77", + "instance_nodebb__nodebb-22368b996ee0e5f11a5189b400b33af3cc8d925a-v4fbcfae8b15e4ce5d132c408bca69ebb9cf146ed", + "instance_protonmail__webclients-01b519cd49e6a24d9a05d2eb97f54e420740072e", + "instance_ansible__ansible-77658704217d5f166404fc67997203c25381cb6e-v390e508d27db7a51eece36bb6d9698b63a5b638a", + "instance_ansible__ansible-cd9c4eb5a6b2bfaf4a6709f001ce3d0c92c1eed2-v0f01c69f1e2528b935359cfe578530722bca2c59", + "instance_protonmail__webclients-ac23d1efa1a6ab7e62724779317ba44c28d78cfd", + "instance_future-architect__vuls-030b2e03525d68d74cb749959aac2d7f3fc0effa", + "instance_nodebb__nodebb-b1f9ad5534bb3a44dab5364f659876a4b7fe34c1-vnan", + "instance_flipt-io__flipt-292fdaca9be39e6a921aaa8874c011d0fdd3e874", + "instance_ansible__ansible-622a493ae03bd5e5cf517d336fc426e9d12208c7-v906c969b551b346ef54a2c0b41e04f632b7b73c2", + "instance_ansible__ansible-9a21e247786ebd294dafafca1105fcd770ff46c6-v67cdaa49f89b34e42b69d5b7830b3c3ad3d8803f", + "instance_flipt-io__flipt-b22f5f02e40b225b6b93fff472914973422e97c6" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1778854176.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.7058823529411765, + "uncertainty": { + "standard_error": { + "value": 0.11391127488845344 + }, + "num_samples": 17 + }, + "details": { + "total_matched_trajectories": "17", + "stopping_reason_count_completed_on_successful_submit": "12", + "stopping_reason_count_repetition_guard": "5", + "avg_total_tokens_target_model": "3731376.35", + "total_total_tokens_target_model": "63433398.0", + "avg_total_tokens_other_models": "302.35", + "total_total_tokens_other_models": "5140.0", + "avg_total_tokens_all_models": "3731678.71", + "total_total_tokens_all_models": "63438538.0", + "avg_turn_count": "90.00", + "total_turn_count": "1530.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "8", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 30000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "8", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/openai/gpt-5.2-2025-12-11/dc897f0c-b7be-4156-ac3f-b06c8aac87d3_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "6b138fa3da1ec2ab085e317a36eebe84913aea6fd59a6592833e33f85f553aa4", + "total_rows": 17 + } +} \ No newline at end of file diff --git a/data/swebenchpro/openai/gpt-5.2-2025-12-11/dc897f0c-b7be-4156-ac3f-b06c8aac87d3_samples.jsonl b/data/swebenchpro/openai/gpt-5.2-2025-12-11/dc897f0c-b7be-4156-ac3f-b06c8aac87d3_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..ce965932fd250941851dc76b4a33f9f1847137de --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.2-2025-12-11/dc897f0c-b7be-4156-ac3f-b06c8aac87d3_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6b138fa3da1ec2ab085e317a36eebe84913aea6fd59a6592833e33f85f553aa4 +size 9408973 diff --git a/data/swebenchpro/openai/gpt-5.2-2025-12-11/e1a9e523-7d30-4bd1-830a-f088abb67129.json b/data/swebenchpro/openai/gpt-5.2-2025-12-11/e1a9e523-7d30-4bd1-830a-f088abb67129.json new file mode 100644 index 0000000000000000000000000000000000000000..cdf21c1bff8d14d034480485b9ae13bec1bc1bf2 --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.2-2025-12-11/e1a9e523-7d30-4bd1-830a-f088abb67129.json @@ -0,0 +1,173 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/openai_gpt-5.2-2025-12-11/1786164438.652146", + "evaluation_timestamp": "1778283817.0", + "retrieved_timestamp": "1786164438.652146", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on swebenchpro/S-adaptive/1e6cb397 for scorer harbor_scorer", + "source_data": { + "dataset_name": "swebenchpro", + "source_type": "hf_dataset", + "samples_number": 61, + "sample_ids": [ + "instance_internetarchive__openlibrary-7cbfb812ef0e1f9716e2d6e85d538a96fcb79d13-vfa6ff903cb27f336e17654595dd900fa943dcd91", + "instance_protonmail__webclients-da91f084c0f532d9cc8ca385a701274d598057b8", + "instance_flipt-io__flipt-b2cd6a6dd73ca91b519015fd5924fde8d17f3f06", + "instance_qutebrowser__qutebrowser-cc360cd4a34a126274c7b51f3b63afbaf3e05a02-v5fc38aaf22415ab0b70567368332beee7955b367", + "instance_flipt-io__flipt-e2bd19dafa7166c96b082fb2a59eb54b4be0d778", + "instance_qutebrowser__qutebrowser-9ed748effa8f3bcd804612d9291da017b514e12f-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_ansible__ansible-1ee70fc272aff6bf3415357c6e13c5de5b928d9b-v1055803c3a812189a1133297f7f5468579283f86", + "instance_element-hq__element-web-75c2c1a572fa45d1ea1d1a96e9e36e303332ecaa-vnan", + "instance_qutebrowser__qutebrowser-f631cd4422744160d9dcf7a0455da532ce973315-v35616345bb8052ea303186706cec663146f0f184", + "instance_ansible__ansible-3db08adbb1cc6aa9941be5e0fc810132c6e1fa4b-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_qutebrowser__qutebrowser-e5340c449f23608803c286da0563b62f58ba25b0-v059c6fdc75567943479b23ebca7c07b5e9a7f34c", + "instance_ansible__ansible-42355d181a11b51ebfc56f6f4b3d9c74e01cb13b-v1055803c3a812189a1133297f7f5468579283f86", + "instance_internetarchive__openlibrary-09865f5fb549694d969f0a8e49b9d204ef1853ca-ve8c8d62a2b60610a3c4631f5f23ed866bada9818", + "instance_internetarchive__openlibrary-c506c1b0b678892af5cb22c1c1dbc35d96787a0a-v0f5aece3601a5b4419f7ccec1dbda2071be28ee4", + "instance_element-hq__element-web-2760bfc8369f1bee640d6d7a7e910783143d4c5f-vnan", + "instance_internetarchive__openlibrary-25858f9f0c165df25742acf8309ce909773f0cdd-v13642507b4fc1f8d234172bf8129942da2c2ca26", + "instance_protonmail__webclients-01ea5214d11e0df8b7170d91bafd34f23cb0f2b1", + "instance_tutao__tutanota-4b4e45949096bb288f2b522f657610e480efa3e8-vee878bb72091875e912c52fc32bc60ec3760227b", + "instance_flipt-io__flipt-05d7234fa582df632f70a7cd10194d61bd7043b9", + "instance_ansible__ansible-a02e22e902a69aeb465f16bf03f7f5a91b2cb828-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_protonmail__webclients-5f0745dd6993bb1430a951c62a49807c6635cd77", + "instance_tutao__tutanota-1ff82aa365763cee2d609c9d19360ad87fdf2ec7-vc4e41fd0029957297843cb9dec4a25c7c756f029", + "instance_internetarchive__openlibrary-30bc73a1395fba2300087c7f307e54bb5372b60a-v76304ecdb3a5954fcf13feb710e8c40fcf24b73c", + "instance_nodebb__nodebb-8ca65b0c78c67c1653487c02d1135e1b702185e1-vf2cf3cbd463b7ad942381f1c6d077626485a1e9e", + "instance_nodebb__nodebb-22368b996ee0e5f11a5189b400b33af3cc8d925a-v4fbcfae8b15e4ce5d132c408bca69ebb9cf146ed", + "instance_protonmail__webclients-2c3559cad02d1090985dba7e8eb5a129144d9811", + "instance_protonmail__webclients-01b519cd49e6a24d9a05d2eb97f54e420740072e", + "instance_ansible__ansible-77658704217d5f166404fc67997203c25381cb6e-v390e508d27db7a51eece36bb6d9698b63a5b638a", + "instance_qutebrowser__qutebrowser-2dd8966fdcf11972062c540b7a787e4d0de8d372-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_ansible__ansible-cd9c4eb5a6b2bfaf4a6709f001ce3d0c92c1eed2-v0f01c69f1e2528b935359cfe578530722bca2c59", + "instance_protonmail__webclients-ac23d1efa1a6ab7e62724779317ba44c28d78cfd", + "instance_future-architect__vuls-030b2e03525d68d74cb749959aac2d7f3fc0effa", + "instance_nodebb__nodebb-b1f9ad5534bb3a44dab5364f659876a4b7fe34c1-vnan", + "instance_flipt-io__flipt-292fdaca9be39e6a921aaa8874c011d0fdd3e874", + "instance_qutebrowser__qutebrowser-ef5ba1a0360b39f9eff027fbdc57f363597c3c3b-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_ansible__ansible-622a493ae03bd5e5cf517d336fc426e9d12208c7-v906c969b551b346ef54a2c0b41e04f632b7b73c2", + "instance_gravitational__teleport-82185f232ae8974258397e121b3bc2ed0c3729ed-v626ec2a48416b10a88641359a169d99e935ff037", + "instance_ansible__ansible-9a21e247786ebd294dafafca1105fcd770ff46c6-v67cdaa49f89b34e42b69d5b7830b3c3ad3d8803f", + "instance_protonmail__webclients-8afd9ce04c8dde9e150e1c2b50d32e7ee2efa3e7", + "instance_flipt-io__flipt-b22f5f02e40b225b6b93fff472914973422e97c6", + "instance_gravitational__teleport-2bb3bbbd8aff1164a2353381cb79e1dc93b90d28-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_ansible__ansible-de5858f48dc9e1ce9117034e0d7e76806f420ca8-v1055803c3a812189a1133297f7f5468579283f86", + "instance_gravitational__teleport-46aa81b1ce96ebb4ebed2ae53fd78cd44a05da6c-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_tutao__tutanota-8513a9e8114a8b42e64f4348335e0f23efa054c4-vee878bb72091875e912c52fc32bc60ec3760227b", + "instance_internetarchive__openlibrary-630221ab686c64e75a2ce253c893c033e4814b2e-v93c53c13d5f9b383ebb411ee7750b49dcd1a34c6", + "instance_internetarchive__openlibrary-7f6b722a10f822171501d027cad60afe53337732-ve8c8d62a2b60610a3c4631f5f23ed866bada9818", + "instance_nodebb__nodebb-a917210c5b2c20637094545401f85783905c074c-vf2cf3cbd463b7ad942381f1c6d077626485a1e9e", + "instance_internetarchive__openlibrary-1351c59fd43689753de1fca32c78d539a116ffc1-v29f82c9cf21d57b242f8d8b0e541525d259e2d63", + "instance_qutebrowser__qutebrowser-a25e8a09873838ca9efefd36ea8a45170bbeb95c-vc2f56a753b62a190ddb23cd330c257b9cf560d12", + "instance_protonmail__webclients-281a6b3f190f323ec2c0630999354fafb84b2880", + "instance_future-architect__vuls-999529a05b202b0fd29c6fca5039a4c47a3766bb", + "instance_element-hq__element-web-1077729a19c0ce902e713cf6fab42c91fb7907f1-vnan", + "instance_future-architect__vuls-be7b9114cc9545e68fb0ee7bc63d7ec53d1a00ad", + "instance_qutebrowser__qutebrowser-c580ebf0801e5a3ecabc54f327498bb753c6d5f2-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_gravitational__teleport-3ff19cf7c41f396ae468797d3aeb61515517edc9-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_future-architect__vuls-aaea15e516ece43978cf98e09e52080478b1d39f", + "instance_protonmail__webclients-f161c10cf7d31abf82e8d64d7a99c9fac5acfa18", + "instance_qutebrowser__qutebrowser-cf06f4e3708f886032d4d2a30108c2fddb042d81-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_internetarchive__openlibrary-3f580a5f244c299d936d73d9e327ba873b6401d9-v0f5aece3601a5b4419f7ccec1dbda2071be28ee4", + "instance_qutebrowser__qutebrowser-0aa57e4f7243024fa4bba8853306691b5dbd77b3-v5149fcda2a9a6fe1d35dfed1bade1444a11ef271", + "instance_future-architect__vuls-50580f6e98eeb36f53f27222f7f4fdfea0b21e8d" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1778283817.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.6967213114754098, + "uncertainty": { + "standard_error": { + "value": 0.05482325145262507 + }, + "num_samples": 195 + }, + "details": { + "total_matched_trajectories": "195", + "stopping_reason_count_repetition_guard": "195", + "avg_total_tokens_target_model": "3044572.69", + "total_total_tokens_target_model": "593691675.0", + "avg_total_tokens_other_models": "1204.69", + "total_total_tokens_other_models": "234914.0", + "avg_total_tokens_all_models": "3045777.38", + "total_total_tokens_all_models": "593926589.0", + "avg_turn_count": "79.84", + "total_turn_count": "15569.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 30000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/openai/gpt-5.2-2025-12-11/e1a9e523-7d30-4bd1-830a-f088abb67129_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "2f52d3798fcd0a8485a968f567a8e7a1a0bf292bcb0c667ac3920c64ec89ca8a", + "total_rows": 195 + } +} \ No newline at end of file diff --git a/data/swebenchpro/openai/gpt-5.2-2025-12-11/e1a9e523-7d30-4bd1-830a-f088abb67129_samples.jsonl b/data/swebenchpro/openai/gpt-5.2-2025-12-11/e1a9e523-7d30-4bd1-830a-f088abb67129_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..e67d755ac3e481983f1fc87a2a40a0c7e76b250b --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.2-2025-12-11/e1a9e523-7d30-4bd1-830a-f088abb67129_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2f52d3798fcd0a8485a968f567a8e7a1a0bf292bcb0c667ac3920c64ec89ca8a +size 105756159 diff --git a/data/swebenchpro/openai/gpt-5.2-2025-12-11/ead22b89-75e7-4edc-9ed8-697e169d09dc.json b/data/swebenchpro/openai/gpt-5.2-2025-12-11/ead22b89-75e7-4edc-9ed8-697e169d09dc.json new file mode 100644 index 0000000000000000000000000000000000000000..e39d2312298005a6aea3bba29e91acd0f957e571 --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.2-2025-12-11/ead22b89-75e7-4edc-9ed8-697e169d09dc.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/openai_gpt-5.2-2025-12-11/1786167167.213741", + "evaluation_timestamp": "1778757246.0", + "retrieved_timestamp": "1786167167.213741", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/openai/gpt-5.2-2025-12-11/ead22b89-75e7-4edc-9ed8-697e169d09dc_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "39bdddf7a8a93a0a8708f529d46c45885358ebefc70489de32b23b3c865abb52", + "total_rows": 112 + } +} \ No newline at end of file diff --git a/data/swebenchpro/openai/gpt-5.2-2025-12-11/ead22b89-75e7-4edc-9ed8-697e169d09dc_samples.jsonl b/data/swebenchpro/openai/gpt-5.2-2025-12-11/ead22b89-75e7-4edc-9ed8-697e169d09dc_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..7682954931f7a3592c7d7b69a01324ac667a00ff --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.2-2025-12-11/ead22b89-75e7-4edc-9ed8-697e169d09dc_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:39bdddf7a8a93a0a8708f529d46c45885358ebefc70489de32b23b3c865abb52 +size 60556729 diff --git a/data/swebenchpro/openai/gpt-5.2-2025-12-11/eeff4639-f50e-40da-9eae-3f4ac3cb1d46.json b/data/swebenchpro/openai/gpt-5.2-2025-12-11/eeff4639-f50e-40da-9eae-3f4ac3cb1d46.json new file mode 100644 index 0000000000000000000000000000000000000000..1b5af1cc9d3998508bcf141a4c28268c56d3785e --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.2-2025-12-11/eeff4639-f50e-40da-9eae-3f4ac3cb1d46.json @@ -0,0 +1,116 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/openai_gpt-5.2-2025-12-11/1786162882.269892", + "evaluation_timestamp": "1778614654.0", + "retrieved_timestamp": "1786162882.269892", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on swebenchpro/S-adaptive+C/+4ep/1bd20ee0 for scorer harbor_scorer", + "source_data": { + "dataset_name": "swebenchpro", + "source_type": "hf_dataset", + "samples_number": 3, + "sample_ids": [ + "instance_qutebrowser__qutebrowser-2e961080a85d660148937ee8f0f6b3445a8f2c01-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_flipt-io__flipt-b68b8960b8a08540d5198d78c665a7eb0bea4008", + "instance_future-architect__vuls-78b52d6a7f480bd610b692de9bf0c86f57332f23" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1778614654.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.4444444444444444, + "uncertainty": { + "standard_error": { + "value": 0.2939723678960657 + }, + "num_samples": 9 + }, + "details": { + "total_matched_trajectories": "9", + "stopping_reason_count_repetition_guard": "6", + "stopping_reason_count_completed_on_successful_submit": "3", + "avg_total_tokens_target_model": "8432103.89", + "total_total_tokens_target_model": "75888935.0", + "avg_total_tokens_other_models": "759.22", + "total_total_tokens_other_models": "6833.0", + "avg_total_tokens_all_models": "8432863.11", + "total_total_tokens_all_models": "75895768.0", + "avg_turn_count": "146.00", + "total_turn_count": "1314.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 30000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/openai/gpt-5.2-2025-12-11/eeff4639-f50e-40da-9eae-3f4ac3cb1d46_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "a25e9ea969440d414552b63935dd66f15041b3d98e34e66f719d94d212bbe056", + "total_rows": 9 + } +} \ No newline at end of file diff --git a/data/swebenchpro/openai/gpt-5.2-2025-12-11/eeff4639-f50e-40da-9eae-3f4ac3cb1d46_samples.jsonl b/data/swebenchpro/openai/gpt-5.2-2025-12-11/eeff4639-f50e-40da-9eae-3f4ac3cb1d46_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..e3e6d73ede3bdef890bbe68c65511433b4b2b406 --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.2-2025-12-11/eeff4639-f50e-40da-9eae-3f4ac3cb1d46_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a25e9ea969440d414552b63935dd66f15041b3d98e34e66f719d94d212bbe056 +size 7808900 diff --git a/data/swebenchpro/openai/gpt-5.2-2025-12-11/f12d5bf6-7ce1-468d-a509-496c6738c78d.json b/data/swebenchpro/openai/gpt-5.2-2025-12-11/f12d5bf6-7ce1-468d-a509-496c6738c78d.json new file mode 100644 index 0000000000000000000000000000000000000000..34ddd55d43e53e5b6d467ce9c330fce6d1837448 --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.2-2025-12-11/f12d5bf6-7ce1-468d-a509-496c6738c78d.json @@ -0,0 +1,131 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/openai_gpt-5.2-2025-12-11/1786164334.797531", + "evaluation_timestamp": "1778518479.0", + "retrieved_timestamp": "1786164334.797531", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on swebenchpro/S-adaptive/+1ep/99aa6efd for scorer harbor_scorer", + "source_data": { + "dataset_name": "swebenchpro", + "source_type": "hf_dataset", + "samples_number": 19, + "sample_ids": [ + "instance_internetarchive__openlibrary-7cbfb812ef0e1f9716e2d6e85d538a96fcb79d13-vfa6ff903cb27f336e17654595dd900fa943dcd91", + "instance_protonmail__webclients-da91f084c0f532d9cc8ca385a701274d598057b8", + "instance_flipt-io__flipt-b2cd6a6dd73ca91b519015fd5924fde8d17f3f06", + "instance_qutebrowser__qutebrowser-cc360cd4a34a126274c7b51f3b63afbaf3e05a02-v5fc38aaf22415ab0b70567368332beee7955b367", + "instance_flipt-io__flipt-e2bd19dafa7166c96b082fb2a59eb54b4be0d778", + "instance_qutebrowser__qutebrowser-9ed748effa8f3bcd804612d9291da017b514e12f-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_ansible__ansible-1ee70fc272aff6bf3415357c6e13c5de5b928d9b-v1055803c3a812189a1133297f7f5468579283f86", + "instance_element-hq__element-web-75c2c1a572fa45d1ea1d1a96e9e36e303332ecaa-vnan", + "instance_qutebrowser__qutebrowser-f631cd4422744160d9dcf7a0455da532ce973315-v35616345bb8052ea303186706cec663146f0f184", + "instance_ansible__ansible-3db08adbb1cc6aa9941be5e0fc810132c6e1fa4b-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_ansible__ansible-42355d181a11b51ebfc56f6f4b3d9c74e01cb13b-v1055803c3a812189a1133297f7f5468579283f86", + "instance_internetarchive__openlibrary-09865f5fb549694d969f0a8e49b9d204ef1853ca-ve8c8d62a2b60610a3c4631f5f23ed866bada9818", + "instance_internetarchive__openlibrary-c506c1b0b678892af5cb22c1c1dbc35d96787a0a-v0f5aece3601a5b4419f7ccec1dbda2071be28ee4", + "instance_element-hq__element-web-2760bfc8369f1bee640d6d7a7e910783143d4c5f-vnan", + "instance_internetarchive__openlibrary-25858f9f0c165df25742acf8309ce909773f0cdd-v13642507b4fc1f8d234172bf8129942da2c2ca26", + "instance_tutao__tutanota-4b4e45949096bb288f2b522f657610e480efa3e8-vee878bb72091875e912c52fc32bc60ec3760227b", + "instance_protonmail__webclients-5f0745dd6993bb1430a951c62a49807c6635cd77", + "instance_tutao__tutanota-1ff82aa365763cee2d609c9d19360ad87fdf2ec7-vc4e41fd0029957297843cb9dec4a25c7c756f029", + "instance_internetarchive__openlibrary-30bc73a1395fba2300087c7f307e54bb5372b60a-v76304ecdb3a5954fcf13feb710e8c40fcf24b73c" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1778518479.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.7777777777777778, + "uncertainty": { + "standard_error": { + "value": 0.1008316903303367 + }, + "num_samples": 18 + }, + "details": { + "total_matched_trajectories": "18", + "stopping_reason_count_repetition_guard": "18", + "avg_total_tokens_target_model": "2774362.94", + "total_total_tokens_target_model": "49938533.0", + "avg_total_tokens_other_models": "1225.67", + "total_total_tokens_other_models": "22062.0", + "avg_total_tokens_all_models": "2775588.61", + "total_total_tokens_all_models": "49960595.0", + "avg_turn_count": "73.17", + "total_turn_count": "1317.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 30000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/openai/gpt-5.2-2025-12-11/f12d5bf6-7ce1-468d-a509-496c6738c78d_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "4ca2cfb78c195e9c7534ba3fa53040f88759acc643040b504676fad882134117", + "total_rows": 18 + } +} \ No newline at end of file diff --git a/data/swebenchpro/openai/gpt-5.2-2025-12-11/f12d5bf6-7ce1-468d-a509-496c6738c78d_samples.jsonl b/data/swebenchpro/openai/gpt-5.2-2025-12-11/f12d5bf6-7ce1-468d-a509-496c6738c78d_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..3fc6094419b439366a0f3f96667113bed086d22a --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.2-2025-12-11/f12d5bf6-7ce1-468d-a509-496c6738c78d_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4ca2cfb78c195e9c7534ba3fa53040f88759acc643040b504676fad882134117 +size 9128169 diff --git a/data/swebenchpro/openai/gpt-5.2-2025-12-11/f4fa2e20-1f48-4fc4-a386-e6cb6aaeeffb.json b/data/swebenchpro/openai/gpt-5.2-2025-12-11/f4fa2e20-1f48-4fc4-a386-e6cb6aaeeffb.json new file mode 100644 index 0000000000000000000000000000000000000000..c03ad9ebbece9be45655996f3b5c93caa01a0be1 --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.2-2025-12-11/f4fa2e20-1f48-4fc4-a386-e6cb6aaeeffb.json @@ -0,0 +1,110 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/openai_gpt-5.2-2025-12-11/1786162858.24213", + "evaluation_timestamp": "1778680678.0", + "retrieved_timestamp": "1786162858.24213", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on swebenchpro/S-adaptive+C/+1ep/52f921f6 for scorer harbor_scorer", + "source_data": { + "dataset_name": "swebenchpro", + "source_type": "hf_dataset", + "samples_number": 1, + "sample_ids": [ + "instance_future-architect__vuls-78b52d6a7f480bd610b692de9bf0c86f57332f23" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1778680678.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.0, + "uncertainty": { + "num_samples": 1 + }, + "details": { + "total_matched_trajectories": "1", + "stopping_reason_count_repetition_guard": "1", + "avg_total_tokens_target_model": "14058218.00", + "total_total_tokens_target_model": "14058218.0", + "avg_total_tokens_other_models": "1420.00", + "total_total_tokens_other_models": "1420.0", + "avg_total_tokens_all_models": "14059638.00", + "total_total_tokens_all_models": "14059638.0", + "avg_turn_count": "250.00", + "total_turn_count": "250.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 30000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/openai/gpt-5.2-2025-12-11/f4fa2e20-1f48-4fc4-a386-e6cb6aaeeffb_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "e43cef11fb24b550f81ef08fd12aedf36bf5ff8433584bb456e5744e4a47b602", + "total_rows": 1 + } +} \ No newline at end of file diff --git a/data/swebenchpro/openai/gpt-5.2-2025-12-11/f4fa2e20-1f48-4fc4-a386-e6cb6aaeeffb_samples.jsonl b/data/swebenchpro/openai/gpt-5.2-2025-12-11/f4fa2e20-1f48-4fc4-a386-e6cb6aaeeffb_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..f2591f75dc59c72c1bf05c3b1bb06a23889c7ccd --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.2-2025-12-11/f4fa2e20-1f48-4fc4-a386-e6cb6aaeeffb_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e43cef11fb24b550f81ef08fd12aedf36bf5ff8433584bb456e5744e4a47b602 +size 1452401 diff --git a/data/swebenchpro/openai/gpt-5.2-2025-12-11/fde1af56-8ae5-477c-a5ca-5a44023d7310.json b/data/swebenchpro/openai/gpt-5.2-2025-12-11/fde1af56-8ae5-477c-a5ca-5a44023d7310.json new file mode 100644 index 0000000000000000000000000000000000000000..1233c44ef29187c4de7d9586fa3f12d11436f0d1 --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.2-2025-12-11/fde1af56-8ae5-477c-a5ca-5a44023d7310.json @@ -0,0 +1,111 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/openai_gpt-5.2-2025-12-11/1786164370.91623", + "evaluation_timestamp": "1776811129.0", + "retrieved_timestamp": "1786164370.91623", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on swebenchpro/S-adaptive/+2ep/a4a47e32 for scorer harbor_scorer", + "source_data": { + "dataset_name": "swebenchpro", + "source_type": "hf_dataset", + "samples_number": 2, + "sample_ids": [ + "instance_future-architect__vuls-78b52d6a7f480bd610b692de9bf0c86f57332f23", + "instance_nodebb__nodebb-84e065752f6d7fbe5c08cbf50cb173ffb866b8fa-vf2cf3cbd463b7ad942381f1c6d077626485a1e9e" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1776811129.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.75, + "uncertainty": { + "standard_error": { + "value": 0.25 + }, + "num_samples": 4 + }, + "details": { + "total_matched_trajectories": "4", + "stopping_reason_count_repetition_guard": "4", + "avg_total_tokens_target_model": "10773919.50", + "total_total_tokens_target_model": "43095678.0", + "avg_total_tokens_other_models": "1693.25", + "total_total_tokens_other_models": "6773.0", + "avg_total_tokens_all_models": "10775612.75", + "total_total_tokens_all_models": "43102451.0", + "avg_turn_count": "195.25", + "total_turn_count": "781.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 30000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/openai/gpt-5.2-2025-12-11/fde1af56-8ae5-477c-a5ca-5a44023d7310_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "052db73e319bf15383b71d07644f8db518422407df0f8c18906d240ab5432d40", + "total_rows": 4 + } +} \ No newline at end of file diff --git a/data/swebenchpro/openai/gpt-5.2-2025-12-11/fde1af56-8ae5-477c-a5ca-5a44023d7310_samples.jsonl b/data/swebenchpro/openai/gpt-5.2-2025-12-11/fde1af56-8ae5-477c-a5ca-5a44023d7310_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..50b483f9a4e35a3ef7e62b1584ebc8e300f67b89 --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.2-2025-12-11/fde1af56-8ae5-477c-a5ca-5a44023d7310_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:052db73e319bf15383b71d07644f8db518422407df0f8c18906d240ab5432d40 +size 4119037 diff --git a/data/swebenchpro/openai/gpt-5.4-2026-03-05/222c4f2b-c130-496b-8f3b-7dacc5432d8a.json b/data/swebenchpro/openai/gpt-5.4-2026-03-05/222c4f2b-c130-496b-8f3b-7dacc5432d8a.json new file mode 100644 index 0000000000000000000000000000000000000000..aaae78ae68a08b119e24efb52b07b6f75e36c37b --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.4-2026-03-05/222c4f2b-c130-496b-8f3b-7dacc5432d8a.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/openai_gpt-5.4-2026-03-05/1786169064.683579", + "evaluation_timestamp": "1777604979.0", + "retrieved_timestamp": "1786169064.683579", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/openai/gpt-5.4-2026-03-05/222c4f2b-c130-496b-8f3b-7dacc5432d8a_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "ab03e701ad517419b6488054e43f239bfd01820133641faebb324e66c59d4859", + "total_rows": 45 + } +} \ No newline at end of file diff --git a/data/swebenchpro/openai/gpt-5.4-2026-03-05/222c4f2b-c130-496b-8f3b-7dacc5432d8a_samples.jsonl b/data/swebenchpro/openai/gpt-5.4-2026-03-05/222c4f2b-c130-496b-8f3b-7dacc5432d8a_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..dc074235380d49a4fe81992017794c43fbe9f291 --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.4-2026-03-05/222c4f2b-c130-496b-8f3b-7dacc5432d8a_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ab03e701ad517419b6488054e43f239bfd01820133641faebb324e66c59d4859 +size 19582323 diff --git a/data/swebenchpro/openai/gpt-5.4-2026-03-05/2bacefd2-0685-4490-ba01-7763c6cf8f4c.json b/data/swebenchpro/openai/gpt-5.4-2026-03-05/2bacefd2-0685-4490-ba01-7763c6cf8f4c.json new file mode 100644 index 0000000000000000000000000000000000000000..2f65e15e811edad005e5481dc50891fdd50740f7 --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.4-2026-03-05/2bacefd2-0685-4490-ba01-7763c6cf8f4c.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/openai_gpt-5.4-2026-03-05/1786167187.061775", + "evaluation_timestamp": "1778005529.0", + "retrieved_timestamp": "1786167187.061775", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/openai/gpt-5.4-2026-03-05/2bacefd2-0685-4490-ba01-7763c6cf8f4c_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "fd9b05b7dbe6f796b418939744cba2892e55b707aeac2f240e9daa5f378ce064", + "total_rows": 24 + } +} \ No newline at end of file diff --git a/data/swebenchpro/openai/gpt-5.4-2026-03-05/2bacefd2-0685-4490-ba01-7763c6cf8f4c_samples.jsonl b/data/swebenchpro/openai/gpt-5.4-2026-03-05/2bacefd2-0685-4490-ba01-7763c6cf8f4c_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..ab92111fcd31a282f56675271bd4a8323419c49d --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.4-2026-03-05/2bacefd2-0685-4490-ba01-7763c6cf8f4c_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fd9b05b7dbe6f796b418939744cba2892e55b707aeac2f240e9daa5f378ce064 +size 25857631 diff --git a/data/swebenchpro/openai/gpt-5.4-2026-03-05/337cfaa2-0e73-461a-a906-5eadd56fb989.json b/data/swebenchpro/openai/gpt-5.4-2026-03-05/337cfaa2-0e73-461a-a906-5eadd56fb989.json new file mode 100644 index 0000000000000000000000000000000000000000..b1379ea693de359580082e0333b8a9763f3aea08 --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.4-2026-03-05/337cfaa2-0e73-461a-a906-5eadd56fb989.json @@ -0,0 +1,149 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/openai_gpt-5.4-2026-03-05/1786164472.314899", + "evaluation_timestamp": "1777484286.0", + "retrieved_timestamp": "1786164472.314899", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on swebenchpro/S-adaptive/+4ep/ca773696 for scorer harbor_scorer", + "source_data": { + "dataset_name": "swebenchpro", + "source_type": "hf_dataset", + "samples_number": 40, + "sample_ids": [ + "instance_qutebrowser__qutebrowser-2e961080a85d660148937ee8f0f6b3445a8f2c01-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_flipt-io__flipt-b68b8960b8a08540d5198d78c665a7eb0bea4008", + "instance_qutebrowser__qutebrowser-01d1d1494411380d97cac14614a829d3a69cecaf-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_future-architect__vuls-78b52d6a7f480bd610b692de9bf0c86f57332f23", + "instance_flipt-io__flipt-e42da21a07a5ae35835ec54f74004ebd58713874", + "instance_nodebb__nodebb-84e065752f6d7fbe5c08cbf50cb173ffb866b8fa-vf2cf3cbd463b7ad942381f1c6d077626485a1e9e", + "instance_element-hq__element-web-f3534b42df3dcfe36dc48bddbf14034085af6d30-vnan", + "instance_navidrome__navidrome-d8e794317f788198227e10fb667e10496b3eb99a", + "instance_future-architect__vuls-f0b3a8b1db98eb1bd32685f1c36c41a99c3452ed", + "instance_future-architect__vuls-6eff6a9329a65cc412e79b8f82444dfa3d0f0b5a", + "instance_internetarchive__openlibrary-fdbc0d8f418333c7e575c40b661b582c301ef7ac-v13642507b4fc1f8d234172bf8129942da2c2ca26", + "instance_ansible__ansible-b2a289dcbb702003377221e25f62c8a3608f0e89-v173091e2e36d38c978002990795f66cfc0af30ad", + "instance_internetarchive__openlibrary-f8cc11d9c1575fdba5ac66aee0befca970da8d64-v13642507b4fc1f8d234172bf8129942da2c2ca26", + "instance_qutebrowser__qutebrowser-e70f5b03187bdd40e8bf70f5f3ead840f52d1f42-v02ad04386d5238fe2d1a1be450df257370de4b6a", + "instance_tutao__tutanota-d1aa0ecec288bfc800cfb9133b087c4f81ad8b38-vbc0d9ba8f0071fbe982809910959a6ff8884dbbf", + "instance_qutebrowser__qutebrowser-e64622cd2df5b521342cf4a62e0d4cb8f8c9ae5a-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_gravitational__teleport-ac2fb2f9b4fd1896b554d3011df23d3d71295779", + "instance_protonmail__webclients-cb8cc309c6968b0a2a5fe4288d0ae0a969ff31e1", + "instance_ansible__ansible-11c1777d56664b1acb56b387a1ad6aeadef1391d-v0f01c69f1e2528b935359cfe578530722bca2c59", + "instance_ansible__ansible-29aea9ff3466e4cd2ed00524b9e56738d568ce8b-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_nodebb__nodebb-70b4a0e2aebebe8f2f559de6680093d96a697b2f-vnan", + "instance_ansible__ansible-e22e103cdf8edc56ff7d9b848a58f94f1471a263-v1055803c3a812189a1133297f7f5468579283f86", + "instance_qutebrowser__qutebrowser-e15d26630934d0b6415ed2295ac42fd570a57620-va0fd88aac89cde702ec1ba84877234da33adce8a", + "instance_ansible__ansible-164881d871964aa64e0f911d03ae270acbad253c-v390e508d27db7a51eece36bb6d9698b63a5b638a", + "instance_ansible__ansible-502270c804c33d3bc963930dc85e0f4ca359674d-v7eee2454f617569fd6889f2211f75bc02a35f9f8", + "instance_internetarchive__openlibrary-9c392b60e2c6fa1d68cb68084b4b4ff04d0cb35c-v2d9a6c849c60ed19fd0858ce9e40b7cc8e097e59", + "instance_element-hq__element-web-b007ea81b2ccd001b00f332bee65070aa7fc00f9-vnan", + "instance_ansible__ansible-811093f0225caa4dd33890933150a81c6a6d5226-v1055803c3a812189a1133297f7f5468579283f86", + "instance_flipt-io__flipt-cd2f3b0a9d4d8b8a6d3d56afab65851ecdc408e8", + "instance_internetarchive__openlibrary-7cbfb812ef0e1f9716e2d6e85d538a96fcb79d13-vfa6ff903cb27f336e17654595dd900fa943dcd91", + "instance_nodebb__nodebb-0f788b8eaa4bba3c142d171fd941d015c53b65fc-v0ec6d6c2baf3cb4797482ce4829bc25cd5716649", + "instance_flipt-io__flipt-0b119520afca1cf25c470ff4288c464d4510b944", + "instance_element-hq__element-web-f63160f38459fb552d00fcc60d4064977a9095a6-vnan", + "instance_ansible__ansible-f327e65d11bb905ed9f15996024f857a95592629-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_gravitational__teleport-73cc189b0e9636d418c4470ecce0d9af5dae2f02-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_internetarchive__openlibrary-3c48b4bb782189e0858e6c3fc7956046cf3e1cfb-v2d9a6c849c60ed19fd0858ce9e40b7cc8e097e59", + "instance_element-hq__element-web-9bf77963ee5e036d54b2a3ca202fbf6378464a5e-vnan", + "instance_future-architect__vuls-73f0adad95c4d227e2ccfa876c85cc95dd065e13", + "instance_protonmail__webclients-da91f084c0f532d9cc8ca385a701274d598057b8", + "instance_qutebrowser__qutebrowser-cc360cd4a34a126274c7b51f3b63afbaf3e05a02-v5fc38aaf22415ab0b70567368332beee7955b367" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1777484286.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.65625, + "uncertainty": { + "standard_error": { + "value": 0.07199509198015355 + }, + "num_samples": 156 + }, + "details": { + "total_matched_trajectories": "155", + "stopping_reason_count_repetition_guard": "155", + "avg_total_tokens_target_model": "1645271.34", + "total_total_tokens_target_model": "255017057.0", + "avg_total_tokens_other_models": "1204.20", + "total_total_tokens_other_models": "186651.0", + "avg_total_tokens_all_models": "1646475.54", + "total_total_tokens_all_models": "255203708.0", + "avg_turn_count": "36.10", + "total_turn_count": "5595.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "32", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 30000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "32", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/openai/gpt-5.4-2026-03-05/337cfaa2-0e73-461a-a906-5eadd56fb989_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "5d5262ebbfa56b5b27e46063f24c3afcad7edf794a7e28f8ec67090279a841b6", + "total_rows": 156 + } +} \ No newline at end of file diff --git a/data/swebenchpro/openai/gpt-5.4-2026-03-05/337cfaa2-0e73-461a-a906-5eadd56fb989_samples.jsonl b/data/swebenchpro/openai/gpt-5.4-2026-03-05/337cfaa2-0e73-461a-a906-5eadd56fb989_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..06778267b25d75c4a71300f58c83e36a1b5c7bcf --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.4-2026-03-05/337cfaa2-0e73-461a-a906-5eadd56fb989_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5d5262ebbfa56b5b27e46063f24c3afcad7edf794a7e28f8ec67090279a841b6 +size 58693696 diff --git a/data/swebenchpro/openai/gpt-5.4-2026-03-05/3d0e4d75-bad7-416f-99f5-edfc77763370.json b/data/swebenchpro/openai/gpt-5.4-2026-03-05/3d0e4d75-bad7-416f-99f5-edfc77763370.json new file mode 100644 index 0000000000000000000000000000000000000000..92c26500ca3f0110f179f02d7c354839027cf62b --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.4-2026-03-05/3d0e4d75-bad7-416f-99f5-edfc77763370.json @@ -0,0 +1,137 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/openai_gpt-5.4-2026-03-05/1786162925.305382", + "evaluation_timestamp": "1778055257.0", + "retrieved_timestamp": "1786162925.305382", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on swebenchpro/S-adaptive+C/+2ep/d1641e77 for scorer harbor_scorer", + "source_data": { + "dataset_name": "swebenchpro", + "source_type": "hf_dataset", + "samples_number": 24, + "sample_ids": [ + "instance_qutebrowser__qutebrowser-01d1d1494411380d97cac14614a829d3a69cecaf-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_future-architect__vuls-78b52d6a7f480bd610b692de9bf0c86f57332f23", + "instance_flipt-io__flipt-e42da21a07a5ae35835ec54f74004ebd58713874", + "instance_nodebb__nodebb-84e065752f6d7fbe5c08cbf50cb173ffb866b8fa-vf2cf3cbd463b7ad942381f1c6d077626485a1e9e", + "instance_element-hq__element-web-f3534b42df3dcfe36dc48bddbf14034085af6d30-vnan", + "instance_navidrome__navidrome-d8e794317f788198227e10fb667e10496b3eb99a", + "instance_future-architect__vuls-f0b3a8b1db98eb1bd32685f1c36c41a99c3452ed", + "instance_internetarchive__openlibrary-fdbc0d8f418333c7e575c40b661b582c301ef7ac-v13642507b4fc1f8d234172bf8129942da2c2ca26", + "instance_ansible__ansible-b2a289dcbb702003377221e25f62c8a3608f0e89-v173091e2e36d38c978002990795f66cfc0af30ad", + "instance_internetarchive__openlibrary-f8cc11d9c1575fdba5ac66aee0befca970da8d64-v13642507b4fc1f8d234172bf8129942da2c2ca26", + "instance_qutebrowser__qutebrowser-e70f5b03187bdd40e8bf70f5f3ead840f52d1f42-v02ad04386d5238fe2d1a1be450df257370de4b6a", + "instance_tutao__tutanota-d1aa0ecec288bfc800cfb9133b087c4f81ad8b38-vbc0d9ba8f0071fbe982809910959a6ff8884dbbf", + "instance_qutebrowser__qutebrowser-e64622cd2df5b521342cf4a62e0d4cb8f8c9ae5a-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_protonmail__webclients-cb8cc309c6968b0a2a5fe4288d0ae0a969ff31e1", + "instance_ansible__ansible-11c1777d56664b1acb56b387a1ad6aeadef1391d-v0f01c69f1e2528b935359cfe578530722bca2c59", + "instance_ansible__ansible-29aea9ff3466e4cd2ed00524b9e56738d568ce8b-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_nodebb__nodebb-70b4a0e2aebebe8f2f559de6680093d96a697b2f-vnan", + "instance_ansible__ansible-e22e103cdf8edc56ff7d9b848a58f94f1471a263-v1055803c3a812189a1133297f7f5468579283f86", + "instance_qutebrowser__qutebrowser-e15d26630934d0b6415ed2295ac42fd570a57620-va0fd88aac89cde702ec1ba84877234da33adce8a", + "instance_ansible__ansible-164881d871964aa64e0f911d03ae270acbad253c-v390e508d27db7a51eece36bb6d9698b63a5b638a", + "instance_ansible__ansible-502270c804c33d3bc963930dc85e0f4ca359674d-v7eee2454f617569fd6889f2211f75bc02a35f9f8", + "instance_internetarchive__openlibrary-9c392b60e2c6fa1d68cb68084b4b4ff04d0cb35c-v2d9a6c849c60ed19fd0858ce9e40b7cc8e097e59", + "instance_element-hq__element-web-b007ea81b2ccd001b00f332bee65070aa7fc00f9-vnan", + "instance_ansible__ansible-811093f0225caa4dd33890933150a81c6a6d5226-v1055803c3a812189a1133297f7f5468579283f86" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1778055257.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.7916666666666666, + "uncertainty": { + "standard_error": { + "value": 0.07915236331504884 + }, + "num_samples": 48 + }, + "details": { + "total_matched_trajectories": "48", + "stopping_reason_count_completed_on_successful_submit": "38", + "stopping_reason_count_repetition_guard": "10", + "avg_total_tokens_target_model": "1527932.67", + "total_total_tokens_target_model": "73340768.0", + "avg_total_tokens_other_models": "358.15", + "total_total_tokens_other_models": "17191.0", + "avg_total_tokens_all_models": "1528290.81", + "total_total_tokens_all_models": "73357959.0", + "avg_turn_count": "32.15", + "total_turn_count": "1543.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 30000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/openai/gpt-5.4-2026-03-05/3d0e4d75-bad7-416f-99f5-edfc77763370_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "6ec59b5a487458e4d67532ed5353a27f074a5ed6f9bb079e83f5cf5b5bd22e78", + "total_rows": 48 + } +} \ No newline at end of file diff --git a/data/swebenchpro/openai/gpt-5.4-2026-03-05/3d0e4d75-bad7-416f-99f5-edfc77763370_samples.jsonl b/data/swebenchpro/openai/gpt-5.4-2026-03-05/3d0e4d75-bad7-416f-99f5-edfc77763370_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..9f21801cb36b551f2f0451b1a5e84e79be20360f --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.4-2026-03-05/3d0e4d75-bad7-416f-99f5-edfc77763370_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6ec59b5a487458e4d67532ed5353a27f074a5ed6f9bb079e83f5cf5b5bd22e78 +size 18409802 diff --git a/data/swebenchpro/openai/gpt-5.4-2026-03-05/5800ca26-64ad-4f0d-9de7-a88f72bab841.json b/data/swebenchpro/openai/gpt-5.4-2026-03-05/5800ca26-64ad-4f0d-9de7-a88f72bab841.json new file mode 100644 index 0000000000000000000000000000000000000000..12d2a7eecbfd507ce80966a9c60ac60f466518e1 --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.4-2026-03-05/5800ca26-64ad-4f0d-9de7-a88f72bab841.json @@ -0,0 +1,165 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/openai_gpt-5.4-2026-03-05/1786162941.023535", + "evaluation_timestamp": "1778068566.0", + "retrieved_timestamp": "1786162941.023535", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on swebenchpro/S-adaptive+C/+3ep/25e5cc24 for scorer harbor_scorer", + "source_data": { + "dataset_name": "swebenchpro", + "source_type": "hf_dataset", + "samples_number": 52, + "sample_ids": [ + "instance_future-architect__vuls-6eff6a9329a65cc412e79b8f82444dfa3d0f0b5a", + "instance_gravitational__teleport-ac2fb2f9b4fd1896b554d3011df23d3d71295779", + "instance_flipt-io__flipt-cd2f3b0a9d4d8b8a6d3d56afab65851ecdc408e8", + "instance_internetarchive__openlibrary-7cbfb812ef0e1f9716e2d6e85d538a96fcb79d13-vfa6ff903cb27f336e17654595dd900fa943dcd91", + "instance_nodebb__nodebb-0f788b8eaa4bba3c142d171fd941d015c53b65fc-v0ec6d6c2baf3cb4797482ce4829bc25cd5716649", + "instance_flipt-io__flipt-0b119520afca1cf25c470ff4288c464d4510b944", + "instance_element-hq__element-web-f63160f38459fb552d00fcc60d4064977a9095a6-vnan", + "instance_ansible__ansible-f327e65d11bb905ed9f15996024f857a95592629-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_gravitational__teleport-73cc189b0e9636d418c4470ecce0d9af5dae2f02-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_internetarchive__openlibrary-3c48b4bb782189e0858e6c3fc7956046cf3e1cfb-v2d9a6c849c60ed19fd0858ce9e40b7cc8e097e59", + "instance_future-architect__vuls-878c25bf5a9c9fd88ac32eb843f5636834d5712d", + "instance_element-hq__element-web-9bf77963ee5e036d54b2a3ca202fbf6378464a5e-vnan", + "instance_qutebrowser__qutebrowser-70248f256f93ed9b1984494d0a1a919ddd774892-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_future-architect__vuls-73f0adad95c4d227e2ccfa876c85cc95dd065e13", + "instance_protonmail__webclients-da91f084c0f532d9cc8ca385a701274d598057b8", + "instance_flipt-io__flipt-b2cd6a6dd73ca91b519015fd5924fde8d17f3f06", + "instance_qutebrowser__qutebrowser-cc360cd4a34a126274c7b51f3b63afbaf3e05a02-v5fc38aaf22415ab0b70567368332beee7955b367", + "instance_flipt-io__flipt-e2bd19dafa7166c96b082fb2a59eb54b4be0d778", + "instance_qutebrowser__qutebrowser-9ed748effa8f3bcd804612d9291da017b514e12f-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_ansible__ansible-1ee70fc272aff6bf3415357c6e13c5de5b928d9b-v1055803c3a812189a1133297f7f5468579283f86", + "instance_element-hq__element-web-75c2c1a572fa45d1ea1d1a96e9e36e303332ecaa-vnan", + "instance_qutebrowser__qutebrowser-f631cd4422744160d9dcf7a0455da532ce973315-v35616345bb8052ea303186706cec663146f0f184", + "instance_ansible__ansible-3db08adbb1cc6aa9941be5e0fc810132c6e1fa4b-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_qutebrowser__qutebrowser-e5340c449f23608803c286da0563b62f58ba25b0-v059c6fdc75567943479b23ebca7c07b5e9a7f34c", + "instance_ansible__ansible-42355d181a11b51ebfc56f6f4b3d9c74e01cb13b-v1055803c3a812189a1133297f7f5468579283f86", + "instance_internetarchive__openlibrary-09865f5fb549694d969f0a8e49b9d204ef1853ca-ve8c8d62a2b60610a3c4631f5f23ed866bada9818", + "instance_internetarchive__openlibrary-c506c1b0b678892af5cb22c1c1dbc35d96787a0a-v0f5aece3601a5b4419f7ccec1dbda2071be28ee4", + "instance_element-hq__element-web-2760bfc8369f1bee640d6d7a7e910783143d4c5f-vnan", + "instance_internetarchive__openlibrary-25858f9f0c165df25742acf8309ce909773f0cdd-v13642507b4fc1f8d234172bf8129942da2c2ca26", + "instance_protonmail__webclients-01ea5214d11e0df8b7170d91bafd34f23cb0f2b1", + "instance_tutao__tutanota-4b4e45949096bb288f2b522f657610e480efa3e8-vee878bb72091875e912c52fc32bc60ec3760227b", + "instance_flipt-io__flipt-05d7234fa582df632f70a7cd10194d61bd7043b9", + "instance_ansible__ansible-a02e22e902a69aeb465f16bf03f7f5a91b2cb828-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_protonmail__webclients-5f0745dd6993bb1430a951c62a49807c6635cd77", + "instance_tutao__tutanota-1ff82aa365763cee2d609c9d19360ad87fdf2ec7-vc4e41fd0029957297843cb9dec4a25c7c756f029", + "instance_internetarchive__openlibrary-30bc73a1395fba2300087c7f307e54bb5372b60a-v76304ecdb3a5954fcf13feb710e8c40fcf24b73c", + "instance_nodebb__nodebb-8ca65b0c78c67c1653487c02d1135e1b702185e1-vf2cf3cbd463b7ad942381f1c6d077626485a1e9e", + "instance_nodebb__nodebb-22368b996ee0e5f11a5189b400b33af3cc8d925a-v4fbcfae8b15e4ce5d132c408bca69ebb9cf146ed", + "instance_protonmail__webclients-2c3559cad02d1090985dba7e8eb5a129144d9811", + "instance_protonmail__webclients-01b519cd49e6a24d9a05d2eb97f54e420740072e", + "instance_ansible__ansible-77658704217d5f166404fc67997203c25381cb6e-v390e508d27db7a51eece36bb6d9698b63a5b638a", + "instance_qutebrowser__qutebrowser-2dd8966fdcf11972062c540b7a787e4d0de8d372-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_ansible__ansible-cd9c4eb5a6b2bfaf4a6709f001ce3d0c92c1eed2-v0f01c69f1e2528b935359cfe578530722bca2c59", + "instance_protonmail__webclients-ac23d1efa1a6ab7e62724779317ba44c28d78cfd", + "instance_future-architect__vuls-030b2e03525d68d74cb749959aac2d7f3fc0effa", + "instance_nodebb__nodebb-b1f9ad5534bb3a44dab5364f659876a4b7fe34c1-vnan", + "instance_flipt-io__flipt-292fdaca9be39e6a921aaa8874c011d0fdd3e874", + "instance_qutebrowser__qutebrowser-ef5ba1a0360b39f9eff027fbdc57f363597c3c3b-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_ansible__ansible-622a493ae03bd5e5cf517d336fc426e9d12208c7-v906c969b551b346ef54a2c0b41e04f632b7b73c2", + "instance_gravitational__teleport-82185f232ae8974258397e121b3bc2ed0c3729ed-v626ec2a48416b10a88641359a169d99e935ff037", + "instance_ansible__ansible-9a21e247786ebd294dafafca1105fcd770ff46c6-v67cdaa49f89b34e42b69d5b7830b3c3ad3d8803f", + "instance_protonmail__webclients-8afd9ce04c8dde9e150e1c2b50d32e7ee2efa3e7" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1778068566.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.7435897435897437, + "uncertainty": { + "standard_error": { + "value": 0.06045424984564429 + }, + "num_samples": 154 + }, + "details": { + "total_matched_trajectories": "154", + "stopping_reason_count_completed_on_successful_submit": "115", + "stopping_reason_count_repetition_guard": "39", + "avg_total_tokens_target_model": "1606769.95", + "total_total_tokens_target_model": "247442573.0", + "avg_total_tokens_other_models": "433.57", + "total_total_tokens_other_models": "66770.0", + "avg_total_tokens_all_models": "1607203.53", + "total_total_tokens_all_models": "247509343.0", + "avg_turn_count": "33.75", + "total_turn_count": "5197.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 30000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/openai/gpt-5.4-2026-03-05/5800ca26-64ad-4f0d-9de7-a88f72bab841_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "c55f26f540ffbdb836f12e3854d608850c470ae9ad0febce8741821956e87090", + "total_rows": 154 + } +} \ No newline at end of file diff --git a/data/swebenchpro/openai/gpt-5.4-2026-03-05/5800ca26-64ad-4f0d-9de7-a88f72bab841_samples.jsonl b/data/swebenchpro/openai/gpt-5.4-2026-03-05/5800ca26-64ad-4f0d-9de7-a88f72bab841_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..fa1b50f965cc62f0bed4f1b8858bc8661f636420 --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.4-2026-03-05/5800ca26-64ad-4f0d-9de7-a88f72bab841_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c55f26f540ffbdb836f12e3854d608850c470ae9ad0febce8741821956e87090 +size 61827256 diff --git a/data/swebenchpro/openai/gpt-5.4-2026-03-05/59369272-6418-42d8-bda5-67a3751f6a52.json b/data/swebenchpro/openai/gpt-5.4-2026-03-05/59369272-6418-42d8-bda5-67a3751f6a52.json new file mode 100644 index 0000000000000000000000000000000000000000..332e4ddea948c5083e86062a27883c0571c34e7f --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.4-2026-03-05/59369272-6418-42d8-bda5-67a3751f6a52.json @@ -0,0 +1,112 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/openai_gpt-5.4-2026-03-05/1786182979.746332", + "evaluation_timestamp": "1774906336.0", + "retrieved_timestamp": "1786182979.746332", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.191.dev111+g7c0c4345,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on swebenchpro for scorer harbor_scorer", + "source_data": { + "dataset_name": "swebenchpro", + "source_type": "hf_dataset", + "samples_number": 731, + "sample_ids": [ + "instance_internetarchive__openlibrary-5de7de19211e71b29b2f2ba3b1dff2fe065d660f-v08d8e8889ec945ab821fb156c04c7d2e2810debb", + "instance_element-hq__element-web-ca8b1b04effb4fec0e1dd3de8e3198eeb364d50e-vnan", + "instance_qutebrowser__qutebrowser-e64622cd2df5b521342cf4a62e0d4cb8f8c9ae5a-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_ansible__ansible-cd473dfb2fdbc97acf3293c134b21cbbcfa89ec3-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_protonmail__webclients-715dbd4e6999499cd2a576a532d8214f75189116" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1774906336.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.9, + "uncertainty": { + "standard_error": { + "value": 0.10000000000000002 + }, + "num_samples": 10 + }, + "details": { + "total_matched_trajectories": "2", + "stopping_reason_count_repetition_guard": "2", + "avg_total_tokens_target_model": "553977.50", + "total_total_tokens_target_model": "1107955.0", + "avg_total_tokens_other_models": "1076.00", + "total_total_tokens_other_models": "2152.0", + "avg_total_tokens_all_models": "555053.50", + "total_total_tokens_all_models": "1110107.0", + "avg_turn_count": "25.00", + "total_turn_count": "50.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\"}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "3", + "attempt_timeout": "1200", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 20000000 + }, + "sandbox": {}, + "max_attempts": 3 + }, + "additional_details": { + "max_retries": "3" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/openai/gpt-5.4-2026-03-05/59369272-6418-42d8-bda5-67a3751f6a52_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "6f595d0b181922d061ca001724ee59e90a95919f91006b42c985a9e55967c8ae", + "total_rows": 10 + } +} \ No newline at end of file diff --git a/data/swebenchpro/openai/gpt-5.4-2026-03-05/59369272-6418-42d8-bda5-67a3751f6a52_samples.jsonl b/data/swebenchpro/openai/gpt-5.4-2026-03-05/59369272-6418-42d8-bda5-67a3751f6a52_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..99a04b8323b45eec23b8d245ae068324823705f8 --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.4-2026-03-05/59369272-6418-42d8-bda5-67a3751f6a52_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6f595d0b181922d061ca001724ee59e90a95919f91006b42c985a9e55967c8ae +size 4412026 diff --git a/data/swebenchpro/openai/gpt-5.4-2026-03-05/5a6fe42d-00c1-4a74-9a9e-8be039d42e92.json b/data/swebenchpro/openai/gpt-5.4-2026-03-05/5a6fe42d-00c1-4a74-9a9e-8be039d42e92.json new file mode 100644 index 0000000000000000000000000000000000000000..719e5b269eb38fe63dcb72bd3a4bb436b4c1dfc5 --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.4-2026-03-05/5a6fe42d-00c1-4a74-9a9e-8be039d42e92.json @@ -0,0 +1,124 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/openai_gpt-5.4-2026-03-05/1786164459.010637", + "evaluation_timestamp": "1777660860.0", + "retrieved_timestamp": "1786164459.010637", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on swebenchpro/S-adaptive/+2ep/328cb201 for scorer harbor_scorer", + "source_data": { + "dataset_name": "swebenchpro", + "source_type": "hf_dataset", + "samples_number": 15, + "sample_ids": [ + "instance_nodebb__nodebb-a917210c5b2c20637094545401f85783905c074c-vf2cf3cbd463b7ad942381f1c6d077626485a1e9e", + "instance_internetarchive__openlibrary-1351c59fd43689753de1fca32c78d539a116ffc1-v29f82c9cf21d57b242f8d8b0e541525d259e2d63", + "instance_qutebrowser__qutebrowser-a25e8a09873838ca9efefd36ea8a45170bbeb95c-vc2f56a753b62a190ddb23cd330c257b9cf560d12", + "instance_protonmail__webclients-281a6b3f190f323ec2c0630999354fafb84b2880", + "instance_future-architect__vuls-999529a05b202b0fd29c6fca5039a4c47a3766bb", + "instance_element-hq__element-web-1077729a19c0ce902e713cf6fab42c91fb7907f1-vnan", + "instance_future-architect__vuls-be7b9114cc9545e68fb0ee7bc63d7ec53d1a00ad", + "instance_qutebrowser__qutebrowser-c580ebf0801e5a3ecabc54f327498bb753c6d5f2-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_gravitational__teleport-3ff19cf7c41f396ae468797d3aeb61515517edc9-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_future-architect__vuls-aaea15e516ece43978cf98e09e52080478b1d39f", + "instance_protonmail__webclients-f161c10cf7d31abf82e8d64d7a99c9fac5acfa18", + "instance_qutebrowser__qutebrowser-cf06f4e3708f886032d4d2a30108c2fddb042d81-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_internetarchive__openlibrary-3f580a5f244c299d936d73d9e327ba873b6401d9-v0f5aece3601a5b4419f7ccec1dbda2071be28ee4", + "instance_qutebrowser__qutebrowser-0aa57e4f7243024fa4bba8853306691b5dbd77b3-v5149fcda2a9a6fe1d35dfed1bade1444a11ef271", + "instance_future-architect__vuls-50580f6e98eeb36f53f27222f7f4fdfea0b21e8d" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1777660860.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.5333333333333333, + "uncertainty": { + "standard_error": { + "value": 0.1333333333333333 + }, + "num_samples": 26 + }, + "details": { + "total_matched_trajectories": "26", + "stopping_reason_count_repetition_guard": "26", + "avg_total_tokens_target_model": "1312933.54", + "total_total_tokens_target_model": "34136272.0", + "avg_total_tokens_other_models": "1327.81", + "total_total_tokens_other_models": "34523.0", + "avg_total_tokens_all_models": "1314261.35", + "total_total_tokens_all_models": "34170795.0", + "avg_turn_count": "32.31", + "total_turn_count": "840.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 30000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/openai/gpt-5.4-2026-03-05/5a6fe42d-00c1-4a74-9a9e-8be039d42e92_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "906143dda26add83d563993c09a9aaad1ae8806c2ce2efda4a7da83a8134058c", + "total_rows": 26 + } +} \ No newline at end of file diff --git a/data/swebenchpro/openai/gpt-5.4-2026-03-05/5a6fe42d-00c1-4a74-9a9e-8be039d42e92_samples.jsonl b/data/swebenchpro/openai/gpt-5.4-2026-03-05/5a6fe42d-00c1-4a74-9a9e-8be039d42e92_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..3bafc51c46d409962f8faad7f93f884b60af6b50 --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.4-2026-03-05/5a6fe42d-00c1-4a74-9a9e-8be039d42e92_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:906143dda26add83d563993c09a9aaad1ae8806c2ce2efda4a7da83a8134058c +size 8619208 diff --git a/data/swebenchpro/openai/gpt-5.4-2026-03-05/5cd57fa2-c06d-4946-8cb6-3efac5755e1e.json b/data/swebenchpro/openai/gpt-5.4-2026-03-05/5cd57fa2-c06d-4946-8cb6-3efac5755e1e.json new file mode 100644 index 0000000000000000000000000000000000000000..039a1f3c737388425a2ef4e261bd692a7f5affd7 --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.4-2026-03-05/5cd57fa2-c06d-4946-8cb6-3efac5755e1e.json @@ -0,0 +1,291 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/openai_gpt-5.4-2026-03-05/1786182742.390567", + "evaluation_timestamp": "1775602861.0", + "retrieved_timestamp": "1786182742.390567", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.196.dev25+gb973b316,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on eval_task for scorer harbor_scorer", + "source_data": { + "dataset_name": "swebenchpro", + "source_type": "hf_dataset", + "samples_number": 731, + "sample_ids": [ + "instance_flipt-io__flipt-e91615cf07966da41756017a7d571f9fc0fdbe80", + "instance_ansible__ansible-8127abbc298cabf04aaa89a478fc5e5e3432a6fc-v30a923fb5c164d6cd18280c02422f75e611e8fb2", + "instance_protonmail__webclients-da91f084c0f532d9cc8ca385a701274d598057b8", + "instance_qutebrowser__qutebrowser-deeb15d6f009b3ca0c3bd503a7cef07462bd16b4-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_internetarchive__openlibrary-5069b09e5f64428dce59b33455c8bb17fe577070-v8717e18970bcdc4e0d2cea3b1527752b21e74866", + "instance_protonmail__webclients-c5a2089ca2bfe9aa1d85a664b8ad87ef843a1c9c", + "instance_qutebrowser__qutebrowser-fd6790fe8c02b144ab2464f1fc8ab3d02ce3c476-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_navidrome__navidrome-eebfbc5381a1e506ff17b5f1371d1ad83d5fd642", + "instance_navidrome__navidrome-812dc2090f20ac4f8ac271b6ed95be5889d1a3ca", + "instance_nodebb__nodebb-0c81642997ea1d827dbd02c311db9d4976112cd4-vf2cf3cbd463b7ad942381f1c6d077626485a1e9e", + "instance_internetarchive__openlibrary-c506c1b0b678892af5cb22c1c1dbc35d96787a0a-v0f5aece3601a5b4419f7ccec1dbda2071be28ee4", + "instance_flipt-io__flipt-5c7037ececb0bead0a8eb56054e224bcd7ac5922", + "instance_qutebrowser__qutebrowser-01d1d1494411380d97cac14614a829d3a69cecaf-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_internetarchive__openlibrary-7f7e53aa4cf74a4f8549a5bcd4810c527e2f6d7e-v13642507b4fc1f8d234172bf8129942da2c2ca26", + "instance_internetarchive__openlibrary-53e02a22972e9253aeded0e1981e6845e1e521fe-vfa6ff903cb27f336e17654595dd900fa943dcd91", + "instance_ansible__ansible-34db57a47f875d11c4068567b9ec7ace174ec4cf-v1055803c3a812189a1133297f7f5468579283f86", + "instance_qutebrowser__qutebrowser-996487c43e4fcc265b541f9eca1e7930e3c5cf05-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_gravitational__teleport-96019ce0be7a2c8e36363f359eb7c943b41dde70", + "instance_protonmail__webclients-df60460f163fd5c34e844ab9015e3176f1ab1ac0", + "instance_future-architect__vuls-ad2edbb8448e2c41a097f1c0b52696c0f6c5924d", + "instance_ansible__ansible-9142be2f6cabbe6597c9254c5bb9186d17036d55-v0f01c69f1e2528b935359cfe578530722bca2c59", + "instance_element-hq__element-web-33299af5c9b7a7ec5a9c31d578d4ec5b18088fb7-vnan", + "instance_element-hq__element-web-772df3021201d9c73835a626df8dcb6334ad9a3e-vnan", + "instance_internetarchive__openlibrary-1be7de788a444f6255e89c10ef6aa608550604a8-v29f82c9cf21d57b242f8d8b0e541525d259e2d63", + "instance_internetarchive__openlibrary-630221ab686c64e75a2ce253c893c033e4814b2e-v93c53c13d5f9b383ebb411ee7750b49dcd1a34c6", + "instance_gravitational__teleport-f432a71a13e698b6e1c4672a2e9e9c1f32d35c12", + "instance_protonmail__webclients-0d0267c4438cf378bda90bc85eed3a3615871ac4", + "instance_internetarchive__openlibrary-dbbd9d539c6d4fd45d5be9662aa19b6d664b5137-v08d8e8889ec945ab821fb156c04c7d2e2810debb", + "instance_flipt-io__flipt-b2170346dc37cf42fda1386cd630f24821ad2ac5", + "instance_element-hq__element-web-aec454dd6feeb93000380523cbb0b3681c0275fd-vnan", + "instance_internetarchive__openlibrary-4a5d2a7d24c9e4c11d3069220c0685b736d5ecde-v13642507b4fc1f8d234172bf8129942da2c2ca26", + "instance_qutebrowser__qutebrowser-ec2dcfce9eee9f808efc17a1b99e227fc4421dea-v5149fcda2a9a6fe1d35dfed1bade1444a11ef271", + "instance_gravitational__teleport-1330415d33a27594c948a36d9d7701f496229e9f", + "instance_protonmail__webclients-3a6790f480309130b5d6332dce6c9d5ccca13ee3", + "instance_flipt-io__flipt-492cc0b158200089dceede3b1aba0ed28df3fb1d", + "instance_future-architect__vuls-edb324c3d9ec3b107bf947f00e38af99d05b3e16", + "instance_ansible__ansible-bec27fb4c0a40c5f8bbcf26a475704227d65ee73-v30a923fb5c164d6cd18280c02422f75e611e8fb2", + "instance_gravitational__teleport-c1b1c6a1541c478d7777a48fca993cc8206c73b9", + "instance_tutao__tutanota-1e516e989b3c0221f4af6b297d9c0e4c43e4adc3-vbc0d9ba8f0071fbe982809910959a6ff8884dbbf", + "instance_flipt-io__flipt-406f9396ad65696d58865b3a6283109cd4eaf40e", + "instance_future-architect__vuls-139f3a81b66c47e6d8f70ce6c4afe7a9196a6ea8", + "instance_flipt-io__flipt-3b2c25ee8a3ac247c3fad13ad8d64ace34ec8ee7", + "instance_gravitational__teleport-0ac7334939981cf85b9591ac295c3816954e287e", + "instance_qutebrowser__qutebrowser-e15d26630934d0b6415ed2295ac42fd570a57620-va0fd88aac89cde702ec1ba84877234da33adce8a", + "instance_ansible__ansible-5e88cd9972f10b66dd97e1ee684c910c6a2dd25e-v906c969b551b346ef54a2c0b41e04f632b7b73c2", + "instance_element-hq__element-web-923ad4323b2006b2b180544429455ffe7d4a6cc3-vnan", + "instance_protonmail__webclients-4feccbc9990980aee26ea29035f8f931d6089895", + "instance_flipt-io__flipt-a0cbc0cb65ae601270bdbe3f5313e2dfd49c80e4", + "instance_gravitational__teleport-b4e7cd3a5e246736d3fe8d6886af55030b232277", + "instance_ansible__ansible-6cc97447aac5816745278f3735af128afb255c81-v0f01c69f1e2528b935359cfe578530722bca2c59", + "instance_ansible__ansible-f327e65d11bb905ed9f15996024f857a95592629-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_qutebrowser__qutebrowser-5fdc83e5da6222fe61163395baaad7ae57fa2cb4-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_navidrome__navidrome-3982ba725883e71d4e3e618c61d5140eeb8d850a", + "instance_qutebrowser__qutebrowser-ebfe9b7aa0c4ba9d451f993e08955004aaec4345-v059c6fdc75567943479b23ebca7c07b5e9a7f34c", + "instance_ansible__ansible-709484969c8a4ffd74b839a673431a8c5caa6457-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_gravitational__teleport-ad41b3c15414b28a6cec8c25424a19bfa7abd0e9-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_protonmail__webclients-2c3559cad02d1090985dba7e8eb5a129144d9811", + "instance_flipt-io__flipt-5af0757e96dec4962a076376d1bedc79de0d4249", + "instance_element-hq__element-web-56c7fc1948923b4b3f3507799e725ac16bcf8018-vnan", + "instance_flipt-io__flipt-381b90f718435c4694380b5fcd0d5cf8e3b5a25a", + "instance_flipt-io__flipt-6fd0f9e2587f14ac1fdd1c229f0bcae0468c8daa", + "instance_ansible__ansible-4c5ce5a1a9e79a845aff4978cfeb72a0d4ecf7d6-v1055803c3a812189a1133297f7f5468579283f86", + "instance_navidrome__navidrome-7073d18b54da7e53274d11c9e2baef1242e8769e", + "instance_internetarchive__openlibrary-0a90f9f0256e4f933523e9842799e39f95ae29ce-v76304ecdb3a5954fcf13feb710e8c40fcf24b73c", + "instance_nodebb__nodebb-be43cd25974681c9743d424238b7536c357dc8d3-vf2cf3cbd463b7ad942381f1c6d077626485a1e9e", + "instance_flipt-io__flipt-05d7234fa582df632f70a7cd10194d61bd7043b9", + "instance_nodebb__nodebb-76c6e30282906ac664f2c9278fc90999b27b1f48-vd59a5728dfc977f44533186ace531248c2917516", + "instance_qutebrowser__qutebrowser-6b320dc18662580e1313d2548fdd6231d2a97e6d-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_future-architect__vuls-01441351c3407abfc21c48a38e28828e1b504e0c", + "instance_ansible__ansible-1a4644ff15355fd696ac5b9d074a566a80fe7ca3-v30a923fb5c164d6cd18280c02422f75e611e8fb2", + "instance_navidrome__navidrome-09ae41a2da66264c60ef307882362d2e2d8d8b89", + "instance_internetarchive__openlibrary-5c6c22f3d2edf2f1b10f5dc335e32cb6a5f40341-v76304ecdb3a5954fcf13feb710e8c40fcf24b73c", + "instance_navidrome__navidrome-bf2bcb12799b21069f137749e0c331f761d1f693", + "instance_flipt-io__flipt-2ce8a0331e8a8f63f2c1b555db8277ffe5aa2e63", + "instance_ansible__ansible-a1569ea4ca6af5480cf0b7b3135f5e12add28a44-v0f01c69f1e2528b935359cfe578530722bca2c59", + "instance_gravitational__teleport-82185f232ae8974258397e121b3bc2ed0c3729ed-v626ec2a48416b10a88641359a169d99e935ff037", + "instance_flipt-io__flipt-72d06db14d58692bfb4d07b1aa745a37b35956f3", + "instance_ansible__ansible-b6290e1d156af608bd79118d209a64a051c55001-v390e508d27db7a51eece36bb6d9698b63a5b638a", + "instance_gravitational__teleport-32bcd71591c234f0d8b091ec01f1f5cbfdc0f13c-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_ansible__ansible-a02e22e902a69aeb465f16bf03f7f5a91b2cb828-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_gravitational__teleport-eda668c30d9d3b56d9c69197b120b01013611186", + "instance_nodebb__nodebb-0f788b8eaa4bba3c142d171fd941d015c53b65fc-v0ec6d6c2baf3cb4797482ce4829bc25cd5716649", + "instance_gravitational__teleport-47530e1fd8bfb84ec096ebcbbc29990f30829655-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_qutebrowser__qutebrowser-e64622cd2df5b521342cf4a62e0d4cb8f8c9ae5a-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_element-hq__element-web-494d9de6f0a94ffb491e74744d2735bce02dc0ab-vnan", + "instance_element-hq__element-web-9bf77963ee5e036d54b2a3ca202fbf6378464a5e-vnan", + "instance_nodebb__nodebb-9c576a0758690f45a6ca03b5884c601e473bf2c1-vd59a5728dfc977f44533186ace531248c2917516", + "instance_internetarchive__openlibrary-25858f9f0c165df25742acf8309ce909773f0cdd-v13642507b4fc1f8d234172bf8129942da2c2ca26", + "instance_protonmail__webclients-f161c10cf7d31abf82e8d64d7a99c9fac5acfa18", + "instance_gravitational__teleport-cb712e3f0b06dadc679f895daef8072cae400c26-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_gravitational__teleport-46a13210519461c7cec8d643bfbe750265775b41", + "instance_ansible__ansible-e40889e7112ae00a21a2c74312b330e67a766cc0-v1055803c3a812189a1133297f7f5468579283f86", + "instance_gravitational__teleport-6eaaf3a27e64f4ef4ef855bd35d7ec338cf17460-v626ec2a48416b10a88641359a169d99e935ff037", + "instance_internetarchive__openlibrary-0d13e6b4bf80bced6c0946b969b9a1b6963f6bce-v76304ecdb3a5954fcf13feb710e8c40fcf24b73c", + "instance_protonmail__webclients-2f2f6c311c6128fe86976950d3c0c2db07b03921", + "instance_ansible__ansible-5260527c4a71bfed99d803e687dd19619423b134-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_qutebrowser__qutebrowser-8d05f0282a271bfd45e614238bd1b555c58b3fc1-v35616345bb8052ea303186706cec663146f0f184", + "instance_qutebrowser__qutebrowser-50efac08f623644a85441bbe02ab9347d2b71a9d-v9f8e9d96c85c85a605e382f1510bd08563afc566", + "instance_flipt-io__flipt-dae029cba7cdb98dfb1a6b416c00d324241e6063", + "instance_nodebb__nodebb-18c45b44613aecd53e9f60457b9812049ab2998d-v0495b863a912fbff5749c67e860612b91825407c", + "instance_future-architect__vuls-e4728e388120b311c4ed469e4f942e0347a2689b-v264a82e2f4818e30f5a25e4da53b27ba119f62b5", + "instance_qutebrowser__qutebrowser-5e0d6dc1483cb3336ea0e3dcbd4fe4aa00fc1742-v5149fcda2a9a6fe1d35dfed1bade1444a11ef271", + "instance_flipt-io__flipt-c12967bc73fdf02054cf3ef8498c05e25f0a18c0", + "instance_qutebrowser__qutebrowser-ff1c025ad3210506fc76e1f604d8c8c27637d88e-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_ansible__ansible-83909bfa22573777e3db5688773bda59721962ad-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_qutebrowser__qutebrowser-1a9e74bfaf9a9db2a510dc14572d33ded6040a57-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_gravitational__teleport-24cafecd8721891092210afc55f6413ab46ca211-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_internetarchive__openlibrary-4b7ea2977be2747496ba792a678940baa985f7ea-v0f5aece3601a5b4419f7ccec1dbda2071be28ee4", + "instance_internetarchive__openlibrary-fad4a40acf5ff5f06cd7441a5c7baf41a7d81fe4-vfa6ff903cb27f336e17654595dd900fa943dcd91", + "instance_qutebrowser__qutebrowser-3fd8e12949b8feda401930574facf09dd4180bba", + "instance_ansible__ansible-164881d871964aa64e0f911d03ae270acbad253c-v390e508d27db7a51eece36bb6d9698b63a5b638a", + "instance_ansible__ansible-fb144c44144f8bd3542e71f5db62b6d322c7bd85-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_element-hq__element-web-582a1b093fc0b77538052f45cbb9c7295f991b51-vnan", + "instance_element-hq__element-web-ec0f940ef0e8e3b61078f145f34dc40d1938e6c5-vnan", + "instance_internetarchive__openlibrary-6fdbbeee4c0a7e976ff3e46fb1d36f4eb110c428-v08d8e8889ec945ab821fb156c04c7d2e2810debb", + "instance_flipt-io__flipt-7161f7b876773a911afdd804b281e52681cb7321", + "instance_navidrome__navidrome-28389fb05e1523564dfc61fa43ed8eb8a10f938c", + "instance_ansible__ansible-5f4e332e3762999d94af27746db29ff1729252c1-v0f01c69f1e2528b935359cfe578530722bca2c59", + "instance_nodebb__nodebb-0e07f3c9bace416cbab078a30eae972868c0a8a3-vf2cf3cbd463b7ad942381f1c6d077626485a1e9e", + "instance_nodebb__nodebb-da0211b1a001d45d73b4c84c6417a4f1b0312575-vf2cf3cbd463b7ad942381f1c6d077626485a1e9e", + "instance_navidrome__navidrome-31799662706fedddf5bcc1a76b50409d1f91d327", + "instance_future-architect__vuls-50580f6e98eeb36f53f27222f7f4fdfea0b21e8d", + "instance_internetarchive__openlibrary-60725705782832a2cb22e17c49697948a42a9d03-v298a7a812ceed28c4c18355a091f1b268fe56d86", + "instance_ansible__ansible-3db08adbb1cc6aa9941be5e0fc810132c6e1fa4b-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_internetarchive__openlibrary-8a5a63af6e0be406aa6c8c9b6d5f28b2f1b6af5a-v0f5aece3601a5b4419f7ccec1dbda2071be28ee4", + "instance_future-architect__vuls-e1df74cbc1a1d1889428b3333a3b2405c4651993", + "instance_internetarchive__openlibrary-123e6e5e1c85b9c07d1e98f70bfc480bc8016890-v2733ff199fb72f0d033a30dc62cb0a4742e3a7f4", + "instance_protonmail__webclients-ae36cb23a1682dcfd69587c1b311ae0227e28f39", + "instance_protonmail__webclients-8be4f6cb9380fcd2e67bcb18cef931ae0d4b869c", + "instance_qutebrowser__qutebrowser-a25e8a09873838ca9efefd36ea8a45170bbeb95c-vc2f56a753b62a190ddb23cd330c257b9cf560d12", + "instance_qutebrowser__qutebrowser-44e64199ed38003253f0296badd4a447645067b6-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_gravitational__teleport-7744f72c6eb631791434b648ba41083b5f6d2278-vce94f93ad1030e3136852817f2423c1b3ac37bc4", + "instance_tutao__tutanota-40e94dee2bcec2b63f362da283123e9df1874cc1-vc4e41fd0029957297843cb9dec4a25c7c756f029", + "instance_future-architect__vuls-c11ba27509f733d7d280bdf661cbbe2e7a99df4c", + "instance_flipt-io__flipt-b2cd6a6dd73ca91b519015fd5924fde8d17f3f06", + "instance_gravitational__teleport-6a14edcf1ff010172fdbac622d0a474ed6af46de", + "instance_flipt-io__flipt-36e62baffae2132f78f9d34dc300a9baa2d7ae0e", + "instance_navidrome__navidrome-c90468b895f6171e33e937ff20dc915c995274f0", + "instance_ansible__ansible-ecea15c508f0e081525be036cf76bbb56dbcdd9d-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_ansible__ansible-e22e103cdf8edc56ff7d9b848a58f94f1471a263-v1055803c3a812189a1133297f7f5468579283f86", + "instance_future-architect__vuls-dc496468b9e9fb73371f9606cdcdb0f8e12e70ca", + "instance_gravitational__teleport-c782838c3a174fdff80cafd8cd3b1aa4dae8beb2", + "instance_internetarchive__openlibrary-91efee627df01e32007abf2d6ebf73f9d9053076-vbee42ad1b72fb23c6a1c874868a720b370983ed2", + "instance_qutebrowser__qutebrowser-52708364b5f91e198defb022d1a5b4b3ebd9b563-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_ansible__ansible-eea46a0d1b99a6dadedbb6a3502d599235fa7ec3-v390e508d27db7a51eece36bb6d9698b63a5b638a", + "instance_protonmail__webclients-c6f65d205c401350a226bb005f42fac1754b0b5b", + "instance_ansible__ansible-cb94c0cc550df9e98f1247bc71d8c2b861c75049-v1055803c3a812189a1133297f7f5468579283f86", + "instance_nodebb__nodebb-05f2236193f407cf8e2072757fbd6bb170bc13f0-vf2cf3cbd463b7ad942381f1c6d077626485a1e9e", + "instance_navidrome__navidrome-55730514ea59d5f1d0b8e3f8745569c29bdbf7b4", + "instance_navidrome__navidrome-0130c6dc13438b48cf0fdfab08a89e357b5517c9", + "instance_flipt-io__flipt-507170da0f7f4da330f6732bffdf11c4df7fc192", + "instance_tutao__tutanota-fe240cbf7f0fdd6744ef7bef8cb61676bcdbb621-vc4e41fd0029957297843cb9dec4a25c7c756f029", + "instance_qutebrowser__qutebrowser-233cb1cc48635130e5602549856a6fa4ab4c087f-v35616345bb8052ea303186706cec663146f0f184", + "instance_internetarchive__openlibrary-f8cc11d9c1575fdba5ac66aee0befca970da8d64-v13642507b4fc1f8d234172bf8129942da2c2ca26", + "instance_qutebrowser__qutebrowser-fec187c2cb53d769c2682b35ca77858a811414a8-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_tutao__tutanota-da4edb7375c10f47f4ed3860a591c5e6557f7b5c-vbc0d9ba8f0071fbe982809910959a6ff8884dbbf", + "instance_ansible__ansible-be2c376ab87e3e872ca21697508f12c6909cf85a-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_future-architect__vuls-457a3a9627fb9a0800d0aecf1d4713fb634a9011", + "instance_internetarchive__openlibrary-d40ec88713dc95ea791b252f92d2f7b75e107440-v13642507b4fc1f8d234172bf8129942da2c2ca26", + "instance_protonmail__webclients-863d524b5717b9d33ce08a0f0535e3fd8e8d1ed8", + "instance_internetarchive__openlibrary-308a35d6999427c02b1dbf5211c033ad3b352556-ve8c8d62a2b60610a3c4631f5f23ed866bada9818", + "instance_flipt-io__flipt-15b76cada1ef29cfa56b0fba36754be36243dded", + "instance_future-architect__vuls-61c39637f2f3809e1b5dad05f0c57c799dce1587", + "instance_ansible__ansible-5d253a13807e884b7ce0b6b57a963a45e2f0322c-v1055803c3a812189a1133297f7f5468579283f86", + "instance_element-hq__element-web-44b98896a79ede48f5ad7ff22619a39d5f6ff03c-vnan", + "instance_qutebrowser__qutebrowser-f91ace96223cac8161c16dd061907e138fe85111-v059c6fdc75567943479b23ebca7c07b5e9a7f34c", + "instance_gravitational__teleport-bb562408da4adeae16e025be65e170959d1ec492-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_gravitational__teleport-78b0d8c72637df1129fb6ff84fc49ef4b5ab1288", + "instance_internetarchive__openlibrary-3f7db6bbbcc7c418b3db72d157c6aed1d45b2ccf-v430f20c722405e462d9ef44dee7d34c41e76fe7a", + "instance_future-architect__vuls-e52fa8d6ed1d23e36f2a86e5d3efe9aa057a1b0d", + "instance_navidrome__navidrome-3bc9e75b2843f91f6a1e9b604e321c2bd4fd442a", + "instance_tutao__tutanota-fb32e5f9d9fc152a00144d56dd0af01760a2d4dc-vc4e41fd0029957297843cb9dec4a25c7c756f029", + "instance_gravitational__teleport-af5e2517de7d18406b614e413aca61c319312171-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_ansible__ansible-a20a52701402a12f91396549df04ac55809f68e9-v1055803c3a812189a1133297f7f5468579283f86", + "instance_protonmail__webclients-1501eb765873b2884b6f1944fd242ecfc9d6b103", + "instance_gravitational__teleport-db89206db6c2969266e664c7c0fb51b70e958b64", + "instance_ansible__ansible-be59caa59bf47ca78a4760eb7ff38568372a8260-v1055803c3a812189a1133297f7f5468579283f86", + "instance_flipt-io__flipt-65581fef4aa807540cb933753d085feb0d7e736f", + "instance_gravitational__teleport-4e1c39639edf1ab494dd7562844c8b277b5cfa18-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_future-architect__vuls-f0b3a8b1db98eb1bd32685f1c36c41a99c3452ed", + "instance_navidrome__navidrome-b65e76293a917ee2dfc5d4b373b1c62e054d0dca", + "instance_element-hq__element-web-33e8edb3d508d6eefb354819ca693b7accc695e7", + "instance_flipt-io__flipt-e50808c03e4b9d25a6a78af9c61a3b1616ea356b", + "instance_future-architect__vuls-e3c27e1817d68248043bd09d63cc31f3344a6f2c", + "instance_element-hq__element-web-ad26925bb6628260cfe0fcf90ec0a8cba381f4a4-vnan", + "instance_nodebb__nodebb-1ea9481af6125ffd6da0592ed439aa62af0bca11-vd59a5728dfc977f44533186ace531248c2917516", + "instance_ansible__ansible-489156378c8e97374a75a544c7c9c2c0dd8146d1-v390e508d27db7a51eece36bb6d9698b63a5b638a", + "instance_internetarchive__openlibrary-bb152d23c004f3d68986877143bb0f83531fe401-ve8c8d62a2b60610a3c4631f5f23ed866bada9818", + "instance_qutebrowser__qutebrowser-0b621cb0ce2b54d3f93d8d41d8ff4257888a87e5-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_qutebrowser__qutebrowser-1af602b258b97aaba69d2585ed499d95e2303ac2-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_future-architect__vuls-e1fab805afcfc92a2a615371d0ec1e667503c254-v264a82e2f4818e30f5a25e4da53b27ba119f62b5", + "instance_internetarchive__openlibrary-7edd1ef09d91fe0b435707633c5cc9af41dedddf-v76304ecdb3a5954fcf13feb710e8c40fcf24b73c", + "instance_qutebrowser__qutebrowser-99029144b5109bb1b2a53964a7c129e009980cd9-va0fd88aac89cde702ec1ba84877234da33adce8a", + "instance_ansible__ansible-12734fa21c08a0ce8c84e533abdc560db2eb1955-v7eee2454f617569fd6889f2211f75bc02a35f9f8", + "instance_protonmail__webclients-7e54526774e577c0ebb58ced7ba8bef349a69fec", + "instance_gravitational__teleport-1a77b7945a022ab86858029d30ac7ad0d5239d00-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_nodebb__nodebb-8ca65b0c78c67c1653487c02d1135e1b702185e1-vf2cf3cbd463b7ad942381f1c6d077626485a1e9e", + "instance_flipt-io__flipt-9f8127f225a86245fa35dca4885c2daef824ee55", + "instance_gravitational__teleport-ac2fb2f9b4fd1896b554d3011df23d3d71295779", + "instance_ansible__ansible-f02a62db509dc7463fab642c9c3458b9bc3476cc-v390e508d27db7a51eece36bb6d9698b63a5b638a" + ], + "additional_details": { + "shuffled": "True" + } + }, + "evaluation_timestamp": "1775602861.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.51, + "uncertainty": { + "standard_error": { + "value": 0.035436970729343695 + }, + "num_samples": 200 + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"If you do not submit an answer, you will receive a score of zero.\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [10, \"Your answer has been saved. Consider whether your previous approach is likely correct or whether a different approach would be more promising. Then either refine your previous answer or try a substantially different approach, and submit when ready.\", \"\"], \"submit\": null, \"on_continue\": null, \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\"}, \"params_passed\": {\"prompt\": [\"If you do not submit an answer, you will receive a score of zero.\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [10, \"Your answer has been saved. Consider whether your previous approach is likely correct or whether a different approach would be more promising. Then either refine your previous answer or try a substantially different approach, and submit when ready.\", \"\"], \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 20000000 + }, + "sandbox": { + "type": "k8s" + } + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/openai/gpt-5.4-2026-03-05/5cd57fa2-c06d-4946-8cb6-3efac5755e1e_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "27497bf6a1378b7d076629aa2372c1cae565369c0e6ce41ddea85cdc34d73fe4", + "total_rows": 200 + } +} \ No newline at end of file diff --git a/data/swebenchpro/openai/gpt-5.4-2026-03-05/5cd57fa2-c06d-4946-8cb6-3efac5755e1e_samples.jsonl b/data/swebenchpro/openai/gpt-5.4-2026-03-05/5cd57fa2-c06d-4946-8cb6-3efac5755e1e_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..cd0db0613929a3cfdda834d1e55f4ce5969bf3ec --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.4-2026-03-05/5cd57fa2-c06d-4946-8cb6-3efac5755e1e_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:27497bf6a1378b7d076629aa2372c1cae565369c0e6ce41ddea85cdc34d73fe4 +size 135900840 diff --git a/data/swebenchpro/openai/gpt-5.4-2026-03-05/6a6ae453-2acd-4b12-b7d3-bfb08d12f56c.json b/data/swebenchpro/openai/gpt-5.4-2026-03-05/6a6ae453-2acd-4b12-b7d3-bfb08d12f56c.json new file mode 100644 index 0000000000000000000000000000000000000000..8448a861cbbbd8fa23a60b983d635e9e349b8d37 --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.4-2026-03-05/6a6ae453-2acd-4b12-b7d3-bfb08d12f56c.json @@ -0,0 +1,113 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/openai_gpt-5.4-2026-03-05/1786164450.219115", + "evaluation_timestamp": "1777544308.0", + "retrieved_timestamp": "1786164450.219115", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on swebenchpro/S-adaptive/+1ep/951bdb97 for scorer harbor_scorer", + "source_data": { + "dataset_name": "swebenchpro", + "source_type": "hf_dataset", + "samples_number": 4, + "sample_ids": [ + "instance_future-architect__vuls-78b52d6a7f480bd610b692de9bf0c86f57332f23", + "instance_ansible__ansible-e22e103cdf8edc56ff7d9b848a58f94f1471a263-v1055803c3a812189a1133297f7f5468579283f86", + "instance_ansible__ansible-811093f0225caa4dd33890933150a81c6a6d5226-v1055803c3a812189a1133297f7f5468579283f86", + "instance_flipt-io__flipt-cd2f3b0a9d4d8b8a6d3d56afab65851ecdc408e8" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1777544308.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.5, + "uncertainty": { + "standard_error": { + "value": 0.28867513459481287 + }, + "num_samples": 4 + }, + "details": { + "total_matched_trajectories": "4", + "stopping_reason_count_repetition_guard": "4", + "avg_total_tokens_target_model": "3286281.25", + "total_total_tokens_target_model": "13145125.0", + "avg_total_tokens_other_models": "1001.75", + "total_total_tokens_other_models": "4007.0", + "avg_total_tokens_all_models": "3287283.00", + "total_total_tokens_all_models": "13149132.0", + "avg_turn_count": "57.75", + "total_turn_count": "231.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "32", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 30000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "32", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/openai/gpt-5.4-2026-03-05/6a6ae453-2acd-4b12-b7d3-bfb08d12f56c_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "34df445b8b1e6ee106b284f7af4c88d40d38432332cde4b29cc9500e7ecc8464", + "total_rows": 4 + } +} \ No newline at end of file diff --git a/data/swebenchpro/openai/gpt-5.4-2026-03-05/6a6ae453-2acd-4b12-b7d3-bfb08d12f56c_samples.jsonl b/data/swebenchpro/openai/gpt-5.4-2026-03-05/6a6ae453-2acd-4b12-b7d3-bfb08d12f56c_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..2af0e9a39ce5d818a3f18ccdcc9ae93a6169a052 --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.4-2026-03-05/6a6ae453-2acd-4b12-b7d3-bfb08d12f56c_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:34df445b8b1e6ee106b284f7af4c88d40d38432332cde4b29cc9500e7ecc8464 +size 3292981 diff --git a/data/swebenchpro/openai/gpt-5.4-2026-03-05/6efe38f4-e002-485f-b258-0b5bb72b5e19.json b/data/swebenchpro/openai/gpt-5.4-2026-03-05/6efe38f4-e002-485f-b258-0b5bb72b5e19.json new file mode 100644 index 0000000000000000000000000000000000000000..48622d31d2c459f9c2e8d4ceedc084ef322a294b --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.4-2026-03-05/6efe38f4-e002-485f-b258-0b5bb72b5e19.json @@ -0,0 +1,116 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/openai_gpt-5.4-2026-03-05/1786162919.95467", + "evaluation_timestamp": "1778228784.0", + "retrieved_timestamp": "1786162919.95467", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on swebenchpro/S-adaptive+C/+1ep/da629de9 for scorer harbor_scorer", + "source_data": { + "dataset_name": "swebenchpro", + "source_type": "hf_dataset", + "samples_number": 3, + "sample_ids": [ + "instance_flipt-io__flipt-b68b8960b8a08540d5198d78c665a7eb0bea4008", + "instance_internetarchive__openlibrary-7cbfb812ef0e1f9716e2d6e85d538a96fcb79d13-vfa6ff903cb27f336e17654595dd900fa943dcd91", + "instance_flipt-io__flipt-05d7234fa582df632f70a7cd10194d61bd7043b9" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1778228784.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.3333333333333333, + "uncertainty": { + "standard_error": { + "value": 0.33333333333333337 + }, + "num_samples": 3 + }, + "details": { + "total_matched_trajectories": "3", + "stopping_reason_count_repetition_guard": "2", + "stopping_reason_count_completed_on_successful_submit": "1", + "avg_total_tokens_target_model": "1933009.00", + "total_total_tokens_target_model": "5799027.0", + "avg_total_tokens_other_models": "1174.67", + "total_total_tokens_other_models": "3524.0", + "avg_total_tokens_all_models": "1934183.67", + "total_total_tokens_all_models": "5802551.0", + "avg_turn_count": "34.33", + "total_turn_count": "103.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 30000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/openai/gpt-5.4-2026-03-05/6efe38f4-e002-485f-b258-0b5bb72b5e19_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "33f17e52cee31a340b56f31ebc7c81580e732092e8cefd5b9ecc39f0cf3c282f", + "total_rows": 3 + } +} \ No newline at end of file diff --git a/data/swebenchpro/openai/gpt-5.4-2026-03-05/6efe38f4-e002-485f-b258-0b5bb72b5e19_samples.jsonl b/data/swebenchpro/openai/gpt-5.4-2026-03-05/6efe38f4-e002-485f-b258-0b5bb72b5e19_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..3e82bc91bbd587f7faf1c08c0476add14581fdce --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.4-2026-03-05/6efe38f4-e002-485f-b258-0b5bb72b5e19_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:33f17e52cee31a340b56f31ebc7c81580e732092e8cefd5b9ecc39f0cf3c282f +size 1124607 diff --git a/data/swebenchpro/openai/gpt-5.4-2026-03-05/7877053f-34d7-421d-a023-97d60d321485.json b/data/swebenchpro/openai/gpt-5.4-2026-03-05/7877053f-34d7-421d-a023-97d60d321485.json new file mode 100644 index 0000000000000000000000000000000000000000..0267b812276657a16b25c1fae0f73769d01bec2d --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.4-2026-03-05/7877053f-34d7-421d-a023-97d60d321485.json @@ -0,0 +1,110 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/openai_gpt-5.4-2026-03-05/1786162959.951366", + "evaluation_timestamp": "1778165698.0", + "retrieved_timestamp": "1786162959.951366", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on swebenchpro/S-adaptive+C/2e5fec05 for scorer harbor_scorer", + "source_data": { + "dataset_name": "swebenchpro", + "source_type": "hf_dataset", + "samples_number": 1, + "sample_ids": [ + "instance_future-architect__vuls-50580f6e98eeb36f53f27222f7f4fdfea0b21e8d" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1778165698.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.0, + "uncertainty": { + "num_samples": 5 + }, + "details": { + "total_matched_trajectories": "5", + "stopping_reason_count_repetition_guard": "5", + "avg_total_tokens_target_model": "1962140.20", + "total_total_tokens_target_model": "9810701.0", + "avg_total_tokens_other_models": "1668.80", + "total_total_tokens_other_models": "8344.0", + "avg_total_tokens_all_models": "1963809.00", + "total_total_tokens_all_models": "9819045.0", + "avg_turn_count": "39.40", + "total_turn_count": "197.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 30000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/openai/gpt-5.4-2026-03-05/7877053f-34d7-421d-a023-97d60d321485_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "bb5920a052fcd735f711f2e8e9a8e7774d15971ac1c826c055015867f6d7d062", + "total_rows": 5 + } +} \ No newline at end of file diff --git a/data/swebenchpro/openai/gpt-5.4-2026-03-05/7877053f-34d7-421d-a023-97d60d321485_samples.jsonl b/data/swebenchpro/openai/gpt-5.4-2026-03-05/7877053f-34d7-421d-a023-97d60d321485_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..db0db163b53c79fe6ea17a59fb89803de1458ea1 --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.4-2026-03-05/7877053f-34d7-421d-a023-97d60d321485_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bb5920a052fcd735f711f2e8e9a8e7774d15971ac1c826c055015867f6d7d062 +size 2109530 diff --git a/data/swebenchpro/openai/gpt-5.4-2026-03-05/9a5d2a80-14c0-4f89-b436-982011c1d6bb.json b/data/swebenchpro/openai/gpt-5.4-2026-03-05/9a5d2a80-14c0-4f89-b436-982011c1d6bb.json new file mode 100644 index 0000000000000000000000000000000000000000..ce87d2bde7e90f44caaf178cfd412950417c8daa --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.4-2026-03-05/9a5d2a80-14c0-4f89-b436-982011c1d6bb.json @@ -0,0 +1,135 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/openai_gpt-5.4-2026-03-05/1786162954.728646", + "evaluation_timestamp": "1778127146.0", + "retrieved_timestamp": "1786162954.728646", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on swebenchpro/S-adaptive+C/+4ep/86d826b8 for scorer harbor_scorer", + "source_data": { + "dataset_name": "swebenchpro", + "source_type": "hf_dataset", + "samples_number": 22, + "sample_ids": [ + "instance_flipt-io__flipt-b68b8960b8a08540d5198d78c665a7eb0bea4008", + "instance_flipt-io__flipt-b22f5f02e40b225b6b93fff472914973422e97c6", + "instance_gravitational__teleport-2bb3bbbd8aff1164a2353381cb79e1dc93b90d28-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_ansible__ansible-de5858f48dc9e1ce9117034e0d7e76806f420ca8-v1055803c3a812189a1133297f7f5468579283f86", + "instance_gravitational__teleport-46aa81b1ce96ebb4ebed2ae53fd78cd44a05da6c-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_tutao__tutanota-8513a9e8114a8b42e64f4348335e0f23efa054c4-vee878bb72091875e912c52fc32bc60ec3760227b", + "instance_internetarchive__openlibrary-630221ab686c64e75a2ce253c893c033e4814b2e-v93c53c13d5f9b383ebb411ee7750b49dcd1a34c6", + "instance_internetarchive__openlibrary-7f6b722a10f822171501d027cad60afe53337732-ve8c8d62a2b60610a3c4631f5f23ed866bada9818", + "instance_nodebb__nodebb-a917210c5b2c20637094545401f85783905c074c-vf2cf3cbd463b7ad942381f1c6d077626485a1e9e", + "instance_internetarchive__openlibrary-1351c59fd43689753de1fca32c78d539a116ffc1-v29f82c9cf21d57b242f8d8b0e541525d259e2d63", + "instance_qutebrowser__qutebrowser-a25e8a09873838ca9efefd36ea8a45170bbeb95c-vc2f56a753b62a190ddb23cd330c257b9cf560d12", + "instance_protonmail__webclients-281a6b3f190f323ec2c0630999354fafb84b2880", + "instance_future-architect__vuls-999529a05b202b0fd29c6fca5039a4c47a3766bb", + "instance_element-hq__element-web-1077729a19c0ce902e713cf6fab42c91fb7907f1-vnan", + "instance_future-architect__vuls-be7b9114cc9545e68fb0ee7bc63d7ec53d1a00ad", + "instance_qutebrowser__qutebrowser-c580ebf0801e5a3ecabc54f327498bb753c6d5f2-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_gravitational__teleport-3ff19cf7c41f396ae468797d3aeb61515517edc9-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_future-architect__vuls-aaea15e516ece43978cf98e09e52080478b1d39f", + "instance_protonmail__webclients-f161c10cf7d31abf82e8d64d7a99c9fac5acfa18", + "instance_qutebrowser__qutebrowser-cf06f4e3708f886032d4d2a30108c2fddb042d81-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_internetarchive__openlibrary-3f580a5f244c299d936d73d9e327ba873b6401d9-v0f5aece3601a5b4419f7ccec1dbda2071be28ee4", + "instance_qutebrowser__qutebrowser-0aa57e4f7243024fa4bba8853306691b5dbd77b3-v5149fcda2a9a6fe1d35dfed1bade1444a11ef271" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1778127146.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.5454545454545454, + "uncertainty": { + "standard_error": { + "value": 0.10091084589558476 + }, + "num_samples": 87 + }, + "details": { + "total_matched_trajectories": "87", + "stopping_reason_count_repetition_guard": "39", + "stopping_reason_count_completed_on_successful_submit": "48", + "avg_total_tokens_target_model": "1864035.70", + "total_total_tokens_target_model": "162171106.0", + "avg_total_tokens_other_models": "710.38", + "total_total_tokens_other_models": "61803.0", + "avg_total_tokens_all_models": "1864746.08", + "total_total_tokens_all_models": "162232909.0", + "avg_turn_count": "37.15", + "total_turn_count": "3232.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 30000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/openai/gpt-5.4-2026-03-05/9a5d2a80-14c0-4f89-b436-982011c1d6bb_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "c95947b05e211ea011a23d3677050eba7ed6e65e3e1ed4e065be197fef0323be", + "total_rows": 87 + } +} \ No newline at end of file diff --git a/data/swebenchpro/openai/gpt-5.4-2026-03-05/9a5d2a80-14c0-4f89-b436-982011c1d6bb_samples.jsonl b/data/swebenchpro/openai/gpt-5.4-2026-03-05/9a5d2a80-14c0-4f89-b436-982011c1d6bb_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..f8f3ef8c0099a1967bbea500a3e414a1ac3450b8 --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.4-2026-03-05/9a5d2a80-14c0-4f89-b436-982011c1d6bb_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c95947b05e211ea011a23d3677050eba7ed6e65e3e1ed4e065be197fef0323be +size 38725094 diff --git a/data/swebenchpro/openai/gpt-5.4-2026-03-05/b3adb0bf-19b8-4514-b355-30ae1d2d52f1.json b/data/swebenchpro/openai/gpt-5.4-2026-03-05/b3adb0bf-19b8-4514-b355-30ae1d2d52f1.json new file mode 100644 index 0000000000000000000000000000000000000000..7d40b5521d72c2776f71fd05e44d7a933c2f0360 --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.4-2026-03-05/b3adb0bf-19b8-4514-b355-30ae1d2d52f1.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/openai_gpt-5.4-2026-03-05/1786169037.344025", + "evaluation_timestamp": "1777386321.0", + "retrieved_timestamp": "1786169037.344025", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/openai/gpt-5.4-2026-03-05/b3adb0bf-19b8-4514-b355-30ae1d2d52f1_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "277d839ada386d145e128f3d42dcbe9a6ddee03660204d6de1e4193a1da5c321", + "total_rows": 40 + } +} \ No newline at end of file diff --git a/data/swebenchpro/openai/gpt-5.4-2026-03-05/b3adb0bf-19b8-4514-b355-30ae1d2d52f1_samples.jsonl b/data/swebenchpro/openai/gpt-5.4-2026-03-05/b3adb0bf-19b8-4514-b355-30ae1d2d52f1_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..3e0ab150605a9e3d5faf6a1c332f42465268effa --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.4-2026-03-05/b3adb0bf-19b8-4514-b355-30ae1d2d52f1_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:277d839ada386d145e128f3d42dcbe9a6ddee03660204d6de1e4193a1da5c321 +size 15907871 diff --git a/data/swebenchpro/openai/gpt-5.4-2026-03-05/b7a74987-dc3d-4f2c-9c9b-70e1d7e6d889.json b/data/swebenchpro/openai/gpt-5.4-2026-03-05/b7a74987-dc3d-4f2c-9c9b-70e1d7e6d889.json new file mode 100644 index 0000000000000000000000000000000000000000..99831f5e718111221c48db1ac8a942a8944a0f64 --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.4-2026-03-05/b7a74987-dc3d-4f2c-9c9b-70e1d7e6d889.json @@ -0,0 +1,114 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/openai_gpt-5.4-2026-03-05/1786164452.726034", + "evaluation_timestamp": "1777645894.0", + "retrieved_timestamp": "1786164452.726034", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on swebenchpro/S-adaptive/+1ep/d46d0ec1 for scorer harbor_scorer", + "source_data": { + "dataset_name": "swebenchpro", + "source_type": "hf_dataset", + "samples_number": 5, + "sample_ids": [ + "instance_ansible__ansible-de5858f48dc9e1ce9117034e0d7e76806f420ca8-v1055803c3a812189a1133297f7f5468579283f86", + "instance_gravitational__teleport-46aa81b1ce96ebb4ebed2ae53fd78cd44a05da6c-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_tutao__tutanota-8513a9e8114a8b42e64f4348335e0f23efa054c4-vee878bb72091875e912c52fc32bc60ec3760227b", + "instance_internetarchive__openlibrary-630221ab686c64e75a2ce253c893c033e4814b2e-v93c53c13d5f9b383ebb411ee7750b49dcd1a34c6", + "instance_internetarchive__openlibrary-7f6b722a10f822171501d027cad60afe53337732-ve8c8d62a2b60610a3c4631f5f23ed866bada9818" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1777645894.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.6, + "uncertainty": { + "standard_error": { + "value": 0.24494897427831783 + }, + "num_samples": 5 + }, + "details": { + "total_matched_trajectories": "5", + "stopping_reason_count_repetition_guard": "5", + "avg_total_tokens_target_model": "2266649.20", + "total_total_tokens_target_model": "11333246.0", + "avg_total_tokens_other_models": "1110.60", + "total_total_tokens_other_models": "5553.0", + "avg_total_tokens_all_models": "2267759.80", + "total_total_tokens_all_models": "11338799.0", + "avg_turn_count": "41.80", + "total_turn_count": "209.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 30000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/openai/gpt-5.4-2026-03-05/b7a74987-dc3d-4f2c-9c9b-70e1d7e6d889_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "8e65cb6a30a0f58bbc79604178806a4fdd0aa201c2a1ba102e4fddf3d1bc1c6a", + "total_rows": 5 + } +} \ No newline at end of file diff --git a/data/swebenchpro/openai/gpt-5.4-2026-03-05/b7a74987-dc3d-4f2c-9c9b-70e1d7e6d889_samples.jsonl b/data/swebenchpro/openai/gpt-5.4-2026-03-05/b7a74987-dc3d-4f2c-9c9b-70e1d7e6d889_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..d5a27a1857d230c8c32df40d012eeea8b87db649 --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.4-2026-03-05/b7a74987-dc3d-4f2c-9c9b-70e1d7e6d889_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8e65cb6a30a0f58bbc79604178806a4fdd0aa201c2a1ba102e4fddf3d1bc1c6a +size 2800392 diff --git a/data/swebenchpro/openai/gpt-5.4-2026-03-05/c15181f1-6943-42c0-8dab-d6566a1ee044.json b/data/swebenchpro/openai/gpt-5.4-2026-03-05/c15181f1-6943-42c0-8dab-d6566a1ee044.json new file mode 100644 index 0000000000000000000000000000000000000000..68885f86655835ad4726ef9f7d2ee3c908dfbba1 --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.4-2026-03-05/c15181f1-6943-42c0-8dab-d6566a1ee044.json @@ -0,0 +1,130 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/openai_gpt-5.4-2026-03-05/1786164447.463511", + "evaluation_timestamp": "1777596448.0", + "retrieved_timestamp": "1786164447.463511", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on swebenchpro/S-adaptive/+1ep/2080fcf4 for scorer harbor_scorer", + "source_data": { + "dataset_name": "swebenchpro", + "source_type": "hf_dataset", + "samples_number": 21, + "sample_ids": [ + "instance_ansible__ansible-3db08adbb1cc6aa9941be5e0fc810132c6e1fa4b-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_protonmail__webclients-5f0745dd6993bb1430a951c62a49807c6635cd77", + "instance_tutao__tutanota-1ff82aa365763cee2d609c9d19360ad87fdf2ec7-vc4e41fd0029957297843cb9dec4a25c7c756f029", + "instance_nodebb__nodebb-8ca65b0c78c67c1653487c02d1135e1b702185e1-vf2cf3cbd463b7ad942381f1c6d077626485a1e9e", + "instance_nodebb__nodebb-22368b996ee0e5f11a5189b400b33af3cc8d925a-v4fbcfae8b15e4ce5d132c408bca69ebb9cf146ed", + "instance_protonmail__webclients-2c3559cad02d1090985dba7e8eb5a129144d9811", + "instance_protonmail__webclients-01b519cd49e6a24d9a05d2eb97f54e420740072e", + "instance_ansible__ansible-77658704217d5f166404fc67997203c25381cb6e-v390e508d27db7a51eece36bb6d9698b63a5b638a", + "instance_qutebrowser__qutebrowser-2dd8966fdcf11972062c540b7a787e4d0de8d372-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_ansible__ansible-cd9c4eb5a6b2bfaf4a6709f001ce3d0c92c1eed2-v0f01c69f1e2528b935359cfe578530722bca2c59", + "instance_protonmail__webclients-ac23d1efa1a6ab7e62724779317ba44c28d78cfd", + "instance_future-architect__vuls-030b2e03525d68d74cb749959aac2d7f3fc0effa", + "instance_nodebb__nodebb-b1f9ad5534bb3a44dab5364f659876a4b7fe34c1-vnan", + "instance_flipt-io__flipt-292fdaca9be39e6a921aaa8874c011d0fdd3e874", + "instance_qutebrowser__qutebrowser-ef5ba1a0360b39f9eff027fbdc57f363597c3c3b-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_ansible__ansible-622a493ae03bd5e5cf517d336fc426e9d12208c7-v906c969b551b346ef54a2c0b41e04f632b7b73c2", + "instance_gravitational__teleport-82185f232ae8974258397e121b3bc2ed0c3729ed-v626ec2a48416b10a88641359a169d99e935ff037", + "instance_ansible__ansible-9a21e247786ebd294dafafca1105fcd770ff46c6-v67cdaa49f89b34e42b69d5b7830b3c3ad3d8803f", + "instance_protonmail__webclients-8afd9ce04c8dde9e150e1c2b50d32e7ee2efa3e7", + "instance_flipt-io__flipt-b22f5f02e40b225b6b93fff472914973422e97c6", + "instance_gravitational__teleport-2bb3bbbd8aff1164a2353381cb79e1dc93b90d28-vee9b09fb20c43af7e520f57e9239bbcf46b7113d" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1777596448.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.7619047619047619, + "uncertainty": { + "standard_error": { + "value": 0.09523809523809523 + }, + "num_samples": 21 + }, + "details": { + "total_matched_trajectories": "21", + "stopping_reason_count_repetition_guard": "21", + "avg_total_tokens_target_model": "1828129.43", + "total_total_tokens_target_model": "38390718.0", + "avg_total_tokens_other_models": "1411.33", + "total_total_tokens_other_models": "29638.0", + "avg_total_tokens_all_models": "1829540.76", + "total_total_tokens_all_models": "38420356.0", + "avg_turn_count": "39.24", + "total_turn_count": "824.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 30000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/openai/gpt-5.4-2026-03-05/c15181f1-6943-42c0-8dab-d6566a1ee044_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "3097125ca7b86e7673440ea3c19566aa11e94e95d3af3c21a9d8f88b0720e7c8", + "total_rows": 21 + } +} \ No newline at end of file diff --git a/data/swebenchpro/openai/gpt-5.4-2026-03-05/c15181f1-6943-42c0-8dab-d6566a1ee044_samples.jsonl b/data/swebenchpro/openai/gpt-5.4-2026-03-05/c15181f1-6943-42c0-8dab-d6566a1ee044_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..39ff4ad0e23fcd21eb0f5aaa65c5b56fdf869392 --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.4-2026-03-05/c15181f1-6943-42c0-8dab-d6566a1ee044_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3097125ca7b86e7673440ea3c19566aa11e94e95d3af3c21a9d8f88b0720e7c8 +size 8740398 diff --git a/data/swebenchpro/openai/gpt-5.4-2026-03-05/cbdff8d4-ca89-4cbc-b8e3-f6dfec020a61.json b/data/swebenchpro/openai/gpt-5.4-2026-03-05/cbdff8d4-ca89-4cbc-b8e3-f6dfec020a61.json new file mode 100644 index 0000000000000000000000000000000000000000..381d7861b582104c770f10d3da72d29d0ebd8f09 --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.4-2026-03-05/cbdff8d4-ca89-4cbc-b8e3-f6dfec020a61.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/openai_gpt-5.4-2026-03-05/1786169088.862982", + "evaluation_timestamp": "1777544308.0", + "retrieved_timestamp": "1786169088.862982", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/openai/gpt-5.4-2026-03-05/cbdff8d4-ca89-4cbc-b8e3-f6dfec020a61_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "65263e243cae67c6030512edaa4c27a0e9b76998e49e9c49adaacb6f6e375e12", + "total_rows": 99 + } +} \ No newline at end of file diff --git a/data/swebenchpro/openai/gpt-5.4-2026-03-05/cbdff8d4-ca89-4cbc-b8e3-f6dfec020a61_samples.jsonl b/data/swebenchpro/openai/gpt-5.4-2026-03-05/cbdff8d4-ca89-4cbc-b8e3-f6dfec020a61_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..cd08c8704ff49b0806befbbfd4d89edd93295df3 --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.4-2026-03-05/cbdff8d4-ca89-4cbc-b8e3-f6dfec020a61_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:65263e243cae67c6030512edaa4c27a0e9b76998e49e9c49adaacb6f6e375e12 +size 40749080 diff --git a/data/swebenchpro/openai/gpt-5.4-2026-03-05/cd858f17-8963-42bb-a4fb-f4bb023e26a0.json b/data/swebenchpro/openai/gpt-5.4-2026-03-05/cd858f17-8963-42bb-a4fb-f4bb023e26a0.json new file mode 100644 index 0000000000000000000000000000000000000000..cd07c9ba6b230ae6b715b9591fcfc82a934ec021 --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.4-2026-03-05/cd858f17-8963-42bb-a4fb-f4bb023e26a0.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/openai_gpt-5.4-2026-03-05/1786167236.470848", + "evaluation_timestamp": "1777899990.0", + "retrieved_timestamp": "1786167236.470848", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/openai/gpt-5.4-2026-03-05/cd858f17-8963-42bb-a4fb-f4bb023e26a0_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "063ccd072b068bb0681567535e3524b2329b42012fec1186510d3e75e0b255c7", + "total_rows": 76 + } +} \ No newline at end of file diff --git a/data/swebenchpro/openai/gpt-5.4-2026-03-05/cd858f17-8963-42bb-a4fb-f4bb023e26a0_samples.jsonl b/data/swebenchpro/openai/gpt-5.4-2026-03-05/cd858f17-8963-42bb-a4fb-f4bb023e26a0_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..7a834afaf48501decb02032a3a80982af518e808 --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.4-2026-03-05/cd858f17-8963-42bb-a4fb-f4bb023e26a0_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:063ccd072b068bb0681567535e3524b2329b42012fec1186510d3e75e0b255c7 +size 102035992 diff --git a/data/swebenchpro/openai/gpt-5.4-2026-03-05/ce84dd29-99b4-4b40-a9c4-76d3ffa92366.json b/data/swebenchpro/openai/gpt-5.4-2026-03-05/ce84dd29-99b4-4b40-a9c4-76d3ffa92366.json new file mode 100644 index 0000000000000000000000000000000000000000..6feaeb5349e176eb16ea941294f0ad09a46432dd --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.4-2026-03-05/ce84dd29-99b4-4b40-a9c4-76d3ffa92366.json @@ -0,0 +1,291 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/openai_gpt-5.4-2026-03-05/1786181992.743573", + "evaluation_timestamp": "1775482110.0", + "retrieved_timestamp": "1786181992.743573", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev7+g741aa0d2,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on eval_task for scorer harbor_scorer", + "source_data": { + "dataset_name": "swebenchpro", + "source_type": "hf_dataset", + "samples_number": 731, + "sample_ids": [ + "instance_flipt-io__flipt-e91615cf07966da41756017a7d571f9fc0fdbe80", + "instance_ansible__ansible-8127abbc298cabf04aaa89a478fc5e5e3432a6fc-v30a923fb5c164d6cd18280c02422f75e611e8fb2", + "instance_protonmail__webclients-da91f084c0f532d9cc8ca385a701274d598057b8", + "instance_qutebrowser__qutebrowser-deeb15d6f009b3ca0c3bd503a7cef07462bd16b4-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_internetarchive__openlibrary-5069b09e5f64428dce59b33455c8bb17fe577070-v8717e18970bcdc4e0d2cea3b1527752b21e74866", + "instance_protonmail__webclients-c5a2089ca2bfe9aa1d85a664b8ad87ef843a1c9c", + "instance_qutebrowser__qutebrowser-fd6790fe8c02b144ab2464f1fc8ab3d02ce3c476-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_navidrome__navidrome-eebfbc5381a1e506ff17b5f1371d1ad83d5fd642", + "instance_navidrome__navidrome-812dc2090f20ac4f8ac271b6ed95be5889d1a3ca", + "instance_nodebb__nodebb-0c81642997ea1d827dbd02c311db9d4976112cd4-vf2cf3cbd463b7ad942381f1c6d077626485a1e9e", + "instance_internetarchive__openlibrary-c506c1b0b678892af5cb22c1c1dbc35d96787a0a-v0f5aece3601a5b4419f7ccec1dbda2071be28ee4", + "instance_flipt-io__flipt-5c7037ececb0bead0a8eb56054e224bcd7ac5922", + "instance_qutebrowser__qutebrowser-01d1d1494411380d97cac14614a829d3a69cecaf-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_internetarchive__openlibrary-7f7e53aa4cf74a4f8549a5bcd4810c527e2f6d7e-v13642507b4fc1f8d234172bf8129942da2c2ca26", + "instance_internetarchive__openlibrary-53e02a22972e9253aeded0e1981e6845e1e521fe-vfa6ff903cb27f336e17654595dd900fa943dcd91", + "instance_ansible__ansible-34db57a47f875d11c4068567b9ec7ace174ec4cf-v1055803c3a812189a1133297f7f5468579283f86", + "instance_qutebrowser__qutebrowser-996487c43e4fcc265b541f9eca1e7930e3c5cf05-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_gravitational__teleport-96019ce0be7a2c8e36363f359eb7c943b41dde70", + "instance_protonmail__webclients-df60460f163fd5c34e844ab9015e3176f1ab1ac0", + "instance_future-architect__vuls-ad2edbb8448e2c41a097f1c0b52696c0f6c5924d", + "instance_ansible__ansible-9142be2f6cabbe6597c9254c5bb9186d17036d55-v0f01c69f1e2528b935359cfe578530722bca2c59", + "instance_element-hq__element-web-33299af5c9b7a7ec5a9c31d578d4ec5b18088fb7-vnan", + "instance_element-hq__element-web-772df3021201d9c73835a626df8dcb6334ad9a3e-vnan", + "instance_internetarchive__openlibrary-1be7de788a444f6255e89c10ef6aa608550604a8-v29f82c9cf21d57b242f8d8b0e541525d259e2d63", + "instance_internetarchive__openlibrary-630221ab686c64e75a2ce253c893c033e4814b2e-v93c53c13d5f9b383ebb411ee7750b49dcd1a34c6", + "instance_gravitational__teleport-f432a71a13e698b6e1c4672a2e9e9c1f32d35c12", + "instance_protonmail__webclients-0d0267c4438cf378bda90bc85eed3a3615871ac4", + "instance_internetarchive__openlibrary-dbbd9d539c6d4fd45d5be9662aa19b6d664b5137-v08d8e8889ec945ab821fb156c04c7d2e2810debb", + "instance_flipt-io__flipt-b2170346dc37cf42fda1386cd630f24821ad2ac5", + "instance_element-hq__element-web-aec454dd6feeb93000380523cbb0b3681c0275fd-vnan", + "instance_internetarchive__openlibrary-4a5d2a7d24c9e4c11d3069220c0685b736d5ecde-v13642507b4fc1f8d234172bf8129942da2c2ca26", + "instance_qutebrowser__qutebrowser-ec2dcfce9eee9f808efc17a1b99e227fc4421dea-v5149fcda2a9a6fe1d35dfed1bade1444a11ef271", + "instance_gravitational__teleport-1330415d33a27594c948a36d9d7701f496229e9f", + "instance_protonmail__webclients-3a6790f480309130b5d6332dce6c9d5ccca13ee3", + "instance_flipt-io__flipt-492cc0b158200089dceede3b1aba0ed28df3fb1d", + "instance_future-architect__vuls-edb324c3d9ec3b107bf947f00e38af99d05b3e16", + "instance_ansible__ansible-bec27fb4c0a40c5f8bbcf26a475704227d65ee73-v30a923fb5c164d6cd18280c02422f75e611e8fb2", + "instance_gravitational__teleport-c1b1c6a1541c478d7777a48fca993cc8206c73b9", + "instance_tutao__tutanota-1e516e989b3c0221f4af6b297d9c0e4c43e4adc3-vbc0d9ba8f0071fbe982809910959a6ff8884dbbf", + "instance_flipt-io__flipt-406f9396ad65696d58865b3a6283109cd4eaf40e", + "instance_future-architect__vuls-139f3a81b66c47e6d8f70ce6c4afe7a9196a6ea8", + "instance_flipt-io__flipt-3b2c25ee8a3ac247c3fad13ad8d64ace34ec8ee7", + "instance_gravitational__teleport-0ac7334939981cf85b9591ac295c3816954e287e", + "instance_qutebrowser__qutebrowser-e15d26630934d0b6415ed2295ac42fd570a57620-va0fd88aac89cde702ec1ba84877234da33adce8a", + "instance_ansible__ansible-5e88cd9972f10b66dd97e1ee684c910c6a2dd25e-v906c969b551b346ef54a2c0b41e04f632b7b73c2", + "instance_element-hq__element-web-923ad4323b2006b2b180544429455ffe7d4a6cc3-vnan", + "instance_protonmail__webclients-4feccbc9990980aee26ea29035f8f931d6089895", + "instance_flipt-io__flipt-a0cbc0cb65ae601270bdbe3f5313e2dfd49c80e4", + "instance_gravitational__teleport-b4e7cd3a5e246736d3fe8d6886af55030b232277", + "instance_ansible__ansible-6cc97447aac5816745278f3735af128afb255c81-v0f01c69f1e2528b935359cfe578530722bca2c59", + "instance_ansible__ansible-f327e65d11bb905ed9f15996024f857a95592629-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_qutebrowser__qutebrowser-5fdc83e5da6222fe61163395baaad7ae57fa2cb4-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_navidrome__navidrome-3982ba725883e71d4e3e618c61d5140eeb8d850a", + "instance_qutebrowser__qutebrowser-ebfe9b7aa0c4ba9d451f993e08955004aaec4345-v059c6fdc75567943479b23ebca7c07b5e9a7f34c", + "instance_ansible__ansible-709484969c8a4ffd74b839a673431a8c5caa6457-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_gravitational__teleport-ad41b3c15414b28a6cec8c25424a19bfa7abd0e9-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_protonmail__webclients-2c3559cad02d1090985dba7e8eb5a129144d9811", + "instance_flipt-io__flipt-5af0757e96dec4962a076376d1bedc79de0d4249", + "instance_element-hq__element-web-56c7fc1948923b4b3f3507799e725ac16bcf8018-vnan", + "instance_flipt-io__flipt-381b90f718435c4694380b5fcd0d5cf8e3b5a25a", + "instance_flipt-io__flipt-6fd0f9e2587f14ac1fdd1c229f0bcae0468c8daa", + "instance_ansible__ansible-4c5ce5a1a9e79a845aff4978cfeb72a0d4ecf7d6-v1055803c3a812189a1133297f7f5468579283f86", + "instance_navidrome__navidrome-7073d18b54da7e53274d11c9e2baef1242e8769e", + "instance_internetarchive__openlibrary-0a90f9f0256e4f933523e9842799e39f95ae29ce-v76304ecdb3a5954fcf13feb710e8c40fcf24b73c", + "instance_nodebb__nodebb-be43cd25974681c9743d424238b7536c357dc8d3-vf2cf3cbd463b7ad942381f1c6d077626485a1e9e", + "instance_flipt-io__flipt-05d7234fa582df632f70a7cd10194d61bd7043b9", + "instance_nodebb__nodebb-76c6e30282906ac664f2c9278fc90999b27b1f48-vd59a5728dfc977f44533186ace531248c2917516", + "instance_qutebrowser__qutebrowser-6b320dc18662580e1313d2548fdd6231d2a97e6d-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_future-architect__vuls-01441351c3407abfc21c48a38e28828e1b504e0c", + "instance_ansible__ansible-1a4644ff15355fd696ac5b9d074a566a80fe7ca3-v30a923fb5c164d6cd18280c02422f75e611e8fb2", + "instance_navidrome__navidrome-09ae41a2da66264c60ef307882362d2e2d8d8b89", + "instance_internetarchive__openlibrary-5c6c22f3d2edf2f1b10f5dc335e32cb6a5f40341-v76304ecdb3a5954fcf13feb710e8c40fcf24b73c", + "instance_navidrome__navidrome-bf2bcb12799b21069f137749e0c331f761d1f693", + "instance_flipt-io__flipt-2ce8a0331e8a8f63f2c1b555db8277ffe5aa2e63", + "instance_ansible__ansible-a1569ea4ca6af5480cf0b7b3135f5e12add28a44-v0f01c69f1e2528b935359cfe578530722bca2c59", + "instance_gravitational__teleport-82185f232ae8974258397e121b3bc2ed0c3729ed-v626ec2a48416b10a88641359a169d99e935ff037", + "instance_flipt-io__flipt-72d06db14d58692bfb4d07b1aa745a37b35956f3", + "instance_ansible__ansible-b6290e1d156af608bd79118d209a64a051c55001-v390e508d27db7a51eece36bb6d9698b63a5b638a", + "instance_gravitational__teleport-32bcd71591c234f0d8b091ec01f1f5cbfdc0f13c-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_ansible__ansible-a02e22e902a69aeb465f16bf03f7f5a91b2cb828-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_gravitational__teleport-eda668c30d9d3b56d9c69197b120b01013611186", + "instance_nodebb__nodebb-0f788b8eaa4bba3c142d171fd941d015c53b65fc-v0ec6d6c2baf3cb4797482ce4829bc25cd5716649", + "instance_gravitational__teleport-47530e1fd8bfb84ec096ebcbbc29990f30829655-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_qutebrowser__qutebrowser-e64622cd2df5b521342cf4a62e0d4cb8f8c9ae5a-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_element-hq__element-web-494d9de6f0a94ffb491e74744d2735bce02dc0ab-vnan", + "instance_element-hq__element-web-9bf77963ee5e036d54b2a3ca202fbf6378464a5e-vnan", + "instance_nodebb__nodebb-9c576a0758690f45a6ca03b5884c601e473bf2c1-vd59a5728dfc977f44533186ace531248c2917516", + "instance_internetarchive__openlibrary-25858f9f0c165df25742acf8309ce909773f0cdd-v13642507b4fc1f8d234172bf8129942da2c2ca26", + "instance_protonmail__webclients-f161c10cf7d31abf82e8d64d7a99c9fac5acfa18", + "instance_gravitational__teleport-cb712e3f0b06dadc679f895daef8072cae400c26-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_gravitational__teleport-46a13210519461c7cec8d643bfbe750265775b41", + "instance_ansible__ansible-e40889e7112ae00a21a2c74312b330e67a766cc0-v1055803c3a812189a1133297f7f5468579283f86", + "instance_gravitational__teleport-6eaaf3a27e64f4ef4ef855bd35d7ec338cf17460-v626ec2a48416b10a88641359a169d99e935ff037", + "instance_internetarchive__openlibrary-0d13e6b4bf80bced6c0946b969b9a1b6963f6bce-v76304ecdb3a5954fcf13feb710e8c40fcf24b73c", + "instance_protonmail__webclients-2f2f6c311c6128fe86976950d3c0c2db07b03921", + "instance_ansible__ansible-5260527c4a71bfed99d803e687dd19619423b134-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_qutebrowser__qutebrowser-8d05f0282a271bfd45e614238bd1b555c58b3fc1-v35616345bb8052ea303186706cec663146f0f184", + "instance_qutebrowser__qutebrowser-50efac08f623644a85441bbe02ab9347d2b71a9d-v9f8e9d96c85c85a605e382f1510bd08563afc566", + "instance_flipt-io__flipt-dae029cba7cdb98dfb1a6b416c00d324241e6063", + "instance_nodebb__nodebb-18c45b44613aecd53e9f60457b9812049ab2998d-v0495b863a912fbff5749c67e860612b91825407c", + "instance_future-architect__vuls-e4728e388120b311c4ed469e4f942e0347a2689b-v264a82e2f4818e30f5a25e4da53b27ba119f62b5", + "instance_qutebrowser__qutebrowser-5e0d6dc1483cb3336ea0e3dcbd4fe4aa00fc1742-v5149fcda2a9a6fe1d35dfed1bade1444a11ef271", + "instance_flipt-io__flipt-c12967bc73fdf02054cf3ef8498c05e25f0a18c0", + "instance_qutebrowser__qutebrowser-ff1c025ad3210506fc76e1f604d8c8c27637d88e-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_ansible__ansible-83909bfa22573777e3db5688773bda59721962ad-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_qutebrowser__qutebrowser-1a9e74bfaf9a9db2a510dc14572d33ded6040a57-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_gravitational__teleport-24cafecd8721891092210afc55f6413ab46ca211-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_internetarchive__openlibrary-4b7ea2977be2747496ba792a678940baa985f7ea-v0f5aece3601a5b4419f7ccec1dbda2071be28ee4", + "instance_internetarchive__openlibrary-fad4a40acf5ff5f06cd7441a5c7baf41a7d81fe4-vfa6ff903cb27f336e17654595dd900fa943dcd91", + "instance_qutebrowser__qutebrowser-3fd8e12949b8feda401930574facf09dd4180bba", + "instance_ansible__ansible-164881d871964aa64e0f911d03ae270acbad253c-v390e508d27db7a51eece36bb6d9698b63a5b638a", + "instance_ansible__ansible-fb144c44144f8bd3542e71f5db62b6d322c7bd85-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_element-hq__element-web-582a1b093fc0b77538052f45cbb9c7295f991b51-vnan", + "instance_element-hq__element-web-ec0f940ef0e8e3b61078f145f34dc40d1938e6c5-vnan", + "instance_internetarchive__openlibrary-6fdbbeee4c0a7e976ff3e46fb1d36f4eb110c428-v08d8e8889ec945ab821fb156c04c7d2e2810debb", + "instance_flipt-io__flipt-7161f7b876773a911afdd804b281e52681cb7321", + "instance_navidrome__navidrome-28389fb05e1523564dfc61fa43ed8eb8a10f938c", + "instance_ansible__ansible-5f4e332e3762999d94af27746db29ff1729252c1-v0f01c69f1e2528b935359cfe578530722bca2c59", + "instance_nodebb__nodebb-0e07f3c9bace416cbab078a30eae972868c0a8a3-vf2cf3cbd463b7ad942381f1c6d077626485a1e9e", + "instance_nodebb__nodebb-da0211b1a001d45d73b4c84c6417a4f1b0312575-vf2cf3cbd463b7ad942381f1c6d077626485a1e9e", + "instance_navidrome__navidrome-31799662706fedddf5bcc1a76b50409d1f91d327", + "instance_future-architect__vuls-50580f6e98eeb36f53f27222f7f4fdfea0b21e8d", + "instance_internetarchive__openlibrary-60725705782832a2cb22e17c49697948a42a9d03-v298a7a812ceed28c4c18355a091f1b268fe56d86", + "instance_ansible__ansible-3db08adbb1cc6aa9941be5e0fc810132c6e1fa4b-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_internetarchive__openlibrary-8a5a63af6e0be406aa6c8c9b6d5f28b2f1b6af5a-v0f5aece3601a5b4419f7ccec1dbda2071be28ee4", + "instance_future-architect__vuls-e1df74cbc1a1d1889428b3333a3b2405c4651993", + "instance_internetarchive__openlibrary-123e6e5e1c85b9c07d1e98f70bfc480bc8016890-v2733ff199fb72f0d033a30dc62cb0a4742e3a7f4", + "instance_protonmail__webclients-ae36cb23a1682dcfd69587c1b311ae0227e28f39", + "instance_protonmail__webclients-8be4f6cb9380fcd2e67bcb18cef931ae0d4b869c", + "instance_qutebrowser__qutebrowser-a25e8a09873838ca9efefd36ea8a45170bbeb95c-vc2f56a753b62a190ddb23cd330c257b9cf560d12", + "instance_qutebrowser__qutebrowser-44e64199ed38003253f0296badd4a447645067b6-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_gravitational__teleport-7744f72c6eb631791434b648ba41083b5f6d2278-vce94f93ad1030e3136852817f2423c1b3ac37bc4", + "instance_tutao__tutanota-40e94dee2bcec2b63f362da283123e9df1874cc1-vc4e41fd0029957297843cb9dec4a25c7c756f029", + "instance_future-architect__vuls-c11ba27509f733d7d280bdf661cbbe2e7a99df4c", + "instance_flipt-io__flipt-b2cd6a6dd73ca91b519015fd5924fde8d17f3f06", + "instance_gravitational__teleport-6a14edcf1ff010172fdbac622d0a474ed6af46de", + "instance_flipt-io__flipt-36e62baffae2132f78f9d34dc300a9baa2d7ae0e", + "instance_navidrome__navidrome-c90468b895f6171e33e937ff20dc915c995274f0", + "instance_ansible__ansible-ecea15c508f0e081525be036cf76bbb56dbcdd9d-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_ansible__ansible-e22e103cdf8edc56ff7d9b848a58f94f1471a263-v1055803c3a812189a1133297f7f5468579283f86", + "instance_future-architect__vuls-dc496468b9e9fb73371f9606cdcdb0f8e12e70ca", + "instance_gravitational__teleport-c782838c3a174fdff80cafd8cd3b1aa4dae8beb2", + "instance_internetarchive__openlibrary-91efee627df01e32007abf2d6ebf73f9d9053076-vbee42ad1b72fb23c6a1c874868a720b370983ed2", + "instance_qutebrowser__qutebrowser-52708364b5f91e198defb022d1a5b4b3ebd9b563-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_ansible__ansible-eea46a0d1b99a6dadedbb6a3502d599235fa7ec3-v390e508d27db7a51eece36bb6d9698b63a5b638a", + "instance_protonmail__webclients-c6f65d205c401350a226bb005f42fac1754b0b5b", + "instance_ansible__ansible-cb94c0cc550df9e98f1247bc71d8c2b861c75049-v1055803c3a812189a1133297f7f5468579283f86", + "instance_nodebb__nodebb-05f2236193f407cf8e2072757fbd6bb170bc13f0-vf2cf3cbd463b7ad942381f1c6d077626485a1e9e", + "instance_navidrome__navidrome-55730514ea59d5f1d0b8e3f8745569c29bdbf7b4", + "instance_navidrome__navidrome-0130c6dc13438b48cf0fdfab08a89e357b5517c9", + "instance_flipt-io__flipt-507170da0f7f4da330f6732bffdf11c4df7fc192", + "instance_tutao__tutanota-fe240cbf7f0fdd6744ef7bef8cb61676bcdbb621-vc4e41fd0029957297843cb9dec4a25c7c756f029", + "instance_qutebrowser__qutebrowser-233cb1cc48635130e5602549856a6fa4ab4c087f-v35616345bb8052ea303186706cec663146f0f184", + "instance_internetarchive__openlibrary-f8cc11d9c1575fdba5ac66aee0befca970da8d64-v13642507b4fc1f8d234172bf8129942da2c2ca26", + "instance_qutebrowser__qutebrowser-fec187c2cb53d769c2682b35ca77858a811414a8-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_tutao__tutanota-da4edb7375c10f47f4ed3860a591c5e6557f7b5c-vbc0d9ba8f0071fbe982809910959a6ff8884dbbf", + "instance_ansible__ansible-be2c376ab87e3e872ca21697508f12c6909cf85a-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_future-architect__vuls-457a3a9627fb9a0800d0aecf1d4713fb634a9011", + "instance_internetarchive__openlibrary-d40ec88713dc95ea791b252f92d2f7b75e107440-v13642507b4fc1f8d234172bf8129942da2c2ca26", + "instance_protonmail__webclients-863d524b5717b9d33ce08a0f0535e3fd8e8d1ed8", + "instance_internetarchive__openlibrary-308a35d6999427c02b1dbf5211c033ad3b352556-ve8c8d62a2b60610a3c4631f5f23ed866bada9818", + "instance_flipt-io__flipt-15b76cada1ef29cfa56b0fba36754be36243dded", + "instance_future-architect__vuls-61c39637f2f3809e1b5dad05f0c57c799dce1587", + "instance_ansible__ansible-5d253a13807e884b7ce0b6b57a963a45e2f0322c-v1055803c3a812189a1133297f7f5468579283f86", + "instance_element-hq__element-web-44b98896a79ede48f5ad7ff22619a39d5f6ff03c-vnan", + "instance_qutebrowser__qutebrowser-f91ace96223cac8161c16dd061907e138fe85111-v059c6fdc75567943479b23ebca7c07b5e9a7f34c", + "instance_gravitational__teleport-bb562408da4adeae16e025be65e170959d1ec492-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_gravitational__teleport-78b0d8c72637df1129fb6ff84fc49ef4b5ab1288", + "instance_internetarchive__openlibrary-3f7db6bbbcc7c418b3db72d157c6aed1d45b2ccf-v430f20c722405e462d9ef44dee7d34c41e76fe7a", + "instance_future-architect__vuls-e52fa8d6ed1d23e36f2a86e5d3efe9aa057a1b0d", + "instance_navidrome__navidrome-3bc9e75b2843f91f6a1e9b604e321c2bd4fd442a", + "instance_tutao__tutanota-fb32e5f9d9fc152a00144d56dd0af01760a2d4dc-vc4e41fd0029957297843cb9dec4a25c7c756f029", + "instance_gravitational__teleport-af5e2517de7d18406b614e413aca61c319312171-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_ansible__ansible-a20a52701402a12f91396549df04ac55809f68e9-v1055803c3a812189a1133297f7f5468579283f86", + "instance_protonmail__webclients-1501eb765873b2884b6f1944fd242ecfc9d6b103", + "instance_gravitational__teleport-db89206db6c2969266e664c7c0fb51b70e958b64", + "instance_ansible__ansible-be59caa59bf47ca78a4760eb7ff38568372a8260-v1055803c3a812189a1133297f7f5468579283f86", + "instance_flipt-io__flipt-65581fef4aa807540cb933753d085feb0d7e736f", + "instance_gravitational__teleport-4e1c39639edf1ab494dd7562844c8b277b5cfa18-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_future-architect__vuls-f0b3a8b1db98eb1bd32685f1c36c41a99c3452ed", + "instance_navidrome__navidrome-b65e76293a917ee2dfc5d4b373b1c62e054d0dca", + "instance_element-hq__element-web-33e8edb3d508d6eefb354819ca693b7accc695e7", + "instance_flipt-io__flipt-e50808c03e4b9d25a6a78af9c61a3b1616ea356b", + "instance_future-architect__vuls-e3c27e1817d68248043bd09d63cc31f3344a6f2c", + "instance_element-hq__element-web-ad26925bb6628260cfe0fcf90ec0a8cba381f4a4-vnan", + "instance_nodebb__nodebb-1ea9481af6125ffd6da0592ed439aa62af0bca11-vd59a5728dfc977f44533186ace531248c2917516", + "instance_ansible__ansible-489156378c8e97374a75a544c7c9c2c0dd8146d1-v390e508d27db7a51eece36bb6d9698b63a5b638a", + "instance_internetarchive__openlibrary-bb152d23c004f3d68986877143bb0f83531fe401-ve8c8d62a2b60610a3c4631f5f23ed866bada9818", + "instance_qutebrowser__qutebrowser-0b621cb0ce2b54d3f93d8d41d8ff4257888a87e5-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_qutebrowser__qutebrowser-1af602b258b97aaba69d2585ed499d95e2303ac2-v2ef375ac784985212b1805e1d0431dc8f1b3c171", + "instance_future-architect__vuls-e1fab805afcfc92a2a615371d0ec1e667503c254-v264a82e2f4818e30f5a25e4da53b27ba119f62b5", + "instance_internetarchive__openlibrary-7edd1ef09d91fe0b435707633c5cc9af41dedddf-v76304ecdb3a5954fcf13feb710e8c40fcf24b73c", + "instance_qutebrowser__qutebrowser-99029144b5109bb1b2a53964a7c129e009980cd9-va0fd88aac89cde702ec1ba84877234da33adce8a", + "instance_ansible__ansible-12734fa21c08a0ce8c84e533abdc560db2eb1955-v7eee2454f617569fd6889f2211f75bc02a35f9f8", + "instance_protonmail__webclients-7e54526774e577c0ebb58ced7ba8bef349a69fec", + "instance_gravitational__teleport-1a77b7945a022ab86858029d30ac7ad0d5239d00-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_nodebb__nodebb-8ca65b0c78c67c1653487c02d1135e1b702185e1-vf2cf3cbd463b7ad942381f1c6d077626485a1e9e", + "instance_flipt-io__flipt-9f8127f225a86245fa35dca4885c2daef824ee55", + "instance_gravitational__teleport-ac2fb2f9b4fd1896b554d3011df23d3d71295779", + "instance_ansible__ansible-f02a62db509dc7463fab642c9c3458b9bc3476cc-v390e508d27db7a51eece36bb6d9698b63a5b638a" + ], + "additional_details": { + "shuffled": "True" + } + }, + "evaluation_timestamp": "1775482110.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.58, + "uncertainty": { + "standard_error": { + "value": 0.0349874349304872 + }, + "num_samples": 200 + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"If you do not submit an answer, you will receive a score of zero.\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"Your submission was incorrect. Please review your work, identify what went wrong, and try again.\", \"to_float\"], \"submit\": null, \"on_continue\": null, \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"If you do not submit an answer, you will receive a score of zero.\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"Your submission was incorrect. Please review your work, identify what went wrong, and try again.\", \"to_float\"], \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 15000000 + }, + "sandbox": { + "type": "k8s" + } + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/openai/gpt-5.4-2026-03-05/ce84dd29-99b4-4b40-a9c4-76d3ffa92366_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "9277b2dc7ddfdda915f821d1d69c703fd58f56d38773bb4a63d5e2b7dec9fc8b", + "total_rows": 200 + } +} \ No newline at end of file diff --git a/data/swebenchpro/openai/gpt-5.4-2026-03-05/ce84dd29-99b4-4b40-a9c4-76d3ffa92366_samples.jsonl b/data/swebenchpro/openai/gpt-5.4-2026-03-05/ce84dd29-99b4-4b40-a9c4-76d3ffa92366_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..2f8b175896b6fb54aff562382a7b28283cd4fa1f --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.4-2026-03-05/ce84dd29-99b4-4b40-a9c4-76d3ffa92366_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9277b2dc7ddfdda915f821d1d69c703fd58f56d38773bb4a63d5e2b7dec9fc8b +size 279681458 diff --git a/data/swebenchpro/openai/gpt-5.4-2026-03-05/d768c3ce-cc49-4970-baf7-7da676a93691.json b/data/swebenchpro/openai/gpt-5.4-2026-03-05/d768c3ce-cc49-4970-baf7-7da676a93691.json new file mode 100644 index 0000000000000000000000000000000000000000..500fa16fcc925224847df03541edaaf3af4cd9d8 --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.4-2026-03-05/d768c3ce-cc49-4970-baf7-7da676a93691.json @@ -0,0 +1,110 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/openai_gpt-5.4-2026-03-05/1786162928.104319", + "evaluation_timestamp": "1777899366.0", + "retrieved_timestamp": "1786162928.104319", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on swebenchpro/S-adaptive+C/+3ep/aacb763d for scorer harbor_scorer", + "source_data": { + "dataset_name": "swebenchpro", + "source_type": "hf_dataset", + "samples_number": 1, + "sample_ids": [ + "instance_qutebrowser__qutebrowser-2e961080a85d660148937ee8f0f6b3445a8f2c01-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1777899366.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 1.0, + "uncertainty": { + "num_samples": 3 + }, + "details": { + "total_matched_trajectories": "3", + "stopping_reason_count_completed_on_successful_submit": "3", + "avg_total_tokens_target_model": "1019214.67", + "total_total_tokens_target_model": "3057644.0", + "avg_total_tokens_other_models": "0.00", + "total_total_tokens_other_models": "0.0", + "avg_total_tokens_all_models": "1019214.67", + "total_total_tokens_all_models": "3057644.0", + "avg_turn_count": "31.33", + "total_turn_count": "94.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 30000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/openai/gpt-5.4-2026-03-05/d768c3ce-cc49-4970-baf7-7da676a93691_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "efbdab81a830097a55fd29bbf8b76609b9899c17a200b06a3b59913fdc0c6f99", + "total_rows": 3 + } +} \ No newline at end of file diff --git a/data/swebenchpro/openai/gpt-5.4-2026-03-05/d768c3ce-cc49-4970-baf7-7da676a93691_samples.jsonl b/data/swebenchpro/openai/gpt-5.4-2026-03-05/d768c3ce-cc49-4970-baf7-7da676a93691_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..717fc4ef18d9b23a822a516feebc5abc99450703 --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.4-2026-03-05/d768c3ce-cc49-4970-baf7-7da676a93691_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:efbdab81a830097a55fd29bbf8b76609b9899c17a200b06a3b59913fdc0c6f99 +size 901602 diff --git a/data/swebenchpro/openai/gpt-5.4-2026-03-05/d960f49b-b1f1-496a-aeb8-1fd50432fbd0.json b/data/swebenchpro/openai/gpt-5.4-2026-03-05/d960f49b-b1f1-496a-aeb8-1fd50432fbd0.json new file mode 100644 index 0000000000000000000000000000000000000000..4d0062458dac3bfe281119880e9888de89e46c9d --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.4-2026-03-05/d960f49b-b1f1-496a-aeb8-1fd50432fbd0.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/openai_gpt-5.4-2026-03-05/1786169052.245338", + "evaluation_timestamp": "1777478369.0", + "retrieved_timestamp": "1786169052.245338", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/openai/gpt-5.4-2026-03-05/d960f49b-b1f1-496a-aeb8-1fd50432fbd0_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "6aea5324f9d2da72bc42773732259dcd0bafc01bec77f532ed2be1429c27d6c6", + "total_rows": 100 + } +} \ No newline at end of file diff --git a/data/swebenchpro/openai/gpt-5.4-2026-03-05/d960f49b-b1f1-496a-aeb8-1fd50432fbd0_samples.jsonl b/data/swebenchpro/openai/gpt-5.4-2026-03-05/d960f49b-b1f1-496a-aeb8-1fd50432fbd0_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..518717c0bd4aac30d386f959d19c345ae1c9767f --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.4-2026-03-05/d960f49b-b1f1-496a-aeb8-1fd50432fbd0_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6aea5324f9d2da72bc42773732259dcd0bafc01bec77f532ed2be1429c27d6c6 +size 42670673 diff --git a/data/swebenchpro/openai/gpt-5.4-2026-03-05/de2797d4-0618-4fe0-8fc4-7ed9bb669868.json b/data/swebenchpro/openai/gpt-5.4-2026-03-05/de2797d4-0618-4fe0-8fc4-7ed9bb669868.json new file mode 100644 index 0000000000000000000000000000000000000000..da9d1817ebad88b24c8ad8b0eb6b3d8a464636ba --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.4-2026-03-05/de2797d4-0618-4fe0-8fc4-7ed9bb669868.json @@ -0,0 +1,114 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/openai_gpt-5.4-2026-03-05/1786164455.402541", + "evaluation_timestamp": "1777656314.0", + "retrieved_timestamp": "1786164455.402541", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on swebenchpro/S-adaptive/+1ep/d46d0ec1 for scorer harbor_scorer", + "source_data": { + "dataset_name": "swebenchpro", + "source_type": "hf_dataset", + "samples_number": 5, + "sample_ids": [ + "instance_ansible__ansible-de5858f48dc9e1ce9117034e0d7e76806f420ca8-v1055803c3a812189a1133297f7f5468579283f86", + "instance_gravitational__teleport-46aa81b1ce96ebb4ebed2ae53fd78cd44a05da6c-vee9b09fb20c43af7e520f57e9239bbcf46b7113d", + "instance_tutao__tutanota-8513a9e8114a8b42e64f4348335e0f23efa054c4-vee878bb72091875e912c52fc32bc60ec3760227b", + "instance_internetarchive__openlibrary-630221ab686c64e75a2ce253c893c033e4814b2e-v93c53c13d5f9b383ebb411ee7750b49dcd1a34c6", + "instance_internetarchive__openlibrary-7f6b722a10f822171501d027cad60afe53337732-ve8c8d62a2b60610a3c4631f5f23ed866bada9818" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1777656314.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.4, + "uncertainty": { + "standard_error": { + "value": 0.24494897427831783 + }, + "num_samples": 5 + }, + "details": { + "total_matched_trajectories": "5", + "stopping_reason_count_repetition_guard": "5", + "avg_total_tokens_target_model": "2883784.60", + "total_total_tokens_target_model": "14418923.0", + "avg_total_tokens_other_models": "1140.00", + "total_total_tokens_other_models": "5700.0", + "avg_total_tokens_all_models": "2884924.60", + "total_total_tokens_all_models": "14424623.0", + "avg_turn_count": "52.60", + "total_turn_count": "263.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 30000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/openai/gpt-5.4-2026-03-05/de2797d4-0618-4fe0-8fc4-7ed9bb669868_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "99c90ee70365e3aefd372d0987c2779a8fbd0feb93f4f38c5dcec736ef82afff", + "total_rows": 5 + } +} \ No newline at end of file diff --git a/data/swebenchpro/openai/gpt-5.4-2026-03-05/de2797d4-0618-4fe0-8fc4-7ed9bb669868_samples.jsonl b/data/swebenchpro/openai/gpt-5.4-2026-03-05/de2797d4-0618-4fe0-8fc4-7ed9bb669868_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..bbf260040a4165db2609be398d386b5b55f315c8 --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.4-2026-03-05/de2797d4-0618-4fe0-8fc4-7ed9bb669868_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:99c90ee70365e3aefd372d0987c2779a8fbd0feb93f4f38c5dcec736ef82afff +size 3322817 diff --git a/data/swebenchpro/openai/gpt-5.4-2026-03-05/ebaf2d2e-0914-4cde-9a01-fa3f21d9f8a3.json b/data/swebenchpro/openai/gpt-5.4-2026-03-05/ebaf2d2e-0914-4cde-9a01-fa3f21d9f8a3.json new file mode 100644 index 0000000000000000000000000000000000000000..8d73a6d4f604e4bf857abbb9df0f04290e37e293 --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.4-2026-03-05/ebaf2d2e-0914-4cde-9a01-fa3f21d9f8a3.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/openai_gpt-5.4-2026-03-05/1786169041.343908", + "evaluation_timestamp": "1777648847.0", + "retrieved_timestamp": "1786169041.343908", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/openai/gpt-5.4-2026-03-05/ebaf2d2e-0914-4cde-9a01-fa3f21d9f8a3_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "38af8bc1c272f78a871360f128fd3cb0b119d48cac1d6287be0f61cbd4935592", + "total_rows": 18 + } +} \ No newline at end of file diff --git a/data/swebenchpro/openai/gpt-5.4-2026-03-05/ebaf2d2e-0914-4cde-9a01-fa3f21d9f8a3_samples.jsonl b/data/swebenchpro/openai/gpt-5.4-2026-03-05/ebaf2d2e-0914-4cde-9a01-fa3f21d9f8a3_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..da83aee68aa47991c9bf1dc02aab9cb1ae70058f --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.4-2026-03-05/ebaf2d2e-0914-4cde-9a01-fa3f21d9f8a3_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:38af8bc1c272f78a871360f128fd3cb0b119d48cac1d6287be0f61cbd4935592 +size 6558551 diff --git a/data/swebenchpro/openai/gpt-5.4-2026-03-05/ecb69d48-91f3-4293-9f66-5dc73f968b05.json b/data/swebenchpro/openai/gpt-5.4-2026-03-05/ecb69d48-91f3-4293-9f66-5dc73f968b05.json new file mode 100644 index 0000000000000000000000000000000000000000..e7ec3b4d92cb36acc9f25594e700bd85d4971d00 --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.4-2026-03-05/ecb69d48-91f3-4293-9f66-5dc73f968b05.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/openai_gpt-5.4-2026-03-05/1786167314.206531", + "evaluation_timestamp": "1777684701.0", + "retrieved_timestamp": "1786167314.206531", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/openai/gpt-5.4-2026-03-05/ecb69d48-91f3-4293-9f66-5dc73f968b05_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "26dab7a9084be6dcb7f84ad41b350510028b505df1515cf9414db9c301e5e175", + "total_rows": 100 + } +} \ No newline at end of file diff --git a/data/swebenchpro/openai/gpt-5.4-2026-03-05/ecb69d48-91f3-4293-9f66-5dc73f968b05_samples.jsonl b/data/swebenchpro/openai/gpt-5.4-2026-03-05/ecb69d48-91f3-4293-9f66-5dc73f968b05_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..9a4eaa58bb56ef9fe204474d40ddedac45ca3a37 --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.4-2026-03-05/ecb69d48-91f3-4293-9f66-5dc73f968b05_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:26dab7a9084be6dcb7f84ad41b350510028b505df1515cf9414db9c301e5e175 +size 152046461 diff --git a/data/swebenchpro/openai/gpt-5.4-2026-03-05/fd5f39f6-c84c-4401-b71f-79d3ebdaf7cc.json b/data/swebenchpro/openai/gpt-5.4-2026-03-05/fd5f39f6-c84c-4401-b71f-79d3ebdaf7cc.json new file mode 100644 index 0000000000000000000000000000000000000000..9d47c073df63eb8f4b34e9098eb8f62da5832ff0 --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.4-2026-03-05/fd5f39f6-c84c-4401-b71f-79d3ebdaf7cc.json @@ -0,0 +1,109 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "swebenchpro/openai_gpt-5.4-2026-03-05/1786182972.397491", + "evaluation_timestamp": "1774909467.0", + "retrieved_timestamp": "1786182972.397491", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.191.dev111+g7c0c4345,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on swebenchpro for scorer harbor_scorer", + "source_data": { + "dataset_name": "swebenchpro", + "source_type": "hf_dataset", + "samples_number": 731, + "sample_ids": [ + "instance_internetarchive__openlibrary-5de7de19211e71b29b2f2ba3b1dff2fe065d660f-v08d8e8889ec945ab821fb156c04c7d2e2810debb", + "instance_element-hq__element-web-ca8b1b04effb4fec0e1dd3de8e3198eeb364d50e-vnan", + "instance_qutebrowser__qutebrowser-e64622cd2df5b521342cf4a62e0d4cb8f8c9ae5a-v363c8a7e5ccdf6968fc7ab84a2053ac78036691d", + "instance_ansible__ansible-cd473dfb2fdbc97acf3293c134b21cbbcfa89ec3-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5", + "instance_protonmail__webclients-715dbd4e6999499cd2a576a532d8214f75189116" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1774909467.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 1.0, + "uncertainty": { + "num_samples": 10 + }, + "details": { + "total_matched_trajectories": "2", + "stopping_reason_count_completed_on_successful_submit": "2", + "avg_total_tokens_target_model": "268948.00", + "total_total_tokens_target_model": "537896.0", + "avg_total_tokens_other_models": "0.00", + "total_total_tokens_other_models": "0.0", + "avg_total_tokens_all_models": "268948.00", + "total_total_tokens_all_models": "537896.0", + "avg_turn_count": "17.00", + "total_turn_count": "34.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\"}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "3", + "attempt_timeout": "1200", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 20000000 + }, + "sandbox": {}, + "max_attempts": 3 + }, + "additional_details": { + "max_retries": "3" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/swebenchpro/openai/gpt-5.4-2026-03-05/fd5f39f6-c84c-4401-b71f-79d3ebdaf7cc_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "deb0f71979ecca29f17d5ee4db58f3479d266a1b2a7f23bb17a93c96ab279fb1", + "total_rows": 10 + } +} \ No newline at end of file diff --git a/data/swebenchpro/openai/gpt-5.4-2026-03-05/fd5f39f6-c84c-4401-b71f-79d3ebdaf7cc_samples.jsonl b/data/swebenchpro/openai/gpt-5.4-2026-03-05/fd5f39f6-c84c-4401-b71f-79d3ebdaf7cc_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..448e3c19230be1c85ad065c0290c82b9fcc39dc5 --- /dev/null +++ b/data/swebenchpro/openai/gpt-5.4-2026-03-05/fd5f39f6-c84c-4401-b71f-79d3ebdaf7cc_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:deb0f71979ecca29f17d5ee4db58f3479d266a1b2a7f23bb17a93c96ab279fb1 +size 3701429 diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/031f80c8-cf9d-4804-ac47-50ad05848cae.json b/data/terminalbench/anthropic/claude-opus-4-20250514/031f80c8-cf9d-4804-ac47-50ad05848cae.json new file mode 100644 index 0000000000000000000000000000000000000000..91d19395217373a57011c266be80a2d8bcafcf99 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/031f80c8-cf9d-4804-ac47-50ad05848cae.json @@ -0,0 +1,119 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-20250514/1786163605.321657", + "evaluation_timestamp": "1777662106.0", + "retrieved_timestamp": "1786163605.321657", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-20250514", + "id": "anthropic/claude-opus-4-20250514", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive+C/+2ep/d32149ed for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 9, + "sample_ids": [ + "sam-cell-seg", + "model-extraction-relu-logits", + "sparql-university", + "sqlite-db-truncate", + "sanitize-git-repo", + "build-pmars", + "sqlite-with-gcov", + "openssl-selfsigned-cert", + "constraints-scheduling" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1777662106.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.7222222222222222, + "uncertainty": { + "standard_error": { + "value": 0.12108052620946315 + }, + "num_samples": 18 + }, + "details": { + "total_matched_trajectories": "18", + "stopping_reason_count_repetition_guard": "8", + "stopping_reason_count_completed_on_successful_submit": "10", + "avg_total_tokens_target_model": "425825.39", + "total_total_tokens_target_model": "7664857.0", + "avg_total_tokens_other_models": "0.00", + "total_total_tokens_other_models": "0.0", + "avg_total_tokens_all_models": "425825.39", + "total_total_tokens_all_models": "7664857.0", + "avg_turn_count": "26.56", + "total_turn_count": "478.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-20250514/031f80c8-cf9d-4804-ac47-50ad05848cae_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "0273331aa491e6a74e899cd6cbc9a4615bad48d01dc119dd54c39c1b613491e9", + "total_rows": 18 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/031f80c8-cf9d-4804-ac47-50ad05848cae_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-20250514/031f80c8-cf9d-4804-ac47-50ad05848cae_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..748b0a62aae1db24e5987aae90b2c7a3cafdf9d6 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/031f80c8-cf9d-4804-ac47-50ad05848cae_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0273331aa491e6a74e899cd6cbc9a4615bad48d01dc119dd54c39c1b613491e9 +size 5548479 diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/33da3f34-3e0c-4d68-a0bb-de877324344e.json b/data/terminalbench/anthropic/claude-opus-4-20250514/33da3f34-3e0c-4d68-a0bb-de877324344e.json new file mode 100644 index 0000000000000000000000000000000000000000..97e01b05ac3ebe75b41c936995eb2438d0f1d3c9 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/33da3f34-3e0c-4d68-a0bb-de877324344e.json @@ -0,0 +1,104 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-20250514/1786166244.268658", + "evaluation_timestamp": "1776601455.0", + "retrieved_timestamp": "1786166244.268658", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-20250514", + "id": "anthropic/claude-opus-4-20250514", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive/+2ep/6ad0425d for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 1, + "sample_ids": [ + "compile-compcert" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1776601455.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.5, + "uncertainty": { + "num_samples": 2 + }, + "details": { + "total_matched_trajectories": "2", + "stopping_reason_count_repetition_guard": "2", + "avg_total_tokens_target_model": "5424652.50", + "total_total_tokens_target_model": "10849305.0", + "avg_total_tokens_other_models": "0.00", + "total_total_tokens_other_models": "0.0", + "avg_total_tokens_all_models": "5424652.50", + "total_total_tokens_all_models": "10849305.0", + "avg_turn_count": "138.00", + "total_turn_count": "276.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "max_connections": "10", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "max_connections": "10" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-20250514/33da3f34-3e0c-4d68-a0bb-de877324344e_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "1abe9e9a0df1325b1d2c4cd812ceef4267d3e8f3ec613d5c58166eabc4173c4e", + "total_rows": 2 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/33da3f34-3e0c-4d68-a0bb-de877324344e_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-20250514/33da3f34-3e0c-4d68-a0bb-de877324344e_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..18a3ce51cd10679a68af2f86868671ecc9693dcc --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/33da3f34-3e0c-4d68-a0bb-de877324344e_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1abe9e9a0df1325b1d2c4cd812ceef4267d3e8f3ec613d5c58166eabc4173c4e +size 1076334 diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/39611402-60d8-4e10-8c08-4e75dff6baa2.json b/data/terminalbench/anthropic/claude-opus-4-20250514/39611402-60d8-4e10-8c08-4e75dff6baa2.json new file mode 100644 index 0000000000000000000000000000000000000000..e1088b0f9629e0bdf5b5232b9c74e3e208cbcf65 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/39611402-60d8-4e10-8c08-4e75dff6baa2.json @@ -0,0 +1,125 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-20250514/1786163589.062499", + "evaluation_timestamp": "1776880341.0", + "retrieved_timestamp": "1786163589.062499", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-20250514", + "id": "anthropic/claude-opus-4-20250514", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive+C/+2ep/29e55879 for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 14, + "sample_ids": [ + "merge-diff-arc-agi-task", + "torch-tensor-parallelism", + "path-tracing", + "caffe-cifar-10", + "custom-memory-heap-crash", + "adaptive-rejection-sampler", + "crack-7z-hash", + "cobol-modernization", + "overfull-hbox", + "hf-model-inference", + "compile-compcert", + "headless-terminal", + "fix-git", + "qemu-startup" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1776880341.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.6428571428571429, + "uncertainty": { + "standard_error": { + "value": 0.12212148775215861 + }, + "num_samples": 28 + }, + "details": { + "total_matched_trajectories": "28", + "stopping_reason_count_completed_on_successful_submit": "17", + "stopping_reason_count_repetition_guard": "9", + "stopping_reason_count_completed_with_submit": "2", + "avg_total_tokens_target_model": "1154716.89", + "total_total_tokens_target_model": "32332073.0", + "avg_total_tokens_other_models": "0.00", + "total_total_tokens_other_models": "0.0", + "avg_total_tokens_all_models": "1154716.89", + "total_total_tokens_all_models": "32332073.0", + "avg_turn_count": "44.50", + "total_turn_count": "1246.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-20250514/39611402-60d8-4e10-8c08-4e75dff6baa2_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "e8ca4963bc446db798af459e05ad5d9267c6b65ccd7268ac517de7d59557558d", + "total_rows": 28 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/39611402-60d8-4e10-8c08-4e75dff6baa2_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-20250514/39611402-60d8-4e10-8c08-4e75dff6baa2_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..8ad46c5e4a5e567f864e36c45340dad973cd5614 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/39611402-60d8-4e10-8c08-4e75dff6baa2_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e8ca4963bc446db798af459e05ad5d9267c6b65ccd7268ac517de7d59557558d +size 7992224 diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/4338a497-8d4c-408a-86b3-86686aa80f6e.json b/data/terminalbench/anthropic/claude-opus-4-20250514/4338a497-8d4c-408a-86b3-86686aa80f6e.json new file mode 100644 index 0000000000000000000000000000000000000000..42748d4c9402d93586f15cfa25420e0d4fcd5e34 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/4338a497-8d4c-408a-86b3-86686aa80f6e.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-20250514/1786171482.163079", + "evaluation_timestamp": "1776629382.0", + "retrieved_timestamp": "1786171482.163079", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-20250514", + "id": "anthropic/claude-opus-4-20250514", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-20250514/4338a497-8d4c-408a-86b3-86686aa80f6e_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "14d8aec61d917364fa62ef5bb6af826a138ba07c30a03e88ddce7643f320871e", + "total_rows": 41 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/4338a497-8d4c-408a-86b3-86686aa80f6e_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-20250514/4338a497-8d4c-408a-86b3-86686aa80f6e_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..03a5360800630c8f15388decd3025f8ce434c27b --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/4338a497-8d4c-408a-86b3-86686aa80f6e_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:14d8aec61d917364fa62ef5bb6af826a138ba07c30a03e88ddce7643f320871e +size 32866492 diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/464ee44c-7ee3-400f-b2e9-c84799a4e942.json b/data/terminalbench/anthropic/claude-opus-4-20250514/464ee44c-7ee3-400f-b2e9-c84799a4e942.json new file mode 100644 index 0000000000000000000000000000000000000000..de39c2e4740ad07a641173ff97290ece530ec116 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/464ee44c-7ee3-400f-b2e9-c84799a4e942.json @@ -0,0 +1,106 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-20250514/1786166251.836979", + "evaluation_timestamp": "1776520455.0", + "retrieved_timestamp": "1786166251.836979", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-20250514", + "id": "anthropic/claude-opus-4-20250514", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive/+2ep/bea77d74 for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 2, + "sample_ids": [ + "regex-chess", + "torch-tensor-parallelism" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1776520455.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.0, + "uncertainty": { + "num_samples": 4 + }, + "details": { + "total_matched_trajectories": "4", + "stopping_reason_count_repetition_guard": "3", + "stopping_reason_count_token_limit": "1", + "avg_total_tokens_target_model": "4572051.00", + "total_total_tokens_target_model": "18288204.0", + "avg_total_tokens_other_models": "389.25", + "total_total_tokens_other_models": "1557.0", + "avg_total_tokens_all_models": "4572440.25", + "total_total_tokens_all_models": "18289761.0", + "avg_turn_count": "60.50", + "total_turn_count": "242.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "max_connections": "10", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "max_connections": "10" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-20250514/464ee44c-7ee3-400f-b2e9-c84799a4e942_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "b487a5caa16c980da70a67af87b422731951f34db7c13b4760f2fe2de462f973", + "total_rows": 4 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/464ee44c-7ee3-400f-b2e9-c84799a4e942_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-20250514/464ee44c-7ee3-400f-b2e9-c84799a4e942_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..ce428e53bb26b7ddc4a9075a4501cd62513443a1 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/464ee44c-7ee3-400f-b2e9-c84799a4e942_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b487a5caa16c980da70a67af87b422731951f34db7c13b4760f2fe2de462f973 +size 2114979 diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/4bb792e5-50f5-48dd-8fec-0afeeeeca10d.json b/data/terminalbench/anthropic/claude-opus-4-20250514/4bb792e5-50f5-48dd-8fec-0afeeeeca10d.json new file mode 100644 index 0000000000000000000000000000000000000000..11f07beaff186cc2685743bebfd16a0b25ff361b --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/4bb792e5-50f5-48dd-8fec-0afeeeeca10d.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-20250514/1786168134.449826", + "evaluation_timestamp": "1776830462.0", + "retrieved_timestamp": "1786168134.449826", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-20250514", + "id": "anthropic/claude-opus-4-20250514", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-20250514/4bb792e5-50f5-48dd-8fec-0afeeeeca10d_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "5fc718586327ef5e1aae7bd9d26103bfad7a456dee31aad679dea9e8f8541f5b", + "total_rows": 17 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/4bb792e5-50f5-48dd-8fec-0afeeeeca10d_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-20250514/4bb792e5-50f5-48dd-8fec-0afeeeeca10d_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..0e97fa931157c02718e97cda62d7225ed0c168f4 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/4bb792e5-50f5-48dd-8fec-0afeeeeca10d_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5fc718586327ef5e1aae7bd9d26103bfad7a456dee31aad679dea9e8f8541f5b +size 6096722 diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/52525f46-134e-49ef-a13f-3867e7b19585.json b/data/terminalbench/anthropic/claude-opus-4-20250514/52525f46-134e-49ef-a13f-3867e7b19585.json new file mode 100644 index 0000000000000000000000000000000000000000..26cf6a2eb0ec91763f5a7e9c8f0fc80ba91b6fe0 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/52525f46-134e-49ef-a13f-3867e7b19585.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-20250514/1786171424.67369", + "evaluation_timestamp": "1776107987.0", + "retrieved_timestamp": "1786171424.67369", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-20250514", + "id": "anthropic/claude-opus-4-20250514", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-20250514/52525f46-134e-49ef-a13f-3867e7b19585_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "b850336c9309c24cedaf1da6b333403c941d70921a9699a867158b799fa7cb7a", + "total_rows": 60 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/52525f46-134e-49ef-a13f-3867e7b19585_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-20250514/52525f46-134e-49ef-a13f-3867e7b19585_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..e98a52e531f0675d9cdd4130293c37ccbc1471e6 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/52525f46-134e-49ef-a13f-3867e7b19585_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b850336c9309c24cedaf1da6b333403c941d70921a9699a867158b799fa7cb7a +size 47302444 diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/54a648b4-2841-4193-8ff1-cd0d2124d40d.json b/data/terminalbench/anthropic/claude-opus-4-20250514/54a648b4-2841-4193-8ff1-cd0d2124d40d.json new file mode 100644 index 0000000000000000000000000000000000000000..ba71b3d972132563e69565b8ead9d0b194e709b3 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/54a648b4-2841-4193-8ff1-cd0d2124d40d.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-20250514/1786171262.203615", + "evaluation_timestamp": "1776359446.0", + "retrieved_timestamp": "1786171262.203615", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-20250514", + "id": "anthropic/claude-opus-4-20250514", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-20250514/54a648b4-2841-4193-8ff1-cd0d2124d40d_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "e9e391baf8d09cb97c6ab78bedb7c81fdaf2928c35d586564eebee5a0fc8d9f4", + "total_rows": 20 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/54a648b4-2841-4193-8ff1-cd0d2124d40d_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-20250514/54a648b4-2841-4193-8ff1-cd0d2124d40d_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..8c9453d88595f055183094d7db14f310be2726b8 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/54a648b4-2841-4193-8ff1-cd0d2124d40d_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e9e391baf8d09cb97c6ab78bedb7c81fdaf2928c35d586564eebee5a0fc8d9f4 +size 16220448 diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/649c3dfc-5f3b-4889-b33b-a4a3d9082171.json b/data/terminalbench/anthropic/claude-opus-4-20250514/649c3dfc-5f3b-4889-b33b-a4a3d9082171.json new file mode 100644 index 0000000000000000000000000000000000000000..2098fdb638d695438468a48a233d14e90e81be95 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/649c3dfc-5f3b-4889-b33b-a4a3d9082171.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-20250514/1786168093.245303", + "evaluation_timestamp": "1777544308.0", + "retrieved_timestamp": "1786168093.245303", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-20250514", + "id": "anthropic/claude-opus-4-20250514", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-20250514/649c3dfc-5f3b-4889-b33b-a4a3d9082171_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "515f582e299c27c11cd64558c53336db06d9341c92d74c449887ecb6fc5ed62a", + "total_rows": 10 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/649c3dfc-5f3b-4889-b33b-a4a3d9082171_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-20250514/649c3dfc-5f3b-4889-b33b-a4a3d9082171_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..e64727c3d7b2414225de24bf141dd2615ec39c6c --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/649c3dfc-5f3b-4889-b33b-a4a3d9082171_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:515f582e299c27c11cd64558c53336db06d9341c92d74c449887ecb6fc5ed62a +size 4022276 diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/65f92667-a5be-4d0b-a491-1aef9f54aa83.json b/data/terminalbench/anthropic/claude-opus-4-20250514/65f92667-a5be-4d0b-a491-1aef9f54aa83.json new file mode 100644 index 0000000000000000000000000000000000000000..b3ef1a822eadfd5e3bd06610891b6704640e366e --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/65f92667-a5be-4d0b-a491-1aef9f54aa83.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-20250514/1786168127.399911", + "evaluation_timestamp": "1776654052.0", + "retrieved_timestamp": "1786168127.399911", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-20250514", + "id": "anthropic/claude-opus-4-20250514", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-20250514/65f92667-a5be-4d0b-a491-1aef9f54aa83_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "3acee3c9a0f1526a1d944c6f162e1746311ed524c7182bbc564efb288eb0271a", + "total_rows": 33 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/65f92667-a5be-4d0b-a491-1aef9f54aa83_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-20250514/65f92667-a5be-4d0b-a491-1aef9f54aa83_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..f444768ec3eab5446d1a8b5cfdd8d14ad9be4f52 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/65f92667-a5be-4d0b-a491-1aef9f54aa83_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3acee3c9a0f1526a1d944c6f162e1746311ed524c7182bbc564efb288eb0271a +size 10854628 diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/660f73cd-d2ac-450f-94c5-e8b94673621f.json b/data/terminalbench/anthropic/claude-opus-4-20250514/660f73cd-d2ac-450f-94c5-e8b94673621f.json new file mode 100644 index 0000000000000000000000000000000000000000..6c3c9005cf45c0ef1ed2e37b9d978115ea4dc7fb --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/660f73cd-d2ac-450f-94c5-e8b94673621f.json @@ -0,0 +1,123 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-20250514/1786163628.337817", + "evaluation_timestamp": "1777369892.0", + "retrieved_timestamp": "1786163628.337817", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-20250514", + "id": "anthropic/claude-opus-4-20250514", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive+C/+4ep/ac62d5e0 for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 11, + "sample_ids": [ + "torch-pipeline-parallelism", + "make-mips-interpreter", + "count-dataset-tokens", + "circuit-fibsqrt", + "mteb-leaderboard", + "mailman", + "raman-fitting", + "extract-elf", + "code-from-image", + "fix-code-vulnerability", + "financial-document-processor" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1777369892.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.38636363636363635, + "uncertainty": { + "standard_error": { + "value": 0.12775881202322906 + }, + "num_samples": 41 + }, + "details": { + "total_matched_trajectories": "41", + "stopping_reason_count_repetition_guard": "25", + "stopping_reason_count_completed_on_successful_submit": "10", + "stopping_reason_count_completed_with_submit": "4", + "stopping_reason_count_token_limit": "2", + "avg_total_tokens_target_model": "2920902.68", + "total_total_tokens_target_model": "119757010.0", + "avg_total_tokens_other_models": "0.00", + "total_total_tokens_other_models": "0.0", + "avg_total_tokens_all_models": "2920902.68", + "total_total_tokens_all_models": "119757010.0", + "avg_turn_count": "72.12", + "total_turn_count": "2957.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-20250514/660f73cd-d2ac-450f-94c5-e8b94673621f_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "05076fe88b25abcc15ac21cf488f083137de93c40f9d3ddec4ef71968e81c90d", + "total_rows": 41 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/660f73cd-d2ac-450f-94c5-e8b94673621f_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-20250514/660f73cd-d2ac-450f-94c5-e8b94673621f_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..222e94ba2a5a6f914284e74814e62755baf6835e --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/660f73cd-d2ac-450f-94c5-e8b94673621f_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:05076fe88b25abcc15ac21cf488f083137de93c40f9d3ddec4ef71968e81c90d +size 24417092 diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/6dedca53-153c-4625-a96f-15992f82f0cc.json b/data/terminalbench/anthropic/claude-opus-4-20250514/6dedca53-153c-4625-a96f-15992f82f0cc.json new file mode 100644 index 0000000000000000000000000000000000000000..ee6ac1a1fb6378910445dafa55d498436402207f --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/6dedca53-153c-4625-a96f-15992f82f0cc.json @@ -0,0 +1,121 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-20250514/1786166213.046722", + "evaluation_timestamp": "1776696546.0", + "retrieved_timestamp": "1786166213.046722", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-20250514", + "id": "anthropic/claude-opus-4-20250514", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive/+1ep/78781612 for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 12, + "sample_ids": [ + "chess-best-move", + "overfull-hbox", + "db-wal-recovery", + "fix-git", + "qemu-startup", + "build-pov-ray", + "git-multibranch", + "video-processing", + "kv-store-grpc", + "make-doom-for-mips", + "tune-mjcf", + "make-mips-interpreter" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1776696546.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.4166666666666667, + "uncertainty": { + "standard_error": { + "value": 0.1486470975026408 + }, + "num_samples": 12 + }, + "details": { + "total_matched_trajectories": "12", + "stopping_reason_count_token_limit": "11", + "stopping_reason_count_repetition_guard": "1", + "avg_total_tokens_target_model": "9491702.17", + "total_total_tokens_target_model": "113900426.0", + "avg_total_tokens_other_models": "30.50", + "total_total_tokens_other_models": "366.0", + "avg_total_tokens_all_models": "9491732.67", + "total_total_tokens_all_models": "113900792.0", + "avg_turn_count": "198.25", + "total_turn_count": "2379.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "3600", + "attempt_timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "3600", + "max_connections": "10" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-20250514/6dedca53-153c-4625-a96f-15992f82f0cc_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "85d91e472dcc9e2cacc9fc39133a1301477796f037ece6aac79a968734597b3c", + "total_rows": 12 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/6dedca53-153c-4625-a96f-15992f82f0cc_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-20250514/6dedca53-153c-4625-a96f-15992f82f0cc_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..3f4c53c3a0654f399fb73128d234e276fbde2707 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/6dedca53-153c-4625-a96f-15992f82f0cc_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:85d91e472dcc9e2cacc9fc39133a1301477796f037ece6aac79a968734597b3c +size 10831145 diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/6fbadcb3-201a-4c85-bf86-9b7e4adc2bf7.json b/data/terminalbench/anthropic/claude-opus-4-20250514/6fbadcb3-201a-4c85-bf86-9b7e4adc2bf7.json new file mode 100644 index 0000000000000000000000000000000000000000..118f01f5b7e593ab92ba363740e65073f15b5a6e --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/6fbadcb3-201a-4c85-bf86-9b7e4adc2bf7.json @@ -0,0 +1,105 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-20250514/1786166215.778602", + "evaluation_timestamp": "1776359447.0", + "retrieved_timestamp": "1786166215.778602", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-20250514", + "id": "anthropic/claude-opus-4-20250514", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive/+1ep/8154f687 for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 2, + "sample_ids": [ + "log-summary-date-ranges", + "multi-source-data-merger" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1776359447.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 1.0, + "uncertainty": { + "num_samples": 2 + }, + "details": { + "total_matched_trajectories": "2", + "stopping_reason_count_repetition_guard": "2", + "avg_total_tokens_target_model": "2192370.00", + "total_total_tokens_target_model": "4384740.0", + "avg_total_tokens_other_models": "622.50", + "total_total_tokens_other_models": "1245.0", + "avg_total_tokens_all_models": "2192992.50", + "total_total_tokens_all_models": "4385985.0", + "avg_turn_count": "76.50", + "total_turn_count": "153.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "max_connections": "10", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "max_connections": "10" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-20250514/6fbadcb3-201a-4c85-bf86-9b7e4adc2bf7_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "11e3e99d866627399df705840e47d5032b8d00fa02f769b96815d748804fc000", + "total_rows": 2 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/6fbadcb3-201a-4c85-bf86-9b7e4adc2bf7_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-20250514/6fbadcb3-201a-4c85-bf86-9b7e4adc2bf7_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..9879932a6337d961adea77aee99c1a85c2aefc8c --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/6fbadcb3-201a-4c85-bf86-9b7e4adc2bf7_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:11e3e99d866627399df705840e47d5032b8d00fa02f769b96815d748804fc000 +size 643419 diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/772f76c4-6c92-4c58-9b00-321e2dc31195.json b/data/terminalbench/anthropic/claude-opus-4-20250514/772f76c4-6c92-4c58-9b00-321e2dc31195.json new file mode 100644 index 0000000000000000000000000000000000000000..378dd4686271782765ab4b5839cc80ae12d03779 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/772f76c4-6c92-4c58-9b00-321e2dc31195.json @@ -0,0 +1,114 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-20250514/1786163635.886171", + "evaluation_timestamp": "1777695145.0", + "retrieved_timestamp": "1786163635.886171", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-20250514", + "id": "anthropic/claude-opus-4-20250514", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive+C/+4ep/9e384388 for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 4, + "sample_ids": [ + "train-fasttext", + "qemu-alpine-ssh", + "rstan-to-pystan", + "dna-assembly" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1777695145.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.3333333333333333, + "uncertainty": { + "standard_error": { + "value": 0.23570226039551584 + }, + "num_samples": 15 + }, + "details": { + "total_matched_trajectories": "15", + "stopping_reason_count_repetition_guard": "13", + "stopping_reason_count_completed_on_successful_submit": "2", + "avg_total_tokens_target_model": "2239048.67", + "total_total_tokens_target_model": "33585730.0", + "avg_total_tokens_other_models": "0.00", + "total_total_tokens_other_models": "0.0", + "avg_total_tokens_all_models": "2239048.67", + "total_total_tokens_all_models": "33585730.0", + "avg_turn_count": "60.60", + "total_turn_count": "909.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-20250514/772f76c4-6c92-4c58-9b00-321e2dc31195_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "0469b94975ffb5a159cb005ca7000378fe082f2697ea39b95b2ecefb603f7108", + "total_rows": 15 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/772f76c4-6c92-4c58-9b00-321e2dc31195_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-20250514/772f76c4-6c92-4c58-9b00-321e2dc31195_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..272e025f7d9d07ced1a3bf4546d79748e3ffc8b5 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/772f76c4-6c92-4c58-9b00-321e2dc31195_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0469b94975ffb5a159cb005ca7000378fe082f2697ea39b95b2ecefb603f7108 +size 10528311 diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/79232d58-51ee-452f-8d46-46a6b7dde25f.json b/data/terminalbench/anthropic/claude-opus-4-20250514/79232d58-51ee-452f-8d46-46a6b7dde25f.json new file mode 100644 index 0000000000000000000000000000000000000000..608f0723c5c10a4dd565527933f307b9146e9243 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/79232d58-51ee-452f-8d46-46a6b7dde25f.json @@ -0,0 +1,107 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-20250514/1786163568.203477", + "evaluation_timestamp": "1777478370.0", + "retrieved_timestamp": "1786163568.203477", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-20250514", + "id": "anthropic/claude-opus-4-20250514", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive+C/+1ep/aee684fa for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 1, + "sample_ids": [ + "mteb-leaderboard" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1777478370.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.0, + "uncertainty": { + "num_samples": 1 + }, + "details": { + "total_matched_trajectories": "1", + "stopping_reason_count_repetition_guard": "1", + "avg_total_tokens_target_model": "3458807.00", + "total_total_tokens_target_model": "3458807.0", + "avg_total_tokens_other_models": "0.00", + "total_total_tokens_other_models": "0.0", + "avg_total_tokens_all_models": "3458807.00", + "total_total_tokens_all_models": "3458807.0", + "avg_turn_count": "108.00", + "total_turn_count": "108.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "32", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "32", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-20250514/79232d58-51ee-452f-8d46-46a6b7dde25f_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "08096bf9e81823f66d2709d4bb04cbc487ff467b899160f78e59f656744eac23", + "total_rows": 1 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/79232d58-51ee-452f-8d46-46a6b7dde25f_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-20250514/79232d58-51ee-452f-8d46-46a6b7dde25f_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..11d990a25406731e7adbd147d94daaeec86cf254 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/79232d58-51ee-452f-8d46-46a6b7dde25f_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:08096bf9e81823f66d2709d4bb04cbc487ff467b899160f78e59f656744eac23 +size 627020 diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/808cd85d-6e1f-4c73-95e7-d3eeac054359.json b/data/terminalbench/anthropic/claude-opus-4-20250514/808cd85d-6e1f-4c73-95e7-d3eeac054359.json new file mode 100644 index 0000000000000000000000000000000000000000..d3e5bd211c5d241450280d088d539d282ae9b6d2 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/808cd85d-6e1f-4c73-95e7-d3eeac054359.json @@ -0,0 +1,113 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-20250514/1786166263.017056", + "evaluation_timestamp": "1776433475.0", + "retrieved_timestamp": "1786166263.017056", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-20250514", + "id": "anthropic/claude-opus-4-20250514", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive/+2ep/6d08c318 for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 6, + "sample_ids": [ + "gpt2-codegolf", + "break-filter-js-from-html", + "write-compressor", + "winning-avg-corewars", + "pytorch-model-cli", + "largest-eigenval" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1776433475.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.25, + "uncertainty": { + "standard_error": { + "value": 0.17078251276599332 + }, + "num_samples": 12 + }, + "details": { + "total_matched_trajectories": "12", + "stopping_reason_count_token_limit": "11", + "stopping_reason_count_repetition_guard": "1", + "avg_total_tokens_target_model": "9773428.67", + "total_total_tokens_target_model": "117281144.0", + "avg_total_tokens_other_models": "56.75", + "total_total_tokens_other_models": "681.0", + "avg_total_tokens_all_models": "9773485.42", + "total_total_tokens_all_models": "117281825.0", + "avg_turn_count": "191.00", + "total_turn_count": "2292.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "max_connections": "10", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "max_connections": "10" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-20250514/808cd85d-6e1f-4c73-95e7-d3eeac054359_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "c8096e7652d935b1eb2c6eef3f2a497e5cc2033075ede2edced8043d224a51dd", + "total_rows": 12 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/808cd85d-6e1f-4c73-95e7-d3eeac054359_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-20250514/808cd85d-6e1f-4c73-95e7-d3eeac054359_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..1129a4701b82bf73d47d10b7eaf1d604a68c4aaa --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/808cd85d-6e1f-4c73-95e7-d3eeac054359_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c8096e7652d935b1eb2c6eef3f2a497e5cc2033075ede2edced8043d224a51dd +size 9774191 diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/84076b0d-728c-467f-b5aa-fa096188f826.json b/data/terminalbench/anthropic/claude-opus-4-20250514/84076b0d-728c-467f-b5aa-fa096188f826.json new file mode 100644 index 0000000000000000000000000000000000000000..00823df8393636ee689ac15da7a4054ecdbfa3d8 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/84076b0d-728c-467f-b5aa-fa096188f826.json @@ -0,0 +1,109 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-20250514/1786166254.94714", + "evaluation_timestamp": "1777304602.0", + "retrieved_timestamp": "1786166254.94714", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-20250514", + "id": "anthropic/claude-opus-4-20250514", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive/+2ep/f32f7983 for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 2, + "sample_ids": [ + "schemelike-metacircular-eval", + "feal-differential-cryptanalysis" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1777304602.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.5, + "uncertainty": { + "num_samples": 4 + }, + "details": { + "total_matched_trajectories": "4", + "stopping_reason_count_repetition_guard": "2", + "stopping_reason_count_token_limit": "2", + "avg_total_tokens_target_model": "7194042.75", + "total_total_tokens_target_model": "28776171.0", + "avg_total_tokens_other_models": "196.50", + "total_total_tokens_other_models": "786.0", + "avg_total_tokens_all_models": "7194239.25", + "total_total_tokens_all_models": "28776957.0", + "avg_turn_count": "144.75", + "total_turn_count": "579.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-20250514/84076b0d-728c-467f-b5aa-fa096188f826_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "2d2e2c53b12b336fce1e25fba2d23f1790947cf36371129ed02dd46320f7db31", + "total_rows": 4 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/84076b0d-728c-467f-b5aa-fa096188f826_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-20250514/84076b0d-728c-467f-b5aa-fa096188f826_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..4051cdda2895f260e626f0f2b4c5e92b2c2d8067 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/84076b0d-728c-467f-b5aa-fa096188f826_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2d2e2c53b12b336fce1e25fba2d23f1790947cf36371129ed02dd46320f7db31 +size 2688468 diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/841b93f3-b336-4602-bb68-f4dd15d18732.json b/data/terminalbench/anthropic/claude-opus-4-20250514/841b93f3-b336-4602-bb68-f4dd15d18732.json new file mode 100644 index 0000000000000000000000000000000000000000..5d9788c9ef7d47aa60239f58da290c3e2cf67fbf --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/841b93f3-b336-4602-bb68-f4dd15d18732.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-20250514/1786168154.380833", + "evaluation_timestamp": "1776830491.0", + "retrieved_timestamp": "1786168154.380833", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-20250514", + "id": "anthropic/claude-opus-4-20250514", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-20250514/841b93f3-b336-4602-bb68-f4dd15d18732_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "36937c49a14c80eea261f4591a8d70484215565c00b866db6a4fd1181e5ff4b9", + "total_rows": 20 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/841b93f3-b336-4602-bb68-f4dd15d18732_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-20250514/841b93f3-b336-4602-bb68-f4dd15d18732_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..c06cb46fe5d40537db7f5a55ff0cecb3d455393d --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/841b93f3-b336-4602-bb68-f4dd15d18732_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:36937c49a14c80eea261f4591a8d70484215565c00b866db6a4fd1181e5ff4b9 +size 4174053 diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/84bdae28-768e-4e4c-a9ac-d1da2e533f1f.json b/data/terminalbench/anthropic/claude-opus-4-20250514/84bdae28-768e-4e4c-a9ac-d1da2e533f1f.json new file mode 100644 index 0000000000000000000000000000000000000000..204992f8d8b8a833fdfa38bedd0bc592bb9d6dea --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/84bdae28-768e-4e4c-a9ac-d1da2e533f1f.json @@ -0,0 +1,107 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-20250514/1786163578.698494", + "evaluation_timestamp": "1777645894.0", + "retrieved_timestamp": "1786163578.698494", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-20250514", + "id": "anthropic/claude-opus-4-20250514", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive+C/+1ep/125be6ed for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 1, + "sample_ids": [ + "feal-differential-cryptanalysis" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1777645894.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 1.0, + "uncertainty": { + "num_samples": 1 + }, + "details": { + "total_matched_trajectories": "1", + "stopping_reason_count_completed_on_successful_submit": "1", + "avg_total_tokens_target_model": "2488529.00", + "total_total_tokens_target_model": "2488529.0", + "avg_total_tokens_other_models": "0.00", + "total_total_tokens_other_models": "0.0", + "avg_total_tokens_all_models": "2488529.00", + "total_total_tokens_all_models": "2488529.0", + "avg_turn_count": "81.00", + "total_turn_count": "81.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-20250514/84bdae28-768e-4e4c-a9ac-d1da2e533f1f_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "e74bf678371d1ebc403d72ad69c67c00e2e09dfe3bd023920e09800d68dedc35", + "total_rows": 1 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/84bdae28-768e-4e4c-a9ac-d1da2e533f1f_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-20250514/84bdae28-768e-4e4c-a9ac-d1da2e533f1f_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..2787d6d5b38d47145bb45e8c69e65070ca21e589 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/84bdae28-768e-4e4c-a9ac-d1da2e533f1f_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e74bf678371d1ebc403d72ad69c67c00e2e09dfe3bd023920e09800d68dedc35 +size 327311 diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/8627243a-d2a2-4a5b-8113-482587a949e2.json b/data/terminalbench/anthropic/claude-opus-4-20250514/8627243a-d2a2-4a5b-8113-482587a949e2.json new file mode 100644 index 0000000000000000000000000000000000000000..40f08d69aeaa216355227dcd7d1049c3a414c90b --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/8627243a-d2a2-4a5b-8113-482587a949e2.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-20250514/1786168170.04127", + "evaluation_timestamp": "1777385218.0", + "retrieved_timestamp": "1786168170.04127", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-20250514", + "id": "anthropic/claude-opus-4-20250514", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-20250514/8627243a-d2a2-4a5b-8113-482587a949e2_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "d951e5a3429326f747414bcd971aeece8333d1a215110bc87c3fab3025537dfc", + "total_rows": 50 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/8627243a-d2a2-4a5b-8113-482587a949e2_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-20250514/8627243a-d2a2-4a5b-8113-482587a949e2_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..39f7a26f5b02b4d30697d5e44f8980d5dfc95e08 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/8627243a-d2a2-4a5b-8113-482587a949e2_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d951e5a3429326f747414bcd971aeece8333d1a215110bc87c3fab3025537dfc +size 19248433 diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/8b58b7e0-0c6f-4103-b3d5-a325bff716c1.json b/data/terminalbench/anthropic/claude-opus-4-20250514/8b58b7e0-0c6f-4103-b3d5-a325bff716c1.json new file mode 100644 index 0000000000000000000000000000000000000000..67f983638f3bf06f10b21d80b5328cc2e91498b3 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/8b58b7e0-0c6f-4103-b3d5-a325bff716c1.json @@ -0,0 +1,105 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-20250514/1786166229.500836", + "evaluation_timestamp": "1776359447.0", + "retrieved_timestamp": "1786166229.500836", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-20250514", + "id": "anthropic/claude-opus-4-20250514", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive/+2ep/84bfae34 for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 1, + "sample_ids": [ + "reshard-c4-data" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1776359447.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.5, + "uncertainty": { + "num_samples": 2 + }, + "details": { + "total_matched_trajectories": "2", + "stopping_reason_count_token_limit": "1", + "stopping_reason_count_repetition_guard": "1", + "avg_total_tokens_target_model": "7230726.00", + "total_total_tokens_target_model": "14461452.0", + "avg_total_tokens_other_models": "1164.50", + "total_total_tokens_other_models": "2329.0", + "avg_total_tokens_all_models": "7231890.50", + "total_total_tokens_all_models": "14463781.0", + "avg_turn_count": "157.00", + "total_turn_count": "314.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "max_connections": "10", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "max_connections": "10" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-20250514/8b58b7e0-0c6f-4103-b3d5-a325bff716c1_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "374841bee7f8732924d9e9e6dbdae7e36fb5883ff930b539221698aa702dc243", + "total_rows": 2 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/8b58b7e0-0c6f-4103-b3d5-a325bff716c1_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-20250514/8b58b7e0-0c6f-4103-b3d5-a325bff716c1_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..f92963b4148e9b5da5225873670a725383c308e2 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/8b58b7e0-0c6f-4103-b3d5-a325bff716c1_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374841bee7f8732924d9e9e6dbdae7e36fb5883ff930b539221698aa702dc243 +size 1564091 diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/8c0b1756-5164-4f3c-85c1-ebe59f3acc8d.json b/data/terminalbench/anthropic/claude-opus-4-20250514/8c0b1756-5164-4f3c-85c1-ebe59f3acc8d.json new file mode 100644 index 0000000000000000000000000000000000000000..ca3476935d6b3c415f596aedd3958d83169b4d3f --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/8c0b1756-5164-4f3c-85c1-ebe59f3acc8d.json @@ -0,0 +1,106 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-20250514/1786166315.05869", + "evaluation_timestamp": "1776629382.0", + "retrieved_timestamp": "1786166315.05869", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-20250514", + "id": "anthropic/claude-opus-4-20250514", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive/+4ep/b897651b for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 1, + "sample_ids": [ + "train-fasttext" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1776629382.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.0, + "uncertainty": { + "num_samples": 2 + }, + "details": { + "total_matched_trajectories": "2", + "stopping_reason_count_repetition_guard": "2", + "avg_total_tokens_target_model": "1709129.00", + "total_total_tokens_target_model": "3418258.0", + "avg_total_tokens_other_models": "0.00", + "total_total_tokens_other_models": "0.0", + "avg_total_tokens_all_models": "1709129.00", + "total_total_tokens_all_models": "3418258.0", + "avg_turn_count": "34.50", + "total_turn_count": "69.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "3600", + "attempt_timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "3600", + "max_connections": "10" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-20250514/8c0b1756-5164-4f3c-85c1-ebe59f3acc8d_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "2fc26abde8491d33f9ad7e616fec344f77c1c34963c320fd4c08fd9c59e52b0a", + "total_rows": 2 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/8c0b1756-5164-4f3c-85c1-ebe59f3acc8d_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-20250514/8c0b1756-5164-4f3c-85c1-ebe59f3acc8d_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..bfe4b2e93ab94401772f3eb6e6bfd47d30183d2c --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/8c0b1756-5164-4f3c-85c1-ebe59f3acc8d_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2fc26abde8491d33f9ad7e616fec344f77c1c34963c320fd4c08fd9c59e52b0a +size 1755609 diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/8c3d1920-216e-4a50-9872-cfb9a2adc3fb.json b/data/terminalbench/anthropic/claude-opus-4-20250514/8c3d1920-216e-4a50-9872-cfb9a2adc3fb.json new file mode 100644 index 0000000000000000000000000000000000000000..63cd5cb43370b2b2c4ba7dc419d6dec8c63b4535 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/8c3d1920-216e-4a50-9872-cfb9a2adc3fb.json @@ -0,0 +1,115 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-20250514/1786166270.90552", + "evaluation_timestamp": "1777025747.0", + "retrieved_timestamp": "1786166270.90552", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-20250514", + "id": "anthropic/claude-opus-4-20250514", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive/+2ep/8faa2289 for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 6, + "sample_ids": [ + "schemelike-metacircular-eval", + "protein-assembly", + "regex-log", + "feal-differential-cryptanalysis", + "polyglot-c-py", + "cancel-async-tasks" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1777025747.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.25, + "uncertainty": { + "standard_error": { + "value": 0.25 + }, + "num_samples": 7 + }, + "details": { + "total_matched_trajectories": "7", + "stopping_reason_count_repetition_guard": "7", + "avg_total_tokens_target_model": "606348.00", + "total_total_tokens_target_model": "4244436.0", + "avg_total_tokens_other_models": "0.00", + "total_total_tokens_other_models": "0.0", + "avg_total_tokens_all_models": "606348.00", + "total_total_tokens_all_models": "4244436.0", + "avg_turn_count": "37.71", + "total_turn_count": "264.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-20250514/8c3d1920-216e-4a50-9872-cfb9a2adc3fb_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "36e0171ca0d16f3710d7c696d9ab47109f4feef88633e90b51c30c140892c324", + "total_rows": 7 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/8c3d1920-216e-4a50-9872-cfb9a2adc3fb_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-20250514/8c3d1920-216e-4a50-9872-cfb9a2adc3fb_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..232cd5e570d0c85af19ceded2462f9daf4fb80d5 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/8c3d1920-216e-4a50-9872-cfb9a2adc3fb_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:36e0171ca0d16f3710d7c696d9ab47109f4feef88633e90b51c30c140892c324 +size 5993221 diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/90995eca-fd67-4997-a9de-05d72f39a1fa.json b/data/terminalbench/anthropic/claude-opus-4-20250514/90995eca-fd67-4997-a9de-05d72f39a1fa.json new file mode 100644 index 0000000000000000000000000000000000000000..1ac378143690e87b5d6c7dd048745da4cbf34945 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/90995eca-fd67-4997-a9de-05d72f39a1fa.json @@ -0,0 +1,107 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-20250514/1786163608.492275", + "evaluation_timestamp": "1776894268.0", + "retrieved_timestamp": "1786163608.492275", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-20250514", + "id": "anthropic/claude-opus-4-20250514", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive+C/+3ep/4ce59f10 for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 1, + "sample_ids": [ + "regex-chess" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1776894268.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.0, + "uncertainty": { + "num_samples": 3 + }, + "details": { + "total_matched_trajectories": "3", + "stopping_reason_count_repetition_guard": "3", + "avg_total_tokens_target_model": "3298104.33", + "total_total_tokens_target_model": "9894313.0", + "avg_total_tokens_other_models": "0.00", + "total_total_tokens_other_models": "0.0", + "avg_total_tokens_all_models": "3298104.33", + "total_total_tokens_all_models": "9894313.0", + "avg_turn_count": "69.67", + "total_turn_count": "209.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-20250514/90995eca-fd67-4997-a9de-05d72f39a1fa_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "04147e810f097a6f87d8bbdc8250ad129e2804e4b474224ddda9c9ea7b3829d3", + "total_rows": 3 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/90995eca-fd67-4997-a9de-05d72f39a1fa_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-20250514/90995eca-fd67-4997-a9de-05d72f39a1fa_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..7e457a12ace4e89a1fceb85e4617f190fed5d609 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/90995eca-fd67-4997-a9de-05d72f39a1fa_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:04147e810f097a6f87d8bbdc8250ad129e2804e4b474224ddda9c9ea7b3829d3 +size 3649753 diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/920ff065-16a7-430d-9236-95953bab507a.json b/data/terminalbench/anthropic/claude-opus-4-20250514/920ff065-16a7-430d-9236-95953bab507a.json new file mode 100644 index 0000000000000000000000000000000000000000..193cfd96375c7679d733dbb5683f9cc9bd318451 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/920ff065-16a7-430d-9236-95953bab507a.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-20250514/1786168108.113446", + "evaluation_timestamp": "1776735889.0", + "retrieved_timestamp": "1786168108.113446", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-20250514", + "id": "anthropic/claude-opus-4-20250514", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-20250514/920ff065-16a7-430d-9236-95953bab507a_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "d3fec44d902a35011f9be399cc9435d3157c7765275f70f7f4bc9efc3a7eeb21", + "total_rows": 14 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/920ff065-16a7-430d-9236-95953bab507a_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-20250514/920ff065-16a7-430d-9236-95953bab507a_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..f668ccf189ba32827aa21d04cce9f969eca0d165 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/920ff065-16a7-430d-9236-95953bab507a_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d3fec44d902a35011f9be399cc9435d3157c7765275f70f7f4bc9efc3a7eeb21 +size 6829684 diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/94390c55-6c05-4be3-b18c-dad6bf7ce09d.json b/data/terminalbench/anthropic/claude-opus-4-20250514/94390c55-6c05-4be3-b18c-dad6bf7ce09d.json new file mode 100644 index 0000000000000000000000000000000000000000..4b265abf7a7a37d2d4f64a0443f646143719cb65 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/94390c55-6c05-4be3-b18c-dad6bf7ce09d.json @@ -0,0 +1,124 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-20250514/1786166239.546195", + "evaluation_timestamp": "1776696546.0", + "retrieved_timestamp": "1786166239.546195", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-20250514", + "id": "anthropic/claude-opus-4-20250514", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive/+2ep/12889ba3 for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 15, + "sample_ids": [ + "compile-compcert", + "qemu-alpine-ssh", + "install-windows-3.11", + "torch-pipeline-parallelism", + "extract-moves-from-video", + "gcode-to-text", + "count-dataset-tokens", + "circuit-fibsqrt", + "mteb-leaderboard", + "mailman", + "raman-fitting", + "extract-elf", + "code-from-image", + "fix-code-vulnerability", + "financial-document-processor" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1776696546.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.03333333333333333, + "uncertainty": { + "standard_error": { + "value": 0.03333333333333333 + }, + "num_samples": 30 + }, + "details": { + "total_matched_trajectories": "30", + "stopping_reason_count_token_limit": "25", + "stopping_reason_count_repetition_guard": "5", + "avg_total_tokens_target_model": "9083875.83", + "total_total_tokens_target_model": "272516275.0", + "avg_total_tokens_other_models": "218.30", + "total_total_tokens_other_models": "6549.0", + "avg_total_tokens_all_models": "9084094.13", + "total_total_tokens_all_models": "272522824.0", + "avg_turn_count": "177.40", + "total_turn_count": "5322.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "3600", + "attempt_timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "3600", + "max_connections": "10" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-20250514/94390c55-6c05-4be3-b18c-dad6bf7ce09d_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "fad9eb1f26cdaeca6712334cd2f4f615a5692079b19f689e9de5434a1ea005d7", + "total_rows": 30 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/94390c55-6c05-4be3-b18c-dad6bf7ce09d_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-20250514/94390c55-6c05-4be3-b18c-dad6bf7ce09d_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..0121b8abd962243dabc4d93413be859016fc375e --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/94390c55-6c05-4be3-b18c-dad6bf7ce09d_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fad9eb1f26cdaeca6712334cd2f4f615a5692079b19f689e9de5434a1ea005d7 +size 27040330 diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/97ae3131-95e1-4bcf-a032-3790310b215b.json b/data/terminalbench/anthropic/claude-opus-4-20250514/97ae3131-95e1-4bcf-a032-3790310b215b.json new file mode 100644 index 0000000000000000000000000000000000000000..151b258aac2dcdb8db17e176a4bbf1b810dc2acd --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/97ae3131-95e1-4bcf-a032-3790310b215b.json @@ -0,0 +1,115 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-20250514/1786166225.057711", + "evaluation_timestamp": "1776520455.0", + "retrieved_timestamp": "1786166225.057711", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-20250514", + "id": "anthropic/claude-opus-4-20250514", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive/+1ep/f005884b for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 8, + "sample_ids": [ + "path-tracing", + "feal-linear-cryptanalysis", + "caffe-cifar-10", + "distribution-search", + "custom-memory-heap-crash", + "adaptive-rejection-sampler", + "crack-7z-hash", + "cobol-modernization" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1776520455.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.375, + "uncertainty": { + "standard_error": { + "value": 0.18298126367784995 + }, + "num_samples": 8 + }, + "details": { + "total_matched_trajectories": "8", + "stopping_reason_count_token_limit": "7", + "stopping_reason_count_repetition_guard": "1", + "avg_total_tokens_target_model": "8813380.88", + "total_total_tokens_target_model": "70507047.0", + "avg_total_tokens_other_models": "15.75", + "total_total_tokens_other_models": "126.0", + "avg_total_tokens_all_models": "8813396.62", + "total_total_tokens_all_models": "70507173.0", + "avg_turn_count": "154.00", + "total_turn_count": "1232.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "max_connections": "10", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "max_connections": "10" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-20250514/97ae3131-95e1-4bcf-a032-3790310b215b_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "49dfad020e272e37099a8d2684a5ee520f572e36dddf91c46007aac287ece2fe", + "total_rows": 8 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/97ae3131-95e1-4bcf-a032-3790310b215b_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-20250514/97ae3131-95e1-4bcf-a032-3790310b215b_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..f14fcb47a3b0704750294102c729dc0ea089eab1 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/97ae3131-95e1-4bcf-a032-3790310b215b_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:49dfad020e272e37099a8d2684a5ee520f572e36dddf91c46007aac287ece2fe +size 7410613 diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/97d1bc44-35d9-4c1b-81eb-5d2cffbb6a9a.json b/data/terminalbench/anthropic/claude-opus-4-20250514/97d1bc44-35d9-4c1b-81eb-5d2cffbb6a9a.json new file mode 100644 index 0000000000000000000000000000000000000000..88e91fcf4fdeb80079136321e908c2302ea4ae1f --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/97d1bc44-35d9-4c1b-81eb-5d2cffbb6a9a.json @@ -0,0 +1,115 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-20250514/1786163595.399809", + "evaluation_timestamp": "1777034999.0", + "retrieved_timestamp": "1786163595.399809", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-20250514", + "id": "anthropic/claude-opus-4-20250514", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive+C/+2ep/c8f118a0 for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 4, + "sample_ids": [ + "configure-git-webserver", + "chess-best-move", + "polyglot-rust-c", + "build-pov-ray" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1777034999.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.5, + "uncertainty": { + "standard_error": { + "value": 0.28867513459481287 + }, + "num_samples": 8 + }, + "details": { + "total_matched_trajectories": "8", + "stopping_reason_count_completed_on_successful_submit": "4", + "stopping_reason_count_repetition_guard": "3", + "stopping_reason_count_token_limit": "1", + "avg_total_tokens_target_model": "2780751.50", + "total_total_tokens_target_model": "22246012.0", + "avg_total_tokens_other_models": "0.00", + "total_total_tokens_other_models": "0.0", + "avg_total_tokens_all_models": "2780751.50", + "total_total_tokens_all_models": "22246012.0", + "avg_turn_count": "81.75", + "total_turn_count": "654.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-20250514/97d1bc44-35d9-4c1b-81eb-5d2cffbb6a9a_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "614e8870e3dad64f6ee5a89600239c114d29ef7fb10e34dac7ef5c30ee4eb0e4", + "total_rows": 8 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/97d1bc44-35d9-4c1b-81eb-5d2cffbb6a9a_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-20250514/97d1bc44-35d9-4c1b-81eb-5d2cffbb6a9a_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..d718c98707f08b31080ff484a9a1e16fa049c5fe --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/97d1bc44-35d9-4c1b-81eb-5d2cffbb6a9a_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:614e8870e3dad64f6ee5a89600239c114d29ef7fb10e34dac7ef5c30ee4eb0e4 +size 3853727 diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/9aa54980-bf88-420e-a468-1241d7670f2a.json b/data/terminalbench/anthropic/claude-opus-4-20250514/9aa54980-bf88-420e-a468-1241d7670f2a.json new file mode 100644 index 0000000000000000000000000000000000000000..f8be8ca63b2f86aa7a30c909b263b958ba6ccf28 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/9aa54980-bf88-420e-a468-1241d7670f2a.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-20250514/1786168114.987115", + "evaluation_timestamp": "1777478369.0", + "retrieved_timestamp": "1786168114.987115", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-20250514", + "id": "anthropic/claude-opus-4-20250514", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-20250514/9aa54980-bf88-420e-a468-1241d7670f2a_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "c94da920a4e79df570fb2050b2ebf9d0d1bd2fcbf380ba8f8349f9e377439628", + "total_rows": 19 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/9aa54980-bf88-420e-a468-1241d7670f2a_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-20250514/9aa54980-bf88-420e-a468-1241d7670f2a_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..a3bd49fb38ac829c4ecabdc7c089d457365e759a --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/9aa54980-bf88-420e-a468-1241d7670f2a_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c94da920a4e79df570fb2050b2ebf9d0d1bd2fcbf380ba8f8349f9e377439628 +size 6942568 diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/9eba157c-2bbb-4495-8b9c-b36d217fd339.json b/data/terminalbench/anthropic/claude-opus-4-20250514/9eba157c-2bbb-4495-8b9c-b36d217fd339.json new file mode 100644 index 0000000000000000000000000000000000000000..a98b0863379e553c2564aef517a3504fd50c153f --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/9eba157c-2bbb-4495-8b9c-b36d217fd339.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-20250514/1786168146.527107", + "evaluation_timestamp": "1776910304.0", + "retrieved_timestamp": "1786168146.527107", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-20250514", + "id": "anthropic/claude-opus-4-20250514", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-20250514/9eba157c-2bbb-4495-8b9c-b36d217fd339_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "cebc10beb83a25352e341b7578fef8f92926d03dbfc1653db31012c03b41e2e9", + "total_rows": 19 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/9eba157c-2bbb-4495-8b9c-b36d217fd339_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-20250514/9eba157c-2bbb-4495-8b9c-b36d217fd339_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..8d96ed398300e5c27bddd10b3b61b3dfd11c87af --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/9eba157c-2bbb-4495-8b9c-b36d217fd339_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cebc10beb83a25352e341b7578fef8f92926d03dbfc1653db31012c03b41e2e9 +size 10450766 diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/9fe46f58-6dff-401d-ba13-10f52b041adb.json b/data/terminalbench/anthropic/claude-opus-4-20250514/9fe46f58-6dff-401d-ba13-10f52b041adb.json new file mode 100644 index 0000000000000000000000000000000000000000..1f82c7382f65beb32dc88f301ec4d29487052175 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/9fe46f58-6dff-401d-ba13-10f52b041adb.json @@ -0,0 +1,107 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-20250514/1786163583.934849", + "evaluation_timestamp": "1777369892.0", + "retrieved_timestamp": "1786163583.934849", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-20250514", + "id": "anthropic/claude-opus-4-20250514", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive+C/+2ep/267c8c83 for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 1, + "sample_ids": [ + "feal-linear-cryptanalysis" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1777369892.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.0, + "uncertainty": { + "num_samples": 2 + }, + "details": { + "total_matched_trajectories": "2", + "stopping_reason_count_repetition_guard": "2", + "avg_total_tokens_target_model": "1260726.50", + "total_total_tokens_target_model": "2521453.0", + "avg_total_tokens_other_models": "0.00", + "total_total_tokens_other_models": "0.0", + "avg_total_tokens_all_models": "1260726.50", + "total_total_tokens_all_models": "2521453.0", + "avg_turn_count": "42.50", + "total_turn_count": "85.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-20250514/9fe46f58-6dff-401d-ba13-10f52b041adb_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "91ced424fc8ce58cf7a4f70a16856ff404fa1e917a541a79e552790b37ec040a", + "total_rows": 2 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/9fe46f58-6dff-401d-ba13-10f52b041adb_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-20250514/9fe46f58-6dff-401d-ba13-10f52b041adb_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..af685da44321eb9ae913d51f6037454dcb6421d9 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/9fe46f58-6dff-401d-ba13-10f52b041adb_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:91ced424fc8ce58cf7a4f70a16856ff404fa1e917a541a79e552790b37ec040a +size 2040001 diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/a0f107de-8e7f-43c1-ab7b-054ff7378912.json b/data/terminalbench/anthropic/claude-opus-4-20250514/a0f107de-8e7f-43c1-ab7b-054ff7378912.json new file mode 100644 index 0000000000000000000000000000000000000000..b57e0a607e377958a44320021d435911d16b9bae --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/a0f107de-8e7f-43c1-ab7b-054ff7378912.json @@ -0,0 +1,120 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-20250514/1786163600.589196", + "evaluation_timestamp": "1777329550.0", + "retrieved_timestamp": "1786163600.589196", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-20250514", + "id": "anthropic/claude-opus-4-20250514", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive+C/+2ep/c8d2dc37 for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 9, + "sample_ids": [ + "feal-linear-cryptanalysis", + "git-multibranch", + "video-processing", + "kv-store-grpc", + "install-windows-3.11", + "make-doom-for-mips", + "tune-mjcf", + "extract-moves-from-video", + "gcode-to-text" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1777329550.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.25, + "uncertainty": { + "standard_error": { + "value": 0.16366341767699427 + }, + "num_samples": 16 + }, + "details": { + "total_matched_trajectories": "16", + "stopping_reason_count_repetition_guard": "10", + "stopping_reason_count_completed_on_successful_submit": "4", + "stopping_reason_count_token_limit": "2", + "avg_total_tokens_target_model": "2768281.00", + "total_total_tokens_target_model": "44292496.0", + "avg_total_tokens_other_models": "0.00", + "total_total_tokens_other_models": "0.0", + "avg_total_tokens_all_models": "2768281.00", + "total_total_tokens_all_models": "44292496.0", + "avg_turn_count": "74.56", + "total_turn_count": "1193.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-20250514/a0f107de-8e7f-43c1-ab7b-054ff7378912_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "706af5d538c6e3e0b30e93fd46ba6fd9641f082876adb0ade46ae7d93367e1ef", + "total_rows": 16 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/a0f107de-8e7f-43c1-ab7b-054ff7378912_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-20250514/a0f107de-8e7f-43c1-ab7b-054ff7378912_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..758931225ae290802f5a224e8d9c45e37d2174c8 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/a0f107de-8e7f-43c1-ab7b-054ff7378912_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:706af5d538c6e3e0b30e93fd46ba6fd9641f082876adb0ade46ae7d93367e1ef +size 8856637 diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/a8803ece-5796-470c-82c9-b8864e156379.json b/data/terminalbench/anthropic/claude-opus-4-20250514/a8803ece-5796-470c-82c9-b8864e156379.json new file mode 100644 index 0000000000000000000000000000000000000000..3b99283d8e7ce78f4265efd124c351d12c99841d --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/a8803ece-5796-470c-82c9-b8864e156379.json @@ -0,0 +1,107 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-20250514/1786163591.735314", + "evaluation_timestamp": "1777478370.0", + "retrieved_timestamp": "1786163591.735314", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-20250514", + "id": "anthropic/claude-opus-4-20250514", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive+C/+2ep/0dd8e472 for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 1, + "sample_ids": [ + "torch-pipeline-parallelism" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1777478370.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.0, + "uncertainty": { + "num_samples": 1 + }, + "details": { + "total_matched_trajectories": "1", + "stopping_reason_count_repetition_guard": "1", + "avg_total_tokens_target_model": "350608.00", + "total_total_tokens_target_model": "350608.0", + "avg_total_tokens_other_models": "0.00", + "total_total_tokens_other_models": "0.0", + "avg_total_tokens_all_models": "350608.00", + "total_total_tokens_all_models": "350608.0", + "avg_turn_count": "22.00", + "total_turn_count": "22.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "32", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "32", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-20250514/a8803ece-5796-470c-82c9-b8864e156379_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "98f0d4d912dd4d4bcd43136ba85e1b9706a6f701cda296ac9d30066156ac5d60", + "total_rows": 1 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/a8803ece-5796-470c-82c9-b8864e156379_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-20250514/a8803ece-5796-470c-82c9-b8864e156379_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..24a48d71d8a777f82f025c004047bdfe76fb4d36 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/a8803ece-5796-470c-82c9-b8864e156379_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:98f0d4d912dd4d4bcd43136ba85e1b9706a6f701cda296ac9d30066156ac5d60 +size 1153436 diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/ab367b19-8fd0-45ec-bae0-3c9d2e542c16.json b/data/terminalbench/anthropic/claude-opus-4-20250514/ab367b19-8fd0-45ec-bae0-3c9d2e542c16.json new file mode 100644 index 0000000000000000000000000000000000000000..d2cad0ed3c80efa43e589e0058c1cbcd21f6bdbb --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/ab367b19-8fd0-45ec-bae0-3c9d2e542c16.json @@ -0,0 +1,112 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-20250514/1786166217.999749", + "evaluation_timestamp": "1776803719.0", + "retrieved_timestamp": "1786166217.999749", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-20250514", + "id": "anthropic/claude-opus-4-20250514", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive/+1ep/fa8b304f for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 2, + "sample_ids": [ + "headless-terminal", + "tune-mjcf" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1776803719.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.5, + "uncertainty": { + "standard_error": { + "value": 0.5 + }, + "num_samples": 2 + }, + "details": { + "total_matched_trajectories": "2", + "stopping_reason_count_repetition_guard": "1", + "stopping_reason_count_token_limit": "1", + "avg_total_tokens_target_model": "6847798.50", + "total_total_tokens_target_model": "13695597.0", + "avg_total_tokens_other_models": "140.00", + "total_total_tokens_other_models": "280.0", + "avg_total_tokens_all_models": "6847938.50", + "total_total_tokens_all_models": "13695877.0", + "avg_turn_count": "150.00", + "total_turn_count": "300.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-20250514/ab367b19-8fd0-45ec-bae0-3c9d2e542c16_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "371e101df70e41eec565a9b09936f85c152733a2c5513910294d0d1c7686e1cc", + "total_rows": 2 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/ab367b19-8fd0-45ec-bae0-3c9d2e542c16_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-20250514/ab367b19-8fd0-45ec-bae0-3c9d2e542c16_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..fb9ce9e8ad57b951e0d06ec0f0617539b5c42dac --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/ab367b19-8fd0-45ec-bae0-3c9d2e542c16_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:371e101df70e41eec565a9b09936f85c152733a2c5513910294d0d1c7686e1cc +size 1353461 diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/acfe3d37-4483-4691-98b7-5eef0546a38c.json b/data/terminalbench/anthropic/claude-opus-4-20250514/acfe3d37-4483-4691-98b7-5eef0546a38c.json new file mode 100644 index 0000000000000000000000000000000000000000..532d248967e274dbb7b5c294a42ea7fe2ccd5577 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/acfe3d37-4483-4691-98b7-5eef0546a38c.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-20250514/1786171522.279305", + "evaluation_timestamp": "1778055257.0", + "retrieved_timestamp": "1786171522.279305", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-20250514", + "id": "anthropic/claude-opus-4-20250514", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-20250514/acfe3d37-4483-4691-98b7-5eef0546a38c_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "dec540ded4f0b49dc5e59fc7e3ae44d93a6dd7c845238cfc4107afe38f215e51", + "total_rows": 10 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/acfe3d37-4483-4691-98b7-5eef0546a38c_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-20250514/acfe3d37-4483-4691-98b7-5eef0546a38c_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..9b524c7e44ea2357d55b23dd86393efc623d99d4 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/acfe3d37-4483-4691-98b7-5eef0546a38c_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dec540ded4f0b49dc5e59fc7e3ae44d93a6dd7c845238cfc4107afe38f215e51 +size 7300473 diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/aff1b49c-31f2-4333-b5a8-19c11a385eb1.json b/data/terminalbench/anthropic/claude-opus-4-20250514/aff1b49c-31f2-4333-b5a8-19c11a385eb1.json new file mode 100644 index 0000000000000000000000000000000000000000..19203e26fb133c8494bbe231a0da598339675529 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/aff1b49c-31f2-4333-b5a8-19c11a385eb1.json @@ -0,0 +1,121 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-20250514/1786163571.112323", + "evaluation_timestamp": "1777586285.0", + "retrieved_timestamp": "1786163571.112323", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-20250514", + "id": "anthropic/claude-opus-4-20250514", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive+C/+1ep/0f0848d3 for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 11, + "sample_ids": [ + "torch-pipeline-parallelism", + "protein-assembly", + "regex-log", + "feal-differential-cryptanalysis", + "polyglot-c-py", + "cancel-async-tasks", + "bn-fit-modify", + "git-leak-recovery", + "build-cython-ext", + "pytorch-model-recovery", + "nginx-request-logging" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1777586285.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.7, + "uncertainty": { + "standard_error": { + "value": 0.15275252316519464 + }, + "num_samples": 10 + }, + "details": { + "total_matched_trajectories": "10", + "stopping_reason_count_completed_on_successful_submit": "6", + "stopping_reason_count_repetition_guard": "4", + "avg_total_tokens_target_model": "424889.60", + "total_total_tokens_target_model": "4248896.0", + "avg_total_tokens_other_models": "0.00", + "total_total_tokens_other_models": "0.0", + "avg_total_tokens_all_models": "424889.60", + "total_total_tokens_all_models": "4248896.0", + "avg_turn_count": "29.30", + "total_turn_count": "293.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-20250514/aff1b49c-31f2-4333-b5a8-19c11a385eb1_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "f359b8aec3ed8b92dee50aabc7e0077a3ba9bfea30f4d79693e8136cbeb62a64", + "total_rows": 10 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/aff1b49c-31f2-4333-b5a8-19c11a385eb1_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-20250514/aff1b49c-31f2-4333-b5a8-19c11a385eb1_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..6e43431256bc57baa3ae0119f8cbbe236a29cea2 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/aff1b49c-31f2-4333-b5a8-19c11a385eb1_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f359b8aec3ed8b92dee50aabc7e0077a3ba9bfea30f4d79693e8136cbeb62a64 +size 3092526 diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/b31ae9f1-f772-43c6-ada5-3c6c7870c163.json b/data/terminalbench/anthropic/claude-opus-4-20250514/b31ae9f1-f772-43c6-ada5-3c6c7870c163.json new file mode 100644 index 0000000000000000000000000000000000000000..91f56ff03e5006d1a0dcb101de82467ef85b8b67 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/b31ae9f1-f772-43c6-ada5-3c6c7870c163.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-20250514/1786168184.899683", + "evaluation_timestamp": "1776711423.0", + "retrieved_timestamp": "1786168184.899683", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-20250514", + "id": "anthropic/claude-opus-4-20250514", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-20250514/b31ae9f1-f772-43c6-ada5-3c6c7870c163_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "571c0ea52b412eb6f696717164904208a2775d94e20ff7053d75a6f4168cde66", + "total_rows": 49 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/b31ae9f1-f772-43c6-ada5-3c6c7870c163_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-20250514/b31ae9f1-f772-43c6-ada5-3c6c7870c163_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..bf423f56f91a6ec296756ab76324f5930c15c2cd --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/b31ae9f1-f772-43c6-ada5-3c6c7870c163_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:571c0ea52b412eb6f696717164904208a2775d94e20ff7053d75a6f4168cde66 +size 22323347 diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/b5852b08-78ed-41e0-8468-631aeab581aa.json b/data/terminalbench/anthropic/claude-opus-4-20250514/b5852b08-78ed-41e0-8468-631aeab581aa.json new file mode 100644 index 0000000000000000000000000000000000000000..7416c3cdfb52b7792c8a1edb85f41ba604558825 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/b5852b08-78ed-41e0-8468-631aeab581aa.json @@ -0,0 +1,122 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-20250514/1786163613.28818", + "evaluation_timestamp": "1777156310.0", + "retrieved_timestamp": "1786163613.28818", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-20250514", + "id": "anthropic/claude-opus-4-20250514", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive+C/+3ep/2afbdcf2 for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 11, + "sample_ids": [ + "feal-linear-cryptanalysis", + "db-wal-recovery", + "schemelike-metacircular-eval", + "git-multibranch", + "video-processing", + "kv-store-grpc", + "install-windows-3.11", + "make-doom-for-mips", + "tune-mjcf", + "extract-moves-from-video", + "gcode-to-text" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1777156310.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.22727272727272727, + "uncertainty": { + "standard_error": { + "value": 0.12364973189759493 + }, + "num_samples": 13 + }, + "details": { + "total_matched_trajectories": "13", + "stopping_reason_count_repetition_guard": "8", + "stopping_reason_count_token_limit": "2", + "stopping_reason_count_completed_on_successful_submit": "3", + "avg_total_tokens_target_model": "2742899.54", + "total_total_tokens_target_model": "35657694.0", + "avg_total_tokens_other_models": "0.00", + "total_total_tokens_other_models": "0.0", + "avg_total_tokens_all_models": "2742899.54", + "total_total_tokens_all_models": "35657694.0", + "avg_turn_count": "68.46", + "total_turn_count": "890.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-20250514/b5852b08-78ed-41e0-8468-631aeab581aa_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "6f214e6a9c7ef30d98d8bfe2150b359258c9f68c327c3abc93c803eb57f46a2f", + "total_rows": 13 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/b5852b08-78ed-41e0-8468-631aeab581aa_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-20250514/b5852b08-78ed-41e0-8468-631aeab581aa_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..5d4c601553ecb645aee6a594710b9608f9abdb7d --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/b5852b08-78ed-41e0-8468-631aeab581aa_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6f214e6a9c7ef30d98d8bfe2150b359258c9f68c327c3abc93c803eb57f46a2f +size 7936514 diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/b6bf121a-635c-4e9e-82b2-529ffc1be05a.json b/data/terminalbench/anthropic/claude-opus-4-20250514/b6bf121a-635c-4e9e-82b2-529ffc1be05a.json new file mode 100644 index 0000000000000000000000000000000000000000..2298b0f0b5c06376973664c2fbad7762cce534cb --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/b6bf121a-635c-4e9e-82b2-529ffc1be05a.json @@ -0,0 +1,110 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-20250514/1786166220.254317", + "evaluation_timestamp": "1777292858.0", + "retrieved_timestamp": "1786166220.254317", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-20250514", + "id": "anthropic/claude-opus-4-20250514", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive/+1ep/b65eb3b6 for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 3, + "sample_ids": [ + "cancel-async-tasks", + "fix-ocaml-gc", + "rstan-to-pystan" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1777292858.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.0, + "uncertainty": { + "num_samples": 3 + }, + "details": { + "total_matched_trajectories": "3", + "stopping_reason_count_repetition_guard": "2", + "stopping_reason_count_token_limit": "1", + "avg_total_tokens_target_model": "5151105.00", + "total_total_tokens_target_model": "15453315.0", + "avg_total_tokens_other_models": "904.67", + "total_total_tokens_other_models": "2714.0", + "avg_total_tokens_all_models": "5152009.67", + "total_total_tokens_all_models": "15456029.0", + "avg_turn_count": "106.67", + "total_turn_count": "320.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-20250514/b6bf121a-635c-4e9e-82b2-529ffc1be05a_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "86d502036c59eccfca900a3b6f2e959edd47fb33ed34aa8f1fc23d36f30c1d48", + "total_rows": 3 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/b6bf121a-635c-4e9e-82b2-529ffc1be05a_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-20250514/b6bf121a-635c-4e9e-82b2-529ffc1be05a_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..efd95497a3afc14b482d9e9180f096a530d3df58 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/b6bf121a-635c-4e9e-82b2-529ffc1be05a_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:86d502036c59eccfca900a3b6f2e959edd47fb33ed34aa8f1fc23d36f30c1d48 +size 1153396 diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/b792e723-6fa1-4474-bad0-579a55516a2f.json b/data/terminalbench/anthropic/claude-opus-4-20250514/b792e723-6fa1-4474-bad0-579a55516a2f.json new file mode 100644 index 0000000000000000000000000000000000000000..55ccea908aa4e141cb9881ac53f68f5af15fbfa2 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/b792e723-6fa1-4474-bad0-579a55516a2f.json @@ -0,0 +1,107 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-20250514/1786166247.370756", + "evaluation_timestamp": "1776803720.0", + "retrieved_timestamp": "1786166247.370756", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-20250514", + "id": "anthropic/claude-opus-4-20250514", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive/+2ep/b897651b for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 1, + "sample_ids": [ + "train-fasttext" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1776803720.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.0, + "uncertainty": { + "num_samples": 2 + }, + "details": { + "total_matched_trajectories": "2", + "stopping_reason_count_repetition_guard": "2", + "avg_total_tokens_target_model": "6404594.50", + "total_total_tokens_target_model": "12809189.0", + "avg_total_tokens_other_models": "734.50", + "total_total_tokens_other_models": "1469.0", + "avg_total_tokens_all_models": "6405329.00", + "total_total_tokens_all_models": "12810658.0", + "avg_turn_count": "114.50", + "total_turn_count": "229.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-20250514/b792e723-6fa1-4474-bad0-579a55516a2f_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "cfe8c7911c6596ddb144737c97424ecacbec2cffda53a91a515e33b752183815", + "total_rows": 2 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/b792e723-6fa1-4474-bad0-579a55516a2f_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-20250514/b792e723-6fa1-4474-bad0-579a55516a2f_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..5b09ed5166d245efbc962f93eef088647d6bac15 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/b792e723-6fa1-4474-bad0-579a55516a2f_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cfe8c7911c6596ddb144737c97424ecacbec2cffda53a91a515e33b752183815 +size 1211960 diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/ba92b465-0ef4-4278-97e6-7ba5a5dafd1a.json b/data/terminalbench/anthropic/claude-opus-4-20250514/ba92b465-0ef4-4278-97e6-7ba5a5dafd1a.json new file mode 100644 index 0000000000000000000000000000000000000000..9c4e40c8bbab55fe0925d14fbe3c780af43d3f5c --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/ba92b465-0ef4-4278-97e6-7ba5a5dafd1a.json @@ -0,0 +1,126 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-20250514/1786163575.705132", + "evaluation_timestamp": "1776880341.0", + "retrieved_timestamp": "1786163575.705132", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-20250514", + "id": "anthropic/claude-opus-4-20250514", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive+C/+1ep/c03ee310 for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 16, + "sample_ids": [ + "gpt2-codegolf", + "llm-inference-batching-scheduler", + "break-filter-js-from-html", + "reshard-c4-data", + "write-compressor", + "winning-avg-corewars", + "pytorch-model-cli", + "password-recovery", + "portfolio-optimization", + "path-tracing-reverse", + "modernize-scientific-stack", + "prove-plus-comm", + "distribution-search", + "mteb-retrieve", + "pypi-server", + "multi-source-data-merger" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1776880341.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.5, + "uncertainty": { + "standard_error": { + "value": 0.12909944487358055 + }, + "num_samples": 16 + }, + "details": { + "total_matched_trajectories": "16", + "stopping_reason_count_repetition_guard": "9", + "stopping_reason_count_completed_on_successful_submit": "7", + "avg_total_tokens_target_model": "1812863.38", + "total_total_tokens_target_model": "29005814.0", + "avg_total_tokens_other_models": "0.00", + "total_total_tokens_other_models": "0.0", + "avg_total_tokens_all_models": "1812863.38", + "total_total_tokens_all_models": "29005814.0", + "avg_turn_count": "56.50", + "total_turn_count": "904.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-20250514/ba92b465-0ef4-4278-97e6-7ba5a5dafd1a_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "287c060f4f9ca9bfa9245a0b2fcb79f487abf7fe272716a3e3cca72eeabe1faa", + "total_rows": 16 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/ba92b465-0ef4-4278-97e6-7ba5a5dafd1a_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-20250514/ba92b465-0ef4-4278-97e6-7ba5a5dafd1a_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..13f67dc803c4f54228f61612e4ce9334f604f619 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/ba92b465-0ef4-4278-97e6-7ba5a5dafd1a_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:287c060f4f9ca9bfa9245a0b2fcb79f487abf7fe272716a3e3cca72eeabe1faa +size 6855980 diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/bcd27f4a-580f-43a7-9dff-abcf6b4191f0.json b/data/terminalbench/anthropic/claude-opus-4-20250514/bcd27f4a-580f-43a7-9dff-abcf6b4191f0.json new file mode 100644 index 0000000000000000000000000000000000000000..6d934b02ebb34aea3a658e77dc6d3ecdba36e7e6 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/bcd27f4a-580f-43a7-9dff-abcf6b4191f0.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-20250514/1786168098.973374", + "evaluation_timestamp": "1777349006.0", + "retrieved_timestamp": "1786168098.973374", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-20250514", + "id": "anthropic/claude-opus-4-20250514", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-20250514/bcd27f4a-580f-43a7-9dff-abcf6b4191f0_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "c278e51ac7d0c1dc5575c2fe377af76bb9c6ccddce8261953361b7bdb02f047b", + "total_rows": 11 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/bcd27f4a-580f-43a7-9dff-abcf6b4191f0_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-20250514/bcd27f4a-580f-43a7-9dff-abcf6b4191f0_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..ee88172b675e788bbac486e27663d797c03dad9f --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/bcd27f4a-580f-43a7-9dff-abcf6b4191f0_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c278e51ac7d0c1dc5575c2fe377af76bb9c6ccddce8261953361b7bdb02f047b +size 5934910 diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/c33d7dff-aa99-45f7-a314-1469817faab2.json b/data/terminalbench/anthropic/claude-opus-4-20250514/c33d7dff-aa99-45f7-a314-1469817faab2.json new file mode 100644 index 0000000000000000000000000000000000000000..bdf1ab87296242c59ac34b1584f0f1c56c8c2dca --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/c33d7dff-aa99-45f7-a314-1469817faab2.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-20250514/1786171515.304505", + "evaluation_timestamp": "1778239213.0", + "retrieved_timestamp": "1786171515.304505", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-20250514", + "id": "anthropic/claude-opus-4-20250514", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-20250514/c33d7dff-aa99-45f7-a314-1469817faab2_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "168f264f577a31fc85d98cb1ac8c5a334000286592409af3f7319b4c7d444b2f", + "total_rows": 5 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/c33d7dff-aa99-45f7-a314-1469817faab2_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-20250514/c33d7dff-aa99-45f7-a314-1469817faab2_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..e4a642b4f4676e8f4a47716d90c84f071fe7de39 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/c33d7dff-aa99-45f7-a314-1469817faab2_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:168f264f577a31fc85d98cb1ac8c5a334000286592409af3f7319b4c7d444b2f +size 4140850 diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/c9bebb70-8c08-4108-b52b-c253d65d22b6.json b/data/terminalbench/anthropic/claude-opus-4-20250514/c9bebb70-8c08-4108-b52b-c253d65d22b6.json new file mode 100644 index 0000000000000000000000000000000000000000..06531806b879581b769e3e870cf349ab79fbb524 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/c9bebb70-8c08-4108-b52b-c253d65d22b6.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-20250514/1786171547.13054", + "evaluation_timestamp": "1776346958.0", + "retrieved_timestamp": "1786171547.13054", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-20250514", + "id": "anthropic/claude-opus-4-20250514", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-20250514/c9bebb70-8c08-4108-b52b-c253d65d22b6_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "e8825ba70166c92a4c117402dbbcc2ef65b7e45701c8d1808f957522609596fb", + "total_rows": 8 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/c9bebb70-8c08-4108-b52b-c253d65d22b6_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-20250514/c9bebb70-8c08-4108-b52b-c253d65d22b6_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..bdf17223248180af17f5e5e2861c13807c7e20e0 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/c9bebb70-8c08-4108-b52b-c253d65d22b6_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e8825ba70166c92a4c117402dbbcc2ef65b7e45701c8d1808f957522609596fb +size 2060898 diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/cb38b345-f4fc-4bcd-86ca-09a0df97e7bd.json b/data/terminalbench/anthropic/claude-opus-4-20250514/cb38b345-f4fc-4bcd-86ca-09a0df97e7bd.json new file mode 100644 index 0000000000000000000000000000000000000000..40b524ff95a3ab8770e63b261fa25ad386874823 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/cb38b345-f4fc-4bcd-86ca-09a0df97e7bd.json @@ -0,0 +1,113 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-20250514/1786166304.785233", + "evaluation_timestamp": "1776361246.0", + "retrieved_timestamp": "1786166304.785233", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-20250514", + "id": "anthropic/claude-opus-4-20250514", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive/+3ep/5c31b3b6 for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 6, + "sample_ids": [ + "merge-diff-arc-agi-task", + "password-recovery", + "prove-plus-comm", + "mteb-retrieve", + "pypi-server", + "polyglot-rust-c" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1776361246.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.5555555555555556, + "uncertainty": { + "standard_error": { + "value": 0.1648044108243481 + }, + "num_samples": 18 + }, + "details": { + "total_matched_trajectories": "18", + "stopping_reason_count_token_limit": "16", + "stopping_reason_count_repetition_guard": "2", + "avg_total_tokens_target_model": "9104675.11", + "total_total_tokens_target_model": "163884152.0", + "avg_total_tokens_other_models": "41.56", + "total_total_tokens_other_models": "748.0", + "avg_total_tokens_all_models": "9104716.67", + "total_total_tokens_all_models": "163884900.0", + "avg_turn_count": "189.17", + "total_turn_count": "3405.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "max_connections": "10", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "max_connections": "10" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-20250514/cb38b345-f4fc-4bcd-86ca-09a0df97e7bd_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "44cba082a4364ace852623e5fa1a8ab8c779c770f98265b6d6043b4282b1edcc", + "total_rows": 18 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/cb38b345-f4fc-4bcd-86ca-09a0df97e7bd_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-20250514/cb38b345-f4fc-4bcd-86ca-09a0df97e7bd_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..593dc56cf3fbbf707701063e3970bba70cbc6cbe --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/cb38b345-f4fc-4bcd-86ca-09a0df97e7bd_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:44cba082a4364ace852623e5fa1a8ab8c779c770f98265b6d6043b4282b1edcc +size 13978606 diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/d126cea7-9a4c-4d8f-9158-a00d93eadb25.json b/data/terminalbench/anthropic/claude-opus-4-20250514/d126cea7-9a4c-4d8f-9158-a00d93eadb25.json new file mode 100644 index 0000000000000000000000000000000000000000..cf150f29bd201ecbcfc59b29865e90c50c3350f6 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/d126cea7-9a4c-4d8f-9158-a00d93eadb25.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-20250514/1786171249.854036", + "evaluation_timestamp": "1776338190.0", + "retrieved_timestamp": "1786171249.854036", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-20250514", + "id": "anthropic/claude-opus-4-20250514", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-20250514/d126cea7-9a4c-4d8f-9158-a00d93eadb25_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "3982623134b793e0aeb0f3171c191d2277a2c6f0424e7ebfee2de59726d58077", + "total_rows": 10 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/d126cea7-9a4c-4d8f-9158-a00d93eadb25_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-20250514/d126cea7-9a4c-4d8f-9158-a00d93eadb25_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..d43d72f7a15179998d9030b47208533bca0421d4 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/d126cea7-9a4c-4d8f-9158-a00d93eadb25_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3982623134b793e0aeb0f3171c191d2277a2c6f0424e7ebfee2de59726d58077 +size 6230515 diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/d1f7f6fc-d9e9-4e7d-b17e-0549d97abf1a.json b/data/terminalbench/anthropic/claude-opus-4-20250514/d1f7f6fc-d9e9-4e7d-b17e-0549d97abf1a.json new file mode 100644 index 0000000000000000000000000000000000000000..3911881ba3a269b4d2e54f060691b293412b4fea --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/d1f7f6fc-d9e9-4e7d-b17e-0549d97abf1a.json @@ -0,0 +1,118 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-20250514/1786166311.202793", + "evaluation_timestamp": "1778265866.0", + "retrieved_timestamp": "1786166311.202793", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-20250514", + "id": "anthropic/claude-opus-4-20250514", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive/+3ep/95715593 for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 6, + "sample_ids": [ + "sparql-university", + "sqlite-with-gcov", + "constraints-scheduling", + "dna-insert", + "vulnerable-secret", + "dna-assembly" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1778265866.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.4444444444444444, + "uncertainty": { + "standard_error": { + "value": 0.1859244503409057 + }, + "num_samples": 18 + }, + "details": { + "total_matched_trajectories": "18", + "stopping_reason_count_repetition_guard": "18", + "avg_total_tokens_target_model": "591228.22", + "total_total_tokens_target_model": "10642108.0", + "avg_total_tokens_other_models": "699.89", + "total_total_tokens_other_models": "12598.0", + "avg_total_tokens_all_models": "591928.11", + "total_total_tokens_all_models": "10654706.0", + "avg_turn_count": "35.33", + "total_turn_count": "636.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-20250514/d1f7f6fc-d9e9-4e7d-b17e-0549d97abf1a_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "8cfa23c221da66b5da18d25bc3fd4816093b42a52b001b259e27c5b46ef3a350", + "total_rows": 18 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/d1f7f6fc-d9e9-4e7d-b17e-0549d97abf1a_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-20250514/d1f7f6fc-d9e9-4e7d-b17e-0549d97abf1a_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..263a0ae4a45299e4f237387a38bca33958b5e902 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/d1f7f6fc-d9e9-4e7d-b17e-0549d97abf1a_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8cfa23c221da66b5da18d25bc3fd4816093b42a52b001b259e27c5b46ef3a350 +size 2651281 diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/d4c64347-2c06-490c-8eac-2ab059eb3080.json b/data/terminalbench/anthropic/claude-opus-4-20250514/d4c64347-2c06-490c-8eac-2ab059eb3080.json new file mode 100644 index 0000000000000000000000000000000000000000..3a34af4911ce71548e15184f44e7c75157e40980 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/d4c64347-2c06-490c-8eac-2ab059eb3080.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-20250514/1786171505.057203", + "evaluation_timestamp": "1776520455.0", + "retrieved_timestamp": "1786171505.057203", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-20250514", + "id": "anthropic/claude-opus-4-20250514", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-20250514/d4c64347-2c06-490c-8eac-2ab059eb3080_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "4ab5676675dc0145ff216d26384f059f126ff2a1a9ea2c71ad2752f4b694bba3", + "total_rows": 30 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/d4c64347-2c06-490c-8eac-2ab059eb3080_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-20250514/d4c64347-2c06-490c-8eac-2ab059eb3080_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..c0d5eee9101b451e5cd2e08434b4ab4229c8389b --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/d4c64347-2c06-490c-8eac-2ab059eb3080_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4ab5676675dc0145ff216d26384f059f126ff2a1a9ea2c71ad2752f4b694bba3 +size 23770759 diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/e13aed8b-641a-459a-b962-9f343df09d1e.json b/data/terminalbench/anthropic/claude-opus-4-20250514/e13aed8b-641a-459a-b962-9f343df09d1e.json new file mode 100644 index 0000000000000000000000000000000000000000..a9a7c47c9af7fc9c4da184cfad635be779a7202c --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/e13aed8b-641a-459a-b962-9f343df09d1e.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-20250514/1786171455.332073", + "evaluation_timestamp": "1776896096.0", + "retrieved_timestamp": "1786171455.332073", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-20250514", + "id": "anthropic/claude-opus-4-20250514", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-20250514/e13aed8b-641a-459a-b962-9f343df09d1e_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "48978c289fee511ef5fc6fb0dd82677f8ba1f2ab10319310a20868dd08c30d28", + "total_rows": 25 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/e13aed8b-641a-459a-b962-9f343df09d1e_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-20250514/e13aed8b-641a-459a-b962-9f343df09d1e_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..61f9013b2de767bf3e1cc157069e4425ee042005 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/e13aed8b-641a-459a-b962-9f343df09d1e_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:48978c289fee511ef5fc6fb0dd82677f8ba1f2ab10319310a20868dd08c30d28 +size 20629461 diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/e8e5838f-7d40-4709-98a5-ca6f47f4a90c.json b/data/terminalbench/anthropic/claude-opus-4-20250514/e8e5838f-7d40-4709-98a5-ca6f47f4a90c.json new file mode 100644 index 0000000000000000000000000000000000000000..86b53bacc91806b3a7012eca21cb2d928b501e2a --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/e8e5838f-7d40-4709-98a5-ca6f47f4a90c.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-20250514/1786171438.490739", + "evaluation_timestamp": "1776417565.0", + "retrieved_timestamp": "1786171438.490739", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-20250514", + "id": "anthropic/claude-opus-4-20250514", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-20250514/e8e5838f-7d40-4709-98a5-ca6f47f4a90c_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "723126f01ff4dbb12dcd83d341f9bb1384a650d2a90e80a77f219a743a76040f", + "total_rows": 6 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/e8e5838f-7d40-4709-98a5-ca6f47f4a90c_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-20250514/e8e5838f-7d40-4709-98a5-ca6f47f4a90c_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..ca181082cc9372012f35f57c755ad64d6058ba62 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/e8e5838f-7d40-4709-98a5-ca6f47f4a90c_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:723126f01ff4dbb12dcd83d341f9bb1384a650d2a90e80a77f219a743a76040f +size 3847549 diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/ed37ddcc-5eff-421f-8e95-44c68b5a0799.json b/data/terminalbench/anthropic/claude-opus-4-20250514/ed37ddcc-5eff-421f-8e95-44c68b5a0799.json new file mode 100644 index 0000000000000000000000000000000000000000..853dc3b950c8b83fdf4e36d154b94dde5040a89f --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/ed37ddcc-5eff-421f-8e95-44c68b5a0799.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-20250514/1786171390.74854", + "evaluation_timestamp": "1776680886.0", + "retrieved_timestamp": "1786171390.74854", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-20250514", + "id": "anthropic/claude-opus-4-20250514", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-20250514/ed37ddcc-5eff-421f-8e95-44c68b5a0799_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "3e9c904c305f338381e99cc1bd683b90e3fb219da2d0e2dd2d542a98aa8b64f0", + "total_rows": 1 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/ed37ddcc-5eff-421f-8e95-44c68b5a0799_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-20250514/ed37ddcc-5eff-421f-8e95-44c68b5a0799_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..35fe9fd9d4c55b7b584018f87f46a5d09e235cb2 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/ed37ddcc-5eff-421f-8e95-44c68b5a0799_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3e9c904c305f338381e99cc1bd683b90e3fb219da2d0e2dd2d542a98aa8b64f0 +size 352124 diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/f02eba4c-73e0-4c3a-b9d0-2ff2fea7c054.json b/data/terminalbench/anthropic/claude-opus-4-20250514/f02eba4c-73e0-4c3a-b9d0-2ff2fea7c054.json new file mode 100644 index 0000000000000000000000000000000000000000..35b75a70a69efd8b993d61ef0088ba05de3a7dea --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/f02eba4c-73e0-4c3a-b9d0-2ff2fea7c054.json @@ -0,0 +1,110 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-20250514/1786166249.296878", + "evaluation_timestamp": "1778265533.0", + "retrieved_timestamp": "1786166249.296878", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-20250514", + "id": "anthropic/claude-opus-4-20250514", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive/+2ep/91a8c639 for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 1, + "sample_ids": [ + "openssl-selfsigned-cert" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1778265533.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 1.0, + "uncertainty": { + "num_samples": 2 + }, + "details": { + "total_matched_trajectories": "2", + "stopping_reason_count_repetition_guard": "2", + "avg_total_tokens_target_model": "160774.50", + "total_total_tokens_target_model": "321549.0", + "avg_total_tokens_other_models": "1118.00", + "total_total_tokens_other_models": "2236.0", + "avg_total_tokens_all_models": "161892.50", + "total_total_tokens_all_models": "323785.0", + "avg_turn_count": "27.50", + "total_turn_count": "55.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-20250514/f02eba4c-73e0-4c3a-b9d0-2ff2fea7c054_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "49cc8acddecadd57ea8ef677c49853cd1e14606514aa2add98802be2800c5757", + "total_rows": 2 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/f02eba4c-73e0-4c3a-b9d0-2ff2fea7c054_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-20250514/f02eba4c-73e0-4c3a-b9d0-2ff2fea7c054_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..cdcfa985855e67f8ea1c431c69d3f14b0045c295 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/f02eba4c-73e0-4c3a-b9d0-2ff2fea7c054_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:49cc8acddecadd57ea8ef677c49853cd1e14606514aa2add98802be2800c5757 +size 164357 diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/f050749f-313e-4a55-b8c2-c7aa3f15a744.json b/data/terminalbench/anthropic/claude-opus-4-20250514/f050749f-313e-4a55-b8c2-c7aa3f15a744.json new file mode 100644 index 0000000000000000000000000000000000000000..fe630d1237d281757d32c0c8081ca75ef6cf404b --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/f050749f-313e-4a55-b8c2-c7aa3f15a744.json @@ -0,0 +1,108 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-20250514/1786163580.971624", + "evaluation_timestamp": "1777297872.0", + "retrieved_timestamp": "1786163580.971624", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-20250514", + "id": "anthropic/claude-opus-4-20250514", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive+C/+1ep/693e10be for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 2, + "sample_ids": [ + "db-wal-recovery", + "schemelike-metacircular-eval" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1777297872.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.0, + "uncertainty": { + "num_samples": 2 + }, + "details": { + "total_matched_trajectories": "2", + "stopping_reason_count_repetition_guard": "2", + "avg_total_tokens_target_model": "3212805.00", + "total_total_tokens_target_model": "6425610.0", + "avg_total_tokens_other_models": "0.00", + "total_total_tokens_other_models": "0.0", + "avg_total_tokens_all_models": "3212805.00", + "total_total_tokens_all_models": "6425610.0", + "avg_turn_count": "99.00", + "total_turn_count": "198.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-20250514/f050749f-313e-4a55-b8c2-c7aa3f15a744_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "9b66d9f4666e363dad656266ceb660745c5573eab3e9714d8ad7989e2fc3b30e", + "total_rows": 2 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/f050749f-313e-4a55-b8c2-c7aa3f15a744_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-20250514/f050749f-313e-4a55-b8c2-c7aa3f15a744_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..e2552a254f1c7c232a911b5538e8686cff51abe5 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/f050749f-313e-4a55-b8c2-c7aa3f15a744_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9b66d9f4666e363dad656266ceb660745c5573eab3e9714d8ad7989e2fc3b30e +size 1396266 diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/f223ab82-2333-4281-9729-a90fb3f0ee91.json b/data/terminalbench/anthropic/claude-opus-4-20250514/f223ab82-2333-4281-9729-a90fb3f0ee91.json new file mode 100644 index 0000000000000000000000000000000000000000..28b2bdadc2320390b633f78764fe739f490057ea --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/f223ab82-2333-4281-9729-a90fb3f0ee91.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-20250514/1786171539.482771", + "evaluation_timestamp": "1776433475.0", + "retrieved_timestamp": "1786171539.482771", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-20250514", + "id": "anthropic/claude-opus-4-20250514", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-20250514/f223ab82-2333-4281-9729-a90fb3f0ee91_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "1e08d2cde87ad5f96fcb1d0350109a25feb8068f426763762d2547c68188ad14", + "total_rows": 30 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/f223ab82-2333-4281-9729-a90fb3f0ee91_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-20250514/f223ab82-2333-4281-9729-a90fb3f0ee91_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..872f8f3ee3f5f1e828269d70045810d0808a986d --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/f223ab82-2333-4281-9729-a90fb3f0ee91_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1e08d2cde87ad5f96fcb1d0350109a25feb8068f426763762d2547c68188ad14 +size 23788202 diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/f3f43d27-a6b1-47ed-a986-11ba73d6f23b.json b/data/terminalbench/anthropic/claude-opus-4-20250514/f3f43d27-a6b1-47ed-a986-11ba73d6f23b.json new file mode 100644 index 0000000000000000000000000000000000000000..5958111dd4ed0f5ae46d4ef71a1ea0e4bc9f350e --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/f3f43d27-a6b1-47ed-a986-11ba73d6f23b.json @@ -0,0 +1,123 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-20250514/1786166288.015596", + "evaluation_timestamp": "1777054709.0", + "retrieved_timestamp": "1786166288.015596", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-20250514", + "id": "anthropic/claude-opus-4-20250514", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive/+3ep/2e9f4554 for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 13, + "sample_ids": [ + "bn-fit-modify", + "fix-ocaml-gc", + "git-leak-recovery", + "build-cython-ext", + "pytorch-model-recovery", + "sam-cell-seg", + "model-extraction-relu-logits", + "nginx-request-logging", + "large-scale-text-editing", + "sqlite-db-truncate", + "sanitize-git-repo", + "build-pmars", + "rstan-to-pystan" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1777054709.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.5, + "uncertainty": { + "standard_error": { + "value": 0.08851416742396573 + }, + "num_samples": 37 + }, + "details": { + "total_matched_trajectories": "37", + "stopping_reason_count_token_limit": "23", + "stopping_reason_count_repetition_guard": "14", + "avg_total_tokens_target_model": "8169854.86", + "total_total_tokens_target_model": "302284630.0", + "avg_total_tokens_other_models": "490.92", + "total_total_tokens_other_models": "18164.0", + "avg_total_tokens_all_models": "8170345.78", + "total_total_tokens_all_models": "302302794.0", + "avg_turn_count": "156.30", + "total_turn_count": "5783.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-20250514/f3f43d27-a6b1-47ed-a986-11ba73d6f23b_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "7d6ac211a217b720ac952a262c7f6078672d68744afae7b4f69443a76d51aa28", + "total_rows": 37 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/f3f43d27-a6b1-47ed-a986-11ba73d6f23b_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-20250514/f3f43d27-a6b1-47ed-a986-11ba73d6f23b_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..6d8afc6cd7332a68ec3d94625320a1d2170fc420 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/f3f43d27-a6b1-47ed-a986-11ba73d6f23b_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7d6ac211a217b720ac952a262c7f6078672d68744afae7b4f69443a76d51aa28 +size 30694517 diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/f401c363-fd4a-4d56-b34a-94aa61a4c6bc.json b/data/terminalbench/anthropic/claude-opus-4-20250514/f401c363-fd4a-4d56-b34a-94aa61a4c6bc.json new file mode 100644 index 0000000000000000000000000000000000000000..b7f0dc0942d00b7f1f65f62fb70c4dd1d08de8d0 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/f401c363-fd4a-4d56-b34a-94aa61a4c6bc.json @@ -0,0 +1,114 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-20250514/1786163617.492517", + "evaluation_timestamp": "1777679979.0", + "retrieved_timestamp": "1786163617.492517", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-20250514", + "id": "anthropic/claude-opus-4-20250514", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive+C/+3ep/e97e538a for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 4, + "sample_ids": [ + "fix-ocaml-gc", + "large-scale-text-editing", + "dna-insert", + "vulnerable-secret" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1777679979.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.75, + "uncertainty": { + "standard_error": { + "value": 0.25 + }, + "num_samples": 12 + }, + "details": { + "total_matched_trajectories": "12", + "stopping_reason_count_repetition_guard": "5", + "stopping_reason_count_completed_on_successful_submit": "7", + "avg_total_tokens_target_model": "831563.58", + "total_total_tokens_target_model": "9978763.0", + "avg_total_tokens_other_models": "0.00", + "total_total_tokens_other_models": "0.0", + "avg_total_tokens_all_models": "831563.58", + "total_total_tokens_all_models": "9978763.0", + "avg_turn_count": "49.92", + "total_turn_count": "599.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-20250514/f401c363-fd4a-4d56-b34a-94aa61a4c6bc_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "5beac9189dd2763a4ad357aea1c756a867cef06cdcaa16dcc85a6879ae526358", + "total_rows": 12 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/f401c363-fd4a-4d56-b34a-94aa61a4c6bc_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-20250514/f401c363-fd4a-4d56-b34a-94aa61a4c6bc_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..5798f7e04314474db46cae94bfbeacfcb0ee9a68 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/f401c363-fd4a-4d56-b34a-94aa61a4c6bc_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5beac9189dd2763a4ad357aea1c756a867cef06cdcaa16dcc85a6879ae526358 +size 4454631 diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/fe40c161-9c3b-4454-9887-163d87c260c8.json b/data/terminalbench/anthropic/claude-opus-4-20250514/fe40c161-9c3b-4454-9887-163d87c260c8.json new file mode 100644 index 0000000000000000000000000000000000000000..bdddceb06e0e7e5149e23a84bb59831a4dfd9e66 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/fe40c161-9c3b-4454-9887-163d87c260c8.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-20250514/1786171378.553127", + "evaluation_timestamp": "1776520455.0", + "retrieved_timestamp": "1786171378.553127", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-20250514", + "id": "anthropic/claude-opus-4-20250514", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-20250514/fe40c161-9c3b-4454-9887-163d87c260c8_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "989eb0ffc94308b53f8251a9a5df1bf01e0b393f14cb41eef1a2d72753ca7291", + "total_rows": 20 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-20250514/fe40c161-9c3b-4454-9887-163d87c260c8_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-20250514/fe40c161-9c3b-4454-9887-163d87c260c8_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..f4d067768306184009130657883d50eb5428b0be --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-20250514/fe40c161-9c3b-4454-9887-163d87c260c8_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:989eb0ffc94308b53f8251a9a5df1bf01e0b393f14cb41eef1a2d72753ca7291 +size 16153618 diff --git a/data/terminalbench/anthropic/claude-opus-4-5-20251101/2f0f1c67-7602-4d19-8457-d9783427bc07.json b/data/terminalbench/anthropic/claude-opus-4-5-20251101/2f0f1c67-7602-4d19-8457-d9783427bc07.json new file mode 100644 index 0000000000000000000000000000000000000000..817c0b12b70b5a3d2196cd458dc8c6c16b32339d --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-5-20251101/2f0f1c67-7602-4d19-8457-d9783427bc07.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-5-20251101/1786170173.335941", + "evaluation_timestamp": "1775646127.0", + "retrieved_timestamp": "1786170173.335941", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-5-20251101", + "id": "anthropic/claude-opus-4-5-20251101", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-5-20251101/2f0f1c67-7602-4d19-8457-d9783427bc07_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "3e1cb0d550857c5a192503b5602f623890fc0fa13e25e41d950de20bb71f9c50", + "total_rows": 7 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-5-20251101/2f0f1c67-7602-4d19-8457-d9783427bc07_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-5-20251101/2f0f1c67-7602-4d19-8457-d9783427bc07_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..52e6ed54155ea863698eadfbea65cd100b5631a1 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-5-20251101/2f0f1c67-7602-4d19-8457-d9783427bc07_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3e1cb0d550857c5a192503b5602f623890fc0fa13e25e41d950de20bb71f9c50 +size 5706909 diff --git a/data/terminalbench/anthropic/claude-opus-4-5-20251101/30e29194-bc13-4e3a-9ee4-f4981e032a8f.json b/data/terminalbench/anthropic/claude-opus-4-5-20251101/30e29194-bc13-4e3a-9ee4-f4981e032a8f.json new file mode 100644 index 0000000000000000000000000000000000000000..6542338bb745072d983c9c309bdcd4a4290d29d7 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-5-20251101/30e29194-bc13-4e3a-9ee4-f4981e032a8f.json @@ -0,0 +1,120 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-5-20251101/1786163483.249122", + "evaluation_timestamp": "1776824322.0", + "retrieved_timestamp": "1786163483.249122", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-5-20251101", + "id": "anthropic/claude-opus-4-5-20251101", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive+C/+2ep/b92cc54a for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 9, + "sample_ids": [ + "model-extraction-relu-logits", + "sparql-university", + "sanitize-git-repo", + "rstan-to-pystan", + "sqlite-with-gcov", + "openssl-selfsigned-cert", + "constraints-scheduling", + "dna-insert", + "vulnerable-secret" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1776824322.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.8888888888888888, + "uncertainty": { + "standard_error": { + "value": 0.07349309197401639 + }, + "num_samples": 18 + }, + "details": { + "total_matched_trajectories": "18", + "stopping_reason_count_completed_on_successful_submit": "15", + "stopping_reason_count_repetition_guard": "2", + "stopping_reason_count_token_limit": "1", + "avg_total_tokens_target_model": "1289517.89", + "total_total_tokens_target_model": "23211322.0", + "avg_total_tokens_other_models": "0.00", + "total_total_tokens_other_models": "0.0", + "avg_total_tokens_all_models": "1289517.89", + "total_total_tokens_all_models": "23211322.0", + "avg_turn_count": "28.28", + "total_turn_count": "509.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-5-20251101/30e29194-bc13-4e3a-9ee4-f4981e032a8f_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "04d575185bd4b0d1341bf5195a4b0c9e5e0ccaa084cbd8be943d9db0a5a8e01d", + "total_rows": 18 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-5-20251101/30e29194-bc13-4e3a-9ee4-f4981e032a8f_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-5-20251101/30e29194-bc13-4e3a-9ee4-f4981e032a8f_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..b62de7ad53158f1ffd953f61fc6188609d7123aa --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-5-20251101/30e29194-bc13-4e3a-9ee4-f4981e032a8f_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:04d575185bd4b0d1341bf5195a4b0c9e5e0ccaa084cbd8be943d9db0a5a8e01d +size 3779308 diff --git a/data/terminalbench/anthropic/claude-opus-4-5-20251101/3e8d84d2-e676-44d5-841d-ac7a4642d7ab.json b/data/terminalbench/anthropic/claude-opus-4-5-20251101/3e8d84d2-e676-44d5-841d-ac7a4642d7ab.json new file mode 100644 index 0000000000000000000000000000000000000000..4003d5cc0320e9a941c333a9f3c6da4ad8739053 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-5-20251101/3e8d84d2-e676-44d5-841d-ac7a4642d7ab.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-5-20251101/1786170162.483102", + "evaluation_timestamp": "1775665083.0", + "retrieved_timestamp": "1786170162.483102", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-5-20251101", + "id": "anthropic/claude-opus-4-5-20251101", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-5-20251101/3e8d84d2-e676-44d5-841d-ac7a4642d7ab_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "eb6a2f908992d959fbacd605d77b759e821320e430103a5da4a24f28a719ec1b", + "total_rows": 47 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-5-20251101/3e8d84d2-e676-44d5-841d-ac7a4642d7ab_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-5-20251101/3e8d84d2-e676-44d5-841d-ac7a4642d7ab_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..457be1209963a9bde89545fdd6b0fb925493c7c7 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-5-20251101/3e8d84d2-e676-44d5-841d-ac7a4642d7ab_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:eb6a2f908992d959fbacd605d77b759e821320e430103a5da4a24f28a719ec1b +size 31380759 diff --git a/data/terminalbench/anthropic/claude-opus-4-5-20251101/4668c113-e926-46ab-86ac-6b21912a4752.json b/data/terminalbench/anthropic/claude-opus-4-5-20251101/4668c113-e926-46ab-86ac-6b21912a4752.json new file mode 100644 index 0000000000000000000000000000000000000000..97672a9034cbd8012b194dfce1d9598d2ac91c80 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-5-20251101/4668c113-e926-46ab-86ac-6b21912a4752.json @@ -0,0 +1,105 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-5-20251101/1786166087.40576", + "evaluation_timestamp": "1775665083.0", + "retrieved_timestamp": "1786166087.40576", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-5-20251101", + "id": "anthropic/claude-opus-4-5-20251101", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive/+1ep for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 2, + "sample_ids": [ + "feal-linear-cryptanalysis", + "torch-pipeline-parallelism" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1775665083.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.5, + "uncertainty": { + "standard_error": { + "value": 0.5 + }, + "num_samples": 2 + }, + "details": { + "total_matched_trajectories": "2", + "stopping_reason_count_repetition_guard": "1", + "stopping_reason_count_token_limit": "1", + "avg_total_tokens_target_model": "8566157.50", + "total_total_tokens_target_model": "17132315.0", + "avg_total_tokens_other_models": "132.00", + "total_total_tokens_other_models": "264.0", + "avg_total_tokens_all_models": "8566289.50", + "total_total_tokens_all_models": "17132579.0", + "avg_turn_count": "164.00", + "total_turn_count": "328.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-5-20251101/4668c113-e926-46ab-86ac-6b21912a4752_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "0f26cba036059d2d73d33244810f063822f8e61830e705d9371e98e4b17328ef", + "total_rows": 2 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-5-20251101/4668c113-e926-46ab-86ac-6b21912a4752_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-5-20251101/4668c113-e926-46ab-86ac-6b21912a4752_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..dc4dec455751160716bc16fc3c64dea50ddc333e --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-5-20251101/4668c113-e926-46ab-86ac-6b21912a4752_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0f26cba036059d2d73d33244810f063822f8e61830e705d9371e98e4b17328ef +size 2203396 diff --git a/data/terminalbench/anthropic/claude-opus-4-5-20251101/5c8d94a8-7f19-431d-ab11-c7d95be6a419.json b/data/terminalbench/anthropic/claude-opus-4-5-20251101/5c8d94a8-7f19-431d-ab11-c7d95be6a419.json new file mode 100644 index 0000000000000000000000000000000000000000..9ec96eb33dfdd53b2d3cf98f842380d62395b813 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-5-20251101/5c8d94a8-7f19-431d-ab11-c7d95be6a419.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-5-20251101/1786170080.14263", + "evaluation_timestamp": "1775558526.0", + "retrieved_timestamp": "1786170080.14263", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.202,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-5-20251101", + "id": "anthropic/claude-opus-4-5-20251101", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-5-20251101/5c8d94a8-7f19-431d-ab11-c7d95be6a419_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "ce5f988ed4c421138c379b10fbbf1fc9b72b0dd9139d326e9814b7f20c292d20", + "total_rows": 20 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-5-20251101/5c8d94a8-7f19-431d-ab11-c7d95be6a419_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-5-20251101/5c8d94a8-7f19-431d-ab11-c7d95be6a419_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..7b568f3ccfd63e021f7405b4d48a341f18e57210 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-5-20251101/5c8d94a8-7f19-431d-ab11-c7d95be6a419_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ce5f988ed4c421138c379b10fbbf1fc9b72b0dd9139d326e9814b7f20c292d20 +size 11264036 diff --git a/data/terminalbench/anthropic/claude-opus-4-5-20251101/610ed0a3-9bbf-4276-8d3b-4a54e16e5768.json b/data/terminalbench/anthropic/claude-opus-4-5-20251101/610ed0a3-9bbf-4276-8d3b-4a54e16e5768.json new file mode 100644 index 0000000000000000000000000000000000000000..5df5505f764338e7c4b11f15f2ed31c218e583ad --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-5-20251101/610ed0a3-9bbf-4276-8d3b-4a54e16e5768.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-5-20251101/1786170136.343602", + "evaluation_timestamp": "1775558526.0", + "retrieved_timestamp": "1786170136.343602", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.202,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-5-20251101", + "id": "anthropic/claude-opus-4-5-20251101", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-5-20251101/610ed0a3-9bbf-4276-8d3b-4a54e16e5768_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "e86514cb61284dbc2ebf885188bfed52d61acc8336ab33b6d4cdaa9053da443c", + "total_rows": 30 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-5-20251101/610ed0a3-9bbf-4276-8d3b-4a54e16e5768_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-5-20251101/610ed0a3-9bbf-4276-8d3b-4a54e16e5768_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..a32103759d1bd214ca4685d9c40ea32c840ec415 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-5-20251101/610ed0a3-9bbf-4276-8d3b-4a54e16e5768_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e86514cb61284dbc2ebf885188bfed52d61acc8336ab33b6d4cdaa9053da443c +size 18810869 diff --git a/data/terminalbench/anthropic/claude-opus-4-5-20251101/6179833e-7c2d-47b6-a9e4-e77b181ce39e.json b/data/terminalbench/anthropic/claude-opus-4-5-20251101/6179833e-7c2d-47b6-a9e4-e77b181ce39e.json new file mode 100644 index 0000000000000000000000000000000000000000..62c42b4efa439e05bf345511d75b66fb6d719f73 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-5-20251101/6179833e-7c2d-47b6-a9e4-e77b181ce39e.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-5-20251101/1786167952.20053", + "evaluation_timestamp": "1776629382.0", + "retrieved_timestamp": "1786167952.20053", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-5-20251101", + "id": "anthropic/claude-opus-4-5-20251101", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-5-20251101/6179833e-7c2d-47b6-a9e4-e77b181ce39e_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "6d6aa8b4f660390a4428f8b3d86802c5e36d30231370f4625f92fe9d5cb9d4ee", + "total_rows": 96 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-5-20251101/6179833e-7c2d-47b6-a9e4-e77b181ce39e_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-5-20251101/6179833e-7c2d-47b6-a9e4-e77b181ce39e_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..053c647edc9689c4189e09dde69c394c7a66280b --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-5-20251101/6179833e-7c2d-47b6-a9e4-e77b181ce39e_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6d6aa8b4f660390a4428f8b3d86802c5e36d30231370f4625f92fe9d5cb9d4ee +size 38537852 diff --git a/data/terminalbench/anthropic/claude-opus-4-5-20251101/6c946a85-5d8d-48f7-a60c-50cb4d749bb9.json b/data/terminalbench/anthropic/claude-opus-4-5-20251101/6c946a85-5d8d-48f7-a60c-50cb4d749bb9.json new file mode 100644 index 0000000000000000000000000000000000000000..6251295f96b75356b71e1553adc5d3357e0a890f --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-5-20251101/6c946a85-5d8d-48f7-a60c-50cb4d749bb9.json @@ -0,0 +1,120 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-5-20251101/1786163443.861293", + "evaluation_timestamp": "1776696546.0", + "retrieved_timestamp": "1786163443.861293", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-5-20251101", + "id": "anthropic/claude-opus-4-5-20251101", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive+C/+1ep/7a40229c for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 10, + "sample_ids": [ + "train-fasttext", + "qemu-alpine-ssh", + "circuit-fibsqrt", + "mailman", + "query-optimize", + "extract-elf", + "protein-assembly", + "code-from-image", + "fix-code-vulnerability", + "financial-document-processor" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1776696546.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.75, + "uncertainty": { + "standard_error": { + "value": 0.16366341767699427 + }, + "num_samples": 8 + }, + "details": { + "total_matched_trajectories": "8", + "stopping_reason_count_completed_on_successful_submit": "6", + "stopping_reason_count_completed_with_submit": "1", + "stopping_reason_count_repetition_guard": "1", + "avg_total_tokens_target_model": "2239717.12", + "total_total_tokens_target_model": "17917737.0", + "avg_total_tokens_other_models": "0.00", + "total_total_tokens_other_models": "0.0", + "avg_total_tokens_all_models": "2239717.12", + "total_total_tokens_all_models": "17917737.0", + "avg_turn_count": "56.75", + "total_turn_count": "454.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "3600", + "attempt_timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "3600", + "max_connections": "10" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-5-20251101/6c946a85-5d8d-48f7-a60c-50cb4d749bb9_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "8125f941820c58c0093a7884e497cc08491633890ed426fa98b1fc5574d1f877", + "total_rows": 8 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-5-20251101/6c946a85-5d8d-48f7-a60c-50cb4d749bb9_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-5-20251101/6c946a85-5d8d-48f7-a60c-50cb4d749bb9_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..941cd04edcd94ab1467446adbf9a192e0ab95d66 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-5-20251101/6c946a85-5d8d-48f7-a60c-50cb4d749bb9_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8125f941820c58c0093a7884e497cc08491633890ed426fa98b1fc5574d1f877 +size 2451471 diff --git a/data/terminalbench/anthropic/claude-opus-4-5-20251101/7fa71351-e315-43be-8322-58b6588d3cbe.json b/data/terminalbench/anthropic/claude-opus-4-5-20251101/7fa71351-e315-43be-8322-58b6588d3cbe.json new file mode 100644 index 0000000000000000000000000000000000000000..5561fee5e2b0451b4065613e20e62c0ea6d86b21 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-5-20251101/7fa71351-e315-43be-8322-58b6588d3cbe.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-5-20251101/1786167911.28626", + "evaluation_timestamp": "1775683252.0", + "retrieved_timestamp": "1786167911.28626", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-5-20251101", + "id": "anthropic/claude-opus-4-5-20251101", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-5-20251101/7fa71351-e315-43be-8322-58b6588d3cbe_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "737eac40b6d85ae9fc1ca0500b8a6471c97f47a07ed41ded888658497725ef2b", + "total_rows": 37 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-5-20251101/7fa71351-e315-43be-8322-58b6588d3cbe_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-5-20251101/7fa71351-e315-43be-8322-58b6588d3cbe_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..43eff7344779446033c48cb75aad1066fa3e4033 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-5-20251101/7fa71351-e315-43be-8322-58b6588d3cbe_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:737eac40b6d85ae9fc1ca0500b8a6471c97f47a07ed41ded888658497725ef2b +size 12337091 diff --git a/data/terminalbench/anthropic/claude-opus-4-5-20251101/95379e7f-9390-40da-9161-119527f215bf.json b/data/terminalbench/anthropic/claude-opus-4-5-20251101/95379e7f-9390-40da-9161-119527f215bf.json new file mode 100644 index 0000000000000000000000000000000000000000..c61194f0b2c5508deb12b5b04d53b3627c2c868b --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-5-20251101/95379e7f-9390-40da-9161-119527f215bf.json @@ -0,0 +1,120 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-5-20251101/1786163447.325428", + "evaluation_timestamp": "1776803719.0", + "retrieved_timestamp": "1786163447.325428", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-5-20251101", + "id": "anthropic/claude-opus-4-5-20251101", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive+C/+1ep/58ffee33 for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 9, + "sample_ids": [ + "train-fasttext", + "qemu-alpine-ssh", + "torch-pipeline-parallelism", + "extract-moves-from-video", + "financial-document-processor", + "polyglot-c-py", + "sam-cell-seg", + "large-scale-text-editing", + "sqlite-db-truncate" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1776803719.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.6666666666666666, + "uncertainty": { + "standard_error": { + "value": 0.16666666666666666 + }, + "num_samples": 9 + }, + "details": { + "total_matched_trajectories": "9", + "stopping_reason_count_completed_on_successful_submit": "4", + "stopping_reason_count_repetition_guard": "4", + "stopping_reason_count_token_limit": "1", + "avg_total_tokens_target_model": "2751763.11", + "total_total_tokens_target_model": "24765868.0", + "avg_total_tokens_other_models": "0.00", + "total_total_tokens_other_models": "0.0", + "avg_total_tokens_all_models": "2751763.11", + "total_total_tokens_all_models": "24765868.0", + "avg_turn_count": "60.56", + "total_turn_count": "545.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-5-20251101/95379e7f-9390-40da-9161-119527f215bf_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "1cf0b230153062df3d06c804e7464a3e0033961faa8278a77bcc082712c56db4", + "total_rows": 9 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-5-20251101/95379e7f-9390-40da-9161-119527f215bf_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-5-20251101/95379e7f-9390-40da-9161-119527f215bf_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..4ea05a0062297c54463f0debf925bfa2319bbe1c --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-5-20251101/95379e7f-9390-40da-9161-119527f215bf_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1cf0b230153062df3d06c804e7464a3e0033961faa8278a77bcc082712c56db4 +size 4508139 diff --git a/data/terminalbench/anthropic/claude-opus-4-5-20251101/9a921621-c252-4318-a386-8964146f2b10.json b/data/terminalbench/anthropic/claude-opus-4-5-20251101/9a921621-c252-4318-a386-8964146f2b10.json new file mode 100644 index 0000000000000000000000000000000000000000..c39225c314cd43e65330a24b7bbbfe9beaaa294c --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-5-20251101/9a921621-c252-4318-a386-8964146f2b10.json @@ -0,0 +1,107 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-5-20251101/1786163441.274439", + "evaluation_timestamp": "1776880341.0", + "retrieved_timestamp": "1786163441.274439", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-5-20251101", + "id": "anthropic/claude-opus-4-5-20251101", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive+C/+1ep/5dd93e71 for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 1, + "sample_ids": [ + "dna-assembly" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1776880341.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 1.0, + "uncertainty": { + "num_samples": 1 + }, + "details": { + "total_matched_trajectories": "1", + "stopping_reason_count_completed_on_successful_submit": "1", + "avg_total_tokens_target_model": "2912943.00", + "total_total_tokens_target_model": "2912943.0", + "avg_total_tokens_other_models": "0.00", + "total_total_tokens_other_models": "0.0", + "avg_total_tokens_all_models": "2912943.00", + "total_total_tokens_all_models": "2912943.0", + "avg_turn_count": "72.00", + "total_turn_count": "72.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-5-20251101/9a921621-c252-4318-a386-8964146f2b10_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "235d6f42de62f30f5196cd51bd841124a654e6c36f1cc1faf90e99db46735f84", + "total_rows": 1 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-5-20251101/9a921621-c252-4318-a386-8964146f2b10_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-5-20251101/9a921621-c252-4318-a386-8964146f2b10_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..d1c4034b22178e2d1a8e6e627ede271a076c8dff --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-5-20251101/9a921621-c252-4318-a386-8964146f2b10_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:235d6f42de62f30f5196cd51bd841124a654e6c36f1cc1faf90e99db46735f84 +size 710069 diff --git a/data/terminalbench/anthropic/claude-opus-4-5-20251101/9da5267c-d47f-48cd-9c78-f9caacb84832.json b/data/terminalbench/anthropic/claude-opus-4-5-20251101/9da5267c-d47f-48cd-9c78-f9caacb84832.json new file mode 100644 index 0000000000000000000000000000000000000000..ba43b3e29577eea6f0a926be42fe60915a6db410 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-5-20251101/9da5267c-d47f-48cd-9c78-f9caacb84832.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-5-20251101/1786167917.901188", + "evaluation_timestamp": "1776680886.0", + "retrieved_timestamp": "1786167917.901188", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-5-20251101", + "id": "anthropic/claude-opus-4-5-20251101", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-5-20251101/9da5267c-d47f-48cd-9c78-f9caacb84832_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "16e5938e937a7ba31bab963509e5ddafb71b6ce1369f63ace030b96980fcaa53", + "total_rows": 7 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-5-20251101/9da5267c-d47f-48cd-9c78-f9caacb84832_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-5-20251101/9da5267c-d47f-48cd-9c78-f9caacb84832_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..6af798a7821b20d921b853226c30a29b2f323323 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-5-20251101/9da5267c-d47f-48cd-9c78-f9caacb84832_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:16e5938e937a7ba31bab963509e5ddafb71b6ce1369f63ace030b96980fcaa53 +size 2002498 diff --git a/data/terminalbench/anthropic/claude-opus-4-5-20251101/a9d62c8f-bef6-4d4c-ad39-ed3dda78bbeb.json b/data/terminalbench/anthropic/claude-opus-4-5-20251101/a9d62c8f-bef6-4d4c-ad39-ed3dda78bbeb.json new file mode 100644 index 0000000000000000000000000000000000000000..a4fecc6666fdf57ad56eeb43523d22a906f278c5 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-5-20251101/a9d62c8f-bef6-4d4c-ad39-ed3dda78bbeb.json @@ -0,0 +1,142 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-5-20251101/1786163527.505561", + "evaluation_timestamp": "1775733996.0", + "retrieved_timestamp": "1786163527.505561", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-5-20251101", + "id": "anthropic/claude-opus-4-5-20251101", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive+C/+8ep for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 37, + "sample_ids": [ + "gpt2-codegolf", + "llm-inference-batching-scheduler", + "break-filter-js-from-html", + "reshard-c4-data", + "write-compressor", + "merge-diff-arc-agi-task", + "winning-avg-corewars", + "log-summary-date-ranges", + "pytorch-model-cli", + "largest-eigenval", + "password-recovery", + "portfolio-optimization", + "path-tracing-reverse", + "regex-chess", + "modernize-scientific-stack", + "torch-tensor-parallelism", + "path-tracing", + "prove-plus-comm", + "feal-linear-cryptanalysis", + "caffe-cifar-10", + "distribution-search", + "mteb-retrieve", + "pypi-server", + "custom-memory-heap-crash", + "adaptive-rejection-sampler", + "multi-source-data-merger", + "configure-git-webserver", + "crack-7z-hash", + "chess-best-move", + "cobol-modernization", + "overfull-hbox", + "polyglot-rust-c", + "hf-model-inference", + "compile-compcert", + "db-wal-recovery", + "headless-terminal", + "fix-git" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1775733996.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.706081081081081, + "uncertainty": { + "standard_error": { + "value": 0.06502342357320744 + }, + "num_samples": 296 + }, + "details": { + "total_matched_trajectories": "185", + "stopping_reason_count_repetition_guard": "35", + "stopping_reason_count_completed_on_successful_submit": "120", + "stopping_reason_count_token_limit": "24", + "stopping_reason_count_completed_with_submit": "6", + "avg_total_tokens_target_model": "2457077.54", + "total_total_tokens_target_model": "454559344.0", + "avg_total_tokens_other_models": "0.00", + "total_total_tokens_other_models": "0.0", + "avg_total_tokens_all_models": "2457077.54", + "total_total_tokens_all_models": "454559344.0", + "avg_turn_count": "55.04", + "total_turn_count": "10182.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-5-20251101/a9d62c8f-bef6-4d4c-ad39-ed3dda78bbeb_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "1a58bf0baffefd277bdb7cd1f036275bad1354eed214e4a91db6eaebb8385b71", + "total_rows": 296 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-5-20251101/a9d62c8f-bef6-4d4c-ad39-ed3dda78bbeb_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-5-20251101/a9d62c8f-bef6-4d4c-ad39-ed3dda78bbeb_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..9ae4413e44d0bcb973e213dae08ca23c55f40382 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-5-20251101/a9d62c8f-bef6-4d4c-ad39-ed3dda78bbeb_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1a58bf0baffefd277bdb7cd1f036275bad1354eed214e4a91db6eaebb8385b71 +size 96147014 diff --git a/data/terminalbench/anthropic/claude-opus-4-5-20251101/b63f8207-3133-422a-b0f4-fad245a3d1d1.json b/data/terminalbench/anthropic/claude-opus-4-5-20251101/b63f8207-3133-422a-b0f4-fad245a3d1d1.json new file mode 100644 index 0000000000000000000000000000000000000000..426b57c7f94b3c30f4fc2b0aa76b3c865304b8fe --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-5-20251101/b63f8207-3133-422a-b0f4-fad245a3d1d1.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-5-20251101/1786170207.854472", + "evaluation_timestamp": "1775604414.0", + "retrieved_timestamp": "1786170207.854472", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.202,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-5-20251101", + "id": "anthropic/claude-opus-4-5-20251101", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-5-20251101/b63f8207-3133-422a-b0f4-fad245a3d1d1_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "dfa0a4bdccd4edf1c425efc942fd0da8f28ff8d5547d90ab185e8ca574b6c662", + "total_rows": 20 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-5-20251101/b63f8207-3133-422a-b0f4-fad245a3d1d1_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-5-20251101/b63f8207-3133-422a-b0f4-fad245a3d1d1_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..9c43775395acc0887dccaec1bfcd505d407c328d --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-5-20251101/b63f8207-3133-422a-b0f4-fad245a3d1d1_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dfa0a4bdccd4edf1c425efc942fd0da8f28ff8d5547d90ab185e8ca574b6c662 +size 17821808 diff --git a/data/terminalbench/anthropic/claude-opus-4-5-20251101/be189a75-82e8-4b0d-8d3c-d7755de08bbd.json b/data/terminalbench/anthropic/claude-opus-4-5-20251101/be189a75-82e8-4b0d-8d3c-d7755de08bbd.json new file mode 100644 index 0000000000000000000000000000000000000000..6f817fb1b2d70096f24256bddca45741c7f55c81 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-5-20251101/be189a75-82e8-4b0d-8d3c-d7755de08bbd.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-5-20251101/1786167972.339002", + "evaluation_timestamp": "1776629382.0", + "retrieved_timestamp": "1786167972.339002", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-5-20251101", + "id": "anthropic/claude-opus-4-5-20251101", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-5-20251101/be189a75-82e8-4b0d-8d3c-d7755de08bbd_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "63750269fb0ea882dca8c6fde41dd4f4cd81f7ab0417066abdfb6b889b2e7452", + "total_rows": 36 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-5-20251101/be189a75-82e8-4b0d-8d3c-d7755de08bbd_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-5-20251101/be189a75-82e8-4b0d-8d3c-d7755de08bbd_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..22faf1ae81354398c700eda137315b251485d08c --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-5-20251101/be189a75-82e8-4b0d-8d3c-d7755de08bbd_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:63750269fb0ea882dca8c6fde41dd4f4cd81f7ab0417066abdfb6b889b2e7452 +size 18420407 diff --git a/data/terminalbench/anthropic/claude-opus-4-5-20251101/c3b0b891-5f15-4a87-8aac-35581f811268.json b/data/terminalbench/anthropic/claude-opus-4-5-20251101/c3b0b891-5f15-4a87-8aac-35581f811268.json new file mode 100644 index 0000000000000000000000000000000000000000..d9a4f5798d1e617207d2a48aff4e6e2bde8d0d91 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-5-20251101/c3b0b891-5f15-4a87-8aac-35581f811268.json @@ -0,0 +1,110 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-5-20251101/1786163486.859185", + "evaluation_timestamp": "1776359447.0", + "retrieved_timestamp": "1786163486.859185", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-5-20251101", + "id": "anthropic/claude-opus-4-5-20251101", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive+C/+4ep/9b933309 for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 3, + "sample_ids": [ + "schemelike-metacircular-eval", + "qemu-startup", + "build-pov-ray" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1776359447.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.6666666666666666, + "uncertainty": { + "standard_error": { + "value": 0.22047927592204924 + }, + "num_samples": 12 + }, + "details": { + "total_matched_trajectories": "12", + "stopping_reason_count_repetition_guard": "4", + "stopping_reason_count_completed_on_successful_submit": "8", + "avg_total_tokens_target_model": "3008239.08", + "total_total_tokens_target_model": "36098869.0", + "avg_total_tokens_other_models": "0.00", + "total_total_tokens_other_models": "0.0", + "avg_total_tokens_all_models": "3008239.08", + "total_total_tokens_all_models": "36098869.0", + "avg_turn_count": "67.08", + "total_turn_count": "805.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "max_connections": "10", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "max_connections": "10" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-5-20251101/c3b0b891-5f15-4a87-8aac-35581f811268_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "eecb1db01fec92e7bc4575da1b68f59251953b1ec000e7c93d03dd3b4ed0af24", + "total_rows": 12 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-5-20251101/c3b0b891-5f15-4a87-8aac-35581f811268_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-5-20251101/c3b0b891-5f15-4a87-8aac-35581f811268_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..ce9ad35434a9dc74bba33c98f355cb04facd790a --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-5-20251101/c3b0b891-5f15-4a87-8aac-35581f811268_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:eecb1db01fec92e7bc4575da1b68f59251953b1ec000e7c93d03dd3b4ed0af24 +size 3674630 diff --git a/data/terminalbench/anthropic/claude-opus-4-5-20251101/c53b696a-b362-4864-a6ae-abef08e75af8.json b/data/terminalbench/anthropic/claude-opus-4-5-20251101/c53b696a-b362-4864-a6ae-abef08e75af8.json new file mode 100644 index 0000000000000000000000000000000000000000..d0508ef965019936c4c24272679c95cb5c6f4789 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-5-20251101/c53b696a-b362-4864-a6ae-abef08e75af8.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-5-20251101/1786168003.233094", + "evaluation_timestamp": "1775078458.0", + "retrieved_timestamp": "1786168003.233094", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.202,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-5-20251101", + "id": "anthropic/claude-opus-4-5-20251101", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-5-20251101/c53b696a-b362-4864-a6ae-abef08e75af8_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "348a16d4faff3450f0292688c2a8bfc147841fbf42070e877a9cc27393b6a473", + "total_rows": 40 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-5-20251101/c53b696a-b362-4864-a6ae-abef08e75af8_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-5-20251101/c53b696a-b362-4864-a6ae-abef08e75af8_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..01e23e64b94231642581a0ac78f8cbd054e2c2a7 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-5-20251101/c53b696a-b362-4864-a6ae-abef08e75af8_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:348a16d4faff3450f0292688c2a8bfc147841fbf42070e877a9cc27393b6a473 +size 11232133 diff --git a/data/terminalbench/anthropic/claude-opus-4-5-20251101/dc74ae38-3fb1-46d4-b321-404e0b1ccc71.json b/data/terminalbench/anthropic/claude-opus-4-5-20251101/dc74ae38-3fb1-46d4-b321-404e0b1ccc71.json new file mode 100644 index 0000000000000000000000000000000000000000..ee02a0c0ecfc05c78875d822ea59d7672c46bf45 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-5-20251101/dc74ae38-3fb1-46d4-b321-404e0b1ccc71.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-5-20251101/1786167899.455796", + "evaluation_timestamp": "1776601454.0", + "retrieved_timestamp": "1786167899.455796", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-5-20251101", + "id": "anthropic/claude-opus-4-5-20251101", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-5-20251101/dc74ae38-3fb1-46d4-b321-404e0b1ccc71_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "758b55a6beaadc319bee3526ae167a98657bfcd824c1f3e3a708251ec17e0c24", + "total_rows": 10 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-5-20251101/dc74ae38-3fb1-46d4-b321-404e0b1ccc71_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-5-20251101/dc74ae38-3fb1-46d4-b321-404e0b1ccc71_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..27e4a6eb153dad6d0165f687efdb2fd3c9ef514b --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-5-20251101/dc74ae38-3fb1-46d4-b321-404e0b1ccc71_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:758b55a6beaadc319bee3526ae167a98657bfcd824c1f3e3a708251ec17e0c24 +size 5805601 diff --git a/data/terminalbench/anthropic/claude-opus-4-5-20251101/e5785f3e-3ef5-4cec-abaa-c46da47ee9a4.json b/data/terminalbench/anthropic/claude-opus-4-5-20251101/e5785f3e-3ef5-4cec-abaa-c46da47ee9a4.json new file mode 100644 index 0000000000000000000000000000000000000000..e96024129dbbb7bfb4cc84f02790a990c81f941d --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-5-20251101/e5785f3e-3ef5-4cec-abaa-c46da47ee9a4.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-5-20251101/1786171074.269757", + "evaluation_timestamp": "1775078458.0", + "retrieved_timestamp": "1786171074.269757", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.202,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-5-20251101", + "id": "anthropic/claude-opus-4-5-20251101", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-5-20251101/e5785f3e-3ef5-4cec-abaa-c46da47ee9a4_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "5fe7faf663e1d4709ab59a5ab624dc7fc971351a9c0671dcaee110c724c4597b", + "total_rows": 560 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-5-20251101/e5785f3e-3ef5-4cec-abaa-c46da47ee9a4_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-5-20251101/e5785f3e-3ef5-4cec-abaa-c46da47ee9a4_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..236a20a3c448599461e344a3f1c967c914e693ae --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-5-20251101/e5785f3e-3ef5-4cec-abaa-c46da47ee9a4_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5fe7faf663e1d4709ab59a5ab624dc7fc971351a9c0671dcaee110c724c4597b +size 374819866 diff --git a/data/terminalbench/anthropic/claude-opus-4-5-20251101/eaac5b30-fdbc-4e06-a35f-0bf7a5b60d15.json b/data/terminalbench/anthropic/claude-opus-4-5-20251101/eaac5b30-fdbc-4e06-a35f-0bf7a5b60d15.json new file mode 100644 index 0000000000000000000000000000000000000000..9e938a4a01a431897a87cb1a611da85871a9c990 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-5-20251101/eaac5b30-fdbc-4e06-a35f-0bf7a5b60d15.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-5-20251101/1786167928.667907", + "evaluation_timestamp": "1776833448.0", + "retrieved_timestamp": "1786167928.667907", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-5-20251101", + "id": "anthropic/claude-opus-4-5-20251101", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-5-20251101/eaac5b30-fdbc-4e06-a35f-0bf7a5b60d15_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "fb037765513b6677d28628bf70d601310b94045d40b10a56f4d74376e07c2431", + "total_rows": 5 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-5-20251101/eaac5b30-fdbc-4e06-a35f-0bf7a5b60d15_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-5-20251101/eaac5b30-fdbc-4e06-a35f-0bf7a5b60d15_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..fb93358451eb05b45b8811b130fdf454eb47ea3c --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-5-20251101/eaac5b30-fdbc-4e06-a35f-0bf7a5b60d15_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fb037765513b6677d28628bf70d601310b94045d40b10a56f4d74376e07c2431 +size 1595139 diff --git a/data/terminalbench/anthropic/claude-opus-4-5-20251101/f21863de-3f2c-437d-abd8-829a26929ca9.json b/data/terminalbench/anthropic/claude-opus-4-5-20251101/f21863de-3f2c-437d-abd8-829a26929ca9.json new file mode 100644 index 0000000000000000000000000000000000000000..f6feba9874d6cbd8fb0c9cfea5d6d97e8db1f8b2 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-5-20251101/f21863de-3f2c-437d-abd8-829a26929ca9.json @@ -0,0 +1,124 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-5-20251101/1786163479.542049", + "evaluation_timestamp": "1776709519.0", + "retrieved_timestamp": "1786163479.542049", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-5-20251101", + "id": "anthropic/claude-opus-4-5-20251101", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive+C/+2ep/59ce501a for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 14, + "sample_ids": [ + "gcode-to-text", + "make-mips-interpreter", + "mteb-leaderboard", + "raman-fitting", + "mcmc-sampling-stan", + "regex-log", + "feal-differential-cryptanalysis", + "cancel-async-tasks", + "bn-fit-modify", + "fix-ocaml-gc", + "git-leak-recovery", + "build-cython-ext", + "pytorch-model-recovery", + "nginx-request-logging" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1776709519.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.75, + "uncertainty": { + "standard_error": { + "value": 0.10149974286764377 + }, + "num_samples": 28 + }, + "details": { + "total_matched_trajectories": "26", + "stopping_reason_count_completed_on_successful_submit": "15", + "stopping_reason_count_repetition_guard": "8", + "stopping_reason_count_token_limit": "3", + "avg_total_tokens_target_model": "2599348.69", + "total_total_tokens_target_model": "67583066.0", + "avg_total_tokens_other_models": "0.00", + "total_total_tokens_other_models": "0.0", + "avg_total_tokens_all_models": "2599348.69", + "total_total_tokens_all_models": "67583066.0", + "avg_turn_count": "52.42", + "total_turn_count": "1363.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "3600", + "attempt_timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "3600", + "max_connections": "10" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-5-20251101/f21863de-3f2c-437d-abd8-829a26929ca9_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "82a7e5a1fe798b5745c388c890ebba845f43a9bd913f9047b1e6db75d259e2a2", + "total_rows": 28 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-5-20251101/f21863de-3f2c-437d-abd8-829a26929ca9_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-5-20251101/f21863de-3f2c-437d-abd8-829a26929ca9_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..bd54b688d10dd0ba511a7d46b74716c4c9d44347 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-5-20251101/f21863de-3f2c-437d-abd8-829a26929ca9_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:82a7e5a1fe798b5745c388c890ebba845f43a9bd913f9047b1e6db75d259e2a2 +size 11200662 diff --git a/data/terminalbench/anthropic/claude-opus-4-5-20251101/f42c4678-d305-4e2a-ae74-66a2ea3a8b01.json b/data/terminalbench/anthropic/claude-opus-4-5-20251101/f42c4678-d305-4e2a-ae74-66a2ea3a8b01.json new file mode 100644 index 0000000000000000000000000000000000000000..000f4949c07a2f35fa518235127551a488d4d9f0 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-5-20251101/f42c4678-d305-4e2a-ae74-66a2ea3a8b01.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-5-20251101/1786167925.136961", + "evaluation_timestamp": "1776696546.0", + "retrieved_timestamp": "1786167925.136961", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-5-20251101", + "id": "anthropic/claude-opus-4-5-20251101", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-5-20251101/f42c4678-d305-4e2a-ae74-66a2ea3a8b01_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "acf2a11aa4ea8fc01a8c7d0ebbc60e1910674765818d373acf1f580a07c823c7", + "total_rows": 19 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-5-20251101/f42c4678-d305-4e2a-ae74-66a2ea3a8b01_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-5-20251101/f42c4678-d305-4e2a-ae74-66a2ea3a8b01_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..a92cd19078d2b7d29e7ab3ea0501e7c2d27d7ea6 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-5-20251101/f42c4678-d305-4e2a-ae74-66a2ea3a8b01_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:acf2a11aa4ea8fc01a8c7d0ebbc60e1910674765818d373acf1f580a07c823c7 +size 6574306 diff --git a/data/terminalbench/anthropic/claude-opus-4-6/01d1acbb-0adf-4e46-9949-871c30f4e54d.json b/data/terminalbench/anthropic/claude-opus-4-6/01d1acbb-0adf-4e46-9949-871c30f4e54d.json new file mode 100644 index 0000000000000000000000000000000000000000..57ade9fb0853967e637d4c158a449fc6fac1c35f --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-6/01d1acbb-0adf-4e46-9949-871c30f4e54d.json @@ -0,0 +1,113 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-6/1786166161.818902", + "evaluation_timestamp": "1778275187.0", + "retrieved_timestamp": "1786166161.818902", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-6", + "id": "anthropic/claude-opus-4-6", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive/+3ep/bf730860 for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 4, + "sample_ids": [ + "bn-fit-modify", + "fix-ocaml-gc", + "git-leak-recovery", + "build-cython-ext" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1778275187.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 1.0, + "uncertainty": { + "num_samples": 12 + }, + "details": { + "total_matched_trajectories": "12", + "stopping_reason_count_repetition_guard": "12", + "avg_total_tokens_target_model": "559660.83", + "total_total_tokens_target_model": "6715930.0", + "avg_total_tokens_other_models": "1016.00", + "total_total_tokens_other_models": "12192.0", + "avg_total_tokens_all_models": "560676.83", + "total_total_tokens_all_models": "6728122.0", + "avg_turn_count": "32.50", + "total_turn_count": "390.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-6/01d1acbb-0adf-4e46-9949-871c30f4e54d_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "6f54787b0071def4a35fe745d9d30e42272d1dd4f5b527b72dc75da704879126", + "total_rows": 12 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-6/01d1acbb-0adf-4e46-9949-871c30f4e54d_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-6/01d1acbb-0adf-4e46-9949-871c30f4e54d_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..2f4289d5cdb0c7955609d11ef2b2db942bbfa194 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-6/01d1acbb-0adf-4e46-9949-871c30f4e54d_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6f54787b0071def4a35fe745d9d30e42272d1dd4f5b527b72dc75da704879126 +size 1422994 diff --git a/data/terminalbench/anthropic/claude-opus-4-6/0999bff4-51c6-4d0e-bf18-b66c11d3f337.json b/data/terminalbench/anthropic/claude-opus-4-6/0999bff4-51c6-4d0e-bf18-b66c11d3f337.json new file mode 100644 index 0000000000000000000000000000000000000000..8a1aeb8df3ae33ad23b7ea0289306abf9da3cf8b --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-6/0999bff4-51c6-4d0e-bf18-b66c11d3f337.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-6/1786168015.763807", + "evaluation_timestamp": "1777478369.0", + "retrieved_timestamp": "1786168015.763807", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-6", + "id": "anthropic/claude-opus-4-6", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-6/0999bff4-51c6-4d0e-bf18-b66c11d3f337_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "cf8f4f2bd1755f27709049ef55b1e3688f4bb0eedbbea6d52b62c9098b9e6e2d", + "total_rows": 18 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-6/0999bff4-51c6-4d0e-bf18-b66c11d3f337_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-6/0999bff4-51c6-4d0e-bf18-b66c11d3f337_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..de969debe9d938005c79e0c285f367412b406c1c --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-6/0999bff4-51c6-4d0e-bf18-b66c11d3f337_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cf8f4f2bd1755f27709049ef55b1e3688f4bb0eedbbea6d52b62c9098b9e6e2d +size 3829197 diff --git a/data/terminalbench/anthropic/claude-opus-4-6/23433e4b-e394-4791-bebd-c33e289e9de4.json b/data/terminalbench/anthropic/claude-opus-4-6/23433e4b-e394-4791-bebd-c33e289e9de4.json new file mode 100644 index 0000000000000000000000000000000000000000..f195fa67d66de1827d9b955691aa4022faa7a0db --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-6/23433e4b-e394-4791-bebd-c33e289e9de4.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-6/1786168007.907781", + "evaluation_timestamp": "1776359446.0", + "retrieved_timestamp": "1786168007.907781", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-6", + "id": "anthropic/claude-opus-4-6", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-6/23433e4b-e394-4791-bebd-c33e289e9de4_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "cc66402a10a08a5ef5b4e9fad6f5469f5ba35b5f913db08bcc8c983404d29217", + "total_rows": 10 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-6/23433e4b-e394-4791-bebd-c33e289e9de4_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-6/23433e4b-e394-4791-bebd-c33e289e9de4_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..2942e7816b4bd79de7c99176529443eb38d04a51 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-6/23433e4b-e394-4791-bebd-c33e289e9de4_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cc66402a10a08a5ef5b4e9fad6f5469f5ba35b5f913db08bcc8c983404d29217 +size 3735052 diff --git a/data/terminalbench/anthropic/claude-opus-4-6/25de5aa0-6c17-4832-867c-529a6177f139.json b/data/terminalbench/anthropic/claude-opus-4-6/25de5aa0-6c17-4832-867c-529a6177f139.json new file mode 100644 index 0000000000000000000000000000000000000000..dc37af6f15a8c4e5b4ff8d4588088e0c956e0075 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-6/25de5aa0-6c17-4832-867c-529a6177f139.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-6/1786168032.766573", + "evaluation_timestamp": "1777313093.0", + "retrieved_timestamp": "1786168032.766573", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-6", + "id": "anthropic/claude-opus-4-6", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-6/25de5aa0-6c17-4832-867c-529a6177f139_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "46caae126bedc3aad9c2969f84228d0c4d4fe2bfa78d9dc0734a065371107e21", + "total_rows": 56 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-6/25de5aa0-6c17-4832-867c-529a6177f139_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-6/25de5aa0-6c17-4832-867c-529a6177f139_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..878e65ad35a11532ad0e1a947bcde4e43b35176a --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-6/25de5aa0-6c17-4832-867c-529a6177f139_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:46caae126bedc3aad9c2969f84228d0c4d4fe2bfa78d9dc0734a065371107e21 +size 26437459 diff --git a/data/terminalbench/anthropic/claude-opus-4-6/27a8e582-639c-4934-9ae6-4e5f94b8d8e1.json b/data/terminalbench/anthropic/claude-opus-4-6/27a8e582-639c-4934-9ae6-4e5f94b8d8e1.json new file mode 100644 index 0000000000000000000000000000000000000000..e50dcefa847d790814739e3530bbaf0b0217c676 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-6/27a8e582-639c-4934-9ae6-4e5f94b8d8e1.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-6/1786171170.135503", + "evaluation_timestamp": "1778241890.0", + "retrieved_timestamp": "1786171170.135503", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-6", + "id": "anthropic/claude-opus-4-6", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-6/27a8e582-639c-4934-9ae6-4e5f94b8d8e1_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "bc283541cc68ed8ae878152f1a6aaad65992f69d7282628e08b2d5a2e949aa7d", + "total_rows": 7 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-6/27a8e582-639c-4934-9ae6-4e5f94b8d8e1_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-6/27a8e582-639c-4934-9ae6-4e5f94b8d8e1_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..13531c793b6cc4e58886061b7794657f3ba302f8 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-6/27a8e582-639c-4934-9ae6-4e5f94b8d8e1_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bc283541cc68ed8ae878152f1a6aaad65992f69d7282628e08b2d5a2e949aa7d +size 4016937 diff --git a/data/terminalbench/anthropic/claude-opus-4-6/2a8ab513-134e-4752-99af-11b0b16ba182.json b/data/terminalbench/anthropic/claude-opus-4-6/2a8ab513-134e-4752-99af-11b0b16ba182.json new file mode 100644 index 0000000000000000000000000000000000000000..5658787b98b526a6a5ede35d6877dafba31a007c --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-6/2a8ab513-134e-4752-99af-11b0b16ba182.json @@ -0,0 +1,117 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-6/1786163554.897374", + "evaluation_timestamp": "1777925324.0", + "retrieved_timestamp": "1786163554.897374", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-6", + "id": "anthropic/claude-opus-4-6", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive+C/+2ep/de7659c6 for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 3, + "sample_ids": [ + "adaptive-rejection-sampler", + "db-wal-recovery", + "video-processing" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1777925324.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.3333333333333333, + "uncertainty": { + "standard_error": { + "value": 0.16666666666666669 + }, + "num_samples": 6 + }, + "details": { + "total_matched_trajectories": "6", + "stopping_reason_count_repetition_guard": "4", + "stopping_reason_count_completed_on_successful_submit": "1", + "stopping_reason_count_token_limit": "1", + "avg_total_tokens_target_model": "6304571.67", + "total_total_tokens_target_model": "37827430.0", + "avg_total_tokens_other_models": "0.00", + "total_total_tokens_other_models": "0.0", + "avg_total_tokens_all_models": "6304571.67", + "total_total_tokens_all_models": "37827430.0", + "avg_turn_count": "103.33", + "total_turn_count": "620.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-6/2a8ab513-134e-4752-99af-11b0b16ba182_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "eea33fe41e7393a113872796f315dfa3d67d57b81336779c0b56bdfe9ada9f17", + "total_rows": 6 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-6/2a8ab513-134e-4752-99af-11b0b16ba182_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-6/2a8ab513-134e-4752-99af-11b0b16ba182_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..bd6b5f3e9bc5454b1d8d8e02e7f897b497df5644 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-6/2a8ab513-134e-4752-99af-11b0b16ba182_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:eea33fe41e7393a113872796f315dfa3d67d57b81336779c0b56bdfe9ada9f17 +size 5800207 diff --git a/data/terminalbench/anthropic/claude-opus-4-6/2ab6f6d0-fc4b-4c1c-b662-740203898b4e.json b/data/terminalbench/anthropic/claude-opus-4-6/2ab6f6d0-fc4b-4c1c-b662-740203898b4e.json new file mode 100644 index 0000000000000000000000000000000000000000..5f2b159dc2f682285ab572cbeaca970c1c9f8f33 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-6/2ab6f6d0-fc4b-4c1c-b662-740203898b4e.json @@ -0,0 +1,118 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-6/1786163564.451743", + "evaluation_timestamp": "1776453766.0", + "retrieved_timestamp": "1786163564.451743", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-6", + "id": "anthropic/claude-opus-4-6", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive+C/+4ep/15ab3444 for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 10, + "sample_ids": [ + "gpt2-codegolf", + "llm-inference-batching-scheduler", + "break-filter-js-from-html", + "reshard-c4-data", + "write-compressor", + "merge-diff-arc-agi-task", + "winning-avg-corewars", + "log-summary-date-ranges", + "pytorch-model-cli", + "largest-eigenval" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1776453766.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.9, + "uncertainty": { + "standard_error": { + "value": 0.07637626158259733 + }, + "num_samples": 40 + }, + "details": { + "total_matched_trajectories": "40", + "stopping_reason_count_completed_on_successful_submit": "34", + "stopping_reason_count_repetition_guard": "5", + "stopping_reason_count_token_limit": "1", + "avg_total_tokens_target_model": "1877752.07", + "total_total_tokens_target_model": "75110083.0", + "avg_total_tokens_other_models": "0.00", + "total_total_tokens_other_models": "0.0", + "avg_total_tokens_all_models": "1877752.07", + "total_total_tokens_all_models": "75110083.0", + "avg_turn_count": "43.12", + "total_turn_count": "1725.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "max_connections": "10", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "max_connections": "10" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-6/2ab6f6d0-fc4b-4c1c-b662-740203898b4e_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "2764590641ec13a14298c0758e58ca8c85d5ac13554a228c3703c8b35309a898", + "total_rows": 40 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-6/2ab6f6d0-fc4b-4c1c-b662-740203898b4e_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-6/2ab6f6d0-fc4b-4c1c-b662-740203898b4e_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..7ef53c06140db9a25060eebefeab998db8e7d315 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-6/2ab6f6d0-fc4b-4c1c-b662-740203898b4e_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2764590641ec13a14298c0758e58ca8c85d5ac13554a228c3703c8b35309a898 +size 16645466 diff --git a/data/terminalbench/anthropic/claude-opus-4-6/2c78f53f-9a22-46fd-84a6-79eff6bb6d3c.json b/data/terminalbench/anthropic/claude-opus-4-6/2c78f53f-9a22-46fd-84a6-79eff6bb6d3c.json new file mode 100644 index 0000000000000000000000000000000000000000..42c67bf91eecbe78641cd77fd104cda5ef85b203 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-6/2c78f53f-9a22-46fd-84a6-79eff6bb6d3c.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-6/1786171235.702277", + "evaluation_timestamp": "1778055257.0", + "retrieved_timestamp": "1786171235.702277", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-6", + "id": "anthropic/claude-opus-4-6", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-6/2c78f53f-9a22-46fd-84a6-79eff6bb6d3c_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "2aaa754b3d435cc8f688d3499cc4d51e81955fe6011172a16731084fcd3a2aac", + "total_rows": 73 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-6/2c78f53f-9a22-46fd-84a6-79eff6bb6d3c_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-6/2c78f53f-9a22-46fd-84a6-79eff6bb6d3c_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..eb025c15c25138426fd53f81b907aa5b2a96e952 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-6/2c78f53f-9a22-46fd-84a6-79eff6bb6d3c_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2aaa754b3d435cc8f688d3499cc4d51e81955fe6011172a16731084fcd3a2aac +size 50999992 diff --git a/data/terminalbench/anthropic/claude-opus-4-6/3116c233-897e-495e-99ac-0e4b0fb333f5.json b/data/terminalbench/anthropic/claude-opus-4-6/3116c233-897e-495e-99ac-0e4b0fb333f5.json new file mode 100644 index 0000000000000000000000000000000000000000..4c307b330b9648a1c6a1905614334ea6328c4121 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-6/3116c233-897e-495e-99ac-0e4b0fb333f5.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-6/1786171147.398309", + "evaluation_timestamp": "1776359446.0", + "retrieved_timestamp": "1786171147.398309", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-6", + "id": "anthropic/claude-opus-4-6", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-6/3116c233-897e-495e-99ac-0e4b0fb333f5_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "1a177fb830aedb4b0feb2d8352b482a7dbd209ff8874c01cb647d3417c403829", + "total_rows": 40 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-6/3116c233-897e-495e-99ac-0e4b0fb333f5_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-6/3116c233-897e-495e-99ac-0e4b0fb333f5_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..d4de33d39887baa8e583a7bb7e5ac9f1c58cc255 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-6/3116c233-897e-495e-99ac-0e4b0fb333f5_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1a177fb830aedb4b0feb2d8352b482a7dbd209ff8874c01cb647d3417c403829 +size 30259018 diff --git a/data/terminalbench/anthropic/claude-opus-4-6/3e3cf1cf-c29c-4193-b024-da08a7b73c4a.json b/data/terminalbench/anthropic/claude-opus-4-6/3e3cf1cf-c29c-4193-b024-da08a7b73c4a.json new file mode 100644 index 0000000000000000000000000000000000000000..a59f0f01d4adc3263718f082f32d758fae2adc7e --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-6/3e3cf1cf-c29c-4193-b024-da08a7b73c4a.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-6/1786171185.968336", + "evaluation_timestamp": "1776520455.0", + "retrieved_timestamp": "1786171185.968336", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-6", + "id": "anthropic/claude-opus-4-6", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-6/3e3cf1cf-c29c-4193-b024-da08a7b73c4a_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "8c9893d01839ce5aca55e0d0c5fb984c321cb6d4029ca75020cf4c6f6fe4de9b", + "total_rows": 30 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-6/3e3cf1cf-c29c-4193-b024-da08a7b73c4a_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-6/3e3cf1cf-c29c-4193-b024-da08a7b73c4a_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..7ee87659753ccf1011d758a216f9336084193720 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-6/3e3cf1cf-c29c-4193-b024-da08a7b73c4a_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8c9893d01839ce5aca55e0d0c5fb984c321cb6d4029ca75020cf4c6f6fe4de9b +size 24824819 diff --git a/data/terminalbench/anthropic/claude-opus-4-6/445a9930-d53a-40de-9cbc-d6a3c8a4724e.json b/data/terminalbench/anthropic/claude-opus-4-6/445a9930-d53a-40de-9cbc-d6a3c8a4724e.json new file mode 100644 index 0000000000000000000000000000000000000000..217b8507e07b1521a3fa6646809dc511e65d6522 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-6/445a9930-d53a-40de-9cbc-d6a3c8a4724e.json @@ -0,0 +1,106 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-6/1786166159.326284", + "evaluation_timestamp": "1776656451.0", + "retrieved_timestamp": "1786166159.326284", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-6", + "id": "anthropic/claude-opus-4-6", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive/+3ep/9e593f48 for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 1, + "sample_ids": [ + "kv-store-grpc" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1776656451.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.6666666666666666, + "uncertainty": { + "num_samples": 3 + }, + "details": { + "total_matched_trajectories": "3", + "stopping_reason_count_repetition_guard": "3", + "avg_total_tokens_target_model": "200665.00", + "total_total_tokens_target_model": "601995.0", + "avg_total_tokens_other_models": "0.00", + "total_total_tokens_other_models": "0.0", + "avg_total_tokens_all_models": "200665.00", + "total_total_tokens_all_models": "601995.0", + "avg_turn_count": "26.33", + "total_turn_count": "79.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "3600", + "attempt_timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "3600", + "max_connections": "10" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-6/445a9930-d53a-40de-9cbc-d6a3c8a4724e_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "e6b4282d2073e0f94a884b1928de9f3a878741751134b5dfbab6088f1a5093d7", + "total_rows": 3 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-6/445a9930-d53a-40de-9cbc-d6a3c8a4724e_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-6/445a9930-d53a-40de-9cbc-d6a3c8a4724e_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..475be35639fc534bdd9d620bc9c6298ec0cf56c1 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-6/445a9930-d53a-40de-9cbc-d6a3c8a4724e_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e6b4282d2073e0f94a884b1928de9f3a878741751134b5dfbab6088f1a5093d7 +size 1723170 diff --git a/data/terminalbench/anthropic/claude-opus-4-6/47ea1867-b22b-4929-9005-6bdbff394aa3.json b/data/terminalbench/anthropic/claude-opus-4-6/47ea1867-b22b-4929-9005-6bdbff394aa3.json new file mode 100644 index 0000000000000000000000000000000000000000..e2827f8b3a49ca36304a2acfc00fbc5e569ec68f --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-6/47ea1867-b22b-4929-9005-6bdbff394aa3.json @@ -0,0 +1,118 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-6/1786166121.106133", + "evaluation_timestamp": "1778229854.0", + "retrieved_timestamp": "1786166121.106133", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-6", + "id": "anthropic/claude-opus-4-6", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive/+1ep/6a5d425e for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 5, + "sample_ids": [ + "regex-log", + "sqlite-db-truncate", + "sanitize-git-repo", + "build-pmars", + "rstan-to-pystan" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1778229854.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.6, + "uncertainty": { + "standard_error": { + "value": 0.24494897427831783 + }, + "num_samples": 5 + }, + "details": { + "total_matched_trajectories": "5", + "stopping_reason_count_repetition_guard": "4", + "stopping_reason_count_token_limit": "1", + "avg_total_tokens_target_model": "4994795.80", + "total_total_tokens_target_model": "24973979.0", + "avg_total_tokens_other_models": "657.80", + "total_total_tokens_other_models": "3289.0", + "avg_total_tokens_all_models": "4995453.60", + "total_total_tokens_all_models": "24977268.0", + "avg_turn_count": "95.00", + "total_turn_count": "475.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-6/47ea1867-b22b-4929-9005-6bdbff394aa3_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "2899972b585d308cb24e0ecb55bfb2aaeff4c9380eb6996f28a00c95ac81496e", + "total_rows": 5 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-6/47ea1867-b22b-4929-9005-6bdbff394aa3_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-6/47ea1867-b22b-4929-9005-6bdbff394aa3_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..4f1e88b3929402ef882d12073be047244c0082a5 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-6/47ea1867-b22b-4929-9005-6bdbff394aa3_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2899972b585d308cb24e0ecb55bfb2aaeff4c9380eb6996f28a00c95ac81496e +size 3486435 diff --git a/data/terminalbench/anthropic/claude-opus-4-6/491b28b9-bada-48b5-a82c-07ae163a9515.json b/data/terminalbench/anthropic/claude-opus-4-6/491b28b9-bada-48b5-a82c-07ae163a9515.json new file mode 100644 index 0000000000000000000000000000000000000000..fca01558df5e068b5cd64cd0c15860291162c5ef --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-6/491b28b9-bada-48b5-a82c-07ae163a9515.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-6/1786171129.433277", + "evaluation_timestamp": "1776680886.0", + "retrieved_timestamp": "1786171129.433277", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-6", + "id": "anthropic/claude-opus-4-6", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-6/491b28b9-bada-48b5-a82c-07ae163a9515_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "f83b40bb3d1ac6d1318b7867463756534590d3ced114382a44efeae4c56e2976", + "total_rows": 2 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-6/491b28b9-bada-48b5-a82c-07ae163a9515_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-6/491b28b9-bada-48b5-a82c-07ae163a9515_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..3678df70a0f3f576c45dbb6b5897b6b5365d0033 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-6/491b28b9-bada-48b5-a82c-07ae163a9515_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f83b40bb3d1ac6d1318b7867463756534590d3ced114382a44efeae4c56e2976 +size 665016 diff --git a/data/terminalbench/anthropic/claude-opus-4-6/5a946188-fdd6-45c2-a1bf-2c1aff177c71.json b/data/terminalbench/anthropic/claude-opus-4-6/5a946188-fdd6-45c2-a1bf-2c1aff177c71.json new file mode 100644 index 0000000000000000000000000000000000000000..1ec32aaee440d8799143a0dcd8fedd21d71d717c --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-6/5a946188-fdd6-45c2-a1bf-2c1aff177c71.json @@ -0,0 +1,114 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-6/1786166196.235479", + "evaluation_timestamp": "1777156177.0", + "retrieved_timestamp": "1786166196.235479", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-6", + "id": "anthropic/claude-opus-4-6", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive/+4ep/9fb02f41 for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 4, + "sample_ids": [ + "extract-elf", + "fix-code-vulnerability", + "financial-document-processor", + "cancel-async-tasks" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1777156177.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.875, + "uncertainty": { + "standard_error": { + "value": 0.07216878364870322 + }, + "num_samples": 16 + }, + "details": { + "total_matched_trajectories": "16", + "stopping_reason_count_token_limit": "11", + "stopping_reason_count_repetition_guard": "5", + "avg_total_tokens_target_model": "7408456.88", + "total_total_tokens_target_model": "118535310.0", + "avg_total_tokens_other_models": "157.50", + "total_total_tokens_other_models": "2520.0", + "avg_total_tokens_all_models": "7408614.38", + "total_total_tokens_all_models": "118537830.0", + "avg_turn_count": "129.62", + "total_turn_count": "2074.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-6/5a946188-fdd6-45c2-a1bf-2c1aff177c71_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "62a471aa38df53d23fd63f15072cedddc3106eb5c7eab1ff4e9f45125af3cb1e", + "total_rows": 16 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-6/5a946188-fdd6-45c2-a1bf-2c1aff177c71_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-6/5a946188-fdd6-45c2-a1bf-2c1aff177c71_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..32d8f177f2e7e171939a377901d38983b96c4e12 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-6/5a946188-fdd6-45c2-a1bf-2c1aff177c71_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:62a471aa38df53d23fd63f15072cedddc3106eb5c7eab1ff4e9f45125af3cb1e +size 13547778 diff --git a/data/terminalbench/anthropic/claude-opus-4-6/6306602b-f609-4a2c-bccc-adda2f5e9beb.json b/data/terminalbench/anthropic/claude-opus-4-6/6306602b-f609-4a2c-bccc-adda2f5e9beb.json new file mode 100644 index 0000000000000000000000000000000000000000..c7bb9e16c6dea2d6de25d4463168fecff2cb6c7c --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-6/6306602b-f609-4a2c-bccc-adda2f5e9beb.json @@ -0,0 +1,110 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-6/1786163551.361569", + "evaluation_timestamp": "1778551875.0", + "retrieved_timestamp": "1786163551.361569", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-6", + "id": "anthropic/claude-opus-4-6", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive+C/+2ep/27966f54 for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 1, + "sample_ids": [ + "code-from-image" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1778551875.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 1.0, + "uncertainty": { + "num_samples": 2 + }, + "details": { + "total_matched_trajectories": "2", + "stopping_reason_count_completed_on_successful_submit": "2", + "avg_total_tokens_target_model": "81085.50", + "total_total_tokens_target_model": "162171.0", + "avg_total_tokens_other_models": "0.00", + "total_total_tokens_other_models": "0.0", + "avg_total_tokens_all_models": "81085.50", + "total_total_tokens_all_models": "162171.0", + "avg_turn_count": "19.00", + "total_turn_count": "38.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-6/6306602b-f609-4a2c-bccc-adda2f5e9beb_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "0deded7fa0d33140a6fe2991eaacbb2fbfbbe885a13aaf4942dda9fd7dbd37ad", + "total_rows": 2 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-6/6306602b-f609-4a2c-bccc-adda2f5e9beb_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-6/6306602b-f609-4a2c-bccc-adda2f5e9beb_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..aa0507cf87e33d978c84b8d5e50cf2ed7c72a191 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-6/6306602b-f609-4a2c-bccc-adda2f5e9beb_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0deded7fa0d33140a6fe2991eaacbb2fbfbbe885a13aaf4942dda9fd7dbd37ad +size 100836 diff --git a/data/terminalbench/anthropic/claude-opus-4-6/6f32cee8-22bd-4339-8cab-a0fd479b2580.json b/data/terminalbench/anthropic/claude-opus-4-6/6f32cee8-22bd-4339-8cab-a0fd479b2580.json new file mode 100644 index 0000000000000000000000000000000000000000..775e46e715ee4f8b7639a596c4c544bc6ebe554a --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-6/6f32cee8-22bd-4339-8cab-a0fd479b2580.json @@ -0,0 +1,121 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-6/1786163558.160091", + "evaluation_timestamp": "1778552100.0", + "retrieved_timestamp": "1786163558.160091", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-6", + "id": "anthropic/claude-opus-4-6", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive+C/+3ep/86d6397e for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 8, + "sample_ids": [ + "regex-log", + "polyglot-c-py", + "cancel-async-tasks", + "bn-fit-modify", + "fix-ocaml-gc", + "git-leak-recovery", + "build-cython-ext", + "pytorch-model-recovery" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1778552100.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.7916666666666666, + "uncertainty": { + "standard_error": { + "value": 0.1399759049560101 + }, + "num_samples": 24 + }, + "details": { + "total_matched_trajectories": "24", + "stopping_reason_count_completed_on_successful_submit": "19", + "stopping_reason_count_repetition_guard": "5", + "avg_total_tokens_target_model": "678407.38", + "total_total_tokens_target_model": "16281777.0", + "avg_total_tokens_other_models": "211.12", + "total_total_tokens_other_models": "5067.0", + "avg_total_tokens_all_models": "678618.50", + "total_total_tokens_all_models": "16286844.0", + "avg_turn_count": "24.83", + "total_turn_count": "596.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-6/6f32cee8-22bd-4339-8cab-a0fd479b2580_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "8e0c430c9aa9764e1f053f63b75a1e7c47c8f3b846d787abccc9a0ac9409b565", + "total_rows": 24 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-6/6f32cee8-22bd-4339-8cab-a0fd479b2580_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-6/6f32cee8-22bd-4339-8cab-a0fd479b2580_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..568bb5b5219946e9700ccf5df3c81287a5e27a1c --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-6/6f32cee8-22bd-4339-8cab-a0fd479b2580_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8e0c430c9aa9764e1f053f63b75a1e7c47c8f3b846d787abccc9a0ac9409b565 +size 4773171 diff --git a/data/terminalbench/anthropic/claude-opus-4-6/7e994e9b-c86c-4722-8d01-01185a15c739.json b/data/terminalbench/anthropic/claude-opus-4-6/7e994e9b-c86c-4722-8d01-01185a15c739.json new file mode 100644 index 0000000000000000000000000000000000000000..b93a23318ed7e9cc0670fcb14decf139293ab34d --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-6/7e994e9b-c86c-4722-8d01-01185a15c739.json @@ -0,0 +1,120 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-6/1786166126.315773", + "evaluation_timestamp": "1778265533.0", + "retrieved_timestamp": "1786166126.315773", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-6", + "id": "anthropic/claude-opus-4-6", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive/+1ep/644f135a for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 7, + "sample_ids": [ + "raman-fitting", + "protein-assembly", + "polyglot-c-py", + "pytorch-model-recovery", + "model-extraction-relu-logits", + "nginx-request-logging", + "sparql-university" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1778265533.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.2857142857142857, + "uncertainty": { + "standard_error": { + "value": 0.18442777839082938 + }, + "num_samples": 7 + }, + "details": { + "total_matched_trajectories": "7", + "stopping_reason_count_repetition_guard": "6", + "stopping_reason_count_completed_no_submit": "1", + "avg_total_tokens_target_model": "601669.29", + "total_total_tokens_target_model": "4211685.0", + "avg_total_tokens_other_models": "945.86", + "total_total_tokens_other_models": "6621.0", + "avg_total_tokens_all_models": "602615.14", + "total_total_tokens_all_models": "4218306.0", + "avg_turn_count": "25.29", + "total_turn_count": "177.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-6/7e994e9b-c86c-4722-8d01-01185a15c739_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "aaaf44ca2e23070e45bd1149a0bf5f4cca65d6c193837ecf89303ec0fb07f3cf", + "total_rows": 7 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-6/7e994e9b-c86c-4722-8d01-01185a15c739_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-6/7e994e9b-c86c-4722-8d01-01185a15c739_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..cf6868f30f9c65a6a2a72be48729a911fa04a09c --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-6/7e994e9b-c86c-4722-8d01-01185a15c739_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:aaaf44ca2e23070e45bd1149a0bf5f4cca65d6c193837ecf89303ec0fb07f3cf +size 916443 diff --git a/data/terminalbench/anthropic/claude-opus-4-6/8b80a10b-5bf6-40f9-a3e0-2e700eca99b1.json b/data/terminalbench/anthropic/claude-opus-4-6/8b80a10b-5bf6-40f9-a3e0-2e700eca99b1.json new file mode 100644 index 0000000000000000000000000000000000000000..f51521989dc9431b7f8db50542c22d1bcd13828b --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-6/8b80a10b-5bf6-40f9-a3e0-2e700eca99b1.json @@ -0,0 +1,110 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-6/1786166112.588193", + "evaluation_timestamp": "1776880341.0", + "retrieved_timestamp": "1786166112.588193", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-6", + "id": "anthropic/claude-opus-4-6", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive/+1ep/b99cff2b for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 3, + "sample_ids": [ + "schemelike-metacircular-eval", + "qemu-alpine-ssh", + "extract-moves-from-video" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1776880341.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 1.0, + "uncertainty": { + "num_samples": 2 + }, + "details": { + "total_matched_trajectories": "2", + "stopping_reason_count_token_limit": "1", + "stopping_reason_count_repetition_guard": "1", + "avg_total_tokens_target_model": "7896175.50", + "total_total_tokens_target_model": "15792351.0", + "avg_total_tokens_other_models": "0.00", + "total_total_tokens_other_models": "0.0", + "avg_total_tokens_all_models": "7896175.50", + "total_total_tokens_all_models": "15792351.0", + "avg_turn_count": "153.00", + "total_turn_count": "306.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-6/8b80a10b-5bf6-40f9-a3e0-2e700eca99b1_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "bcf947964b17a2fcb2c058a1b28c5d8e3356cfe12a5ddb297cd8fb4f81a5e71c", + "total_rows": 2 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-6/8b80a10b-5bf6-40f9-a3e0-2e700eca99b1_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-6/8b80a10b-5bf6-40f9-a3e0-2e700eca99b1_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..c181a34b444aa2034d673bbf4a9a9d0b6affe347 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-6/8b80a10b-5bf6-40f9-a3e0-2e700eca99b1_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bcf947964b17a2fcb2c058a1b28c5d8e3356cfe12a5ddb297cd8fb4f81a5e71c +size 2125080 diff --git a/data/terminalbench/anthropic/claude-opus-4-6/8bba1b56-bf71-4eb0-9b81-f310266d6933.json b/data/terminalbench/anthropic/claude-opus-4-6/8bba1b56-bf71-4eb0-9b81-f310266d6933.json new file mode 100644 index 0000000000000000000000000000000000000000..452cb93349a4c6713c91f4af6e6cd2b80dd3a057 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-6/8bba1b56-bf71-4eb0-9b81-f310266d6933.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-6/1786168011.406239", + "evaluation_timestamp": "1777542076.0", + "retrieved_timestamp": "1786168011.406239", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-6", + "id": "anthropic/claude-opus-4-6", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-6/8bba1b56-bf71-4eb0-9b81-f310266d6933_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "9967296a6620d547464f8f5c3c50b26960a8fc25ef6836a1987d4547ef23464f", + "total_rows": 1 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-6/8bba1b56-bf71-4eb0-9b81-f310266d6933_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-6/8bba1b56-bf71-4eb0-9b81-f310266d6933_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..7f501c5da1ca28bee7a10a2884481b5b9814f6d9 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-6/8bba1b56-bf71-4eb0-9b81-f310266d6933_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9967296a6620d547464f8f5c3c50b26960a8fc25ef6836a1987d4547ef23464f +size 30272 diff --git a/data/terminalbench/anthropic/claude-opus-4-6/8e822ce1-c26b-4db5-a509-97dacdc0db31.json b/data/terminalbench/anthropic/claude-opus-4-6/8e822ce1-c26b-4db5-a509-97dacdc0db31.json new file mode 100644 index 0000000000000000000000000000000000000000..6ef700270a93163100eac838632dfb5123d72409 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-6/8e822ce1-c26b-4db5-a509-97dacdc0db31.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-6/1786171156.908764", + "evaluation_timestamp": "1776747747.0", + "retrieved_timestamp": "1786171156.908764", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-6", + "id": "anthropic/claude-opus-4-6", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-6/8e822ce1-c26b-4db5-a509-97dacdc0db31_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "6a5b8fa19798d1867fbbaf2429634ba60d37ab6b0ab2f2f6394d4c93c6a43ce7", + "total_rows": 4 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-6/8e822ce1-c26b-4db5-a509-97dacdc0db31_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-6/8e822ce1-c26b-4db5-a509-97dacdc0db31_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..46cc715fe607a4cd85b57b01f237c120ee91ab84 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-6/8e822ce1-c26b-4db5-a509-97dacdc0db31_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6a5b8fa19798d1867fbbaf2429634ba60d37ab6b0ab2f2f6394d4c93c6a43ce7 +size 3595867 diff --git a/data/terminalbench/anthropic/claude-opus-4-6/8f8171ee-8187-426d-ba4c-006f3b9d4595.json b/data/terminalbench/anthropic/claude-opus-4-6/8f8171ee-8187-426d-ba4c-006f3b9d4595.json new file mode 100644 index 0000000000000000000000000000000000000000..9dbb97cffc248cf789aa4c9ec18e68f6e3fd9680 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-6/8f8171ee-8187-426d-ba4c-006f3b9d4595.json @@ -0,0 +1,110 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-6/1786163536.957196", + "evaluation_timestamp": "1778238200.0", + "retrieved_timestamp": "1786163536.957196", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-6", + "id": "anthropic/claude-opus-4-6", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive+C/+1ep/aee684fa for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 1, + "sample_ids": [ + "mteb-leaderboard" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1778238200.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.0, + "uncertainty": { + "num_samples": 1 + }, + "details": { + "total_matched_trajectories": "1", + "stopping_reason_count_token_limit": "1", + "avg_total_tokens_target_model": "10060573.00", + "total_total_tokens_target_model": "10060573.0", + "avg_total_tokens_other_models": "0.00", + "total_total_tokens_other_models": "0.0", + "avg_total_tokens_all_models": "10060573.00", + "total_total_tokens_all_models": "10060573.0", + "avg_turn_count": "117.00", + "total_turn_count": "117.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-6/8f8171ee-8187-426d-ba4c-006f3b9d4595_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "fa3aa0e664d6635ab21cad643dab18f4694e6cfee6c1526ef1a6b681a56f2638", + "total_rows": 1 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-6/8f8171ee-8187-426d-ba4c-006f3b9d4595_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-6/8f8171ee-8187-426d-ba4c-006f3b9d4595_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..076c0f9e56719cbbfba56499693ea2ff690fbd6d --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-6/8f8171ee-8187-426d-ba4c-006f3b9d4595_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fa3aa0e664d6635ab21cad643dab18f4694e6cfee6c1526ef1a6b681a56f2638 +size 935189 diff --git a/data/terminalbench/anthropic/claude-opus-4-6/90e339a0-a193-4482-9094-dc87a33668c0.json b/data/terminalbench/anthropic/claude-opus-4-6/90e339a0-a193-4482-9094-dc87a33668c0.json new file mode 100644 index 0000000000000000000000000000000000000000..8ddad01cef1559f7741f9b029322a61175a47596 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-6/90e339a0-a193-4482-9094-dc87a33668c0.json @@ -0,0 +1,132 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-6/1786163541.200124", + "evaluation_timestamp": "1778055257.0", + "retrieved_timestamp": "1786163541.200124", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-6", + "id": "anthropic/claude-opus-4-6", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive+C/+1ep/e063ed2e for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 16, + "sample_ids": [ + "overfull-hbox", + "train-fasttext", + "install-windows-3.11", + "make-doom-for-mips", + "extract-moves-from-video", + "make-mips-interpreter", + "count-dataset-tokens", + "circuit-fibsqrt", + "mteb-leaderboard", + "mailman", + "raman-fitting", + "extract-elf", + "protein-assembly", + "fix-code-vulnerability", + "financial-document-processor", + "feal-differential-cryptanalysis" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1778055257.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.4666666666666667, + "uncertainty": { + "standard_error": { + "value": 0.1333333333333333 + }, + "num_samples": 15 + }, + "details": { + "total_matched_trajectories": "15", + "stopping_reason_count_completed_on_successful_submit": "5", + "stopping_reason_count_repetition_guard": "4", + "stopping_reason_count_token_limit": "4", + "stopping_reason_count_completed_with_submit": "1", + "stopping_reason_count_completed_no_submit": "1", + "avg_total_tokens_target_model": "4250319.13", + "total_total_tokens_target_model": "63754787.0", + "avg_total_tokens_other_models": "0.00", + "total_total_tokens_other_models": "0.0", + "avg_total_tokens_all_models": "4250319.13", + "total_total_tokens_all_models": "63754787.0", + "avg_turn_count": "76.27", + "total_turn_count": "1144.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-6/90e339a0-a193-4482-9094-dc87a33668c0_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "6ffb55410aca99527e90babfffed377fc9b02a7f8f8ffe122e36087ba2b294a4", + "total_rows": 15 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-6/90e339a0-a193-4482-9094-dc87a33668c0_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-6/90e339a0-a193-4482-9094-dc87a33668c0_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..d2692ac966834bb03edf0268b14c695d81d95f90 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-6/90e339a0-a193-4482-9094-dc87a33668c0_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6ffb55410aca99527e90babfffed377fc9b02a7f8f8ffe122e36087ba2b294a4 +size 7918169 diff --git a/data/terminalbench/anthropic/claude-opus-4-6/965981b0-c781-49cb-98a9-d02241417e50.json b/data/terminalbench/anthropic/claude-opus-4-6/965981b0-c781-49cb-98a9-d02241417e50.json new file mode 100644 index 0000000000000000000000000000000000000000..274871e9609c102f53372417c829e1c570649dbb --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-6/965981b0-c781-49cb-98a9-d02241417e50.json @@ -0,0 +1,122 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-6/1786166187.931665", + "evaluation_timestamp": "1776696546.0", + "retrieved_timestamp": "1786166187.931665", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-6", + "id": "anthropic/claude-opus-4-6", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive/+4ep/86ad83bf for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 13, + "sample_ids": [ + "install-windows-3.11", + "torch-pipeline-parallelism", + "tune-mjcf", + "extract-moves-from-video", + "gcode-to-text", + "make-mips-interpreter", + "count-dataset-tokens", + "circuit-fibsqrt", + "mteb-leaderboard", + "mailman", + "query-optimize", + "code-from-image", + "feal-differential-cryptanalysis" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1776696546.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.5208333333333334, + "uncertainty": { + "standard_error": { + "value": 0.12857881873101273 + }, + "num_samples": 47 + }, + "details": { + "total_matched_trajectories": "47", + "stopping_reason_count_repetition_guard": "14", + "stopping_reason_count_token_limit": "33", + "avg_total_tokens_target_model": "8728698.70", + "total_total_tokens_target_model": "410248839.0", + "avg_total_tokens_other_models": "116.68", + "total_total_tokens_other_models": "5484.0", + "avg_total_tokens_all_models": "8728815.38", + "total_total_tokens_all_models": "410254323.0", + "avg_turn_count": "144.04", + "total_turn_count": "6770.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "3600", + "attempt_timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "3600", + "max_connections": "10" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-6/965981b0-c781-49cb-98a9-d02241417e50_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "849a11286cca4bee939a57ecd77fb631da7ea969766a5b6d2ed3bc919539bc84", + "total_rows": 47 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-6/965981b0-c781-49cb-98a9-d02241417e50_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-6/965981b0-c781-49cb-98a9-d02241417e50_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..3df9aa7d702551d80b2933ac2b5b067182a388bd --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-6/965981b0-c781-49cb-98a9-d02241417e50_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:849a11286cca4bee939a57ecd77fb631da7ea969766a5b6d2ed3bc919539bc84 +size 44065748 diff --git a/data/terminalbench/anthropic/claude-opus-4-6/96769cb9-54bb-4f8a-8556-a4515af86232.json b/data/terminalbench/anthropic/claude-opus-4-6/96769cb9-54bb-4f8a-8556-a4515af86232.json new file mode 100644 index 0000000000000000000000000000000000000000..9faf7d597ec9d1cf7c4ec464bdb33ee7ac16cce6 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-6/96769cb9-54bb-4f8a-8556-a4515af86232.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-6/1786171124.361812", + "evaluation_timestamp": "1776661808.0", + "retrieved_timestamp": "1786171124.361812", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-6", + "id": "anthropic/claude-opus-4-6", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-6/96769cb9-54bb-4f8a-8556-a4515af86232_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "6ab8ae0755b1221cbc2e04354dcb9a1ff45cb46abbd2f7dc097c3202a7851730", + "total_rows": 10 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-6/96769cb9-54bb-4f8a-8556-a4515af86232_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-6/96769cb9-54bb-4f8a-8556-a4515af86232_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..2e407e3b7f4418e1a166dc69feaa95fb0dc3185e --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-6/96769cb9-54bb-4f8a-8556-a4515af86232_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6ab8ae0755b1221cbc2e04354dcb9a1ff45cb46abbd2f7dc097c3202a7851730 +size 8648708 diff --git a/data/terminalbench/anthropic/claude-opus-4-6/9e05f025-1283-4687-a549-68fc38f9f540.json b/data/terminalbench/anthropic/claude-opus-4-6/9e05f025-1283-4687-a549-68fc38f9f540.json new file mode 100644 index 0000000000000000000000000000000000000000..399efbcebf043bb1ae3183fda6647e2c0f8b648c --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-6/9e05f025-1283-4687-a549-68fc38f9f540.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-6/1786171161.539572", + "evaluation_timestamp": "1776417565.0", + "retrieved_timestamp": "1786171161.539572", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-6", + "id": "anthropic/claude-opus-4-6", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-6/9e05f025-1283-4687-a549-68fc38f9f540_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "cac00098d16f9fa37403f58681caacbe7da1c9619be3c48248e569f8f806ecf8", + "total_rows": 5 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-6/9e05f025-1283-4687-a549-68fc38f9f540_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-6/9e05f025-1283-4687-a549-68fc38f9f540_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..c20fa8ecff5efdea733257a843ecc055802c7bf4 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-6/9e05f025-1283-4687-a549-68fc38f9f540_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cac00098d16f9fa37403f58681caacbe7da1c9619be3c48248e569f8f806ecf8 +size 3668631 diff --git a/data/terminalbench/anthropic/claude-opus-4-6/9f5e699a-256e-4005-ae54-ce52b3253b81.json b/data/terminalbench/anthropic/claude-opus-4-6/9f5e699a-256e-4005-ae54-ce52b3253b81.json new file mode 100644 index 0000000000000000000000000000000000000000..66447e35da8bd85bc83a78de750d465cbe6501c9 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-6/9f5e699a-256e-4005-ae54-ce52b3253b81.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-6/1786168039.887658", + "evaluation_timestamp": "1777645894.0", + "retrieved_timestamp": "1786168039.887658", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-6", + "id": "anthropic/claude-opus-4-6", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-6/9f5e699a-256e-4005-ae54-ce52b3253b81_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "be1a1220ef1c48b44144971be8eb7b316b606e9eedfb7a979293b204d42cd800", + "total_rows": 6 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-6/9f5e699a-256e-4005-ae54-ce52b3253b81_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-6/9f5e699a-256e-4005-ae54-ce52b3253b81_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..fc1991471e6808b205d5b3aa0d6fcb991dfc8045 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-6/9f5e699a-256e-4005-ae54-ce52b3253b81_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:be1a1220ef1c48b44144971be8eb7b316b606e9eedfb7a979293b204d42cd800 +size 2310760 diff --git a/data/terminalbench/anthropic/claude-opus-4-6/a4a8981f-662e-4d46-84d2-6b5638d308db.json b/data/terminalbench/anthropic/claude-opus-4-6/a4a8981f-662e-4d46-84d2-6b5638d308db.json new file mode 100644 index 0000000000000000000000000000000000000000..14378b0d6a85cdaacb29f1e5fcbb23bdd5ed6bea --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-6/a4a8981f-662e-4d46-84d2-6b5638d308db.json @@ -0,0 +1,111 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-6/1786163546.662132", + "evaluation_timestamp": "1778005529.0", + "retrieved_timestamp": "1786163546.662132", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-6", + "id": "anthropic/claude-opus-4-6", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive+C/+1ep/97d1f448 for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 2, + "sample_ids": [ + "overfull-hbox", + "polyglot-rust-c" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1778005529.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.0, + "uncertainty": { + "num_samples": 1 + }, + "details": { + "total_matched_trajectories": "1", + "stopping_reason_count_repetition_guard": "1", + "avg_total_tokens_target_model": "657684.00", + "total_total_tokens_target_model": "657684.0", + "avg_total_tokens_other_models": "0.00", + "total_total_tokens_other_models": "0.0", + "avg_total_tokens_all_models": "657684.00", + "total_total_tokens_all_models": "657684.0", + "avg_turn_count": "35.00", + "total_turn_count": "35.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-6/a4a8981f-662e-4d46-84d2-6b5638d308db_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "03f096b09f62fb941993514ce26643301d5f5baae282d76301ab3d9e096652f0", + "total_rows": 1 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-6/a4a8981f-662e-4d46-84d2-6b5638d308db_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-6/a4a8981f-662e-4d46-84d2-6b5638d308db_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..89275a203b42bd83c828c5a136fa32e684a17363 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-6/a4a8981f-662e-4d46-84d2-6b5638d308db_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:03f096b09f62fb941993514ce26643301d5f5baae282d76301ab3d9e096652f0 +size 1061381 diff --git a/data/terminalbench/anthropic/claude-opus-4-6/a8229212-bb9f-4814-811d-02792b1b58e8.json b/data/terminalbench/anthropic/claude-opus-4-6/a8229212-bb9f-4814-811d-02792b1b58e8.json new file mode 100644 index 0000000000000000000000000000000000000000..f541197b6487b56ffc6591486bf9764f6a0906fc --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-6/a8229212-bb9f-4814-811d-02792b1b58e8.json @@ -0,0 +1,131 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-6/1786163544.483399", + "evaluation_timestamp": "1777899366.0", + "retrieved_timestamp": "1786163544.483399", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-6", + "id": "anthropic/claude-opus-4-6", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive+C/+1ep/0010ad56 for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 17, + "sample_ids": [ + "path-tracing", + "distribution-search", + "pypi-server", + "custom-memory-heap-crash", + "multi-source-data-merger", + "crack-7z-hash", + "chess-best-move", + "cobol-modernization", + "overfull-hbox", + "polyglot-rust-c", + "hf-model-inference", + "headless-terminal", + "fix-git", + "qemu-startup", + "build-pov-ray", + "git-multibranch", + "kv-store-grpc" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1777899366.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.8666666666666667, + "uncertainty": { + "standard_error": { + "value": 0.09085135251589957 + }, + "num_samples": 15 + }, + "details": { + "total_matched_trajectories": "15", + "stopping_reason_count_completed_on_successful_submit": "13", + "stopping_reason_count_repetition_guard": "1", + "stopping_reason_count_completed_with_submit": "1", + "avg_total_tokens_target_model": "575008.47", + "total_total_tokens_target_model": "8625127.0", + "avg_total_tokens_other_models": "0.00", + "total_total_tokens_other_models": "0.0", + "avg_total_tokens_all_models": "575008.47", + "total_total_tokens_all_models": "8625127.0", + "avg_turn_count": "31.13", + "total_turn_count": "467.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-6/a8229212-bb9f-4814-811d-02792b1b58e8_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "f579c557355b47a8a967615476fd93e7d7007b7339c22d2f3797e1b501e08f77", + "total_rows": 15 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-6/a8229212-bb9f-4814-811d-02792b1b58e8_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-6/a8229212-bb9f-4814-811d-02792b1b58e8_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..db208887d89685384bf7e389317b665c8ed1bafd --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-6/a8229212-bb9f-4814-811d-02792b1b58e8_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f579c557355b47a8a967615476fd93e7d7007b7339c22d2f3797e1b501e08f77 +size 2281766 diff --git a/data/terminalbench/anthropic/claude-opus-4-6/af9c3aca-31cb-48c0-9f56-998243f31d85.json b/data/terminalbench/anthropic/claude-opus-4-6/af9c3aca-31cb-48c0-9f56-998243f31d85.json new file mode 100644 index 0000000000000000000000000000000000000000..da506f9509a361ed13357c5906af96fe93fd77a4 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-6/af9c3aca-31cb-48c0-9f56-998243f31d85.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-6/1786168062.931616", + "evaluation_timestamp": "1777369892.0", + "retrieved_timestamp": "1786168062.931616", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-6", + "id": "anthropic/claude-opus-4-6", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-6/af9c3aca-31cb-48c0-9f56-998243f31d85_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "32e5ea25329113530af19456c2e41332b66ec07a1b972f817182b3fabfbd2d24", + "total_rows": 84 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-6/af9c3aca-31cb-48c0-9f56-998243f31d85_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-6/af9c3aca-31cb-48c0-9f56-998243f31d85_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..1312a93386cb684349cd4db1d5331137bf192734 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-6/af9c3aca-31cb-48c0-9f56-998243f31d85_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:32e5ea25329113530af19456c2e41332b66ec07a1b972f817182b3fabfbd2d24 +size 36515578 diff --git a/data/terminalbench/anthropic/claude-opus-4-6/bc5014f2-5cc9-45f7-9624-8b79a10dca32.json b/data/terminalbench/anthropic/claude-opus-4-6/bc5014f2-5cc9-45f7-9624-8b79a10dca32.json new file mode 100644 index 0000000000000000000000000000000000000000..eee7709ef8cf5f906251df7d5544279764b0bf3e --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-6/bc5014f2-5cc9-45f7-9624-8b79a10dca32.json @@ -0,0 +1,117 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-6/1786166152.766201", + "evaluation_timestamp": "1776629382.0", + "retrieved_timestamp": "1786166152.766201", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-6", + "id": "anthropic/claude-opus-4-6", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive/+2ep/a38ab61b for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 8, + "sample_ids": [ + "db-wal-recovery", + "schemelike-metacircular-eval", + "qemu-startup", + "git-multibranch", + "train-fasttext", + "video-processing", + "qemu-alpine-ssh", + "make-doom-for-mips" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1776629382.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.375, + "uncertainty": { + "standard_error": { + "value": 0.15669579263200217 + }, + "num_samples": 14 + }, + "details": { + "total_matched_trajectories": "14", + "stopping_reason_count_repetition_guard": "1", + "stopping_reason_count_token_limit": "13", + "avg_total_tokens_target_model": "9628286.36", + "total_total_tokens_target_model": "134796009.0", + "avg_total_tokens_other_models": "53.00", + "total_total_tokens_other_models": "742.0", + "avg_total_tokens_all_models": "9628339.36", + "total_total_tokens_all_models": "134796751.0", + "avg_turn_count": "158.00", + "total_turn_count": "2212.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "3600", + "attempt_timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "3600", + "max_connections": "10" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-6/bc5014f2-5cc9-45f7-9624-8b79a10dca32_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "33fb51f8133dc410ac6e3f9d789b718c60ae0944effb438511dc7e9d09a65eca", + "total_rows": 14 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-6/bc5014f2-5cc9-45f7-9624-8b79a10dca32_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-6/bc5014f2-5cc9-45f7-9624-8b79a10dca32_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..6e1a801d4a497a369ec3402f22a0913bfc3ccf9c --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-6/bc5014f2-5cc9-45f7-9624-8b79a10dca32_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:33fb51f8133dc410ac6e3f9d789b718c60ae0944effb438511dc7e9d09a65eca +size 13857806 diff --git a/data/terminalbench/anthropic/claude-opus-4-6/c24eb3ed-8d35-46ff-a707-90f5583e5f82.json b/data/terminalbench/anthropic/claude-opus-4-6/c24eb3ed-8d35-46ff-a707-90f5583e5f82.json new file mode 100644 index 0000000000000000000000000000000000000000..cab300b3bfdc3cb27b320b71433d16cfc4c9adab --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-6/c24eb3ed-8d35-46ff-a707-90f5583e5f82.json @@ -0,0 +1,120 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-6/1786166155.828946", + "evaluation_timestamp": "1778267760.0", + "retrieved_timestamp": "1786166155.828946", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-6", + "id": "anthropic/claude-opus-4-6", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive/+2ep/78105260 for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 8, + "sample_ids": [ + "sam-cell-seg", + "large-scale-text-editing", + "sqlite-with-gcov", + "openssl-selfsigned-cert", + "constraints-scheduling", + "dna-insert", + "vulnerable-secret", + "dna-assembly" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1778267760.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.625, + "uncertainty": { + "standard_error": { + "value": 0.15669579263200217 + }, + "num_samples": 15 + }, + "details": { + "total_matched_trajectories": "15", + "stopping_reason_count_repetition_guard": "15", + "avg_total_tokens_target_model": "615522.73", + "total_total_tokens_target_model": "9232841.0", + "avg_total_tokens_other_models": "703.27", + "total_total_tokens_other_models": "10549.0", + "avg_total_tokens_all_models": "616226.00", + "total_total_tokens_all_models": "9243390.0", + "avg_turn_count": "28.67", + "total_turn_count": "430.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-6/c24eb3ed-8d35-46ff-a707-90f5583e5f82_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "91d7f9f5eeff479f69de890ba3bc6718757308d50768041c797a24a87f79cf20", + "total_rows": 15 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-6/c24eb3ed-8d35-46ff-a707-90f5583e5f82_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-6/c24eb3ed-8d35-46ff-a707-90f5583e5f82_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..d4edca1b47823430c5574421a2891c2150745eda --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-6/c24eb3ed-8d35-46ff-a707-90f5583e5f82_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:91d7f9f5eeff479f69de890ba3bc6718757308d50768041c797a24a87f79cf20 +size 2156079 diff --git a/data/terminalbench/anthropic/claude-opus-4-6/cd3dfe30-4762-49a5-885d-1959b1c2543a.json b/data/terminalbench/anthropic/claude-opus-4-6/cd3dfe30-4762-49a5-885d-1959b1c2543a.json new file mode 100644 index 0000000000000000000000000000000000000000..fd449fa64df350ed1114f69a757e45b5d2b147e1 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-6/cd3dfe30-4762-49a5-885d-1959b1c2543a.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-6/1786171204.011645", + "evaluation_timestamp": "1776433475.0", + "retrieved_timestamp": "1786171204.011645", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-6", + "id": "anthropic/claude-opus-4-6", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-6/cd3dfe30-4762-49a5-885d-1959b1c2543a_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "371ca9f1adaa6db445333f6d021638d45ebd480d9c3a245caff0d5eb94494235", + "total_rows": 50 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-6/cd3dfe30-4762-49a5-885d-1959b1c2543a_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-6/cd3dfe30-4762-49a5-885d-1959b1c2543a_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..4c099890befce7b29cd4602c26bba2bee79f06f9 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-6/cd3dfe30-4762-49a5-885d-1959b1c2543a_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:371ca9f1adaa6db445333f6d021638d45ebd480d9c3a245caff0d5eb94494235 +size 41511494 diff --git a/data/terminalbench/anthropic/claude-opus-4-6/d15bbfb0-df31-4904-847a-30da2525dd4e.json b/data/terminalbench/anthropic/claude-opus-4-6/d15bbfb0-df31-4904-847a-30da2525dd4e.json new file mode 100644 index 0000000000000000000000000000000000000000..353c8cbce32dd53d7c4156607816df9e1931a9d3 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-6/d15bbfb0-df31-4904-847a-30da2525dd4e.json @@ -0,0 +1,107 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-6/1786166092.238347", + "evaluation_timestamp": "1777025747.0", + "retrieved_timestamp": "1786166092.238347", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-6", + "id": "anthropic/claude-opus-4-6", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive/+1ep/dc014c9a for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 1, + "sample_ids": [ + "schemelike-metacircular-eval" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1777025747.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 1.0, + "uncertainty": { + "num_samples": 1 + }, + "details": { + "total_matched_trajectories": "1", + "stopping_reason_count_token_limit": "1", + "avg_total_tokens_target_model": "10014609.00", + "total_total_tokens_target_model": "10014609.0", + "avg_total_tokens_other_models": "0.00", + "total_total_tokens_other_models": "0.0", + "avg_total_tokens_all_models": "10014609.00", + "total_total_tokens_all_models": "10014609.0", + "avg_turn_count": "136.00", + "total_turn_count": "136.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-6/d15bbfb0-df31-4904-847a-30da2525dd4e_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "5bb95129b0b25c3631b9dec5c00aa0146a54c854b961c91c608eed2aa25941d2", + "total_rows": 1 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-6/d15bbfb0-df31-4904-847a-30da2525dd4e_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-6/d15bbfb0-df31-4904-847a-30da2525dd4e_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..6cbcde1599432e796323de45a34c12736f4ea034 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-6/d15bbfb0-df31-4904-847a-30da2525dd4e_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5bb95129b0b25c3631b9dec5c00aa0146a54c854b961c91c608eed2aa25941d2 +size 1728400 diff --git a/data/terminalbench/anthropic/claude-opus-4-6/d4c920e6-02ee-43af-a973-a8aed1a6a8ff.json b/data/terminalbench/anthropic/claude-opus-4-6/d4c920e6-02ee-43af-a973-a8aed1a6a8ff.json new file mode 100644 index 0000000000000000000000000000000000000000..3ba1328ce9230ddd6c4084f3254197995d4085aa --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-6/d4c920e6-02ee-43af-a973-a8aed1a6a8ff.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-6/1786168083.852329", + "evaluation_timestamp": "1777369892.0", + "retrieved_timestamp": "1786168083.852329", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-6", + "id": "anthropic/claude-opus-4-6", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-6/d4c920e6-02ee-43af-a973-a8aed1a6a8ff_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "b76b69b1d308bcc6f287903a4c0417eb835a80962e8a9498f14bba9ccbc26cc2", + "total_rows": 92 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-6/d4c920e6-02ee-43af-a973-a8aed1a6a8ff_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-6/d4c920e6-02ee-43af-a973-a8aed1a6a8ff_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..80ecc639325f7440ed55c73ab50c8bec251c9a0b --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-6/d4c920e6-02ee-43af-a973-a8aed1a6a8ff_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b76b69b1d308bcc6f287903a4c0417eb835a80962e8a9498f14bba9ccbc26cc2 +size 28695774 diff --git a/data/terminalbench/anthropic/claude-opus-4-6/d83f9977-5e02-417f-8e32-a0dc93197aea.json b/data/terminalbench/anthropic/claude-opus-4-6/d83f9977-5e02-417f-8e32-a0dc93197aea.json new file mode 100644 index 0000000000000000000000000000000000000000..1a2da8a62558f45c63294c611ae7ab38996816c4 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-6/d83f9977-5e02-417f-8e32-a0dc93197aea.json @@ -0,0 +1,110 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-6/1786166108.439693", + "evaluation_timestamp": "1778513006.0", + "retrieved_timestamp": "1786166108.439693", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-6", + "id": "anthropic/claude-opus-4-6", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive/+1ep/5dd93e71 for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 1, + "sample_ids": [ + "dna-assembly" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1778513006.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 1.0, + "uncertainty": { + "num_samples": 1 + }, + "details": { + "total_matched_trajectories": "1", + "stopping_reason_count_repetition_guard": "1", + "avg_total_tokens_target_model": "2806143.00", + "total_total_tokens_target_model": "2806143.0", + "avg_total_tokens_other_models": "1689.00", + "total_total_tokens_other_models": "1689.0", + "avg_total_tokens_all_models": "2807832.00", + "total_total_tokens_all_models": "2807832.0", + "avg_turn_count": "55.00", + "total_turn_count": "55.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-6/d83f9977-5e02-417f-8e32-a0dc93197aea_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "5eceee65f0cb02fb1342785507fc5100233afeaf0b92dcb4e9c14f7a6749eaec", + "total_rows": 1 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-6/d83f9977-5e02-417f-8e32-a0dc93197aea_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-6/d83f9977-5e02-417f-8e32-a0dc93197aea_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..2fccd5254e8ab2cfa52a5efcb4172235c87c25d8 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-6/d83f9977-5e02-417f-8e32-a0dc93197aea_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5eceee65f0cb02fb1342785507fc5100233afeaf0b92dcb4e9c14f7a6749eaec +size 709051 diff --git a/data/terminalbench/anthropic/claude-opus-4-6/da52db97-3694-4ef3-b9bb-27947533dbb6.json b/data/terminalbench/anthropic/claude-opus-4-6/da52db97-3694-4ef3-b9bb-27947533dbb6.json new file mode 100644 index 0000000000000000000000000000000000000000..683ed68b83820de130daef9eb48bf3bbd28459aa --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-6/da52db97-3694-4ef3-b9bb-27947533dbb6.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-6/1786171165.357802", + "evaluation_timestamp": "1776346958.0", + "retrieved_timestamp": "1786171165.357802", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-6", + "id": "anthropic/claude-opus-4-6", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-6/da52db97-3694-4ef3-b9bb-27947533dbb6_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "7c684a7a29d7e3fcb52af678a22f50973de14a01de71f431c060c76c3283de04", + "total_rows": 6 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-6/da52db97-3694-4ef3-b9bb-27947533dbb6_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-6/da52db97-3694-4ef3-b9bb-27947533dbb6_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..df87c8002ec722ada94b1af003a5199ebffbdcf1 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-6/da52db97-3694-4ef3-b9bb-27947533dbb6_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7c684a7a29d7e3fcb52af678a22f50973de14a01de71f431c060c76c3283de04 +size 2989520 diff --git a/data/terminalbench/anthropic/claude-opus-4-6/db3668f4-ae68-4c82-afc3-dc5a2f86a1aa.json b/data/terminalbench/anthropic/claude-opus-4-6/db3668f4-ae68-4c82-afc3-dc5a2f86a1aa.json new file mode 100644 index 0000000000000000000000000000000000000000..51be560c78ffe04cc35f901cc0c7c5e3957ca318 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-6/db3668f4-ae68-4c82-afc3-dc5a2f86a1aa.json @@ -0,0 +1,122 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-6/1786166144.645421", + "evaluation_timestamp": "1776520455.0", + "retrieved_timestamp": "1786166144.645421", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-6", + "id": "anthropic/claude-opus-4-6", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive/+2ep/4fede33d for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 15, + "sample_ids": [ + "llm-inference-batching-scheduler", + "reshard-c4-data", + "write-compressor", + "pytorch-model-cli", + "password-recovery", + "portfolio-optimization", + "regex-chess", + "path-tracing", + "prove-plus-comm", + "caffe-cifar-10", + "distribution-search", + "mteb-retrieve", + "pypi-server", + "adaptive-rejection-sampler", + "multi-source-data-merger" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1776520455.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.8, + "uncertainty": { + "standard_error": { + "value": 0.0951189731211342 + }, + "num_samples": 30 + }, + "details": { + "total_matched_trajectories": "30", + "stopping_reason_count_repetition_guard": "11", + "stopping_reason_count_token_limit": "19", + "avg_total_tokens_target_model": "8154786.00", + "total_total_tokens_target_model": "244643580.0", + "avg_total_tokens_other_models": "150.27", + "total_total_tokens_other_models": "4508.0", + "avg_total_tokens_all_models": "8154936.27", + "total_total_tokens_all_models": "244648088.0", + "avg_turn_count": "137.17", + "total_turn_count": "4115.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "max_connections": "10", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "max_connections": "10" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-6/db3668f4-ae68-4c82-afc3-dc5a2f86a1aa_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "bbb8389ab923d2f24b9232ec13940ac226b84fb2d7b205cb596eced8a28479e0", + "total_rows": 30 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-6/db3668f4-ae68-4c82-afc3-dc5a2f86a1aa_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-6/db3668f4-ae68-4c82-afc3-dc5a2f86a1aa_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..deb7d34242f4e42b11f2f88c26c73c4dc8b7b7f4 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-6/db3668f4-ae68-4c82-afc3-dc5a2f86a1aa_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bbb8389ab923d2f24b9232ec13940ac226b84fb2d7b205cb596eced8a28479e0 +size 24942765 diff --git a/data/terminalbench/anthropic/claude-opus-4-6/e4e37724-0583-44c0-bb81-85c0a5cb01c7.json b/data/terminalbench/anthropic/claude-opus-4-6/e4e37724-0583-44c0-bb81-85c0a5cb01c7.json new file mode 100644 index 0000000000000000000000000000000000000000..87d9f0df1a38b92bed6eb84d4c2fcb4923b999ff --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-6/e4e37724-0583-44c0-bb81-85c0a5cb01c7.json @@ -0,0 +1,113 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-6/1786166204.713978", + "evaluation_timestamp": "1776359447.0", + "retrieved_timestamp": "1786166204.713978", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-6", + "id": "anthropic/claude-opus-4-6", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive/+4ep/598f3ed0 for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 6, + "sample_ids": [ + "log-summary-date-ranges", + "path-tracing-reverse", + "modernize-scientific-stack", + "torch-tensor-parallelism", + "feal-linear-cryptanalysis", + "custom-memory-heap-crash" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1776359447.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.6666666666666666, + "uncertainty": { + "standard_error": { + "value": 0.16666666666666669 + }, + "num_samples": 24 + }, + "details": { + "total_matched_trajectories": "24", + "stopping_reason_count_repetition_guard": "15", + "stopping_reason_count_token_limit": "9", + "avg_total_tokens_target_model": "6701067.67", + "total_total_tokens_target_model": "160825624.0", + "avg_total_tokens_other_models": "288.62", + "total_total_tokens_other_models": "6927.0", + "avg_total_tokens_all_models": "6701356.29", + "total_total_tokens_all_models": "160832551.0", + "avg_turn_count": "110.83", + "total_turn_count": "2660.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "max_connections": "10", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "max_connections": "10" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-6/e4e37724-0583-44c0-bb81-85c0a5cb01c7_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "1627ac59d9473648eb8bd2eb6b757f71eedc88c2fea79eab89c4a7e81ad4baa6", + "total_rows": 24 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-6/e4e37724-0583-44c0-bb81-85c0a5cb01c7_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-6/e4e37724-0583-44c0-bb81-85c0a5cb01c7_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..d3357e6ff5515a34e43884acec48988e942baa13 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-6/e4e37724-0583-44c0-bb81-85c0a5cb01c7_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1627ac59d9473648eb8bd2eb6b757f71eedc88c2fea79eab89c4a7e81ad4baa6 +size 17970045 diff --git a/data/terminalbench/anthropic/claude-opus-4-6/e63053a7-a1b6-41a8-8917-2f2d534d3931.json b/data/terminalbench/anthropic/claude-opus-4-6/e63053a7-a1b6-41a8-8917-2f2d534d3931.json new file mode 100644 index 0000000000000000000000000000000000000000..752f2893d938d081a04c59fdfa95c52acf174961 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-6/e63053a7-a1b6-41a8-8917-2f2d534d3931.json @@ -0,0 +1,112 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-6/1786166167.358123", + "evaluation_timestamp": "1776433475.0", + "retrieved_timestamp": "1786166167.358123", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-6", + "id": "anthropic/claude-opus-4-6", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive/+3ep/bd45a2f0 for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 5, + "sample_ids": [ + "gpt2-codegolf", + "break-filter-js-from-html", + "merge-diff-arc-agi-task", + "winning-avg-corewars", + "largest-eigenval" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1776433475.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.8, + "uncertainty": { + "standard_error": { + "value": 0.19999999999999998 + }, + "num_samples": 15 + }, + "details": { + "total_matched_trajectories": "15", + "stopping_reason_count_repetition_guard": "5", + "stopping_reason_count_token_limit": "10", + "avg_total_tokens_target_model": "8773215.73", + "total_total_tokens_target_model": "131598236.0", + "avg_total_tokens_other_models": "132.87", + "total_total_tokens_other_models": "1993.0", + "avg_total_tokens_all_models": "8773348.60", + "total_total_tokens_all_models": "131600229.0", + "avg_turn_count": "158.53", + "total_turn_count": "2378.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "max_connections": "10", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "max_connections": "10" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-6/e63053a7-a1b6-41a8-8917-2f2d534d3931_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "18f0f80423d689022d760bd01643ec8d88d14ec71e0d34429d0263ea060296a8", + "total_rows": 15 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-6/e63053a7-a1b6-41a8-8917-2f2d534d3931_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-6/e63053a7-a1b6-41a8-8917-2f2d534d3931_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..63c4800b79478787f048e7f33a50c937f571959d --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-6/e63053a7-a1b6-41a8-8917-2f2d534d3931_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:18f0f80423d689022d760bd01643ec8d88d14ec71e0d34429d0263ea060296a8 +size 12663572 diff --git a/data/terminalbench/anthropic/claude-opus-4-6/eaaa74c4-c541-4c6d-b921-ea1136a529aa.json b/data/terminalbench/anthropic/claude-opus-4-6/eaaa74c4-c541-4c6d-b921-ea1136a529aa.json new file mode 100644 index 0000000000000000000000000000000000000000..67513716482c47916465586ee134eeafa6e9041c --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-6/eaaa74c4-c541-4c6d-b921-ea1136a529aa.json @@ -0,0 +1,115 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-6/1786163549.425704", + "evaluation_timestamp": "1777478370.0", + "retrieved_timestamp": "1786163549.425704", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-6", + "id": "anthropic/claude-opus-4-6", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive+C/+1ep/be03a748 for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 8, + "sample_ids": [ + "schemelike-metacircular-eval", + "qemu-alpine-ssh", + "large-scale-text-editing", + "sqlite-with-gcov", + "constraints-scheduling", + "dna-insert", + "vulnerable-secret", + "dna-assembly" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1777478370.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 1.0, + "uncertainty": { + "num_samples": 8 + }, + "details": { + "total_matched_trajectories": "8", + "stopping_reason_count_completed_on_successful_submit": "7", + "stopping_reason_count_token_limit": "1", + "avg_total_tokens_target_model": "2447466.38", + "total_total_tokens_target_model": "19579731.0", + "avg_total_tokens_other_models": "0.00", + "total_total_tokens_other_models": "0.0", + "avg_total_tokens_all_models": "2447466.38", + "total_total_tokens_all_models": "19579731.0", + "avg_turn_count": "50.50", + "total_turn_count": "404.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "32", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "32", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-6/eaaa74c4-c541-4c6d-b921-ea1136a529aa_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "b7faf7446f686c34571bbd63d63f6cc7eb011a589aad105956ea721d90af88c8", + "total_rows": 8 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-6/eaaa74c4-c541-4c6d-b921-ea1136a529aa_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-6/eaaa74c4-c541-4c6d-b921-ea1136a529aa_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..6e1d50bb439df161b793e60573a3809fc878ef53 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-6/eaaa74c4-c541-4c6d-b921-ea1136a529aa_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b7faf7446f686c34571bbd63d63f6cc7eb011a589aad105956ea721d90af88c8 +size 2833673 diff --git a/data/terminalbench/anthropic/claude-opus-4-6/ef90bb9d-3b17-4109-b2ab-439b6b9a574d.json b/data/terminalbench/anthropic/claude-opus-4-6/ef90bb9d-3b17-4109-b2ab-439b6b9a574d.json new file mode 100644 index 0000000000000000000000000000000000000000..e2eb7d4e6eeb0802a15771441f2cf4a8fc741ebf --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-6/ef90bb9d-3b17-4109-b2ab-439b6b9a574d.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-6/1786168022.113588", + "evaluation_timestamp": "1777954003.0", + "retrieved_timestamp": "1786168022.113588", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-6", + "id": "anthropic/claude-opus-4-6", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-6/ef90bb9d-3b17-4109-b2ab-439b6b9a574d_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "3b99c3219b3e61855237622eab99bab3bbd03fd40f5785560c37e8d96b0139ad", + "total_rows": 20 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-6/ef90bb9d-3b17-4109-b2ab-439b6b9a574d_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-6/ef90bb9d-3b17-4109-b2ab-439b6b9a574d_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..9b34560b5a3469607f386ca182c3e8f6c1cd5636 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-6/ef90bb9d-3b17-4109-b2ab-439b6b9a574d_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3b99c3219b3e61855237622eab99bab3bbd03fd40f5785560c37e8d96b0139ad +size 9666402 diff --git a/data/terminalbench/anthropic/claude-opus-4-6/f6c893f3-76d7-4c1b-b27f-959c6db1f07f.json b/data/terminalbench/anthropic/claude-opus-4-6/f6c893f3-76d7-4c1b-b27f-959c6db1f07f.json new file mode 100644 index 0000000000000000000000000000000000000000..8436e26f4c054c64ec529a8ab11fa59c745d95d2 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-6/f6c893f3-76d7-4c1b-b27f-959c6db1f07f.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-6/1786168069.007875", + "evaluation_timestamp": "1777544308.0", + "retrieved_timestamp": "1786168069.007875", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-6", + "id": "anthropic/claude-opus-4-6", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-6/f6c893f3-76d7-4c1b-b27f-959c6db1f07f_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "6f403f8f11cddd06cda8d1a9677defcf9f7259fcc2d33e5398b4ed5b87143f6e", + "total_rows": 8 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-6/f6c893f3-76d7-4c1b-b27f-959c6db1f07f_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-6/f6c893f3-76d7-4c1b-b27f-959c6db1f07f_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..06ea32e152b7a180625c9fb34aa4daa79c2a7d21 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-6/f6c893f3-76d7-4c1b-b27f-959c6db1f07f_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6f403f8f11cddd06cda8d1a9677defcf9f7259fcc2d33e5398b4ed5b87143f6e +size 1853843 diff --git a/data/terminalbench/anthropic/claude-opus-4-6/fa3b966f-6af2-4072-9265-f2c73e40f01e.json b/data/terminalbench/anthropic/claude-opus-4-6/fa3b966f-6af2-4072-9265-f2c73e40f01e.json new file mode 100644 index 0000000000000000000000000000000000000000..acd510c6fbd88cb2396e2d79db011a234f9dc865 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-6/fa3b966f-6af2-4072-9265-f2c73e40f01e.json @@ -0,0 +1,118 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-6/1786166100.784737", + "evaluation_timestamp": "1776601455.0", + "retrieved_timestamp": "1786166100.784737", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-6", + "id": "anthropic/claude-opus-4-6", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive/+1ep/48bf0a83 for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 11, + "sample_ids": [ + "configure-git-webserver", + "crack-7z-hash", + "chess-best-move", + "cobol-modernization", + "overfull-hbox", + "polyglot-rust-c", + "hf-model-inference", + "compile-compcert", + "headless-terminal", + "fix-git", + "build-pov-ray" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1776601455.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.7272727272727273, + "uncertainty": { + "standard_error": { + "value": 0.14083575804390605 + }, + "num_samples": 11 + }, + "details": { + "total_matched_trajectories": "11", + "stopping_reason_count_token_limit": "9", + "stopping_reason_count_repetition_guard": "2", + "avg_total_tokens_target_model": "8732237.00", + "total_total_tokens_target_model": "96054607.0", + "avg_total_tokens_other_models": "35.55", + "total_total_tokens_other_models": "391.0", + "avg_total_tokens_all_models": "8732272.55", + "total_total_tokens_all_models": "96054998.0", + "avg_turn_count": "158.64", + "total_turn_count": "1745.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "max_connections": "10", + "reasoning_effort": "\"xhigh\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "max_connections": "10" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-6/fa3b966f-6af2-4072-9265-f2c73e40f01e_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "a3020a5de44e9b556e87350ace499554e338731288e7870cc3b3810361a70ee4", + "total_rows": 11 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-6/fa3b966f-6af2-4072-9265-f2c73e40f01e_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-6/fa3b966f-6af2-4072-9265-f2c73e40f01e_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..ba53a703c10dacaaf9367cb2bc82a47e12d0b41c --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-6/fa3b966f-6af2-4072-9265-f2c73e40f01e_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a3020a5de44e9b556e87350ace499554e338731288e7870cc3b3810361a70ee4 +size 8705774 diff --git a/data/terminalbench/anthropic/claude-opus-4-6/fd5990bd-61ab-42fa-9595-5e7120bad2ff.json b/data/terminalbench/anthropic/claude-opus-4-6/fd5990bd-61ab-42fa-9595-5e7120bad2ff.json new file mode 100644 index 0000000000000000000000000000000000000000..d375cd3a8739e55517242a11c7b230132ec56581 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-6/fd5990bd-61ab-42fa-9595-5e7120bad2ff.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/anthropic_claude-opus-4-6/1786168046.043515", + "evaluation_timestamp": "1778011122.0", + "retrieved_timestamp": "1786168046.043515", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "anthropic/claude-opus-4-6", + "id": "anthropic/claude-opus-4-6", + "developer": "anthropic", + "inference_platform": "anthropic", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/anthropic/claude-opus-4-6/fd5990bd-61ab-42fa-9595-5e7120bad2ff_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "dfe85f0dc279b6254b95f1b39fe37ec1c2aca9b25539890fc79df719527b29b3", + "total_rows": 23 + } +} \ No newline at end of file diff --git a/data/terminalbench/anthropic/claude-opus-4-6/fd5990bd-61ab-42fa-9595-5e7120bad2ff_samples.jsonl b/data/terminalbench/anthropic/claude-opus-4-6/fd5990bd-61ab-42fa-9595-5e7120bad2ff_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..73e0c35079f24f7433e3d40b663afabe40c9ee78 --- /dev/null +++ b/data/terminalbench/anthropic/claude-opus-4-6/fd5990bd-61ab-42fa-9595-5e7120bad2ff_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dfe85f0dc279b6254b95f1b39fe37ec1c2aca9b25539890fc79df719527b29b3 +size 12242643 diff --git a/data/terminalbench/openai/gpt-5-2025-08-07/01e0f1f1-880d-4f51-9488-67326e3f2b92.json b/data/terminalbench/openai/gpt-5-2025-08-07/01e0f1f1-880d-4f51-9488-67326e3f2b92.json new file mode 100644 index 0000000000000000000000000000000000000000..0860138546931a9803ba2a39944bb8d880920a05 --- /dev/null +++ b/data/terminalbench/openai/gpt-5-2025-08-07/01e0f1f1-880d-4f51-9488-67326e3f2b92.json @@ -0,0 +1,107 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5-2025-08-07/1786163379.462347", + "evaluation_timestamp": "1777025747.0", + "retrieved_timestamp": "1786163379.462347", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5-2025-08-07", + "id": "openai/gpt-5-2025-08-07", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive+C/+1ep/0dd8e472 for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 1, + "sample_ids": [ + "torch-pipeline-parallelism" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1777025747.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.0, + "uncertainty": { + "num_samples": 1 + }, + "details": { + "total_matched_trajectories": "1", + "stopping_reason_count_repetition_guard": "1", + "avg_total_tokens_target_model": "740219.00", + "total_total_tokens_target_model": "740219.0", + "avg_total_tokens_other_models": "0.00", + "total_total_tokens_other_models": "0.0", + "avg_total_tokens_all_models": "740219.00", + "total_total_tokens_all_models": "740219.0", + "avg_turn_count": "27.00", + "total_turn_count": "27.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5-2025-08-07/01e0f1f1-880d-4f51-9488-67326e3f2b92_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "82fe7493913e1019c956b1124313eb6038141c92435af993321e6414273b07b8", + "total_rows": 1 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5-2025-08-07/01e0f1f1-880d-4f51-9488-67326e3f2b92_samples.jsonl b/data/terminalbench/openai/gpt-5-2025-08-07/01e0f1f1-880d-4f51-9488-67326e3f2b92_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..2097b518e306d0be699c4562d0835c01de0a03f7 --- /dev/null +++ b/data/terminalbench/openai/gpt-5-2025-08-07/01e0f1f1-880d-4f51-9488-67326e3f2b92_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:82fe7493913e1019c956b1124313eb6038141c92435af993321e6414273b07b8 +size 1627166 diff --git a/data/terminalbench/openai/gpt-5-2025-08-07/0219d612-660d-4de3-8b39-146f57e3c7c3.json b/data/terminalbench/openai/gpt-5-2025-08-07/0219d612-660d-4de3-8b39-146f57e3c7c3.json new file mode 100644 index 0000000000000000000000000000000000000000..81ebe6d80b330a3a17f583687bcdc825764f86bb --- /dev/null +++ b/data/terminalbench/openai/gpt-5-2025-08-07/0219d612-660d-4de3-8b39-146f57e3c7c3.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5-2025-08-07/1786167861.036392", + "evaluation_timestamp": "1775665083.0", + "retrieved_timestamp": "1786167861.036392", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5-2025-08-07", + "id": "openai/gpt-5-2025-08-07", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5-2025-08-07/0219d612-660d-4de3-8b39-146f57e3c7c3_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "3ba12cdbb6585676e861806f648045e5bc4dbd3aa7163a8ab8d21113e3eb1f57", + "total_rows": 56 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5-2025-08-07/0219d612-660d-4de3-8b39-146f57e3c7c3_samples.jsonl b/data/terminalbench/openai/gpt-5-2025-08-07/0219d612-660d-4de3-8b39-146f57e3c7c3_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..08f88d09d1288dfac06152e073f9337f333b571b --- /dev/null +++ b/data/terminalbench/openai/gpt-5-2025-08-07/0219d612-660d-4de3-8b39-146f57e3c7c3_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ba12cdbb6585676e861806f648045e5bc4dbd3aa7163a8ab8d21113e3eb1f57 +size 27907959 diff --git a/data/terminalbench/openai/gpt-5-2025-08-07/09c15f65-4873-4744-a868-d78ccd005655.json b/data/terminalbench/openai/gpt-5-2025-08-07/09c15f65-4873-4744-a868-d78ccd005655.json new file mode 100644 index 0000000000000000000000000000000000000000..d262181ef15e4032b4fb18933826e717d9b581bd --- /dev/null +++ b/data/terminalbench/openai/gpt-5-2025-08-07/09c15f65-4873-4744-a868-d78ccd005655.json @@ -0,0 +1,120 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5-2025-08-07/1786163401.366006", + "evaluation_timestamp": "1777031202.0", + "retrieved_timestamp": "1786163401.366006", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5-2025-08-07", + "id": "openai/gpt-5-2025-08-07", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive+C/+2ep/b8ba6a27 for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 8, + "sample_ids": [ + "install-windows-3.11", + "make-mips-interpreter", + "mteb-leaderboard", + "mailman", + "extract-elf", + "code-from-image", + "fix-code-vulnerability", + "regex-log" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1777031202.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.3125, + "uncertainty": { + "standard_error": { + "value": 0.16194961738586655 + }, + "num_samples": 16 + }, + "details": { + "total_matched_trajectories": "16", + "stopping_reason_count_completed_on_successful_submit": "5", + "stopping_reason_count_repetition_guard": "5", + "stopping_reason_count_token_limit": "5", + "stopping_reason_count_completed_with_submit": "1", + "avg_total_tokens_target_model": "3829295.50", + "total_total_tokens_target_model": "61268728.0", + "avg_total_tokens_other_models": "4340.50", + "total_total_tokens_other_models": "69448.0", + "avg_total_tokens_all_models": "3833636.00", + "total_total_tokens_all_models": "61338176.0", + "avg_turn_count": "59.81", + "total_turn_count": "957.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5-2025-08-07/09c15f65-4873-4744-a868-d78ccd005655_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "82188580f90af261aff07257df1ad1abcb973ffe18496e29ad394cb0b7811def", + "total_rows": 16 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5-2025-08-07/09c15f65-4873-4744-a868-d78ccd005655_samples.jsonl b/data/terminalbench/openai/gpt-5-2025-08-07/09c15f65-4873-4744-a868-d78ccd005655_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..c6decf1cdea10b947af7cc4ba28471c3186b7228 --- /dev/null +++ b/data/terminalbench/openai/gpt-5-2025-08-07/09c15f65-4873-4744-a868-d78ccd005655_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:82188580f90af261aff07257df1ad1abcb973ffe18496e29ad394cb0b7811def +size 9292715 diff --git a/data/terminalbench/openai/gpt-5-2025-08-07/18690467-7e27-4737-9377-9c30a574e884.json b/data/terminalbench/openai/gpt-5-2025-08-07/18690467-7e27-4737-9377-9c30a574e884.json new file mode 100644 index 0000000000000000000000000000000000000000..bd4ccb385c114a3892965e9af651202d9229e191 --- /dev/null +++ b/data/terminalbench/openai/gpt-5-2025-08-07/18690467-7e27-4737-9377-9c30a574e884.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5-2025-08-07/1786169702.680641", + "evaluation_timestamp": "1775646127.0", + "retrieved_timestamp": "1786169702.680641", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5-2025-08-07", + "id": "openai/gpt-5-2025-08-07", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5-2025-08-07/18690467-7e27-4737-9377-9c30a574e884_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "2e92d5eb45306a6a316406309a016f76fa81a1b97a73fc3d59d7386da60b7923", + "total_rows": 1 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5-2025-08-07/18690467-7e27-4737-9377-9c30a574e884_samples.jsonl b/data/terminalbench/openai/gpt-5-2025-08-07/18690467-7e27-4737-9377-9c30a574e884_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..1dd2d80ad9189d229dd1bf62fd9c519647edfb0b --- /dev/null +++ b/data/terminalbench/openai/gpt-5-2025-08-07/18690467-7e27-4737-9377-9c30a574e884_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2e92d5eb45306a6a316406309a016f76fa81a1b97a73fc3d59d7386da60b7923 +size 306766 diff --git a/data/terminalbench/openai/gpt-5-2025-08-07/1eb8e28b-1cec-408a-921f-a1e7b57b4b1d.json b/data/terminalbench/openai/gpt-5-2025-08-07/1eb8e28b-1cec-408a-921f-a1e7b57b4b1d.json new file mode 100644 index 0000000000000000000000000000000000000000..a7ee57a1a92e8754b2ca95bd1dd0b8d83b404468 --- /dev/null +++ b/data/terminalbench/openai/gpt-5-2025-08-07/1eb8e28b-1cec-408a-921f-a1e7b57b4b1d.json @@ -0,0 +1,119 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5-2025-08-07/1786163415.650051", + "evaluation_timestamp": "1776696546.0", + "retrieved_timestamp": "1786163415.650051", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5-2025-08-07", + "id": "openai/gpt-5-2025-08-07", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive+C/+3ep/ac8e9b79 for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 8, + "sample_ids": [ + "qemu-alpine-ssh", + "kv-store-grpc", + "install-windows-3.11", + "make-doom-for-mips", + "torch-pipeline-parallelism", + "tune-mjcf", + "extract-moves-from-video", + "count-dataset-tokens" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1776696546.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.42857142857142855, + "uncertainty": { + "standard_error": { + "value": 0.2020305089104421 + }, + "num_samples": 14 + }, + "details": { + "total_matched_trajectories": "14", + "stopping_reason_count_completed_on_successful_submit": "9", + "stopping_reason_count_repetition_guard": "2", + "stopping_reason_count_token_limit": "3", + "avg_total_tokens_target_model": "3124845.29", + "total_total_tokens_target_model": "43747834.0", + "avg_total_tokens_other_models": "157.86", + "total_total_tokens_other_models": "2210.0", + "avg_total_tokens_all_models": "3125003.14", + "total_total_tokens_all_models": "43750044.0", + "avg_turn_count": "58.50", + "total_turn_count": "819.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "3600", + "attempt_timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5-2025-08-07/1eb8e28b-1cec-408a-921f-a1e7b57b4b1d_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "48600701fdeb661613433560248d16141d69bdab7f7380de2fa398b8ff3a8b4a", + "total_rows": 14 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5-2025-08-07/1eb8e28b-1cec-408a-921f-a1e7b57b4b1d_samples.jsonl b/data/terminalbench/openai/gpt-5-2025-08-07/1eb8e28b-1cec-408a-921f-a1e7b57b4b1d_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..001ee6626718518e9995ca7044ce75f91c8a46dd --- /dev/null +++ b/data/terminalbench/openai/gpt-5-2025-08-07/1eb8e28b-1cec-408a-921f-a1e7b57b4b1d_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:48600701fdeb661613433560248d16141d69bdab7f7380de2fa398b8ff3a8b4a +size 5677064 diff --git a/data/terminalbench/openai/gpt-5-2025-08-07/22e7fc84-96fb-4a1c-8e18-2904e08a4fb8.json b/data/terminalbench/openai/gpt-5-2025-08-07/22e7fc84-96fb-4a1c-8e18-2904e08a4fb8.json new file mode 100644 index 0000000000000000000000000000000000000000..b1d7fdd252d9ef630aa682b54f16ad728eae191c --- /dev/null +++ b/data/terminalbench/openai/gpt-5-2025-08-07/22e7fc84-96fb-4a1c-8e18-2904e08a4fb8.json @@ -0,0 +1,111 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5-2025-08-07/1786163412.269089", + "evaluation_timestamp": "1775749914.0", + "retrieved_timestamp": "1786163412.269089", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5-2025-08-07", + "id": "openai/gpt-5-2025-08-07", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive+C/+3ep for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 4, + "sample_ids": [ + "write-compressor", + "log-summary-date-ranges", + "pytorch-model-cli", + "largest-eigenval" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1775749914.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.5833333333333333, + "uncertainty": { + "standard_error": { + "value": 0.20971762320196524 + }, + "num_samples": 12 + }, + "details": { + "total_matched_trajectories": "12", + "stopping_reason_count_completed_on_successful_submit": "7", + "stopping_reason_count_repetition_guard": "3", + "stopping_reason_count_token_limit": "2", + "avg_total_tokens_target_model": "4225884.67", + "total_total_tokens_target_model": "50710616.0", + "avg_total_tokens_other_models": "137.67", + "total_total_tokens_other_models": "1652.0", + "avg_total_tokens_all_models": "4226022.33", + "total_total_tokens_all_models": "50712268.0", + "avg_turn_count": "64.50", + "total_turn_count": "774.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5-2025-08-07/22e7fc84-96fb-4a1c-8e18-2904e08a4fb8_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "172d5544a4430730e08820e20d84dfc7d094b8595a757a60eba4fc70c9fdfab9", + "total_rows": 12 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5-2025-08-07/22e7fc84-96fb-4a1c-8e18-2904e08a4fb8_samples.jsonl b/data/terminalbench/openai/gpt-5-2025-08-07/22e7fc84-96fb-4a1c-8e18-2904e08a4fb8_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..58f265566dcf43f77d81b32abd8941733d300bba --- /dev/null +++ b/data/terminalbench/openai/gpt-5-2025-08-07/22e7fc84-96fb-4a1c-8e18-2904e08a4fb8_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:172d5544a4430730e08820e20d84dfc7d094b8595a757a60eba4fc70c9fdfab9 +size 8316373 diff --git a/data/terminalbench/openai/gpt-5-2025-08-07/2a6ccfb6-9bad-4602-9a65-e9b8151b9a8e.json b/data/terminalbench/openai/gpt-5-2025-08-07/2a6ccfb6-9bad-4602-9a65-e9b8151b9a8e.json new file mode 100644 index 0000000000000000000000000000000000000000..96cbb671740e8aa1247989981a29703f4cbb339f --- /dev/null +++ b/data/terminalbench/openai/gpt-5-2025-08-07/2a6ccfb6-9bad-4602-9a65-e9b8151b9a8e.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5-2025-08-07/1786169674.384439", + "evaluation_timestamp": "1776346958.0", + "retrieved_timestamp": "1786169674.384439", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5-2025-08-07", + "id": "openai/gpt-5-2025-08-07", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5-2025-08-07/2a6ccfb6-9bad-4602-9a65-e9b8151b9a8e_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "50354d09d17f4bf0f306436d5d87e39e08ee64849f2001ca640f400adc8d67d0", + "total_rows": 17 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5-2025-08-07/2a6ccfb6-9bad-4602-9a65-e9b8151b9a8e_samples.jsonl b/data/terminalbench/openai/gpt-5-2025-08-07/2a6ccfb6-9bad-4602-9a65-e9b8151b9a8e_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..de49ab8df843f35affbfc6dbe0936e0f88b06779 --- /dev/null +++ b/data/terminalbench/openai/gpt-5-2025-08-07/2a6ccfb6-9bad-4602-9a65-e9b8151b9a8e_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:50354d09d17f4bf0f306436d5d87e39e08ee64849f2001ca640f400adc8d67d0 +size 3489431 diff --git a/data/terminalbench/openai/gpt-5-2025-08-07/32cd6f30-bdb3-4bbc-a833-26d8be3a30e8.json b/data/terminalbench/openai/gpt-5-2025-08-07/32cd6f30-bdb3-4bbc-a833-26d8be3a30e8.json new file mode 100644 index 0000000000000000000000000000000000000000..cd24695cd50014c830a1cd315cc19afb28862072 --- /dev/null +++ b/data/terminalbench/openai/gpt-5-2025-08-07/32cd6f30-bdb3-4bbc-a833-26d8be3a30e8.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5-2025-08-07/1786167844.340169", + "evaluation_timestamp": "1776359446.0", + "retrieved_timestamp": "1786167844.340169", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5-2025-08-07", + "id": "openai/gpt-5-2025-08-07", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5-2025-08-07/32cd6f30-bdb3-4bbc-a833-26d8be3a30e8_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "1dc8b6654a1e66990021223fe69e053776c3a3a9d5d3f291394ddbca018abdb6", + "total_rows": 50 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5-2025-08-07/32cd6f30-bdb3-4bbc-a833-26d8be3a30e8_samples.jsonl b/data/terminalbench/openai/gpt-5-2025-08-07/32cd6f30-bdb3-4bbc-a833-26d8be3a30e8_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..d92c6aff31f1a2b5c92f0fd3f6bf7bbe28731eef --- /dev/null +++ b/data/terminalbench/openai/gpt-5-2025-08-07/32cd6f30-bdb3-4bbc-a833-26d8be3a30e8_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1dc8b6654a1e66990021223fe69e053776c3a3a9d5d3f291394ddbca018abdb6 +size 19511279 diff --git a/data/terminalbench/openai/gpt-5-2025-08-07/44f978a8-d165-4771-82f2-ad5c53c55638.json b/data/terminalbench/openai/gpt-5-2025-08-07/44f978a8-d165-4771-82f2-ad5c53c55638.json new file mode 100644 index 0000000000000000000000000000000000000000..8d29f84eec3136c0eaea6cce291244d2cabe2213 --- /dev/null +++ b/data/terminalbench/openai/gpt-5-2025-08-07/44f978a8-d165-4771-82f2-ad5c53c55638.json @@ -0,0 +1,122 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5-2025-08-07/1786163393.655864", + "evaluation_timestamp": "1776433475.0", + "retrieved_timestamp": "1786163393.655864", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5-2025-08-07", + "id": "openai/gpt-5-2025-08-07", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive+C/+2ep/b52a3250 for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 13, + "sample_ids": [ + "regex-chess", + "torch-tensor-parallelism", + "path-tracing", + "distribution-search", + "mteb-retrieve", + "pypi-server", + "custom-memory-heap-crash", + "adaptive-rejection-sampler", + "multi-source-data-merger", + "configure-git-webserver", + "crack-7z-hash", + "cobol-modernization", + "hf-model-inference" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1776433475.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.6538461538461539, + "uncertainty": { + "standard_error": { + "value": 0.1313711636661487 + }, + "num_samples": 26 + }, + "details": { + "total_matched_trajectories": "26", + "stopping_reason_count_completed_on_successful_submit": "17", + "stopping_reason_count_repetition_guard": "6", + "stopping_reason_count_token_limit": "3", + "avg_total_tokens_target_model": "2176833.08", + "total_total_tokens_target_model": "56597660.0", + "avg_total_tokens_other_models": "707.85", + "total_total_tokens_other_models": "18404.0", + "avg_total_tokens_all_models": "2177540.92", + "total_total_tokens_all_models": "56616064.0", + "avg_turn_count": "39.04", + "total_turn_count": "1015.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5-2025-08-07/44f978a8-d165-4771-82f2-ad5c53c55638_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "65175a410d8b70caca6339751f5e862a02cc6a8aadb3e7a6c79815c7ebc841ac", + "total_rows": 26 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5-2025-08-07/44f978a8-d165-4771-82f2-ad5c53c55638_samples.jsonl b/data/terminalbench/openai/gpt-5-2025-08-07/44f978a8-d165-4771-82f2-ad5c53c55638_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..3e2c251042feb89df180c08491242cfd5013a47d --- /dev/null +++ b/data/terminalbench/openai/gpt-5-2025-08-07/44f978a8-d165-4771-82f2-ad5c53c55638_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:65175a410d8b70caca6339751f5e862a02cc6a8aadb3e7a6c79815c7ebc841ac +size 11171601 diff --git a/data/terminalbench/openai/gpt-5-2025-08-07/45f84b03-712c-4831-9955-409b0218499a.json b/data/terminalbench/openai/gpt-5-2025-08-07/45f84b03-712c-4831-9955-409b0218499a.json new file mode 100644 index 0000000000000000000000000000000000000000..1d5ee79bd1b061b5bf46134cc4c9ea35c5b01d51 --- /dev/null +++ b/data/terminalbench/openai/gpt-5-2025-08-07/45f84b03-712c-4831-9955-409b0218499a.json @@ -0,0 +1,115 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5-2025-08-07/1786166019.954488", + "evaluation_timestamp": "1775604415.0", + "retrieved_timestamp": "1786166019.954488", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.202,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5-2025-08-07", + "id": "openai/gpt-5-2025-08-07", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive/+2ep for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 9, + "sample_ids": [ + "mteb-retrieve", + "headless-terminal", + "qemu-startup", + "install-windows-3.11", + "make-mips-interpreter", + "mteb-leaderboard", + "fix-ocaml-gc", + "pytorch-model-recovery", + "model-extraction-relu-logits" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1775604415.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.3888888888888889, + "uncertainty": { + "standard_error": { + "value": 0.16197088596792503 + }, + "num_samples": 18 + }, + "details": { + "total_matched_trajectories": "18", + "stopping_reason_count_repetition_guard": "17", + "stopping_reason_count_token_limit": "1", + "avg_total_tokens_target_model": "2340536.17", + "total_total_tokens_target_model": "42129651.0", + "avg_total_tokens_other_models": "1299.78", + "total_total_tokens_other_models": "23396.0", + "avg_total_tokens_all_models": "2341835.94", + "total_total_tokens_all_models": "42153047.0", + "avg_turn_count": "48.83", + "total_turn_count": "879.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5-2025-08-07/45f84b03-712c-4831-9955-409b0218499a_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "2d602ee0621a0c7afe18642b2598f1e665291b7d1af11c049d14f020c95b1af3", + "total_rows": 18 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5-2025-08-07/45f84b03-712c-4831-9955-409b0218499a_samples.jsonl b/data/terminalbench/openai/gpt-5-2025-08-07/45f84b03-712c-4831-9955-409b0218499a_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..e3028bcfeef3c3462c63e61e0f53f0abadad1a6c --- /dev/null +++ b/data/terminalbench/openai/gpt-5-2025-08-07/45f84b03-712c-4831-9955-409b0218499a_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2d602ee0621a0c7afe18642b2598f1e665291b7d1af11c049d14f020c95b1af3 +size 6172690 diff --git a/data/terminalbench/openai/gpt-5-2025-08-07/5023298d-d25e-48f9-8a9b-cd6703103459.json b/data/terminalbench/openai/gpt-5-2025-08-07/5023298d-d25e-48f9-8a9b-cd6703103459.json new file mode 100644 index 0000000000000000000000000000000000000000..a73726ca0477d2b5ee2f1212b3fc3ffcc049fda8 --- /dev/null +++ b/data/terminalbench/openai/gpt-5-2025-08-07/5023298d-d25e-48f9-8a9b-cd6703103459.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5-2025-08-07/1786169695.925003", + "evaluation_timestamp": "1776338190.0", + "retrieved_timestamp": "1786169695.925003", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5-2025-08-07", + "id": "openai/gpt-5-2025-08-07", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5-2025-08-07/5023298d-d25e-48f9-8a9b-cd6703103459_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "f18e693e9b4e8ec636e56381e76737f0bb258f995e4a5cd0b9e772ae7aaace1b", + "total_rows": 9 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5-2025-08-07/5023298d-d25e-48f9-8a9b-cd6703103459_samples.jsonl b/data/terminalbench/openai/gpt-5-2025-08-07/5023298d-d25e-48f9-8a9b-cd6703103459_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..9caf1d817d23015a483ebed3fc7e0323b2da388e --- /dev/null +++ b/data/terminalbench/openai/gpt-5-2025-08-07/5023298d-d25e-48f9-8a9b-cd6703103459_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f18e693e9b4e8ec636e56381e76737f0bb258f995e4a5cd0b9e772ae7aaace1b +size 1979117 diff --git a/data/terminalbench/openai/gpt-5-2025-08-07/56d6aa42-d987-44c1-a8b5-1af82450f2b0.json b/data/terminalbench/openai/gpt-5-2025-08-07/56d6aa42-d987-44c1-a8b5-1af82450f2b0.json new file mode 100644 index 0000000000000000000000000000000000000000..bcc3da6a831803c8e1801b0595e93d3c13081afb --- /dev/null +++ b/data/terminalbench/openai/gpt-5-2025-08-07/56d6aa42-d987-44c1-a8b5-1af82450f2b0.json @@ -0,0 +1,116 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5-2025-08-07/1786163389.48556", + "evaluation_timestamp": "1776433475.0", + "retrieved_timestamp": "1786163389.48556", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5-2025-08-07", + "id": "openai/gpt-5-2025-08-07", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive+C/+1ep/83690c9f for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 7, + "sample_ids": [ + "password-recovery", + "portfolio-optimization", + "path-tracing-reverse", + "modernize-scientific-stack", + "prove-plus-comm", + "feal-linear-cryptanalysis", + "caffe-cifar-10" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1776433475.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.7142857142857143, + "uncertainty": { + "standard_error": { + "value": 0.18442777839082936 + }, + "num_samples": 7 + }, + "details": { + "total_matched_trajectories": "7", + "stopping_reason_count_repetition_guard": "1", + "stopping_reason_count_completed_on_successful_submit": "5", + "stopping_reason_count_token_limit": "1", + "avg_total_tokens_target_model": "1995965.71", + "total_total_tokens_target_model": "13971760.0", + "avg_total_tokens_other_models": "29.00", + "total_total_tokens_other_models": "203.0", + "avg_total_tokens_all_models": "1995994.71", + "total_total_tokens_all_models": "13971963.0", + "avg_turn_count": "36.71", + "total_turn_count": "257.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5-2025-08-07/56d6aa42-d987-44c1-a8b5-1af82450f2b0_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "e25c622b38111c88b07ddbc3ee404e9fcdaadb6761d11fac68e283a9939c5871", + "total_rows": 7 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5-2025-08-07/56d6aa42-d987-44c1-a8b5-1af82450f2b0_samples.jsonl b/data/terminalbench/openai/gpt-5-2025-08-07/56d6aa42-d987-44c1-a8b5-1af82450f2b0_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..4e4628fff6daa2459d3aa564de62a940522505af --- /dev/null +++ b/data/terminalbench/openai/gpt-5-2025-08-07/56d6aa42-d987-44c1-a8b5-1af82450f2b0_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e25c622b38111c88b07ddbc3ee404e9fcdaadb6761d11fac68e283a9939c5871 +size 2855624 diff --git a/data/terminalbench/openai/gpt-5-2025-08-07/5fcf1cba-2979-4d5f-9393-f8ef36d622a7.json b/data/terminalbench/openai/gpt-5-2025-08-07/5fcf1cba-2979-4d5f-9393-f8ef36d622a7.json new file mode 100644 index 0000000000000000000000000000000000000000..cbe9e239f6dc4accf8a6d2067c40725c376ebfe4 --- /dev/null +++ b/data/terminalbench/openai/gpt-5-2025-08-07/5fcf1cba-2979-4d5f-9393-f8ef36d622a7.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5-2025-08-07/1786167824.252351", + "evaluation_timestamp": "1776680886.0", + "retrieved_timestamp": "1786167824.252351", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5-2025-08-07", + "id": "openai/gpt-5-2025-08-07", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5-2025-08-07/5fcf1cba-2979-4d5f-9393-f8ef36d622a7_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "b5321b6d654ba3fd7aa8b7fa6296791c15978a177df3d839cf5e396ec43cf6da", + "total_rows": 8 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5-2025-08-07/5fcf1cba-2979-4d5f-9393-f8ef36d622a7_samples.jsonl b/data/terminalbench/openai/gpt-5-2025-08-07/5fcf1cba-2979-4d5f-9393-f8ef36d622a7_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..2bc9a0dda86fe605dca0566a1e863e58bf1fd626 --- /dev/null +++ b/data/terminalbench/openai/gpt-5-2025-08-07/5fcf1cba-2979-4d5f-9393-f8ef36d622a7_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b5321b6d654ba3fd7aa8b7fa6296791c15978a177df3d839cf5e396ec43cf6da +size 5020921 diff --git a/data/terminalbench/openai/gpt-5-2025-08-07/66fef12a-897d-416c-a1a3-4b76bac126eb.json b/data/terminalbench/openai/gpt-5-2025-08-07/66fef12a-897d-416c-a1a3-4b76bac126eb.json new file mode 100644 index 0000000000000000000000000000000000000000..36f3c699e59b2d4083368a2d7358be4376afb460 --- /dev/null +++ b/data/terminalbench/openai/gpt-5-2025-08-07/66fef12a-897d-416c-a1a3-4b76bac126eb.json @@ -0,0 +1,107 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5-2025-08-07/1786163426.128664", + "evaluation_timestamp": "1777096559.0", + "retrieved_timestamp": "1786163426.128664", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5-2025-08-07", + "id": "openai/gpt-5-2025-08-07", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive+C/+4ep/d0aae35e for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 1, + "sample_ids": [ + "cancel-async-tasks" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1777096559.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 1.0, + "uncertainty": { + "num_samples": 4 + }, + "details": { + "total_matched_trajectories": "4", + "stopping_reason_count_completed_on_successful_submit": "4", + "avg_total_tokens_target_model": "82949.00", + "total_total_tokens_target_model": "331796.0", + "avg_total_tokens_other_models": "0.00", + "total_total_tokens_other_models": "0.0", + "avg_total_tokens_all_models": "82949.00", + "total_total_tokens_all_models": "331796.0", + "avg_turn_count": "7.50", + "total_turn_count": "30.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5-2025-08-07/66fef12a-897d-416c-a1a3-4b76bac126eb_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "68cf53072ab104c62fe04e0e92bacaabdae941bc3e0a36256bc681eac13ee847", + "total_rows": 4 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5-2025-08-07/66fef12a-897d-416c-a1a3-4b76bac126eb_samples.jsonl b/data/terminalbench/openai/gpt-5-2025-08-07/66fef12a-897d-416c-a1a3-4b76bac126eb_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..8082945a9a221d8a4a04288056a3550d473d9cad --- /dev/null +++ b/data/terminalbench/openai/gpt-5-2025-08-07/66fef12a-897d-416c-a1a3-4b76bac126eb_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68cf53072ab104c62fe04e0e92bacaabdae941bc3e0a36256bc681eac13ee847 +size 522475 diff --git a/data/terminalbench/openai/gpt-5-2025-08-07/6c1348bd-176b-4500-a462-f2f5a6129f3e.json b/data/terminalbench/openai/gpt-5-2025-08-07/6c1348bd-176b-4500-a462-f2f5a6129f3e.json new file mode 100644 index 0000000000000000000000000000000000000000..ebb82d632970311c6f3ae678a1f69001c3a95696 --- /dev/null +++ b/data/terminalbench/openai/gpt-5-2025-08-07/6c1348bd-176b-4500-a462-f2f5a6129f3e.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5-2025-08-07/1786170018.35964", + "evaluation_timestamp": "1775078458.0", + "retrieved_timestamp": "1786170018.35964", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.202,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5-2025-08-07", + "id": "openai/gpt-5-2025-08-07", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5-2025-08-07/6c1348bd-176b-4500-a462-f2f5a6129f3e_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "14fc78f5a90e4906f8d23a5f4a7848c3c066d9a7a9000334dffafb989b8f46be", + "total_rows": 790 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5-2025-08-07/6c1348bd-176b-4500-a462-f2f5a6129f3e_samples.jsonl b/data/terminalbench/openai/gpt-5-2025-08-07/6c1348bd-176b-4500-a462-f2f5a6129f3e_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..de3b00e04fdcae5555adbe50279ffa97a14e4f17 --- /dev/null +++ b/data/terminalbench/openai/gpt-5-2025-08-07/6c1348bd-176b-4500-a462-f2f5a6129f3e_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:14fc78f5a90e4906f8d23a5f4a7848c3c066d9a7a9000334dffafb989b8f46be +size 244307479 diff --git a/data/terminalbench/openai/gpt-5-2025-08-07/6f6bedce-1198-438f-95c7-625d0796f7c0.json b/data/terminalbench/openai/gpt-5-2025-08-07/6f6bedce-1198-438f-95c7-625d0796f7c0.json new file mode 100644 index 0000000000000000000000000000000000000000..0eda6b32b0ae0e7b2f8e8d6ce82ba64e00b3da63 --- /dev/null +++ b/data/terminalbench/openai/gpt-5-2025-08-07/6f6bedce-1198-438f-95c7-625d0796f7c0.json @@ -0,0 +1,105 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5-2025-08-07/1786166034.806852", + "evaluation_timestamp": "1775604415.0", + "retrieved_timestamp": "1786166034.806852", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.202,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5-2025-08-07", + "id": "openai/gpt-5-2025-08-07", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive/+3ep for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 2, + "sample_ids": [ + "schemelike-metacircular-eval", + "train-fasttext" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1775604415.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.0, + "uncertainty": { + "num_samples": 6 + }, + "details": { + "total_matched_trajectories": "6", + "stopping_reason_count_repetition_guard": "4", + "stopping_reason_count_token_limit": "2", + "avg_total_tokens_target_model": "5784482.67", + "total_total_tokens_target_model": "34706896.0", + "avg_total_tokens_other_models": "5584.50", + "total_total_tokens_other_models": "33507.0", + "avg_total_tokens_all_models": "5790067.17", + "total_total_tokens_all_models": "34740403.0", + "avg_turn_count": "95.50", + "total_turn_count": "573.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5-2025-08-07/6f6bedce-1198-438f-95c7-625d0796f7c0_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "3bd506b62af5cbea3ef6bc24cb79cd36a78d180207f56bc20e972ed3ff84024b", + "total_rows": 6 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5-2025-08-07/6f6bedce-1198-438f-95c7-625d0796f7c0_samples.jsonl b/data/terminalbench/openai/gpt-5-2025-08-07/6f6bedce-1198-438f-95c7-625d0796f7c0_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..3472ea75ddc59606045627940be7dde8fec7d2dd --- /dev/null +++ b/data/terminalbench/openai/gpt-5-2025-08-07/6f6bedce-1198-438f-95c7-625d0796f7c0_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3bd506b62af5cbea3ef6bc24cb79cd36a78d180207f56bc20e972ed3ff84024b +size 3796570 diff --git a/data/terminalbench/openai/gpt-5-2025-08-07/7e3ad108-527b-4fb4-9289-b52c2eb92ad4.json b/data/terminalbench/openai/gpt-5-2025-08-07/7e3ad108-527b-4fb4-9289-b52c2eb92ad4.json new file mode 100644 index 0000000000000000000000000000000000000000..fd399bf14f84a7012814c0cac2bad84fb25b8369 --- /dev/null +++ b/data/terminalbench/openai/gpt-5-2025-08-07/7e3ad108-527b-4fb4-9289-b52c2eb92ad4.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5-2025-08-07/1786167832.799597", + "evaluation_timestamp": "1776724749.0", + "retrieved_timestamp": "1786167832.799597", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5-2025-08-07", + "id": "openai/gpt-5-2025-08-07", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5-2025-08-07/7e3ad108-527b-4fb4-9289-b52c2eb92ad4_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "84f5bf0e84bc984126187df98396e46172e0b22f3b3b9f340035ee01b7d442f5", + "total_rows": 24 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5-2025-08-07/7e3ad108-527b-4fb4-9289-b52c2eb92ad4_samples.jsonl b/data/terminalbench/openai/gpt-5-2025-08-07/7e3ad108-527b-4fb4-9289-b52c2eb92ad4_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..47646581da23a8cfb93f9a607b03436281f5cdd1 --- /dev/null +++ b/data/terminalbench/openai/gpt-5-2025-08-07/7e3ad108-527b-4fb4-9289-b52c2eb92ad4_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:84f5bf0e84bc984126187df98396e46172e0b22f3b3b9f340035ee01b7d442f5 +size 14530645 diff --git a/data/terminalbench/openai/gpt-5-2025-08-07/897b3671-5ded-41f7-a150-a7bd0f2b4d66.json b/data/terminalbench/openai/gpt-5-2025-08-07/897b3671-5ded-41f7-a150-a7bd0f2b4d66.json new file mode 100644 index 0000000000000000000000000000000000000000..6cb8a1c5ff71d0f9489c1b320e413ec45f0042da --- /dev/null +++ b/data/terminalbench/openai/gpt-5-2025-08-07/897b3671-5ded-41f7-a150-a7bd0f2b4d66.json @@ -0,0 +1,111 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5-2025-08-07/1786163428.924556", + "evaluation_timestamp": "1777917637.0", + "retrieved_timestamp": "1786163428.924556", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5-2025-08-07", + "id": "openai/gpt-5-2025-08-07", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive+C/+4ep/64ae8c02 for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 1, + "sample_ids": [ + "sam-cell-seg" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1777917637.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.0, + "uncertainty": { + "num_samples": 4 + }, + "details": { + "total_matched_trajectories": "4", + "stopping_reason_count_repetition_guard": "3", + "stopping_reason_count_token_limit": "1", + "avg_total_tokens_target_model": "5580786.75", + "total_total_tokens_target_model": "22323147.0", + "avg_total_tokens_other_models": "1466.75", + "total_total_tokens_other_models": "5867.0", + "avg_total_tokens_all_models": "5582253.50", + "total_total_tokens_all_models": "22329014.0", + "avg_turn_count": "91.25", + "total_turn_count": "365.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5-2025-08-07/897b3671-5ded-41f7-a150-a7bd0f2b4d66_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "49218c2dad5e2f7f80bf61e82e15140479e2df171135c479cb427c0bec229fe6", + "total_rows": 4 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5-2025-08-07/897b3671-5ded-41f7-a150-a7bd0f2b4d66_samples.jsonl b/data/terminalbench/openai/gpt-5-2025-08-07/897b3671-5ded-41f7-a150-a7bd0f2b4d66_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..6227093114bfa2b6dff5c00f025dc416fcdeadd4 --- /dev/null +++ b/data/terminalbench/openai/gpt-5-2025-08-07/897b3671-5ded-41f7-a150-a7bd0f2b4d66_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:49218c2dad5e2f7f80bf61e82e15140479e2df171135c479cb427c0bec229fe6 +size 3754397 diff --git a/data/terminalbench/openai/gpt-5-2025-08-07/8dda7514-efda-4be6-a914-977d22a76cc4.json b/data/terminalbench/openai/gpt-5-2025-08-07/8dda7514-efda-4be6-a914-977d22a76cc4.json new file mode 100644 index 0000000000000000000000000000000000000000..87b156d9ee5940d9bc682257fa637d6413185fae --- /dev/null +++ b/data/terminalbench/openai/gpt-5-2025-08-07/8dda7514-efda-4be6-a914-977d22a76cc4.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5-2025-08-07/1786169692.00385", + "evaluation_timestamp": "1775558526.0", + "retrieved_timestamp": "1786169692.00385", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.202,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5-2025-08-07", + "id": "openai/gpt-5-2025-08-07", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5-2025-08-07/8dda7514-efda-4be6-a914-977d22a76cc4_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "3a7d4d23636c15c4c21ac0b86a272b9ab9f20b656120ca10eb7a3aa99234c76a", + "total_rows": 30 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5-2025-08-07/8dda7514-efda-4be6-a914-977d22a76cc4_samples.jsonl b/data/terminalbench/openai/gpt-5-2025-08-07/8dda7514-efda-4be6-a914-977d22a76cc4_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..3198e38ed9dcefc761831157f2c7603e04becb24 --- /dev/null +++ b/data/terminalbench/openai/gpt-5-2025-08-07/8dda7514-efda-4be6-a914-977d22a76cc4_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3a7d4d23636c15c4c21ac0b86a272b9ab9f20b656120ca10eb7a3aa99234c76a +size 10281971 diff --git a/data/terminalbench/openai/gpt-5-2025-08-07/9638f5a9-5e9d-4997-86b8-96ecaeae6de7.json b/data/terminalbench/openai/gpt-5-2025-08-07/9638f5a9-5e9d-4997-86b8-96ecaeae6de7.json new file mode 100644 index 0000000000000000000000000000000000000000..4215af09adae3f5362f87e497844a5357574849e --- /dev/null +++ b/data/terminalbench/openai/gpt-5-2025-08-07/9638f5a9-5e9d-4997-86b8-96ecaeae6de7.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5-2025-08-07/1786167814.219861", + "evaluation_timestamp": "1776648141.0", + "retrieved_timestamp": "1786167814.219861", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5-2025-08-07", + "id": "openai/gpt-5-2025-08-07", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5-2025-08-07/9638f5a9-5e9d-4997-86b8-96ecaeae6de7_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "643b204291382ef0eababaad6accc7c9016dd6fe265fa8a6ba966ebe25253cfd", + "total_rows": 26 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5-2025-08-07/9638f5a9-5e9d-4997-86b8-96ecaeae6de7_samples.jsonl b/data/terminalbench/openai/gpt-5-2025-08-07/9638f5a9-5e9d-4997-86b8-96ecaeae6de7_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..aed702783cf7109dc540f497b7a7d390cac5cdd3 --- /dev/null +++ b/data/terminalbench/openai/gpt-5-2025-08-07/9638f5a9-5e9d-4997-86b8-96ecaeae6de7_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:643b204291382ef0eababaad6accc7c9016dd6fe265fa8a6ba966ebe25253cfd +size 12266470 diff --git a/data/terminalbench/openai/gpt-5-2025-08-07/a0789cfd-ed3b-4a53-bf37-9eba06e78969.json b/data/terminalbench/openai/gpt-5-2025-08-07/a0789cfd-ed3b-4a53-bf37-9eba06e78969.json new file mode 100644 index 0000000000000000000000000000000000000000..5209698341d0597a2d735c285aebf6873df87ceb --- /dev/null +++ b/data/terminalbench/openai/gpt-5-2025-08-07/a0789cfd-ed3b-4a53-bf37-9eba06e78969.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5-2025-08-07/1786169700.619448", + "evaluation_timestamp": "1775604414.0", + "retrieved_timestamp": "1786169700.619448", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.202,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5-2025-08-07", + "id": "openai/gpt-5-2025-08-07", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5-2025-08-07/a0789cfd-ed3b-4a53-bf37-9eba06e78969_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "d23902680bbb6e93bbd6676088c4fc79a8c1aca814367a163bac8e9254505a0b", + "total_rows": 3 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5-2025-08-07/a0789cfd-ed3b-4a53-bf37-9eba06e78969_samples.jsonl b/data/terminalbench/openai/gpt-5-2025-08-07/a0789cfd-ed3b-4a53-bf37-9eba06e78969_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..3c0c6cd83425d6076b83e6a03e17504ad8b1ad77 --- /dev/null +++ b/data/terminalbench/openai/gpt-5-2025-08-07/a0789cfd-ed3b-4a53-bf37-9eba06e78969_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d23902680bbb6e93bbd6676088c4fc79a8c1aca814367a163bac8e9254505a0b +size 666641 diff --git a/data/terminalbench/openai/gpt-5-2025-08-07/a89282be-b674-488e-9d64-a89928508219.json b/data/terminalbench/openai/gpt-5-2025-08-07/a89282be-b674-488e-9d64-a89928508219.json new file mode 100644 index 0000000000000000000000000000000000000000..f08b7026b0c0106b69ce3e19e2893e0529fe8a4b --- /dev/null +++ b/data/terminalbench/openai/gpt-5-2025-08-07/a89282be-b674-488e-9d64-a89928508219.json @@ -0,0 +1,107 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5-2025-08-07/1786163381.377906", + "evaluation_timestamp": "1777292858.0", + "retrieved_timestamp": "1786163381.377906", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5-2025-08-07", + "id": "openai/gpt-5-2025-08-07", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive+C/+1ep/347739d1 for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 1, + "sample_ids": [ + "circuit-fibsqrt" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1777292858.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 1.0, + "uncertainty": { + "num_samples": 1 + }, + "details": { + "total_matched_trajectories": "1", + "stopping_reason_count_completed_on_successful_submit": "1", + "avg_total_tokens_target_model": "5028745.00", + "total_total_tokens_target_model": "5028745.0", + "avg_total_tokens_other_models": "0.00", + "total_total_tokens_other_models": "0.0", + "avg_total_tokens_all_models": "5028745.00", + "total_total_tokens_all_models": "5028745.0", + "avg_turn_count": "91.00", + "total_turn_count": "91.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5-2025-08-07/a89282be-b674-488e-9d64-a89928508219_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "7c594780f2a9151fb967762aca1cc32b5c9c03b045aeea4bcab06ce950f2808e", + "total_rows": 1 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5-2025-08-07/a89282be-b674-488e-9d64-a89928508219_samples.jsonl b/data/terminalbench/openai/gpt-5-2025-08-07/a89282be-b674-488e-9d64-a89928508219_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..e2a3e210a5943bfafc529e1c103ed32f080aff09 --- /dev/null +++ b/data/terminalbench/openai/gpt-5-2025-08-07/a89282be-b674-488e-9d64-a89928508219_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7c594780f2a9151fb967762aca1cc32b5c9c03b045aeea4bcab06ce950f2808e +size 609818 diff --git a/data/terminalbench/openai/gpt-5-2025-08-07/aa66b933-896a-42de-87d2-8cb5dc5cf317.json b/data/terminalbench/openai/gpt-5-2025-08-07/aa66b933-896a-42de-87d2-8cb5dc5cf317.json new file mode 100644 index 0000000000000000000000000000000000000000..e480b5adca424a830266cbe1097d999f03a87d32 --- /dev/null +++ b/data/terminalbench/openai/gpt-5-2025-08-07/aa66b933-896a-42de-87d2-8cb5dc5cf317.json @@ -0,0 +1,107 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5-2025-08-07/1786163377.24002", + "evaluation_timestamp": "1776814698.0", + "retrieved_timestamp": "1786163377.24002", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5-2025-08-07", + "id": "openai/gpt-5-2025-08-07", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive+C/+1ep/d0f1e342 for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 1, + "sample_ids": [ + "make-doom-for-mips" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1776814698.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.0, + "uncertainty": { + "num_samples": 1 + }, + "details": { + "total_matched_trajectories": "1", + "stopping_reason_count_repetition_guard": "1", + "avg_total_tokens_target_model": "9849666.00", + "total_total_tokens_target_model": "9849666.0", + "avg_total_tokens_other_models": "3346.00", + "total_total_tokens_other_models": "3346.0", + "avg_total_tokens_all_models": "9853012.00", + "total_total_tokens_all_models": "9853012.0", + "avg_turn_count": "139.00", + "total_turn_count": "139.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5-2025-08-07/aa66b933-896a-42de-87d2-8cb5dc5cf317_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "1b8bc8fae05da02d6f9c013dcec07c5f4856bd02ff68949a8ce11b3d28e584f0", + "total_rows": 1 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5-2025-08-07/aa66b933-896a-42de-87d2-8cb5dc5cf317_samples.jsonl b/data/terminalbench/openai/gpt-5-2025-08-07/aa66b933-896a-42de-87d2-8cb5dc5cf317_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..01d0237838f6d92cec54ef7ae5a81e7a891e4164 --- /dev/null +++ b/data/terminalbench/openai/gpt-5-2025-08-07/aa66b933-896a-42de-87d2-8cb5dc5cf317_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1b8bc8fae05da02d6f9c013dcec07c5f4856bd02ff68949a8ce11b3d28e584f0 +size 926871 diff --git a/data/terminalbench/openai/gpt-5-2025-08-07/b2a34d12-4da2-4413-99f7-0383217110a3.json b/data/terminalbench/openai/gpt-5-2025-08-07/b2a34d12-4da2-4413-99f7-0383217110a3.json new file mode 100644 index 0000000000000000000000000000000000000000..7c8b83f486570dc5efe57645f03e5f183c5a7453 --- /dev/null +++ b/data/terminalbench/openai/gpt-5-2025-08-07/b2a34d12-4da2-4413-99f7-0383217110a3.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5-2025-08-07/1786167891.348793", + "evaluation_timestamp": "1775078458.0", + "retrieved_timestamp": "1786167891.348793", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.202,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5-2025-08-07", + "id": "openai/gpt-5-2025-08-07", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5-2025-08-07/b2a34d12-4da2-4413-99f7-0383217110a3_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "c278b874b20c2936bd4128c71f31eec0cb0edf5bb76bc93468c35da6891642db", + "total_rows": 40 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5-2025-08-07/b2a34d12-4da2-4413-99f7-0383217110a3_samples.jsonl b/data/terminalbench/openai/gpt-5-2025-08-07/b2a34d12-4da2-4413-99f7-0383217110a3_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..7e4964da3dab8dc548b4247c967ac2025499e0c0 --- /dev/null +++ b/data/terminalbench/openai/gpt-5-2025-08-07/b2a34d12-4da2-4413-99f7-0383217110a3_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c278b874b20c2936bd4128c71f31eec0cb0edf5bb76bc93468c35da6891642db +size 18840202 diff --git a/data/terminalbench/openai/gpt-5-2025-08-07/b3aa8e91-9d8b-47ca-b05c-99a777adebce.json b/data/terminalbench/openai/gpt-5-2025-08-07/b3aa8e91-9d8b-47ca-b05c-99a777adebce.json new file mode 100644 index 0000000000000000000000000000000000000000..39a0dc697e211a6a893d8e23d3b81b6d410fc398 --- /dev/null +++ b/data/terminalbench/openai/gpt-5-2025-08-07/b3aa8e91-9d8b-47ca-b05c-99a777adebce.json @@ -0,0 +1,113 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5-2025-08-07/1786163383.646007", + "evaluation_timestamp": "1776886986.0", + "retrieved_timestamp": "1786163383.646007", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5-2025-08-07", + "id": "openai/gpt-5-2025-08-07", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive+C/+1ep/04c03b4a for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 4, + "sample_ids": [ + "video-processing", + "qemu-alpine-ssh", + "torch-pipeline-parallelism", + "extract-moves-from-video" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1776886986.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.3333333333333333, + "uncertainty": { + "standard_error": { + "value": 0.33333333333333337 + }, + "num_samples": 3 + }, + "details": { + "total_matched_trajectories": "3", + "stopping_reason_count_repetition_guard": "3", + "avg_total_tokens_target_model": "4258619.67", + "total_total_tokens_target_model": "12775859.0", + "avg_total_tokens_other_models": "0.00", + "total_total_tokens_other_models": "0.0", + "avg_total_tokens_all_models": "4258619.67", + "total_total_tokens_all_models": "12775859.0", + "avg_turn_count": "98.33", + "total_turn_count": "295.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5-2025-08-07/b3aa8e91-9d8b-47ca-b05c-99a777adebce_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "ddff9d52692df7f60665f616ad08c6c8e16e0560106d2391bb34d6b6a66112f9", + "total_rows": 3 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5-2025-08-07/b3aa8e91-9d8b-47ca-b05c-99a777adebce_samples.jsonl b/data/terminalbench/openai/gpt-5-2025-08-07/b3aa8e91-9d8b-47ca-b05c-99a777adebce_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..9818f900de0ac552692e30d1e9a4fec170631af2 --- /dev/null +++ b/data/terminalbench/openai/gpt-5-2025-08-07/b3aa8e91-9d8b-47ca-b05c-99a777adebce_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ddff9d52692df7f60665f616ad08c6c8e16e0560106d2391bb34d6b6a66112f9 +size 2592256 diff --git a/data/terminalbench/openai/gpt-5-2025-08-07/cc7e182d-0116-4998-a3ad-5f08815bc798.json b/data/terminalbench/openai/gpt-5-2025-08-07/cc7e182d-0116-4998-a3ad-5f08815bc798.json new file mode 100644 index 0000000000000000000000000000000000000000..991e3eb8e7a74360db08ac87ed25fc28ee3d0bd6 --- /dev/null +++ b/data/terminalbench/openai/gpt-5-2025-08-07/cc7e182d-0116-4998-a3ad-5f08815bc798.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5-2025-08-07/1786167800.204707", + "evaluation_timestamp": "1775646127.0", + "retrieved_timestamp": "1786167800.204707", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5-2025-08-07", + "id": "openai/gpt-5-2025-08-07", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5-2025-08-07/cc7e182d-0116-4998-a3ad-5f08815bc798_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "1b85bc31e471bf509d88210709eec9859ee791083b85ab17d598fa611fb560bf", + "total_rows": 10 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5-2025-08-07/cc7e182d-0116-4998-a3ad-5f08815bc798_samples.jsonl b/data/terminalbench/openai/gpt-5-2025-08-07/cc7e182d-0116-4998-a3ad-5f08815bc798_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..ebdc82578dc02c26db8732129e4b291696862bb6 --- /dev/null +++ b/data/terminalbench/openai/gpt-5-2025-08-07/cc7e182d-0116-4998-a3ad-5f08815bc798_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1b85bc31e471bf509d88210709eec9859ee791083b85ab17d598fa611fb560bf +size 4196690 diff --git a/data/terminalbench/openai/gpt-5-2025-08-07/d94d31f1-f344-423f-a2d9-20da4ddbf198.json b/data/terminalbench/openai/gpt-5-2025-08-07/d94d31f1-f344-423f-a2d9-20da4ddbf198.json new file mode 100644 index 0000000000000000000000000000000000000000..ec28b72b09384d071586e7f79bf34a8729bb6317 --- /dev/null +++ b/data/terminalbench/openai/gpt-5-2025-08-07/d94d31f1-f344-423f-a2d9-20da4ddbf198.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5-2025-08-07/1786167818.918794", + "evaluation_timestamp": "1776417565.0", + "retrieved_timestamp": "1786167818.918794", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5-2025-08-07", + "id": "openai/gpt-5-2025-08-07", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5-2025-08-07/d94d31f1-f344-423f-a2d9-20da4ddbf198_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "9d528eaec897232c111aeeb53f4599fe5d264d2780acfd4be4d1bec96462c879", + "total_rows": 7 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5-2025-08-07/d94d31f1-f344-423f-a2d9-20da4ddbf198_samples.jsonl b/data/terminalbench/openai/gpt-5-2025-08-07/d94d31f1-f344-423f-a2d9-20da4ddbf198_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..4fab786a391e9c8bff942ec37e5babce765c0585 --- /dev/null +++ b/data/terminalbench/openai/gpt-5-2025-08-07/d94d31f1-f344-423f-a2d9-20da4ddbf198_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9d528eaec897232c111aeeb53f4599fe5d264d2780acfd4be4d1bec96462c879 +size 1630039 diff --git a/data/terminalbench/openai/gpt-5-2025-08-07/d9b2b67f-7927-40b1-8b3e-0b81663b8902.json b/data/terminalbench/openai/gpt-5-2025-08-07/d9b2b67f-7927-40b1-8b3e-0b81663b8902.json new file mode 100644 index 0000000000000000000000000000000000000000..1d5c20f509deea840d912859c5ab2aa578e342bd --- /dev/null +++ b/data/terminalbench/openai/gpt-5-2025-08-07/d9b2b67f-7927-40b1-8b3e-0b81663b8902.json @@ -0,0 +1,132 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5-2025-08-07/1786163437.391471", + "evaluation_timestamp": "1777478370.0", + "retrieved_timestamp": "1786163437.391471", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5-2025-08-07", + "id": "openai/gpt-5-2025-08-07", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive+C/ba38d0ac for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 21, + "sample_ids": [ + "raman-fitting", + "bn-fit-modify", + "fix-ocaml-gc", + "git-leak-recovery", + "build-cython-ext", + "pytorch-model-recovery", + "sam-cell-seg", + "model-extraction-relu-logits", + "nginx-request-logging", + "sparql-university", + "large-scale-text-editing", + "sqlite-db-truncate", + "sanitize-git-repo", + "build-pmars", + "rstan-to-pystan", + "sqlite-with-gcov", + "openssl-selfsigned-cert", + "constraints-scheduling", + "dna-insert", + "vulnerable-secret", + "dna-assembly" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1777478370.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.7452380952380953, + "uncertainty": { + "standard_error": { + "value": 0.08436772980351875 + }, + "num_samples": 96 + }, + "details": { + "total_matched_trajectories": "96", + "stopping_reason_count_completed_on_successful_submit": "75", + "stopping_reason_count_repetition_guard": "16", + "stopping_reason_count_token_limit": "5", + "avg_total_tokens_target_model": "1396763.97", + "total_total_tokens_target_model": "134089341.0", + "avg_total_tokens_other_models": "397.83", + "total_total_tokens_other_models": "38192.0", + "avg_total_tokens_all_models": "1397161.80", + "total_total_tokens_all_models": "134127533.0", + "avg_turn_count": "33.40", + "total_turn_count": "3206.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "32", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "32", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5-2025-08-07/d9b2b67f-7927-40b1-8b3e-0b81663b8902_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "1dd6dae8b3fd4e5057688763bda621a5f230a5723d7330c1b9893788169798d8", + "total_rows": 96 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5-2025-08-07/d9b2b67f-7927-40b1-8b3e-0b81663b8902_samples.jsonl b/data/terminalbench/openai/gpt-5-2025-08-07/d9b2b67f-7927-40b1-8b3e-0b81663b8902_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..36a6bbb8e98cad3eac43527d662b769bcab619cd --- /dev/null +++ b/data/terminalbench/openai/gpt-5-2025-08-07/d9b2b67f-7927-40b1-8b3e-0b81663b8902_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1dd6dae8b3fd4e5057688763bda621a5f230a5723d7330c1b9893788169798d8 +size 29573317 diff --git a/data/terminalbench/openai/gpt-5-2025-08-07/dee9cf4e-e5f2-48bc-abee-1a5618d34644.json b/data/terminalbench/openai/gpt-5-2025-08-07/dee9cf4e-e5f2-48bc-abee-1a5618d34644.json new file mode 100644 index 0000000000000000000000000000000000000000..627b80cb8f18f050d5169d477f5737589eb78d9e --- /dev/null +++ b/data/terminalbench/openai/gpt-5-2025-08-07/dee9cf4e-e5f2-48bc-abee-1a5618d34644.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5-2025-08-07/1786167805.277803", + "evaluation_timestamp": "1776433475.0", + "retrieved_timestamp": "1786167805.277803", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5-2025-08-07", + "id": "openai/gpt-5-2025-08-07", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5-2025-08-07/dee9cf4e-e5f2-48bc-abee-1a5618d34644_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "b2df7ab977559d5bf7bb9079fe45a62e3b1d495dc71dab496dd27e5f4c3bd26f", + "total_rows": 10 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5-2025-08-07/dee9cf4e-e5f2-48bc-abee-1a5618d34644_samples.jsonl b/data/terminalbench/openai/gpt-5-2025-08-07/dee9cf4e-e5f2-48bc-abee-1a5618d34644_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..b75f16530ffc2551077e85e0e646030d99c37cbd --- /dev/null +++ b/data/terminalbench/openai/gpt-5-2025-08-07/dee9cf4e-e5f2-48bc-abee-1a5618d34644_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b2df7ab977559d5bf7bb9079fe45a62e3b1d495dc71dab496dd27e5f4c3bd26f +size 5286104 diff --git a/data/terminalbench/openai/gpt-5-2025-08-07/e05b6194-2024-4d44-9de5-0269a0b034c1.json b/data/terminalbench/openai/gpt-5-2025-08-07/e05b6194-2024-4d44-9de5-0269a0b034c1.json new file mode 100644 index 0000000000000000000000000000000000000000..e88b5295262765b62f6ecf0308da4113d00047a9 --- /dev/null +++ b/data/terminalbench/openai/gpt-5-2025-08-07/e05b6194-2024-4d44-9de5-0269a0b034c1.json @@ -0,0 +1,119 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5-2025-08-07/1786163422.900722", + "evaluation_timestamp": "1777059332.0", + "retrieved_timestamp": "1786163422.900722", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5-2025-08-07", + "id": "openai/gpt-5-2025-08-07", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive+C/+3ep/c816f58a for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 8, + "sample_ids": [ + "db-wal-recovery", + "train-fasttext", + "gcode-to-text", + "circuit-fibsqrt", + "protein-assembly", + "financial-document-processor", + "feal-differential-cryptanalysis", + "polyglot-c-py" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1777059332.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.5, + "uncertainty": { + "standard_error": { + "value": 0.16666666666666666 + }, + "num_samples": 23 + }, + "details": { + "total_matched_trajectories": "23", + "stopping_reason_count_completed_on_successful_submit": "11", + "stopping_reason_count_token_limit": "6", + "stopping_reason_count_repetition_guard": "6", + "avg_total_tokens_target_model": "4649287.26", + "total_total_tokens_target_model": "106933607.0", + "avg_total_tokens_other_models": "994.43", + "total_total_tokens_other_models": "22872.0", + "avg_total_tokens_all_models": "4650281.70", + "total_total_tokens_all_models": "106956479.0", + "avg_turn_count": "84.83", + "total_turn_count": "1951.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5-2025-08-07/e05b6194-2024-4d44-9de5-0269a0b034c1_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "f234a541d8f762f65c2184b61424ee2f8762372a88f9e458dd806249ffc6139c", + "total_rows": 23 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5-2025-08-07/e05b6194-2024-4d44-9de5-0269a0b034c1_samples.jsonl b/data/terminalbench/openai/gpt-5-2025-08-07/e05b6194-2024-4d44-9de5-0269a0b034c1_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..56a91b5e93833516b649131debecdcb0cef262e0 --- /dev/null +++ b/data/terminalbench/openai/gpt-5-2025-08-07/e05b6194-2024-4d44-9de5-0269a0b034c1_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f234a541d8f762f65c2184b61424ee2f8762372a88f9e458dd806249ffc6139c +size 25042402 diff --git a/data/terminalbench/openai/gpt-5-2025-08-07/e15888b0-30ef-4996-b854-5a0c8a1a23fe.json b/data/terminalbench/openai/gpt-5-2025-08-07/e15888b0-30ef-4996-b854-5a0c8a1a23fe.json new file mode 100644 index 0000000000000000000000000000000000000000..943e7fd8fbaeb7111f12a2bacb59e6bb459b1758 --- /dev/null +++ b/data/terminalbench/openai/gpt-5-2025-08-07/e15888b0-30ef-4996-b854-5a0c8a1a23fe.json @@ -0,0 +1,120 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5-2025-08-07/1786163407.728069", + "evaluation_timestamp": "1776441651.0", + "retrieved_timestamp": "1786163407.728069", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5-2025-08-07", + "id": "openai/gpt-5-2025-08-07", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive+C/+3ep/8facad4a for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 10, + "sample_ids": [ + "chess-best-move", + "overfull-hbox", + "polyglot-rust-c", + "compile-compcert", + "headless-terminal", + "fix-git", + "schemelike-metacircular-eval", + "qemu-startup", + "build-pov-ray", + "git-multibranch" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1776441651.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.4666666666666666, + "uncertainty": { + "standard_error": { + "value": 0.1507184440694504 + }, + "num_samples": 30 + }, + "details": { + "total_matched_trajectories": "30", + "stopping_reason_count_completed_on_successful_submit": "14", + "stopping_reason_count_token_limit": "5", + "stopping_reason_count_repetition_guard": "8", + "stopping_reason_count_completed_with_submit": "3", + "avg_total_tokens_target_model": "2737001.53", + "total_total_tokens_target_model": "82110046.0", + "avg_total_tokens_other_models": "1488.53", + "total_total_tokens_other_models": "44656.0", + "avg_total_tokens_all_models": "2738490.07", + "total_total_tokens_all_models": "82154702.0", + "avg_turn_count": "53.83", + "total_turn_count": "1615.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5-2025-08-07/e15888b0-30ef-4996-b854-5a0c8a1a23fe_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "dc022bbd505276e8e024b68dc4f79e09d216d314f8ee8f1809b899bde7f3cfb0", + "total_rows": 30 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5-2025-08-07/e15888b0-30ef-4996-b854-5a0c8a1a23fe_samples.jsonl b/data/terminalbench/openai/gpt-5-2025-08-07/e15888b0-30ef-4996-b854-5a0c8a1a23fe_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..8aa8c843543875d08fe86124637913cca7824c27 --- /dev/null +++ b/data/terminalbench/openai/gpt-5-2025-08-07/e15888b0-30ef-4996-b854-5a0c8a1a23fe_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dc022bbd505276e8e024b68dc4f79e09d216d314f8ee8f1809b899bde7f3cfb0 +size 16259908 diff --git a/data/terminalbench/openai/gpt-5-2025-08-07/f3f34002-4fe8-4888-8f9a-ad84dd5617e6.json b/data/terminalbench/openai/gpt-5-2025-08-07/f3f34002-4fe8-4888-8f9a-ad84dd5617e6.json new file mode 100644 index 0000000000000000000000000000000000000000..b7b2b116e246c23ba5000cbc9a006959333043b7 --- /dev/null +++ b/data/terminalbench/openai/gpt-5-2025-08-07/f3f34002-4fe8-4888-8f9a-ad84dd5617e6.json @@ -0,0 +1,119 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5-2025-08-07/1786163386.762868", + "evaluation_timestamp": "1777899366.0", + "retrieved_timestamp": "1786163386.762868", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5-2025-08-07", + "id": "openai/gpt-5-2025-08-07", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive+C/+1ep/49f2397b for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 5, + "sample_ids": [ + "raman-fitting", + "fix-ocaml-gc", + "pytorch-model-recovery", + "model-extraction-relu-logits", + "rstan-to-pystan" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1777899366.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.4, + "uncertainty": { + "standard_error": { + "value": 0.24494897427831783 + }, + "num_samples": 5 + }, + "details": { + "total_matched_trajectories": "5", + "stopping_reason_count_token_limit": "2", + "stopping_reason_count_completed_on_successful_submit": "2", + "stopping_reason_count_repetition_guard": "1", + "avg_total_tokens_target_model": "6257855.80", + "total_total_tokens_target_model": "31289279.0", + "avg_total_tokens_other_models": "357.20", + "total_total_tokens_other_models": "1786.0", + "avg_total_tokens_all_models": "6258213.00", + "total_total_tokens_all_models": "31291065.0", + "avg_turn_count": "93.60", + "total_turn_count": "468.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5-2025-08-07/f3f34002-4fe8-4888-8f9a-ad84dd5617e6_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "c36c4fd26dadcf70fada5f9147e8bd9da353f545ed2fc6114ca4931997263dbc", + "total_rows": 5 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5-2025-08-07/f3f34002-4fe8-4888-8f9a-ad84dd5617e6_samples.jsonl b/data/terminalbench/openai/gpt-5-2025-08-07/f3f34002-4fe8-4888-8f9a-ad84dd5617e6_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..0c977204191d6627e6fcaa8b63e78af5b2d6ec20 --- /dev/null +++ b/data/terminalbench/openai/gpt-5-2025-08-07/f3f34002-4fe8-4888-8f9a-ad84dd5617e6_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c36c4fd26dadcf70fada5f9147e8bd9da353f545ed2fc6114ca4931997263dbc +size 4687983 diff --git a/data/terminalbench/openai/gpt-5-2025-08-07/fa41b9f4-c82c-4a66-9c54-5ef246089c2b.json b/data/terminalbench/openai/gpt-5-2025-08-07/fa41b9f4-c82c-4a66-9c54-5ef246089c2b.json new file mode 100644 index 0000000000000000000000000000000000000000..7e50ea93fbead2d39aff2f30e9744b5988bb7eed --- /dev/null +++ b/data/terminalbench/openai/gpt-5-2025-08-07/fa41b9f4-c82c-4a66-9c54-5ef246089c2b.json @@ -0,0 +1,112 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5-2025-08-07/1786163397.150764", + "evaluation_timestamp": "1775733996.0", + "retrieved_timestamp": "1786163397.150764", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5-2025-08-07", + "id": "openai/gpt-5-2025-08-07", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive+C/+2ep for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 6, + "sample_ids": [ + "gpt2-codegolf", + "llm-inference-batching-scheduler", + "break-filter-js-from-html", + "reshard-c4-data", + "merge-diff-arc-agi-task", + "winning-avg-corewars" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1775733996.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.8333333333333334, + "uncertainty": { + "standard_error": { + "value": 0.16666666666666669 + }, + "num_samples": 12 + }, + "details": { + "total_matched_trajectories": "12", + "stopping_reason_count_completed_on_successful_submit": "10", + "stopping_reason_count_repetition_guard": "2", + "avg_total_tokens_target_model": "938448.50", + "total_total_tokens_target_model": "11261382.0", + "avg_total_tokens_other_models": "2208.33", + "total_total_tokens_other_models": "26500.0", + "avg_total_tokens_all_models": "940656.83", + "total_total_tokens_all_models": "11287882.0", + "avg_turn_count": "31.67", + "total_turn_count": "380.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5-2025-08-07/fa41b9f4-c82c-4a66-9c54-5ef246089c2b_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "973954ec7353c198f14d6b23a53e0467d4d70123213a37661542d20827fe2da8", + "total_rows": 12 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5-2025-08-07/fa41b9f4-c82c-4a66-9c54-5ef246089c2b_samples.jsonl b/data/terminalbench/openai/gpt-5-2025-08-07/fa41b9f4-c82c-4a66-9c54-5ef246089c2b_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..9af9b3c1a98dc465339ad963c7a36ec8f198f59a --- /dev/null +++ b/data/terminalbench/openai/gpt-5-2025-08-07/fa41b9f4-c82c-4a66-9c54-5ef246089c2b_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:973954ec7353c198f14d6b23a53e0467d4d70123213a37661542d20827fe2da8 +size 2949859 diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/00851d8b-4107-415b-a7c0-9b5cca0f2e12.json b/data/terminalbench/openai/gpt-5.2-2025-12-11/00851d8b-4107-415b-a7c0-9b5cca0f2e12.json new file mode 100644 index 0000000000000000000000000000000000000000..4a651943a3ecd81db9c8fa577661b044319a2b13 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/00851d8b-4107-415b-a7c0-9b5cca0f2e12.json @@ -0,0 +1,107 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5.2-2025-12-11/1786165943.933854", + "evaluation_timestamp": "1775604415.0", + "retrieved_timestamp": "1786165943.933854", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.202,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive/+8ep for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 2, + "sample_ids": [ + "gpt2-codegolf", + "break-filter-js-from-html" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1775604415.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.5, + "uncertainty": { + "standard_error": { + "value": 0.5 + }, + "num_samples": 16 + }, + "details": { + "total_matched_trajectories": "6", + "stopping_reason_count_repetition_guard": "6", + "avg_total_tokens_target_model": "249622.83", + "total_total_tokens_target_model": "1497737.0", + "avg_total_tokens_other_models": "1443.17", + "total_total_tokens_other_models": "8659.0", + "avg_total_tokens_all_models": "251066.00", + "total_total_tokens_all_models": "1506396.0", + "avg_turn_count": "20.83", + "total_turn_count": "125.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5.2-2025-12-11/00851d8b-4107-415b-a7c0-9b5cca0f2e12_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "1457b44c6c7caa5334fa01bf8aa644c54a013a69c16aa2c0e5d7bb97db35b2e4", + "total_rows": 16 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/00851d8b-4107-415b-a7c0-9b5cca0f2e12_samples.jsonl b/data/terminalbench/openai/gpt-5.2-2025-12-11/00851d8b-4107-415b-a7c0-9b5cca0f2e12_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..e89a9e609539f23343317fab71ccc49f9687c66b --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/00851d8b-4107-415b-a7c0-9b5cca0f2e12_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1457b44c6c7caa5334fa01bf8aa644c54a013a69c16aa2c0e5d7bb97db35b2e4 +size 1943984 diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/058a3267-a35d-45c5-87ac-df269ba55bc6.json b/data/terminalbench/openai/gpt-5.2-2025-12-11/058a3267-a35d-45c5-87ac-df269ba55bc6.json new file mode 100644 index 0000000000000000000000000000000000000000..34fd9eaa36f095e2a28934da7d6807a74470a773 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/058a3267-a35d-45c5-87ac-df269ba55bc6.json @@ -0,0 +1,107 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5.2-2025-12-11/1786165855.147164", + "evaluation_timestamp": "1775646127.0", + "retrieved_timestamp": "1786165855.147164", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive/+2ep for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 2, + "sample_ids": [ + "gpt2-codegolf", + "break-filter-js-from-html" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1775646127.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.5, + "uncertainty": { + "standard_error": { + "value": 0.5 + }, + "num_samples": 4 + }, + "details": { + "total_matched_trajectories": "4", + "stopping_reason_count_repetition_guard": "4", + "avg_total_tokens_target_model": "284904.25", + "total_total_tokens_target_model": "1139617.0", + "avg_total_tokens_other_models": "1257.75", + "total_total_tokens_other_models": "5031.0", + "avg_total_tokens_all_models": "286162.00", + "total_total_tokens_all_models": "1144648.0", + "avg_turn_count": "24.50", + "total_turn_count": "98.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5.2-2025-12-11/058a3267-a35d-45c5-87ac-df269ba55bc6_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "380f45bb85505b2214abe31c3f6a932e0d6caf0c04e84596aa7a69e77bf037d0", + "total_rows": 4 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/058a3267-a35d-45c5-87ac-df269ba55bc6_samples.jsonl b/data/terminalbench/openai/gpt-5.2-2025-12-11/058a3267-a35d-45c5-87ac-df269ba55bc6_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..00c07a600748dca8e0f56b8c816ffe3f3789b8e4 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/058a3267-a35d-45c5-87ac-df269ba55bc6_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:380f45bb85505b2214abe31c3f6a932e0d6caf0c04e84596aa7a69e77bf037d0 +size 633990 diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/06bf8dce-af04-4d82-afc7-3e6681911b97.json b/data/terminalbench/openai/gpt-5.2-2025-12-11/06bf8dce-af04-4d82-afc7-3e6681911b97.json new file mode 100644 index 0000000000000000000000000000000000000000..1991b6211464c62b11f234550949bb4583969a1f --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/06bf8dce-af04-4d82-afc7-3e6681911b97.json @@ -0,0 +1,112 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5.2-2025-12-11/1786163286.310868", + "evaluation_timestamp": "1777305978.0", + "retrieved_timestamp": "1786163286.310868", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive+C/+1ep/80c65c46 for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 5, + "sample_ids": [ + "db-wal-recovery", + "schemelike-metacircular-eval", + "install-windows-3.11", + "extract-moves-from-video", + "circuit-fibsqrt" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1777305978.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.0, + "uncertainty": { + "num_samples": 3 + }, + "details": { + "total_matched_trajectories": "3", + "stopping_reason_count_repetition_guard": "2", + "stopping_reason_count_token_limit": "1", + "avg_total_tokens_target_model": "5427172.33", + "total_total_tokens_target_model": "16281517.0", + "avg_total_tokens_other_models": "0.00", + "total_total_tokens_other_models": "0.0", + "avg_total_tokens_all_models": "5427172.33", + "total_total_tokens_all_models": "16281517.0", + "avg_turn_count": "83.67", + "total_turn_count": "251.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5.2-2025-12-11/06bf8dce-af04-4d82-afc7-3e6681911b97_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "68b2a91ae221141b7c16c94be2a4e382a51b1a4f96d7806ee0c0e2559af66439", + "total_rows": 3 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/06bf8dce-af04-4d82-afc7-3e6681911b97_samples.jsonl b/data/terminalbench/openai/gpt-5.2-2025-12-11/06bf8dce-af04-4d82-afc7-3e6681911b97_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..11b88da5726befc4938d72fbc71b546e1c36e804 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/06bf8dce-af04-4d82-afc7-3e6681911b97_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68b2a91ae221141b7c16c94be2a4e382a51b1a4f96d7806ee0c0e2559af66439 +size 5227233 diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/07e32d05-7934-41cc-9436-f315ceeba5c6.json b/data/terminalbench/openai/gpt-5.2-2025-12-11/07e32d05-7934-41cc-9436-f315ceeba5c6.json new file mode 100644 index 0000000000000000000000000000000000000000..ca245a4b1dbec5e596e990ef918210210db426f9 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/07e32d05-7934-41cc-9436-f315ceeba5c6.json @@ -0,0 +1,119 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5.2-2025-12-11/1786165931.209953", + "evaluation_timestamp": "1777064074.0", + "retrieved_timestamp": "1786165931.209953", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive/+4ep/ea3ff254 for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 10, + "sample_ids": [ + "winning-avg-corewars", + "mteb-retrieve", + "polyglot-rust-c", + "compile-compcert", + "db-wal-recovery", + "headless-terminal", + "fix-git", + "schemelike-metacircular-eval", + "qemu-startup", + "build-pov-ray" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1777064074.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.5166666666666666, + "uncertainty": { + "standard_error": { + "value": 0.135400640077266 + }, + "num_samples": 33 + }, + "details": { + "total_matched_trajectories": "33", + "stopping_reason_count_repetition_guard": "33", + "avg_total_tokens_target_model": "1841730.27", + "total_total_tokens_target_model": "60777099.0", + "avg_total_tokens_other_models": "1427.61", + "total_total_tokens_other_models": "47111.0", + "avg_total_tokens_all_models": "1843157.88", + "total_total_tokens_all_models": "60824210.0", + "avg_turn_count": "62.33", + "total_turn_count": "2057.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5.2-2025-12-11/07e32d05-7934-41cc-9436-f315ceeba5c6_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "2f7bff26e2489c58b34b4c34b60b8125a5154c09b0630fc06896536a52d49345", + "total_rows": 33 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/07e32d05-7934-41cc-9436-f315ceeba5c6_samples.jsonl b/data/terminalbench/openai/gpt-5.2-2025-12-11/07e32d05-7934-41cc-9436-f315ceeba5c6_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..0d317fc8f480f1cf1ab40ff47e8dd5be5d04c21d --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/07e32d05-7934-41cc-9436-f315ceeba5c6_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2f7bff26e2489c58b34b4c34b60b8125a5154c09b0630fc06896536a52d49345 +size 14092343 diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/12746957-da07-4e94-861b-0c3ac5108a19.json b/data/terminalbench/openai/gpt-5.2-2025-12-11/12746957-da07-4e94-861b-0c3ac5108a19.json new file mode 100644 index 0000000000000000000000000000000000000000..114d61c8f0b2c3f69dd38da67942f78fe3b102e2 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/12746957-da07-4e94-861b-0c3ac5108a19.json @@ -0,0 +1,126 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5.2-2025-12-11/1786163242.067184", + "evaluation_timestamp": "1777025747.0", + "retrieved_timestamp": "1786163242.067184", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive+C/+1ep/eed60920 for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 14, + "sample_ids": [ + "db-wal-recovery", + "train-fasttext", + "install-windows-3.11", + "torch-pipeline-parallelism", + "extract-moves-from-video", + "gcode-to-text", + "make-mips-interpreter", + "count-dataset-tokens", + "circuit-fibsqrt", + "mteb-leaderboard", + "mailman", + "extract-elf", + "protein-assembly", + "code-from-image" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1777025747.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.7, + "uncertainty": { + "standard_error": { + "value": 0.15275252316519464 + }, + "num_samples": 10 + }, + "details": { + "total_matched_trajectories": "10", + "stopping_reason_count_completed_on_successful_submit": "6", + "stopping_reason_count_token_limit": "2", + "stopping_reason_count_completed_with_submit": "1", + "stopping_reason_count_repetition_guard": "1", + "avg_total_tokens_target_model": "3821266.70", + "total_total_tokens_target_model": "38212667.0", + "avg_total_tokens_other_models": "0.00", + "total_total_tokens_other_models": "0.0", + "avg_total_tokens_all_models": "3821266.70", + "total_total_tokens_all_models": "38212667.0", + "avg_turn_count": "74.10", + "total_turn_count": "741.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5.2-2025-12-11/12746957-da07-4e94-861b-0c3ac5108a19_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "bb3272e9dd71aa52263a1cf9715ae5a93fb41cd78c1acdc7ac2c8bcf0f2fd569", + "total_rows": 10 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/12746957-da07-4e94-861b-0c3ac5108a19_samples.jsonl b/data/terminalbench/openai/gpt-5.2-2025-12-11/12746957-da07-4e94-861b-0c3ac5108a19_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..7d76ee7065cd034816e5c465ff5c01672e4a7cd1 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/12746957-da07-4e94-861b-0c3ac5108a19_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bb3272e9dd71aa52263a1cf9715ae5a93fb41cd78c1acdc7ac2c8bcf0f2fd569 +size 6435658 diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/130b183b-84ca-409c-8e21-7a4dd57395db.json b/data/terminalbench/openai/gpt-5.2-2025-12-11/130b183b-84ca-409c-8e21-7a4dd57395db.json new file mode 100644 index 0000000000000000000000000000000000000000..3d38f622d3527cc6842ec7bf586a994e89c59ec1 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/130b183b-84ca-409c-8e21-7a4dd57395db.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5.2-2025-12-11/1786169574.027464", + "evaluation_timestamp": "1777773860.0", + "retrieved_timestamp": "1786169574.027464", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5.2-2025-12-11/130b183b-84ca-409c-8e21-7a4dd57395db_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "dbb32bde7fc16d9f0a83e7b6acbd08401fca0189da61c16f3d591f6fbff2d918", + "total_rows": 48 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/130b183b-84ca-409c-8e21-7a4dd57395db_samples.jsonl b/data/terminalbench/openai/gpt-5.2-2025-12-11/130b183b-84ca-409c-8e21-7a4dd57395db_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..c3363c51b5c410fbbeb6a5745bec40bc1481ee8d --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/130b183b-84ca-409c-8e21-7a4dd57395db_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dbb32bde7fc16d9f0a83e7b6acbd08401fca0189da61c16f3d591f6fbff2d918 +size 21180678 diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/1c374c1a-b5d5-402d-b319-5f9c501541bc.json b/data/terminalbench/openai/gpt-5.2-2025-12-11/1c374c1a-b5d5-402d-b319-5f9c501541bc.json new file mode 100644 index 0000000000000000000000000000000000000000..1f79396f3835455c75778b36fd1fab2111c92fd8 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/1c374c1a-b5d5-402d-b319-5f9c501541bc.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5.2-2025-12-11/1786167685.35886", + "evaluation_timestamp": "1776915105.0", + "retrieved_timestamp": "1786167685.35886", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5.2-2025-12-11/1c374c1a-b5d5-402d-b319-5f9c501541bc_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "d3e45c22a47f621604c5f252a7ad7d7e0e78a2be14c95708931b82dc164f5133", + "total_rows": 3 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/1c374c1a-b5d5-402d-b319-5f9c501541bc_samples.jsonl b/data/terminalbench/openai/gpt-5.2-2025-12-11/1c374c1a-b5d5-402d-b319-5f9c501541bc_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..b9cbb3c436a37f7cd7159f06a48743190f5500a4 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/1c374c1a-b5d5-402d-b319-5f9c501541bc_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d3e45c22a47f621604c5f252a7ad7d7e0e78a2be14c95708931b82dc164f5133 +size 1006485 diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/23618492-9620-4f2f-b599-7c9c2bb64a8f.json b/data/terminalbench/openai/gpt-5.2-2025-12-11/23618492-9620-4f2f-b599-7c9c2bb64a8f.json new file mode 100644 index 0000000000000000000000000000000000000000..9fdbf3e061ca12c281403d84dac4eab78dbbfae0 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/23618492-9620-4f2f-b599-7c9c2bb64a8f.json @@ -0,0 +1,108 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5.2-2025-12-11/1786163280.225762", + "evaluation_timestamp": "1777375735.0", + "retrieved_timestamp": "1786163280.225762", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive+C/+1ep/090282b0 for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 2, + "sample_ids": [ + "install-windows-3.11", + "circuit-fibsqrt" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1777375735.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.0, + "uncertainty": { + "num_samples": 1 + }, + "details": { + "total_matched_trajectories": "1", + "stopping_reason_count_token_limit": "1", + "avg_total_tokens_target_model": "10020650.00", + "total_total_tokens_target_model": "10020650.0", + "avg_total_tokens_other_models": "0.00", + "total_total_tokens_other_models": "0.0", + "avg_total_tokens_all_models": "10020650.00", + "total_total_tokens_all_models": "10020650.0", + "avg_turn_count": "181.00", + "total_turn_count": "181.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5.2-2025-12-11/23618492-9620-4f2f-b599-7c9c2bb64a8f_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "43aae274406eb560db919731455dfc360b5a7e4ad4e9171ef4bcda54eab1bb04", + "total_rows": 1 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/23618492-9620-4f2f-b599-7c9c2bb64a8f_samples.jsonl b/data/terminalbench/openai/gpt-5.2-2025-12-11/23618492-9620-4f2f-b599-7c9c2bb64a8f_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..1b737e7b731ad594b0e7e49fe72d6b1cb6583141 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/23618492-9620-4f2f-b599-7c9c2bb64a8f_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:43aae274406eb560db919731455dfc360b5a7e4ad4e9171ef4bcda54eab1bb04 +size 787839 diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/28aa6438-2327-4fca-96a7-2b5dfbdaf02e.json b/data/terminalbench/openai/gpt-5.2-2025-12-11/28aa6438-2327-4fca-96a7-2b5dfbdaf02e.json new file mode 100644 index 0000000000000000000000000000000000000000..cffb99f2a09410205b1353e383be464d9095c573 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/28aa6438-2327-4fca-96a7-2b5dfbdaf02e.json @@ -0,0 +1,112 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5.2-2025-12-11/1786165865.701376", + "evaluation_timestamp": "1777336645.0", + "retrieved_timestamp": "1786165865.701376", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive/+2ep/528a494b for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 2, + "sample_ids": [ + "polyglot-rust-c", + "schemelike-metacircular-eval" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1777336645.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.25, + "uncertainty": { + "standard_error": { + "value": 0.25 + }, + "num_samples": 4 + }, + "details": { + "total_matched_trajectories": "4", + "stopping_reason_count_repetition_guard": "3", + "stopping_reason_count_token_limit": "1", + "avg_total_tokens_target_model": "4624092.75", + "total_total_tokens_target_model": "18496371.0", + "avg_total_tokens_other_models": "6018.75", + "total_total_tokens_other_models": "24075.0", + "avg_total_tokens_all_models": "4630111.50", + "total_total_tokens_all_models": "18520446.0", + "avg_turn_count": "89.00", + "total_turn_count": "356.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5.2-2025-12-11/28aa6438-2327-4fca-96a7-2b5dfbdaf02e_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "e629292b28af40bb47947410161b27cf5fd18f6fbff2d3dcffb8fcc812759d53", + "total_rows": 4 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/28aa6438-2327-4fca-96a7-2b5dfbdaf02e_samples.jsonl b/data/terminalbench/openai/gpt-5.2-2025-12-11/28aa6438-2327-4fca-96a7-2b5dfbdaf02e_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..ee9c0b956ad2f2b33785f70c52a4859c2e1c391f --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/28aa6438-2327-4fca-96a7-2b5dfbdaf02e_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e629292b28af40bb47947410161b27cf5fd18f6fbff2d3dcffb8fcc812759d53 +size 4506196 diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/2b5738de-5e82-43a3-8841-9e291f386542.json b/data/terminalbench/openai/gpt-5.2-2025-12-11/2b5738de-5e82-43a3-8841-9e291f386542.json new file mode 100644 index 0000000000000000000000000000000000000000..bd51758ef56c953a57e8fdc39cde9c7aa73665f4 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/2b5738de-5e82-43a3-8841-9e291f386542.json @@ -0,0 +1,109 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5.2-2025-12-11/1786165939.770081", + "evaluation_timestamp": "1776359447.0", + "retrieved_timestamp": "1786165939.770081", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive/+4ep/92226305 for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 2, + "sample_ids": [ + "regex-chess", + "modernize-scientific-stack" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1776359447.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.625, + "uncertainty": { + "standard_error": { + "value": 0.37499999999999994 + }, + "num_samples": 8 + }, + "details": { + "total_matched_trajectories": "8", + "stopping_reason_count_repetition_guard": "8", + "avg_total_tokens_target_model": "2686087.88", + "total_total_tokens_target_model": "21488703.0", + "avg_total_tokens_other_models": "447.88", + "total_total_tokens_other_models": "3583.0", + "avg_total_tokens_all_models": "2686535.75", + "total_total_tokens_all_models": "21492286.0", + "avg_turn_count": "60.38", + "total_turn_count": "483.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5.2-2025-12-11/2b5738de-5e82-43a3-8841-9e291f386542_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "0f419f9f23421c689ebadb80405b38a0fcad875c61704f539e70970985ca5956", + "total_rows": 8 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/2b5738de-5e82-43a3-8841-9e291f386542_samples.jsonl b/data/terminalbench/openai/gpt-5.2-2025-12-11/2b5738de-5e82-43a3-8841-9e291f386542_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..7d808ecb70c464b927f66300a0934cb249f2cc5c --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/2b5738de-5e82-43a3-8841-9e291f386542_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0f419f9f23421c689ebadb80405b38a0fcad875c61704f539e70970985ca5956 +size 5211651 diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/33a224c8-f66a-4757-a7d6-616df98dd138.json b/data/terminalbench/openai/gpt-5.2-2025-12-11/33a224c8-f66a-4757-a7d6-616df98dd138.json new file mode 100644 index 0000000000000000000000000000000000000000..0c256e6435c9f224a7b246c97e6c437f2f194ee5 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/33a224c8-f66a-4757-a7d6-616df98dd138.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5.2-2025-12-11/1786169587.195833", + "evaluation_timestamp": "1778237142.0", + "retrieved_timestamp": "1786169587.195833", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5.2-2025-12-11/33a224c8-f66a-4757-a7d6-616df98dd138_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "3be71bcb90731ce708fb885306eddbf766917d153705794fed0153f8e5795744", + "total_rows": 10 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/33a224c8-f66a-4757-a7d6-616df98dd138_samples.jsonl b/data/terminalbench/openai/gpt-5.2-2025-12-11/33a224c8-f66a-4757-a7d6-616df98dd138_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..5c93b235fba22014042a07d385d135043bf2219c --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/33a224c8-f66a-4757-a7d6-616df98dd138_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3be71bcb90731ce708fb885306eddbf766917d153705794fed0153f8e5795744 +size 3660562 diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/34aeb2cf-992f-4783-bcb4-11363412b0b2.json b/data/terminalbench/openai/gpt-5.2-2025-12-11/34aeb2cf-992f-4783-bcb4-11363412b0b2.json new file mode 100644 index 0000000000000000000000000000000000000000..55f8fcb139763e9c6ad38be077ee7e0a137c38ce --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/34aeb2cf-992f-4783-bcb4-11363412b0b2.json @@ -0,0 +1,116 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5.2-2025-12-11/1786165843.692254", + "evaluation_timestamp": "1776433475.0", + "retrieved_timestamp": "1786165843.692254", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive/+1ep/6b2ba410 for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 9, + "sample_ids": [ + "largest-eigenval", + "path-tracing", + "feal-linear-cryptanalysis", + "caffe-cifar-10", + "distribution-search", + "adaptive-rejection-sampler", + "configure-git-webserver", + "crack-7z-hash", + "cobol-modernization" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1776433475.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.8888888888888888, + "uncertainty": { + "standard_error": { + "value": 0.11111111111111112 + }, + "num_samples": 9 + }, + "details": { + "total_matched_trajectories": "9", + "stopping_reason_count_repetition_guard": "9", + "avg_total_tokens_target_model": "1698139.78", + "total_total_tokens_target_model": "15283258.0", + "avg_total_tokens_other_models": "1970.44", + "total_total_tokens_other_models": "17734.0", + "avg_total_tokens_all_models": "1700110.22", + "total_total_tokens_all_models": "15300992.0", + "avg_turn_count": "60.11", + "total_turn_count": "541.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5.2-2025-12-11/34aeb2cf-992f-4783-bcb4-11363412b0b2_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "5b554ff6ddb6a21282b6689f45a6b90b3e152b76233c24d1d307ec1198846957", + "total_rows": 9 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/34aeb2cf-992f-4783-bcb4-11363412b0b2_samples.jsonl b/data/terminalbench/openai/gpt-5.2-2025-12-11/34aeb2cf-992f-4783-bcb4-11363412b0b2_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..d32aa31844fc602e8afc3f99dd061c2d41d89c59 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/34aeb2cf-992f-4783-bcb4-11363412b0b2_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5b554ff6ddb6a21282b6689f45a6b90b3e152b76233c24d1d307ec1198846957 +size 3154514 diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/3542742e-1837-4289-9239-fc571337705a.json b/data/terminalbench/openai/gpt-5.2-2025-12-11/3542742e-1837-4289-9239-fc571337705a.json new file mode 100644 index 0000000000000000000000000000000000000000..3c76f36404080658d9a37d3d78d789bf74042392 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/3542742e-1837-4289-9239-fc571337705a.json @@ -0,0 +1,103 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5.2-2025-12-11/1786163336.121831", + "evaluation_timestamp": "1775733996.0", + "retrieved_timestamp": "1786163336.121831", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive+C/+6ep for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 1, + "sample_ids": [ + "mteb-retrieve" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1775733996.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 1.0, + "uncertainty": { + "num_samples": 6 + }, + "details": { + "total_matched_trajectories": "5", + "stopping_reason_count_completed_on_successful_submit": "5", + "avg_total_tokens_target_model": "170477.60", + "total_total_tokens_target_model": "852388.0", + "avg_total_tokens_other_models": "0.00", + "total_total_tokens_other_models": "0.0", + "avg_total_tokens_all_models": "170477.60", + "total_total_tokens_all_models": "852388.0", + "avg_turn_count": "23.80", + "total_turn_count": "119.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5.2-2025-12-11/3542742e-1837-4289-9239-fc571337705a_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "86e71d1606a1e0c95f77a59096c301637ec87c7a858387e6278fa5fddd246e85", + "total_rows": 6 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/3542742e-1837-4289-9239-fc571337705a_samples.jsonl b/data/terminalbench/openai/gpt-5.2-2025-12-11/3542742e-1837-4289-9239-fc571337705a_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..c417f2af78e5ba4f09b01839bb17a004779c43f2 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/3542742e-1837-4289-9239-fc571337705a_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:86e71d1606a1e0c95f77a59096c301637ec87c7a858387e6278fa5fddd246e85 +size 642704 diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/3566071f-755a-4254-bf30-73aa905cd1a4.json b/data/terminalbench/openai/gpt-5.2-2025-12-11/3566071f-755a-4254-bf30-73aa905cd1a4.json new file mode 100644 index 0000000000000000000000000000000000000000..b7e87116500794d6786601d5997680dc4f4ff976 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/3566071f-755a-4254-bf30-73aa905cd1a4.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5.2-2025-12-11/1786167545.740997", + "evaluation_timestamp": "1776520455.0", + "retrieved_timestamp": "1786167545.740997", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5.2-2025-12-11/3566071f-755a-4254-bf30-73aa905cd1a4_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "6f74e37775666577288f501abe06bd0b25fe3c771f280b14db2cd19c043bcfae", + "total_rows": 40 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/3566071f-755a-4254-bf30-73aa905cd1a4_samples.jsonl b/data/terminalbench/openai/gpt-5.2-2025-12-11/3566071f-755a-4254-bf30-73aa905cd1a4_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..e2aadf8f9f3451405a1155202c504bb4faf2d757 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/3566071f-755a-4254-bf30-73aa905cd1a4_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6f74e37775666577288f501abe06bd0b25fe3c771f280b14db2cd19c043bcfae +size 27676490 diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/356f3654-ec22-4844-b618-79ba76843210.json b/data/terminalbench/openai/gpt-5.2-2025-12-11/356f3654-ec22-4844-b618-79ba76843210.json new file mode 100644 index 0000000000000000000000000000000000000000..7f77a450dbaca1c575fb669afeb6ce6cc9a683b2 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/356f3654-ec22-4844-b618-79ba76843210.json @@ -0,0 +1,107 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5.2-2025-12-11/1786163291.582611", + "evaluation_timestamp": "1777026254.0", + "retrieved_timestamp": "1786163291.582611", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive+C/+2ep/dc014c9a for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 1, + "sample_ids": [ + "schemelike-metacircular-eval" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1777026254.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 1.0, + "uncertainty": { + "num_samples": 1 + }, + "details": { + "total_matched_trajectories": "1", + "stopping_reason_count_repetition_guard": "1", + "avg_total_tokens_target_model": "5560868.00", + "total_total_tokens_target_model": "5560868.0", + "avg_total_tokens_other_models": "0.00", + "total_total_tokens_other_models": "0.0", + "avg_total_tokens_all_models": "5560868.00", + "total_total_tokens_all_models": "5560868.0", + "avg_turn_count": "84.00", + "total_turn_count": "84.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5.2-2025-12-11/356f3654-ec22-4844-b618-79ba76843210_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "4e6203a0c0ba9d1a70d112c31173593b9e82c5a77f408531dbddb77c58c16f16", + "total_rows": 1 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/356f3654-ec22-4844-b618-79ba76843210_samples.jsonl b/data/terminalbench/openai/gpt-5.2-2025-12-11/356f3654-ec22-4844-b618-79ba76843210_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..62fb1983b17fc7e3884a998c3bd5404c85ce298d --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/356f3654-ec22-4844-b618-79ba76843210_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4e6203a0c0ba9d1a70d112c31173593b9e82c5a77f408531dbddb77c58c16f16 +size 1024408 diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/42c34bf8-54fd-4284-869e-667981a6ade8.json b/data/terminalbench/openai/gpt-5.2-2025-12-11/42c34bf8-54fd-4284-869e-667981a6ade8.json new file mode 100644 index 0000000000000000000000000000000000000000..43a38f6d1cfe49601e9769911ab00d7a3351f5db --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/42c34bf8-54fd-4284-869e-667981a6ade8.json @@ -0,0 +1,109 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5.2-2025-12-11/1786163282.497887", + "evaluation_timestamp": "1777656314.0", + "retrieved_timestamp": "1786163282.497887", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive+C/+1ep/21f6c3e0 for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 2, + "sample_ids": [ + "polyglot-rust-c", + "polyglot-c-py" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1777656314.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.0, + "uncertainty": { + "num_samples": 2 + }, + "details": { + "total_matched_trajectories": "2", + "stopping_reason_count_completed_with_submit": "1", + "stopping_reason_count_repetition_guard": "1", + "avg_total_tokens_target_model": "837277.50", + "total_total_tokens_target_model": "1674555.0", + "avg_total_tokens_other_models": "0.00", + "total_total_tokens_other_models": "0.0", + "avg_total_tokens_all_models": "837277.50", + "total_total_tokens_all_models": "1674555.0", + "avg_turn_count": "28.50", + "total_turn_count": "57.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5.2-2025-12-11/42c34bf8-54fd-4284-869e-667981a6ade8_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "0ca571b31a7f64e6015a41c81949226845a8f8084dc40d8d51c909d5b54687ca", + "total_rows": 2 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/42c34bf8-54fd-4284-869e-667981a6ade8_samples.jsonl b/data/terminalbench/openai/gpt-5.2-2025-12-11/42c34bf8-54fd-4284-869e-667981a6ade8_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..d906affb0a3e8b7beaf5df80f1fb23676b97d7c5 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/42c34bf8-54fd-4284-869e-667981a6ade8_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0ca571b31a7f64e6015a41c81949226845a8f8084dc40d8d51c909d5b54687ca +size 2332745 diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/456a9b39-d5bb-443e-96dc-c277ce3bc267.json b/data/terminalbench/openai/gpt-5.2-2025-12-11/456a9b39-d5bb-443e-96dc-c277ce3bc267.json new file mode 100644 index 0000000000000000000000000000000000000000..d84661799605defb7b0103489788a740c96d0881 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/456a9b39-d5bb-443e-96dc-c277ce3bc267.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5.2-2025-12-11/1786167554.113567", + "evaluation_timestamp": "1776820114.0", + "retrieved_timestamp": "1786167554.113567", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5.2-2025-12-11/456a9b39-d5bb-443e-96dc-c277ce3bc267_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "df463d402e7d9a4ea7e7d312b5adfd9afd9b1488f62b61551d10fe47115a79bf", + "total_rows": 15 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/456a9b39-d5bb-443e-96dc-c277ce3bc267_samples.jsonl b/data/terminalbench/openai/gpt-5.2-2025-12-11/456a9b39-d5bb-443e-96dc-c277ce3bc267_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..0eb4eb0b73929f01d0709b309596218749f51ee3 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/456a9b39-d5bb-443e-96dc-c277ce3bc267_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:df463d402e7d9a4ea7e7d312b5adfd9afd9b1488f62b61551d10fe47115a79bf +size 7496922 diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/459d6561-b021-4414-a51f-d8f5dec4d12b.json b/data/terminalbench/openai/gpt-5.2-2025-12-11/459d6561-b021-4414-a51f-d8f5dec4d12b.json new file mode 100644 index 0000000000000000000000000000000000000000..93c79214697028eccfb40461508a78987147bc76 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/459d6561-b021-4414-a51f-d8f5dec4d12b.json @@ -0,0 +1,111 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5.2-2025-12-11/1786163320.806701", + "evaluation_timestamp": "1778275254.0", + "retrieved_timestamp": "1786163320.806701", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive+C/5dd93e71 for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 1, + "sample_ids": [ + "dna-assembly" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1778275254.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.6666666666666666, + "uncertainty": { + "num_samples": 3 + }, + "details": { + "total_matched_trajectories": "3", + "stopping_reason_count_completed_on_successful_submit": "2", + "stopping_reason_count_repetition_guard": "1", + "avg_total_tokens_target_model": "2368045.00", + "total_total_tokens_target_model": "7104135.0", + "avg_total_tokens_other_models": "244.00", + "total_total_tokens_other_models": "732.0", + "avg_total_tokens_all_models": "2368289.00", + "total_total_tokens_all_models": "7104867.0", + "avg_turn_count": "53.00", + "total_turn_count": "159.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5.2-2025-12-11/459d6561-b021-4414-a51f-d8f5dec4d12b_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "bf5e28c0354f17095d474f163f547c063e0e3c27582e880041b1dd01cc91cc10", + "total_rows": 3 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/459d6561-b021-4414-a51f-d8f5dec4d12b_samples.jsonl b/data/terminalbench/openai/gpt-5.2-2025-12-11/459d6561-b021-4414-a51f-d8f5dec4d12b_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..1190efac9aaa48ca731d2a543278006dc628248c --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/459d6561-b021-4414-a51f-d8f5dec4d12b_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf5e28c0354f17095d474f163f547c063e0e3c27582e880041b1dd01cc91cc10 +size 2112084 diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/4c5e1f8e-2d24-4d40-ab10-e7a1eac04601.json b/data/terminalbench/openai/gpt-5.2-2025-12-11/4c5e1f8e-2d24-4d40-ab10-e7a1eac04601.json new file mode 100644 index 0000000000000000000000000000000000000000..4153f89032126e922867b10ac545a1a2e9e47ddc --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/4c5e1f8e-2d24-4d40-ab10-e7a1eac04601.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5.2-2025-12-11/1786167566.361191", + "evaluation_timestamp": "1777405572.0", + "retrieved_timestamp": "1786167566.361191", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5.2-2025-12-11/4c5e1f8e-2d24-4d40-ab10-e7a1eac04601_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "e8778b4bdaede8ff4e5c3c57b77be66bc0272807901b143fd6214935740a2806", + "total_rows": 24 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/4c5e1f8e-2d24-4d40-ab10-e7a1eac04601_samples.jsonl b/data/terminalbench/openai/gpt-5.2-2025-12-11/4c5e1f8e-2d24-4d40-ab10-e7a1eac04601_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..418e6dda8c8fa14871213b48204bef525b2dc490 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/4c5e1f8e-2d24-4d40-ab10-e7a1eac04601_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e8778b4bdaede8ff4e5c3c57b77be66bc0272807901b143fd6214935740a2806 +size 3607697 diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/5046f7c7-f495-4147-828d-d82de9dcfc3e.json b/data/terminalbench/openai/gpt-5.2-2025-12-11/5046f7c7-f495-4147-828d-d82de9dcfc3e.json new file mode 100644 index 0000000000000000000000000000000000000000..ad9c5eb1899d6f4baa7ac8b6d4281cc56be59681 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/5046f7c7-f495-4147-828d-d82de9dcfc3e.json @@ -0,0 +1,110 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5.2-2025-12-11/1786163313.433639", + "evaluation_timestamp": "1777933072.0", + "retrieved_timestamp": "1786163313.433639", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive+C/+3ep/4ce59f10 for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 1, + "sample_ids": [ + "regex-chess" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1777933072.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 1.0, + "uncertainty": { + "num_samples": 1 + }, + "details": { + "total_matched_trajectories": "1", + "stopping_reason_count_completed_on_successful_submit": "1", + "avg_total_tokens_target_model": "4031419.00", + "total_total_tokens_target_model": "4031419.0", + "avg_total_tokens_other_models": "0.00", + "total_total_tokens_other_models": "0.0", + "avg_total_tokens_all_models": "4031419.00", + "total_total_tokens_all_models": "4031419.0", + "avg_turn_count": "52.00", + "total_turn_count": "52.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5.2-2025-12-11/5046f7c7-f495-4147-828d-d82de9dcfc3e_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "272214152d05c48fc197a3e184d18daeb2e33fc98a0ba9093d8cf429bbdda83e", + "total_rows": 1 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/5046f7c7-f495-4147-828d-d82de9dcfc3e_samples.jsonl b/data/terminalbench/openai/gpt-5.2-2025-12-11/5046f7c7-f495-4147-828d-d82de9dcfc3e_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..64ba14b03a8a047ae56053988cc390d07a6be6c7 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/5046f7c7-f495-4147-828d-d82de9dcfc3e_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:272214152d05c48fc197a3e184d18daeb2e33fc98a0ba9093d8cf429bbdda83e +size 627174 diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/56a77f6e-0c3d-4336-ae52-82bd4f35c1b1.json b/data/terminalbench/openai/gpt-5.2-2025-12-11/56a77f6e-0c3d-4336-ae52-82bd4f35c1b1.json new file mode 100644 index 0000000000000000000000000000000000000000..76f87d14458fb6d0b88ebe76622059a2dabbe2c6 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/56a77f6e-0c3d-4336-ae52-82bd4f35c1b1.json @@ -0,0 +1,113 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5.2-2025-12-11/1786165837.442543", + "evaluation_timestamp": "1777292858.0", + "retrieved_timestamp": "1786165837.442543", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive/+1ep/153ad3ad for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 3, + "sample_ids": [ + "db-wal-recovery", + "qemu-startup", + "build-pov-ray" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1777292858.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.6666666666666666, + "uncertainty": { + "standard_error": { + "value": 0.33333333333333337 + }, + "num_samples": 3 + }, + "details": { + "total_matched_trajectories": "3", + "stopping_reason_count_repetition_guard": "2", + "stopping_reason_count_token_limit": "1", + "avg_total_tokens_target_model": "4642677.00", + "total_total_tokens_target_model": "13928031.0", + "avg_total_tokens_other_models": "393.33", + "total_total_tokens_other_models": "1180.0", + "avg_total_tokens_all_models": "4643070.33", + "total_total_tokens_all_models": "13929211.0", + "avg_turn_count": "127.67", + "total_turn_count": "383.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5.2-2025-12-11/56a77f6e-0c3d-4336-ae52-82bd4f35c1b1_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "9c2414f478276641ea1869125f6294b38516e5cc7b7babc74053261ab118c650", + "total_rows": 3 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/56a77f6e-0c3d-4336-ae52-82bd4f35c1b1_samples.jsonl b/data/terminalbench/openai/gpt-5.2-2025-12-11/56a77f6e-0c3d-4336-ae52-82bd4f35c1b1_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..52fe0c8fc1dcffb8877411c9d0c55296dc7bbc0f --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/56a77f6e-0c3d-4336-ae52-82bd4f35c1b1_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9c2414f478276641ea1869125f6294b38516e5cc7b7babc74053261ab118c650 +size 2139391 diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/5bea36d5-00f3-46be-bad3-d1b94b71f44e.json b/data/terminalbench/openai/gpt-5.2-2025-12-11/5bea36d5-00f3-46be-bad3-d1b94b71f44e.json new file mode 100644 index 0000000000000000000000000000000000000000..2f35843d3e075077df0bb4707af0dbee7ca7e616 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/5bea36d5-00f3-46be-bad3-d1b94b71f44e.json @@ -0,0 +1,107 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5.2-2025-12-11/1786163311.580521", + "evaluation_timestamp": "1777479198.0", + "retrieved_timestamp": "1786163311.580521", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive+C/+3ep/038c4ae3 for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 1, + "sample_ids": [ + "regex-log" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1777479198.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 1.0, + "uncertainty": { + "num_samples": 3 + }, + "details": { + "total_matched_trajectories": "3", + "stopping_reason_count_completed_on_successful_submit": "3", + "avg_total_tokens_target_model": "20149.67", + "total_total_tokens_target_model": "60449.0", + "avg_total_tokens_other_models": "0.00", + "total_total_tokens_other_models": "0.0", + "avg_total_tokens_all_models": "20149.67", + "total_total_tokens_all_models": "60449.0", + "avg_turn_count": "4.33", + "total_turn_count": "13.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "32", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "32", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5.2-2025-12-11/5bea36d5-00f3-46be-bad3-d1b94b71f44e_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "967929d7f0678c364cc0ddace3ff72ec87638fc5296e55d3273143addac1d52d", + "total_rows": 3 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/5bea36d5-00f3-46be-bad3-d1b94b71f44e_samples.jsonl b/data/terminalbench/openai/gpt-5.2-2025-12-11/5bea36d5-00f3-46be-bad3-d1b94b71f44e_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..13e716fac4308de5a8966da2a20ca72309d39580 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/5bea36d5-00f3-46be-bad3-d1b94b71f44e_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:967929d7f0678c364cc0ddace3ff72ec87638fc5296e55d3273143addac1d52d +size 93026 diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/68736bf0-2914-4fab-aaa5-c81200c8d2a5.json b/data/terminalbench/openai/gpt-5.2-2025-12-11/68736bf0-2914-4fab-aaa5-c81200c8d2a5.json new file mode 100644 index 0000000000000000000000000000000000000000..5cddde3b1173281a8e990f108875b596a812effc --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/68736bf0-2914-4fab-aaa5-c81200c8d2a5.json @@ -0,0 +1,105 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5.2-2025-12-11/1786165849.592497", + "evaluation_timestamp": "1776433475.0", + "retrieved_timestamp": "1786165849.592497", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive/+2ep/e4058c62 for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 1, + "sample_ids": [ + "overfull-hbox" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1776433475.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.5, + "uncertainty": { + "num_samples": 2 + }, + "details": { + "total_matched_trajectories": "2", + "stopping_reason_count_repetition_guard": "2", + "avg_total_tokens_target_model": "686027.00", + "total_total_tokens_target_model": "1372054.0", + "avg_total_tokens_other_models": "287.00", + "total_total_tokens_other_models": "574.0", + "avg_total_tokens_all_models": "686314.00", + "total_total_tokens_all_models": "1372628.0", + "avg_turn_count": "38.00", + "total_turn_count": "76.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5.2-2025-12-11/68736bf0-2914-4fab-aaa5-c81200c8d2a5_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "daf1181303a386a10b5cbf1221a3f93aaf5366d324582c048d0f994fce9fe327", + "total_rows": 2 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/68736bf0-2914-4fab-aaa5-c81200c8d2a5_samples.jsonl b/data/terminalbench/openai/gpt-5.2-2025-12-11/68736bf0-2914-4fab-aaa5-c81200c8d2a5_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..36e0c77c55465402cc15618c8c400aea364bd5a5 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/68736bf0-2914-4fab-aaa5-c81200c8d2a5_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:daf1181303a386a10b5cbf1221a3f93aaf5366d324582c048d0f994fce9fe327 +size 593582 diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/701d089a-3b5c-4a2e-a41b-1511edd9072d.json b/data/terminalbench/openai/gpt-5.2-2025-12-11/701d089a-3b5c-4a2e-a41b-1511edd9072d.json new file mode 100644 index 0000000000000000000000000000000000000000..fe60c1c61f3597822ce842b083ce41cbde78bb86 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/701d089a-3b5c-4a2e-a41b-1511edd9072d.json @@ -0,0 +1,149 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5.2-2025-12-11/1786165906.883267", + "evaluation_timestamp": "1778067146.0", + "retrieved_timestamp": "1786165906.883267", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive/+3ep/34de3b1f for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 36, + "sample_ids": [ + "train-fasttext", + "torch-pipeline-parallelism", + "extract-moves-from-video", + "gcode-to-text", + "make-mips-interpreter", + "count-dataset-tokens", + "mteb-leaderboard", + "mailman", + "raman-fitting", + "extract-elf", + "protein-assembly", + "code-from-image", + "financial-document-processor", + "regex-log", + "feal-differential-cryptanalysis", + "polyglot-c-py", + "cancel-async-tasks", + "bn-fit-modify", + "fix-ocaml-gc", + "git-leak-recovery", + "build-cython-ext", + "pytorch-model-recovery", + "sam-cell-seg", + "model-extraction-relu-logits", + "nginx-request-logging", + "sparql-university", + "large-scale-text-editing", + "sqlite-db-truncate", + "sanitize-git-repo", + "build-pmars", + "rstan-to-pystan", + "sqlite-with-gcov", + "openssl-selfsigned-cert", + "constraints-scheduling", + "vulnerable-secret", + "dna-assembly" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1778067146.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.6157407407407408, + "uncertainty": { + "standard_error": { + "value": 0.07418559726433295 + }, + "num_samples": 105 + }, + "details": { + "total_matched_trajectories": "105", + "stopping_reason_count_repetition_guard": "92", + "stopping_reason_count_token_limit": "13", + "avg_total_tokens_target_model": "2838488.17", + "total_total_tokens_target_model": "298041258.0", + "avg_total_tokens_other_models": "1306.94", + "total_total_tokens_other_models": "137229.0", + "avg_total_tokens_all_models": "2839795.11", + "total_total_tokens_all_models": "298178487.0", + "avg_turn_count": "69.52", + "total_turn_count": "7300.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5.2-2025-12-11/701d089a-3b5c-4a2e-a41b-1511edd9072d_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "64da121d4f1b05517f51d08864d09d81439e2913dd99952c6d049c7e4361d040", + "total_rows": 105 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/701d089a-3b5c-4a2e-a41b-1511edd9072d_samples.jsonl b/data/terminalbench/openai/gpt-5.2-2025-12-11/701d089a-3b5c-4a2e-a41b-1511edd9072d_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..b74412382db338859f82e100ea462c0c5e90698f --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/701d089a-3b5c-4a2e-a41b-1511edd9072d_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:64da121d4f1b05517f51d08864d09d81439e2913dd99952c6d049c7e4361d040 +size 56909899 diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/717f4103-da5b-4c60-999c-0facc8c60048.json b/data/terminalbench/openai/gpt-5.2-2025-12-11/717f4103-da5b-4c60-999c-0facc8c60048.json new file mode 100644 index 0000000000000000000000000000000000000000..4e1727055c6346d46fe0ead2b6cd74f107685dcc --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/717f4103-da5b-4c60-999c-0facc8c60048.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5.2-2025-12-11/1786167518.469567", + "evaluation_timestamp": "1776601454.0", + "retrieved_timestamp": "1786167518.469567", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5.2-2025-12-11/717f4103-da5b-4c60-999c-0facc8c60048_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "e4d7eeb3eeffb9bf523c16c8f740d8e2eced2864399ccf861973f86a8de3e536", + "total_rows": 10 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/717f4103-da5b-4c60-999c-0facc8c60048_samples.jsonl b/data/terminalbench/openai/gpt-5.2-2025-12-11/717f4103-da5b-4c60-999c-0facc8c60048_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..ee0c257c6391063e29839e9c5f070c4d8a2f0df8 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/717f4103-da5b-4c60-999c-0facc8c60048_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e4d7eeb3eeffb9bf523c16c8f740d8e2eced2864399ccf861973f86a8de3e536 +size 2868966 diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/728837e4-89f4-4bd1-83bf-17af32d9ddeb.json b/data/terminalbench/openai/gpt-5.2-2025-12-11/728837e4-89f4-4bd1-83bf-17af32d9ddeb.json new file mode 100644 index 0000000000000000000000000000000000000000..7141c5712e7028533e70c3bd856f22629d902c99 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/728837e4-89f4-4bd1-83bf-17af32d9ddeb.json @@ -0,0 +1,107 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5.2-2025-12-11/1786163298.619223", + "evaluation_timestamp": "1777591126.0", + "retrieved_timestamp": "1786163298.619223", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive+C/+2ep/ff75906e for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 1, + "sample_ids": [ + "polyglot-rust-c" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1777591126.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.0, + "uncertainty": { + "num_samples": 1 + }, + "details": { + "total_matched_trajectories": "1", + "stopping_reason_count_repetition_guard": "1", + "avg_total_tokens_target_model": "951980.00", + "total_total_tokens_target_model": "951980.0", + "avg_total_tokens_other_models": "0.00", + "total_total_tokens_other_models": "0.0", + "avg_total_tokens_all_models": "951980.00", + "total_total_tokens_all_models": "951980.0", + "avg_turn_count": "33.00", + "total_turn_count": "33.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5.2-2025-12-11/728837e4-89f4-4bd1-83bf-17af32d9ddeb_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "1557d12b0510a2f6d37c014176c7184106c1672e84ffb15618c2fbbc4db4bd31", + "total_rows": 1 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/728837e4-89f4-4bd1-83bf-17af32d9ddeb_samples.jsonl b/data/terminalbench/openai/gpt-5.2-2025-12-11/728837e4-89f4-4bd1-83bf-17af32d9ddeb_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..d62fa7647399e6a796d3c6542e1db551fb48398a --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/728837e4-89f4-4bd1-83bf-17af32d9ddeb_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1557d12b0510a2f6d37c014176c7184106c1672e84ffb15618c2fbbc4db4bd31 +size 1283010 diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/7f537bc5-4574-4533-9584-fc31b7b25f3b.json b/data/terminalbench/openai/gpt-5.2-2025-12-11/7f537bc5-4574-4533-9584-fc31b7b25f3b.json new file mode 100644 index 0000000000000000000000000000000000000000..6825cfc9d8d99f375ca4d5ac624297859888de07 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/7f537bc5-4574-4533-9584-fc31b7b25f3b.json @@ -0,0 +1,108 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5.2-2025-12-11/1786163305.559068", + "evaluation_timestamp": "1777042005.0", + "retrieved_timestamp": "1786163305.559068", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive+C/+3ep/f552c61b for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 1, + "sample_ids": [ + "raman-fitting" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1777042005.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.0, + "uncertainty": { + "num_samples": 3 + }, + "details": { + "total_matched_trajectories": "3", + "stopping_reason_count_repetition_guard": "2", + "stopping_reason_count_token_limit": "1", + "avg_total_tokens_target_model": "9444174.00", + "total_total_tokens_target_model": "28332522.0", + "avg_total_tokens_other_models": "0.00", + "total_total_tokens_other_models": "0.0", + "avg_total_tokens_all_models": "9444174.00", + "total_total_tokens_all_models": "28332522.0", + "avg_turn_count": "155.67", + "total_turn_count": "467.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5.2-2025-12-11/7f537bc5-4574-4533-9584-fc31b7b25f3b_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "f522bc37fdbe47bb1623be71c1ef888e26cafd81e0c83c5bce8d46d4ebb1f91b", + "total_rows": 3 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/7f537bc5-4574-4533-9584-fc31b7b25f3b_samples.jsonl b/data/terminalbench/openai/gpt-5.2-2025-12-11/7f537bc5-4574-4533-9584-fc31b7b25f3b_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..b0da778144cf31343f09f55485de168d6231ad26 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/7f537bc5-4574-4533-9584-fc31b7b25f3b_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f522bc37fdbe47bb1623be71c1ef888e26cafd81e0c83c5bce8d46d4ebb1f91b +size 3859232 diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/813ca58e-b129-483a-8bd2-48ee257387eb.json b/data/terminalbench/openai/gpt-5.2-2025-12-11/813ca58e-b129-483a-8bd2-48ee257387eb.json new file mode 100644 index 0000000000000000000000000000000000000000..35fdd82a5aa0fed280f8f39b6b4ca09d22e54781 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/813ca58e-b129-483a-8bd2-48ee257387eb.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5.2-2025-12-11/1786167702.811763", + "evaluation_timestamp": "1777544308.0", + "retrieved_timestamp": "1786167702.811763", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5.2-2025-12-11/813ca58e-b129-483a-8bd2-48ee257387eb_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "5c3227dba9813ca127f9106ff897f66afc35b612c8ca6f9233754b41c46eade1", + "total_rows": 6 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/813ca58e-b129-483a-8bd2-48ee257387eb_samples.jsonl b/data/terminalbench/openai/gpt-5.2-2025-12-11/813ca58e-b129-483a-8bd2-48ee257387eb_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..29e863168bce5aa97a2fb5c7d180f4789eb25d92 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/813ca58e-b129-483a-8bd2-48ee257387eb_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5c3227dba9813ca127f9106ff897f66afc35b612c8ca6f9233754b41c46eade1 +size 4134574 diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/84d3addc-c3c2-4e7c-b745-eba5104888e4.json b/data/terminalbench/openai/gpt-5.2-2025-12-11/84d3addc-c3c2-4e7c-b745-eba5104888e4.json new file mode 100644 index 0000000000000000000000000000000000000000..e7c39dd8f8b7b6f6cb15a95f81cdefed61b13328 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/84d3addc-c3c2-4e7c-b745-eba5104888e4.json @@ -0,0 +1,107 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5.2-2025-12-11/1786163238.710667", + "evaluation_timestamp": "1777478370.0", + "retrieved_timestamp": "1786163238.710667", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive+C/+1ep/347739d1 for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 1, + "sample_ids": [ + "circuit-fibsqrt" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1777478370.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 1.0, + "uncertainty": { + "num_samples": 1 + }, + "details": { + "total_matched_trajectories": "1", + "stopping_reason_count_completed_on_successful_submit": "1", + "avg_total_tokens_target_model": "355910.00", + "total_total_tokens_target_model": "355910.0", + "avg_total_tokens_other_models": "0.00", + "total_total_tokens_other_models": "0.0", + "avg_total_tokens_all_models": "355910.00", + "total_total_tokens_all_models": "355910.0", + "avg_turn_count": "14.00", + "total_turn_count": "14.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "32", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "32", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5.2-2025-12-11/84d3addc-c3c2-4e7c-b745-eba5104888e4_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "a030048e2363061d279518547f8214ccfbfadd3ac38ef1304fb348a95b8e5d67", + "total_rows": 1 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/84d3addc-c3c2-4e7c-b745-eba5104888e4_samples.jsonl b/data/terminalbench/openai/gpt-5.2-2025-12-11/84d3addc-c3c2-4e7c-b745-eba5104888e4_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..8f05ba0890228c906efcab4d0a40f6278b076caa --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/84d3addc-c3c2-4e7c-b745-eba5104888e4_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a030048e2363061d279518547f8214ccfbfadd3ac38ef1304fb348a95b8e5d67 +size 212143 diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/84eb4274-c753-46bd-9937-d33614c47da1.json b/data/terminalbench/openai/gpt-5.2-2025-12-11/84eb4274-c753-46bd-9937-d33614c47da1.json new file mode 100644 index 0000000000000000000000000000000000000000..583110a400f619352e7781bb274c56e40ffb39ba --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/84eb4274-c753-46bd-9937-d33614c47da1.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5.2-2025-12-11/1786169551.62213", + "evaluation_timestamp": "1775558526.0", + "retrieved_timestamp": "1786169551.62213", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.202,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5.2-2025-12-11/84eb4274-c753-46bd-9937-d33614c47da1_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "7197f3246da86ad9789dc5d09ab63e1945689faf6c685110ebc4d9f4b5147e58", + "total_rows": 30 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/84eb4274-c753-46bd-9937-d33614c47da1_samples.jsonl b/data/terminalbench/openai/gpt-5.2-2025-12-11/84eb4274-c753-46bd-9937-d33614c47da1_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..4323486a3dba20a15e3b4e303c8ec543e6d7074d --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/84eb4274-c753-46bd-9937-d33614c47da1_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7197f3246da86ad9789dc5d09ab63e1945689faf6c685110ebc4d9f4b5147e58 +size 9882383 diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/899cf972-5ebe-453b-a20a-70960d687ba5.json b/data/terminalbench/openai/gpt-5.2-2025-12-11/899cf972-5ebe-453b-a20a-70960d687ba5.json new file mode 100644 index 0000000000000000000000000000000000000000..fb790029e7b78dfa8c3fc1a3576bd2c698f39838 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/899cf972-5ebe-453b-a20a-70960d687ba5.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5.2-2025-12-11/1786169582.525473", + "evaluation_timestamp": "1776359447.0", + "retrieved_timestamp": "1786169582.525473", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5.2-2025-12-11/899cf972-5ebe-453b-a20a-70960d687ba5_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "67c52602684af11fb01617655e44af4240200c4d319e239e414396b0a20b0f56", + "total_rows": 52 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/899cf972-5ebe-453b-a20a-70960d687ba5_samples.jsonl b/data/terminalbench/openai/gpt-5.2-2025-12-11/899cf972-5ebe-453b-a20a-70960d687ba5_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..63f2526de3ce7892539129c6a6c2189a5b8cf32a --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/899cf972-5ebe-453b-a20a-70960d687ba5_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:67c52602684af11fb01617655e44af4240200c4d319e239e414396b0a20b0f56 +size 19973076 diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/8b583422-f923-4bef-88f0-74246fea0006.json b/data/terminalbench/openai/gpt-5.2-2025-12-11/8b583422-f923-4bef-88f0-74246fea0006.json new file mode 100644 index 0000000000000000000000000000000000000000..3a7e0872177b2fb8f66528491e2ede3fb15ef9de --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/8b583422-f923-4bef-88f0-74246fea0006.json @@ -0,0 +1,111 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5.2-2025-12-11/1786163302.743854", + "evaluation_timestamp": "1778532794.0", + "retrieved_timestamp": "1786163302.743854", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive+C/+2ep/cbfb50a6 for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 2, + "sample_ids": [ + "regex-chess", + "dna-assembly" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1778532794.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.0, + "uncertainty": { + "num_samples": 2 + }, + "details": { + "total_matched_trajectories": "2", + "stopping_reason_count_repetition_guard": "2", + "avg_total_tokens_target_model": "2738893.50", + "total_total_tokens_target_model": "5477787.0", + "avg_total_tokens_other_models": "726.50", + "total_total_tokens_other_models": "1453.0", + "avg_total_tokens_all_models": "2739620.00", + "total_total_tokens_all_models": "5479240.0", + "avg_turn_count": "66.00", + "total_turn_count": "132.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5.2-2025-12-11/8b583422-f923-4bef-88f0-74246fea0006_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "a687bbc26ee84f6d894e90186c9dde6e3b518b03aef7c06e70ac0c00ead1ecff", + "total_rows": 2 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/8b583422-f923-4bef-88f0-74246fea0006_samples.jsonl b/data/terminalbench/openai/gpt-5.2-2025-12-11/8b583422-f923-4bef-88f0-74246fea0006_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..d73161ad92541a87b6e6c2b20e955973ac648c12 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/8b583422-f923-4bef-88f0-74246fea0006_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a687bbc26ee84f6d894e90186c9dde6e3b518b03aef7c06e70ac0c00ead1ecff +size 1762076 diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/919e5d73-2559-4a05-8ef2-2e958e1999b8.json b/data/terminalbench/openai/gpt-5.2-2025-12-11/919e5d73-2559-4a05-8ef2-2e958e1999b8.json new file mode 100644 index 0000000000000000000000000000000000000000..f5df3ef79ef818dfdd51c3848929aae1376c0fe5 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/919e5d73-2559-4a05-8ef2-2e958e1999b8.json @@ -0,0 +1,110 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5.2-2025-12-11/1786165851.713069", + "evaluation_timestamp": "1778268121.0", + "retrieved_timestamp": "1786165851.713069", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive/+2ep/4b5ebcb3 for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 1, + "sample_ids": [ + "dna-insert" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1778268121.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.0, + "uncertainty": { + "num_samples": 2 + }, + "details": { + "total_matched_trajectories": "2", + "stopping_reason_count_repetition_guard": "2", + "avg_total_tokens_target_model": "742446.50", + "total_total_tokens_target_model": "1484893.0", + "avg_total_tokens_other_models": "515.00", + "total_total_tokens_other_models": "1030.0", + "avg_total_tokens_all_models": "742961.50", + "total_total_tokens_all_models": "1485923.0", + "avg_turn_count": "45.00", + "total_turn_count": "90.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5.2-2025-12-11/919e5d73-2559-4a05-8ef2-2e958e1999b8_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "945345f863f0ef0211795f441df00b0c6d35d54cf546d1138e3f24efa27fc62c", + "total_rows": 2 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/919e5d73-2559-4a05-8ef2-2e958e1999b8_samples.jsonl b/data/terminalbench/openai/gpt-5.2-2025-12-11/919e5d73-2559-4a05-8ef2-2e958e1999b8_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..917caa665cf6a6632131ee7af863570ea16d9b01 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/919e5d73-2559-4a05-8ef2-2e958e1999b8_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:945345f863f0ef0211795f441df00b0c6d35d54cf546d1138e3f24efa27fc62c +size 689908 diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/9c29263f-cb46-4462-8894-e641b1d521e1.json b/data/terminalbench/openai/gpt-5.2-2025-12-11/9c29263f-cb46-4462-8894-e641b1d521e1.json new file mode 100644 index 0000000000000000000000000000000000000000..b82c66183eaf365d58e6b1f4b972c2a597e6419d --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/9c29263f-cb46-4462-8894-e641b1d521e1.json @@ -0,0 +1,109 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5.2-2025-12-11/1786163246.204185", + "evaluation_timestamp": "1776880341.0", + "retrieved_timestamp": "1786163246.204185", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive+C/+1ep/a2aacf5a for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 2, + "sample_ids": [ + "qemu-alpine-ssh", + "tune-mjcf" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1776880341.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 1.0, + "uncertainty": { + "num_samples": 2 + }, + "details": { + "total_matched_trajectories": "2", + "stopping_reason_count_repetition_guard": "1", + "stopping_reason_count_completed_on_successful_submit": "1", + "avg_total_tokens_target_model": "4609698.50", + "total_total_tokens_target_model": "9219397.0", + "avg_total_tokens_other_models": "0.00", + "total_total_tokens_other_models": "0.0", + "avg_total_tokens_all_models": "4609698.50", + "total_total_tokens_all_models": "9219397.0", + "avg_turn_count": "94.50", + "total_turn_count": "189.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5.2-2025-12-11/9c29263f-cb46-4462-8894-e641b1d521e1_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "1ae47e8aa1e64dbfafd883a720efbed7e94bfe52b0bc6d0f01a61d1d959ebb04", + "total_rows": 2 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/9c29263f-cb46-4462-8894-e641b1d521e1_samples.jsonl b/data/terminalbench/openai/gpt-5.2-2025-12-11/9c29263f-cb46-4462-8894-e641b1d521e1_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..6dfd2a57b5d82a5f631e2616b23663f031c0caba --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/9c29263f-cb46-4462-8894-e641b1d521e1_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1ae47e8aa1e64dbfafd883a720efbed7e94bfe52b0bc6d0f01a61d1d959ebb04 +size 1182466 diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/9d84e69b-b778-4261-a593-c5d9550d6093.json b/data/terminalbench/openai/gpt-5.2-2025-12-11/9d84e69b-b778-4261-a593-c5d9550d6093.json new file mode 100644 index 0000000000000000000000000000000000000000..6656c48a9b1d866d1f0fb515c1eea251f3ee639e --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/9d84e69b-b778-4261-a593-c5d9550d6093.json @@ -0,0 +1,110 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5.2-2025-12-11/1786163244.245718", + "evaluation_timestamp": "1777899366.0", + "retrieved_timestamp": "1786163244.245718", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive+C/+1ep/4b5ebcb3 for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 1, + "sample_ids": [ + "dna-insert" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1777899366.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 1.0, + "uncertainty": { + "num_samples": 1 + }, + "details": { + "total_matched_trajectories": "1", + "stopping_reason_count_completed_on_successful_submit": "1", + "avg_total_tokens_target_model": "1030021.00", + "total_total_tokens_target_model": "1030021.0", + "avg_total_tokens_other_models": "0.00", + "total_total_tokens_other_models": "0.0", + "avg_total_tokens_all_models": "1030021.00", + "total_total_tokens_all_models": "1030021.0", + "avg_turn_count": "49.00", + "total_turn_count": "49.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5.2-2025-12-11/9d84e69b-b778-4261-a593-c5d9550d6093_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "c27d498885be2b787660a2acf83786552a9c47b4a27e82495a3c24f66267bd90", + "total_rows": 1 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/9d84e69b-b778-4261-a593-c5d9550d6093_samples.jsonl b/data/terminalbench/openai/gpt-5.2-2025-12-11/9d84e69b-b778-4261-a593-c5d9550d6093_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..e3afbc5a740c40fdc43e69cc682b59bfffae08d0 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/9d84e69b-b778-4261-a593-c5d9550d6093_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c27d498885be2b787660a2acf83786552a9c47b4a27e82495a3c24f66267bd90 +size 362146 diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/a59d1226-d931-49eb-a269-14eac2350366.json b/data/terminalbench/openai/gpt-5.2-2025-12-11/a59d1226-d931-49eb-a269-14eac2350366.json new file mode 100644 index 0000000000000000000000000000000000000000..2e17b9dd0f74244bdc706d695f690aae15ea29cc --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/a59d1226-d931-49eb-a269-14eac2350366.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5.2-2025-12-11/1786167722.995791", + "evaluation_timestamp": "1775078458.0", + "retrieved_timestamp": "1786167722.995791", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.202,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5.2-2025-12-11/a59d1226-d931-49eb-a269-14eac2350366_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "f7b3a67e709a52a8634490544d45d206cad7d85e96f5c53ecd0cc618103a10e8", + "total_rows": 20 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/a59d1226-d931-49eb-a269-14eac2350366_samples.jsonl b/data/terminalbench/openai/gpt-5.2-2025-12-11/a59d1226-d931-49eb-a269-14eac2350366_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..63daa898cf3f740109d21cf049ac81486fcc2fa2 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/a59d1226-d931-49eb-a269-14eac2350366_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f7b3a67e709a52a8634490544d45d206cad7d85e96f5c53ecd0cc618103a10e8 +size 10046829 diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/a86d9a33-c73d-4136-a909-fc6fc9c0ddd2.json b/data/terminalbench/openai/gpt-5.2-2025-12-11/a86d9a33-c73d-4136-a909-fc6fc9c0ddd2.json new file mode 100644 index 0000000000000000000000000000000000000000..0cc114a254d0911b90a78a9b3844012b67f48ff3 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/a86d9a33-c73d-4136-a909-fc6fc9c0ddd2.json @@ -0,0 +1,123 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5.2-2025-12-11/1786163295.961888", + "evaluation_timestamp": "1776810534.0", + "retrieved_timestamp": "1786163295.961888", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive+C/+2ep/3a452b80 for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 11, + "sample_ids": [ + "overfull-hbox", + "compile-compcert", + "headless-terminal", + "fix-git", + "qemu-startup", + "build-pov-ray", + "git-multibranch", + "video-processing", + "qemu-alpine-ssh", + "kv-store-grpc", + "make-doom-for-mips" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1776810534.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.55, + "uncertainty": { + "standard_error": { + "value": 0.15723301886761007 + }, + "num_samples": 19 + }, + "details": { + "total_matched_trajectories": "19", + "stopping_reason_count_completed_on_successful_submit": "10", + "stopping_reason_count_token_limit": "6", + "stopping_reason_count_completed_with_submit": "2", + "stopping_reason_count_repetition_guard": "1", + "avg_total_tokens_target_model": "3582172.26", + "total_total_tokens_target_model": "68061273.0", + "avg_total_tokens_other_models": "80.05", + "total_total_tokens_other_models": "1521.0", + "avg_total_tokens_all_models": "3582252.32", + "total_total_tokens_all_models": "68062794.0", + "avg_turn_count": "73.68", + "total_turn_count": "1400.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5.2-2025-12-11/a86d9a33-c73d-4136-a909-fc6fc9c0ddd2_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "cb0d679362a2ca663000cab61e30f0781a2b10a84d358ec7ce2e888ff6038540", + "total_rows": 19 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/a86d9a33-c73d-4136-a909-fc6fc9c0ddd2_samples.jsonl b/data/terminalbench/openai/gpt-5.2-2025-12-11/a86d9a33-c73d-4136-a909-fc6fc9c0ddd2_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..dde4839b099fd2b3e5d45c4441ef16bb6e8eaf42 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/a86d9a33-c73d-4136-a909-fc6fc9c0ddd2_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cb0d679362a2ca663000cab61e30f0781a2b10a84d358ec7ce2e888ff6038540 +size 10105908 diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/a9cb715a-cc77-41a3-997a-cc143137072e.json b/data/terminalbench/openai/gpt-5.2-2025-12-11/a9cb715a-cc77-41a3-997a-cc143137072e.json new file mode 100644 index 0000000000000000000000000000000000000000..4dd54723129a7f3592e7601749efd4922f5cab57 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/a9cb715a-cc77-41a3-997a-cc143137072e.json @@ -0,0 +1,111 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5.2-2025-12-11/1786163316.358004", + "evaluation_timestamp": "1777596544.0", + "retrieved_timestamp": "1786163316.358004", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive+C/+3ep/6a64b126 for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 2, + "sample_ids": [ + "polyglot-c-py", + "dna-insert" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1777596544.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.5, + "uncertainty": { + "standard_error": { + "value": 0.5 + }, + "num_samples": 3 + }, + "details": { + "total_matched_trajectories": "3", + "stopping_reason_count_repetition_guard": "3", + "avg_total_tokens_target_model": "624663.67", + "total_total_tokens_target_model": "1873991.0", + "avg_total_tokens_other_models": "0.00", + "total_total_tokens_other_models": "0.0", + "avg_total_tokens_all_models": "624663.67", + "total_total_tokens_all_models": "1873991.0", + "avg_turn_count": "28.33", + "total_turn_count": "85.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5.2-2025-12-11/a9cb715a-cc77-41a3-997a-cc143137072e_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "726d225ea1d296bba2099594f193b9b26ae2510f0d9131522ef279dd767b17a5", + "total_rows": 3 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/a9cb715a-cc77-41a3-997a-cc143137072e_samples.jsonl b/data/terminalbench/openai/gpt-5.2-2025-12-11/a9cb715a-cc77-41a3-997a-cc143137072e_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..adf7601d083c5d3a9e411cfcabafd855e5c70c1b --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/a9cb715a-cc77-41a3-997a-cc143137072e_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:726d225ea1d296bba2099594f193b9b26ae2510f0d9131522ef279dd767b17a5 +size 7701979 diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/acc0c53c-11dd-47cc-bab1-055b2bb3083c.json b/data/terminalbench/openai/gpt-5.2-2025-12-11/acc0c53c-11dd-47cc-bab1-055b2bb3083c.json new file mode 100644 index 0000000000000000000000000000000000000000..1252ab494cc1d2cc09ffa14a061e85de7796131e --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/acc0c53c-11dd-47cc-bab1-055b2bb3083c.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5.2-2025-12-11/1786167520.715579", + "evaluation_timestamp": "1777542076.0", + "retrieved_timestamp": "1786167520.715579", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5.2-2025-12-11/acc0c53c-11dd-47cc-bab1-055b2bb3083c_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "8b60df2c58b95f5cfda6d3bcff3f70a9e9091a917fa96711cdea6e1be1682015", + "total_rows": 1 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/acc0c53c-11dd-47cc-bab1-055b2bb3083c_samples.jsonl b/data/terminalbench/openai/gpt-5.2-2025-12-11/acc0c53c-11dd-47cc-bab1-055b2bb3083c_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..bdbd0bdc6982c9268941fb2b44542a95a88c261a --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/acc0c53c-11dd-47cc-bab1-055b2bb3083c_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8b60df2c58b95f5cfda6d3bcff3f70a9e9091a917fa96711cdea6e1be1682015 +size 130460 diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/acc36583-f686-4da6-881e-c88cf520559c.json b/data/terminalbench/openai/gpt-5.2-2025-12-11/acc36583-f686-4da6-881e-c88cf520559c.json new file mode 100644 index 0000000000000000000000000000000000000000..a7b87f297b97d08ffd28f53515eff2e92dbd65d2 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/acc36583-f686-4da6-881e-c88cf520559c.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5.2-2025-12-11/1786167561.725675", + "evaluation_timestamp": "1776885748.0", + "retrieved_timestamp": "1786167561.725675", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5.2-2025-12-11/acc36583-f686-4da6-881e-c88cf520559c_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "450f4b989a7dff7a8f215f0865d418689396738a75dfb973cf06a76ae156eb6e", + "total_rows": 15 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/acc36583-f686-4da6-881e-c88cf520559c_samples.jsonl b/data/terminalbench/openai/gpt-5.2-2025-12-11/acc36583-f686-4da6-881e-c88cf520559c_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..ddf1960a712cbd1b0c8c46978b468d1d83d17c93 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/acc36583-f686-4da6-881e-c88cf520559c_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:450f4b989a7dff7a8f215f0865d418689396738a75dfb973cf06a76ae156eb6e +size 10652322 diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/af95ff8c-fcd2-41b7-b68a-ba0434d2a6f4.json b/data/terminalbench/openai/gpt-5.2-2025-12-11/af95ff8c-fcd2-41b7-b68a-ba0434d2a6f4.json new file mode 100644 index 0000000000000000000000000000000000000000..ff1e029998713521068ec7a5b996926edb88f9f6 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/af95ff8c-fcd2-41b7-b68a-ba0434d2a6f4.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5.2-2025-12-11/1786167529.406689", + "evaluation_timestamp": "1775604415.0", + "retrieved_timestamp": "1786167529.406689", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.202,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5.2-2025-12-11/af95ff8c-fcd2-41b7-b68a-ba0434d2a6f4_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "ddf40b96aa24c3765429149a995f1b3521dc5459591b864698b48f6780262b87", + "total_rows": 20 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/af95ff8c-fcd2-41b7-b68a-ba0434d2a6f4_samples.jsonl b/data/terminalbench/openai/gpt-5.2-2025-12-11/af95ff8c-fcd2-41b7-b68a-ba0434d2a6f4_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..808ce2c273e362bb3d9973dde5313f71702e4e2d --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/af95ff8c-fcd2-41b7-b68a-ba0434d2a6f4_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ddf40b96aa24c3765429149a995f1b3521dc5459591b864698b48f6780262b87 +size 5838382 diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/af9e61d7-0f8e-45d6-81c0-6728afd7c122.json b/data/terminalbench/openai/gpt-5.2-2025-12-11/af9e61d7-0f8e-45d6-81c0-6728afd7c122.json new file mode 100644 index 0000000000000000000000000000000000000000..5204bced879381cfd1386429f585fb6f7ecbeb5e --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/af9e61d7-0f8e-45d6-81c0-6728afd7c122.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5.2-2025-12-11/1786169563.568876", + "evaluation_timestamp": "1777940349.0", + "retrieved_timestamp": "1786169563.568876", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5.2-2025-12-11/af9e61d7-0f8e-45d6-81c0-6728afd7c122_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "c6c01d98374fe1e4d4e4c1d4219cdca388c369b26b7ca5356f08fa2aa24ce1b0", + "total_rows": 40 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/af9e61d7-0f8e-45d6-81c0-6728afd7c122_samples.jsonl b/data/terminalbench/openai/gpt-5.2-2025-12-11/af9e61d7-0f8e-45d6-81c0-6728afd7c122_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..ce1766b6ab50a181758511b15726a914bfa7c6ea --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/af9e61d7-0f8e-45d6-81c0-6728afd7c122_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c6c01d98374fe1e4d4e4c1d4219cdca388c369b26b7ca5356f08fa2aa24ce1b0 +size 19286564 diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/b17c8c23-5c7a-4a0d-9833-1c6e1df00cd3.json b/data/terminalbench/openai/gpt-5.2-2025-12-11/b17c8c23-5c7a-4a0d-9833-1c6e1df00cd3.json new file mode 100644 index 0000000000000000000000000000000000000000..0db72ff7f1b1a55ca799a3307b02a467ac8de1ea --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/b17c8c23-5c7a-4a0d-9833-1c6e1df00cd3.json @@ -0,0 +1,107 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5.2-2025-12-11/1786163300.628465", + "evaluation_timestamp": "1777663098.0", + "retrieved_timestamp": "1786163300.628465", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive+C/+2ep/4b5ebcb3 for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 1, + "sample_ids": [ + "dna-insert" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1777663098.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 1.0, + "uncertainty": { + "num_samples": 1 + }, + "details": { + "total_matched_trajectories": "1", + "stopping_reason_count_completed_on_successful_submit": "1", + "avg_total_tokens_target_model": "1436476.00", + "total_total_tokens_target_model": "1436476.0", + "avg_total_tokens_other_models": "0.00", + "total_total_tokens_other_models": "0.0", + "avg_total_tokens_all_models": "1436476.00", + "total_total_tokens_all_models": "1436476.0", + "avg_turn_count": "52.00", + "total_turn_count": "52.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5.2-2025-12-11/b17c8c23-5c7a-4a0d-9833-1c6e1df00cd3_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "7552a9ab6c1f7d0a944b88ce53e782ed7d7400be204193d44394e14e16207dba", + "total_rows": 1 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/b17c8c23-5c7a-4a0d-9833-1c6e1df00cd3_samples.jsonl b/data/terminalbench/openai/gpt-5.2-2025-12-11/b17c8c23-5c7a-4a0d-9833-1c6e1df00cd3_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..1ade122f15275d219a1fa4f3cf7bf86bd6fedb85 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/b17c8c23-5c7a-4a0d-9833-1c6e1df00cd3_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7552a9ab6c1f7d0a944b88ce53e782ed7d7400be204193d44394e14e16207dba +size 777822 diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/b1d606d8-2aa2-4a43-90f0-545afe59c3dc.json b/data/terminalbench/openai/gpt-5.2-2025-12-11/b1d606d8-2aa2-4a43-90f0-545afe59c3dc.json new file mode 100644 index 0000000000000000000000000000000000000000..814bbaa81ae3b980f071d7b84f0ea95e38555721 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/b1d606d8-2aa2-4a43-90f0-545afe59c3dc.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5.2-2025-12-11/1786167696.106894", + "evaluation_timestamp": "1777478369.0", + "retrieved_timestamp": "1786167696.106894", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5.2-2025-12-11/b1d606d8-2aa2-4a43-90f0-545afe59c3dc_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "c93329a71b190a688b2c8c71db0d831ba973ae353bf18bb79c61e4642c22d4f9", + "total_rows": 80 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/b1d606d8-2aa2-4a43-90f0-545afe59c3dc_samples.jsonl b/data/terminalbench/openai/gpt-5.2-2025-12-11/b1d606d8-2aa2-4a43-90f0-545afe59c3dc_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..7e866a2eea3fe9e2c6e92edac719d092e86fd3c8 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/b1d606d8-2aa2-4a43-90f0-545afe59c3dc_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c93329a71b190a688b2c8c71db0d831ba973ae353bf18bb79c61e4642c22d4f9 +size 18695419 diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/b2913c0c-57d4-40cb-8b0c-b9c7a88dc05d.json b/data/terminalbench/openai/gpt-5.2-2025-12-11/b2913c0c-57d4-40cb-8b0c-b9c7a88dc05d.json new file mode 100644 index 0000000000000000000000000000000000000000..6681427b07ad9d6e2fa6e34d5df20063dca259f1 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/b2913c0c-57d4-40cb-8b0c-b9c7a88dc05d.json @@ -0,0 +1,118 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5.2-2025-12-11/1786165918.364942", + "evaluation_timestamp": "1777910667.0", + "retrieved_timestamp": "1786165918.364942", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive/+3ep/d985c697 for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 5, + "sample_ids": [ + "git-multibranch", + "video-processing", + "qemu-alpine-ssh", + "kv-store-grpc", + "install-windows-3.11" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1777910667.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.26666666666666666, + "uncertainty": { + "standard_error": { + "value": 0.19436506316151003 + }, + "num_samples": 13 + }, + "details": { + "total_matched_trajectories": "13", + "stopping_reason_count_token_limit": "3", + "stopping_reason_count_repetition_guard": "10", + "avg_total_tokens_target_model": "3710262.85", + "total_total_tokens_target_model": "48233417.0", + "avg_total_tokens_other_models": "1092.38", + "total_total_tokens_other_models": "14201.0", + "avg_total_tokens_all_models": "3711355.23", + "total_total_tokens_all_models": "48247618.0", + "avg_turn_count": "83.15", + "total_turn_count": "1081.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5.2-2025-12-11/b2913c0c-57d4-40cb-8b0c-b9c7a88dc05d_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "242b6e113adf12ebbbefa3ad82e5b5fce15176049f2dea547b45aaa4f4bb63cb", + "total_rows": 13 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/b2913c0c-57d4-40cb-8b0c-b9c7a88dc05d_samples.jsonl b/data/terminalbench/openai/gpt-5.2-2025-12-11/b2913c0c-57d4-40cb-8b0c-b9c7a88dc05d_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..a205ff20dff5dfd8f5ec009c7c8887222f4692d3 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/b2913c0c-57d4-40cb-8b0c-b9c7a88dc05d_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:242b6e113adf12ebbbefa3ad82e5b5fce15176049f2dea547b45aaa4f4bb63cb +size 7105238 diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/c346b295-471d-444a-836b-48e0c4e1e92e.json b/data/terminalbench/openai/gpt-5.2-2025-12-11/c346b295-471d-444a-836b-48e0c4e1e92e.json new file mode 100644 index 0000000000000000000000000000000000000000..64a3be5337f983d80623ddfa73933c82777657da --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/c346b295-471d-444a-836b-48e0c4e1e92e.json @@ -0,0 +1,107 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5.2-2025-12-11/1786163318.605942", + "evaluation_timestamp": "1777680998.0", + "retrieved_timestamp": "1786163318.605942", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive+C/+4ep/4ce59f10 for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 1, + "sample_ids": [ + "regex-chess" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1777680998.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 1.0, + "uncertainty": { + "num_samples": 1 + }, + "details": { + "total_matched_trajectories": "1", + "stopping_reason_count_completed_on_successful_submit": "1", + "avg_total_tokens_target_model": "2335272.00", + "total_total_tokens_target_model": "2335272.0", + "avg_total_tokens_other_models": "0.00", + "total_total_tokens_other_models": "0.0", + "avg_total_tokens_all_models": "2335272.00", + "total_total_tokens_all_models": "2335272.0", + "avg_turn_count": "35.00", + "total_turn_count": "35.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5.2-2025-12-11/c346b295-471d-444a-836b-48e0c4e1e92e_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "ef12d5dd11786185322c54b80b36aa4c30f8b99722657ecf0843c05c30339ab8", + "total_rows": 1 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/c346b295-471d-444a-836b-48e0c4e1e92e_samples.jsonl b/data/terminalbench/openai/gpt-5.2-2025-12-11/c346b295-471d-444a-836b-48e0c4e1e92e_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..84c10520010bf55e0614566ba742ace561d19a4a --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/c346b295-471d-444a-836b-48e0c4e1e92e_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ef12d5dd11786185322c54b80b36aa4c30f8b99722657ecf0843c05c30339ab8 +size 632890 diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/c388a35c-0104-4d3f-b278-995b7d665271.json b/data/terminalbench/openai/gpt-5.2-2025-12-11/c388a35c-0104-4d3f-b278-995b7d665271.json new file mode 100644 index 0000000000000000000000000000000000000000..9258ca14ad04771e9c2870812d0c1956e97b706d --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/c388a35c-0104-4d3f-b278-995b7d665271.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5.2-2025-12-11/1786169540.708852", + "evaluation_timestamp": "1776417565.0", + "retrieved_timestamp": "1786169540.708852", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5.2-2025-12-11/c388a35c-0104-4d3f-b278-995b7d665271_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "3a02a3c1470c9f55f15f1a309b31fd68462ea143bcd2ab712e81105a4ab6b7ca", + "total_rows": 5 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/c388a35c-0104-4d3f-b278-995b7d665271_samples.jsonl b/data/terminalbench/openai/gpt-5.2-2025-12-11/c388a35c-0104-4d3f-b278-995b7d665271_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..2de00a28c3675fdcef948872954540d77fc65832 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/c388a35c-0104-4d3f-b278-995b7d665271_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3a02a3c1470c9f55f15f1a309b31fd68462ea143bcd2ab712e81105a4ab6b7ca +size 2126456 diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/c6132929-0a3f-4abf-98d8-8a02c735006c.json b/data/terminalbench/openai/gpt-5.2-2025-12-11/c6132929-0a3f-4abf-98d8-8a02c735006c.json new file mode 100644 index 0000000000000000000000000000000000000000..63354423232a77f62ab605a521eba5bb7b67663d --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/c6132929-0a3f-4abf-98d8-8a02c735006c.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5.2-2025-12-11/1786169531.627377", + "evaluation_timestamp": "1776429050.0", + "retrieved_timestamp": "1786169531.627377", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5.2-2025-12-11/c6132929-0a3f-4abf-98d8-8a02c735006c_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "46505cf0d2f1a95362362337672db0dc28eacf6b79e5c88a7818611365ae1608", + "total_rows": 10 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/c6132929-0a3f-4abf-98d8-8a02c735006c_samples.jsonl b/data/terminalbench/openai/gpt-5.2-2025-12-11/c6132929-0a3f-4abf-98d8-8a02c735006c_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..fc3d10fd579cafe7985501d1d9da5605c8792714 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/c6132929-0a3f-4abf-98d8-8a02c735006c_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:46505cf0d2f1a95362362337672db0dc28eacf6b79e5c88a7818611365ae1608 +size 2935290 diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/c6538def-f42c-4502-a792-11b70cc6b5c1.json b/data/terminalbench/openai/gpt-5.2-2025-12-11/c6538def-f42c-4502-a792-11b70cc6b5c1.json new file mode 100644 index 0000000000000000000000000000000000000000..30862f31a5ea1ec4a34653de2f92a0a48d920b85 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/c6538def-f42c-4502-a792-11b70cc6b5c1.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5.2-2025-12-11/1786169536.066458", + "evaluation_timestamp": "1776433475.0", + "retrieved_timestamp": "1786169536.066458", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5.2-2025-12-11/c6538def-f42c-4502-a792-11b70cc6b5c1_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "485f8f9801e9696bdf1dd3e7b280febeea518ac999a2e48b3f8044bf27ef1816", + "total_rows": 10 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/c6538def-f42c-4502-a792-11b70cc6b5c1_samples.jsonl b/data/terminalbench/openai/gpt-5.2-2025-12-11/c6538def-f42c-4502-a792-11b70cc6b5c1_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..87b1e9f2376a2055b29ab65993b09609691d3895 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/c6538def-f42c-4502-a792-11b70cc6b5c1_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:485f8f9801e9696bdf1dd3e7b280febeea518ac999a2e48b3f8044bf27ef1816 +size 5074707 diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/c9b3078c-e706-44ea-a8af-471a6a0baeaf.json b/data/terminalbench/openai/gpt-5.2-2025-12-11/c9b3078c-e706-44ea-a8af-471a6a0baeaf.json new file mode 100644 index 0000000000000000000000000000000000000000..067fca15e10647d549182763c2dc18388eff4399 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/c9b3078c-e706-44ea-a8af-471a6a0baeaf.json @@ -0,0 +1,110 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5.2-2025-12-11/1786165828.44699", + "evaluation_timestamp": "1778265533.0", + "retrieved_timestamp": "1786165828.44699", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive/+1ep/5dd93e71 for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 1, + "sample_ids": [ + "dna-assembly" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1778265533.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.0, + "uncertainty": { + "num_samples": 1 + }, + "details": { + "total_matched_trajectories": "1", + "stopping_reason_count_repetition_guard": "1", + "avg_total_tokens_target_model": "3159557.00", + "total_total_tokens_target_model": "3159557.0", + "avg_total_tokens_other_models": "784.00", + "total_total_tokens_other_models": "784.0", + "avg_total_tokens_all_models": "3160341.00", + "total_total_tokens_all_models": "3160341.0", + "avg_turn_count": "89.00", + "total_turn_count": "89.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5.2-2025-12-11/c9b3078c-e706-44ea-a8af-471a6a0baeaf_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "7319242b148d0325c5a25d9a93d25aa5f88e970771ba10e5cd090e8654d088cc", + "total_rows": 1 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/c9b3078c-e706-44ea-a8af-471a6a0baeaf_samples.jsonl b/data/terminalbench/openai/gpt-5.2-2025-12-11/c9b3078c-e706-44ea-a8af-471a6a0baeaf_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..2808bfa5fd81ef52258b7c1e4b07d0e435094ecd --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/c9b3078c-e706-44ea-a8af-471a6a0baeaf_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7319242b148d0325c5a25d9a93d25aa5f88e970771ba10e5cd090e8654d088cc +size 807178 diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/caac4753-8b73-4834-88bc-ee03f161b4a6.json b/data/terminalbench/openai/gpt-5.2-2025-12-11/caac4753-8b73-4834-88bc-ee03f161b4a6.json new file mode 100644 index 0000000000000000000000000000000000000000..ff35190d4a92820cf11ecb8e3179a490ba7fd44d --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/caac4753-8b73-4834-88bc-ee03f161b4a6.json @@ -0,0 +1,110 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5.2-2025-12-11/1786165833.970374", + "evaluation_timestamp": "1780423668.0", + "retrieved_timestamp": "1786165833.970374", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive/+1ep/347739d1 for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 1, + "sample_ids": [ + "circuit-fibsqrt" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1780423668.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 1.0, + "uncertainty": { + "num_samples": 1 + }, + "details": { + "total_matched_trajectories": "1", + "stopping_reason_count_repetition_guard": "1", + "avg_total_tokens_target_model": "359124.00", + "total_total_tokens_target_model": "359124.0", + "avg_total_tokens_other_models": "755.00", + "total_total_tokens_other_models": "755.0", + "avg_total_tokens_all_models": "359879.00", + "total_total_tokens_all_models": "359879.0", + "avg_turn_count": "17.00", + "total_turn_count": "17.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "3", + "timeout": "1800", + "attempt_timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {}, + "max_attempts": 3 + }, + "additional_details": { + "max_retries": "3", + "timeout": "1800", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5.2-2025-12-11/caac4753-8b73-4834-88bc-ee03f161b4a6_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "1bd5a94e8dec45bffb03d3d473d8af9c6c6adb068cc76cc1c322a3601dc34cef", + "total_rows": 1 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/caac4753-8b73-4834-88bc-ee03f161b4a6_samples.jsonl b/data/terminalbench/openai/gpt-5.2-2025-12-11/caac4753-8b73-4834-88bc-ee03f161b4a6_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..3eca336d4832a7fa51397f6db6488ff89d822a57 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/caac4753-8b73-4834-88bc-ee03f161b4a6_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1bd5a94e8dec45bffb03d3d473d8af9c6c6adb068cc76cc1c322a3601dc34cef +size 231862 diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/cbc8da76-3b4a-4623-9caf-732662f88e81.json b/data/terminalbench/openai/gpt-5.2-2025-12-11/cbc8da76-3b4a-4623-9caf-732662f88e81.json new file mode 100644 index 0000000000000000000000000000000000000000..48288441a6ad113a5826c12283d5610ecaaae8cc --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/cbc8da76-3b4a-4623-9caf-732662f88e81.json @@ -0,0 +1,126 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5.2-2025-12-11/1786163332.161671", + "evaluation_timestamp": "1775733996.0", + "retrieved_timestamp": "1786163332.161671", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive+C/+5ep for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 19, + "sample_ids": [ + "gpt2-codegolf", + "llm-inference-batching-scheduler", + "break-filter-js-from-html", + "write-compressor", + "merge-diff-arc-agi-task", + "winning-avg-corewars", + "log-summary-date-ranges", + "pytorch-model-cli", + "largest-eigenval", + "password-recovery", + "portfolio-optimization", + "path-tracing-reverse", + "modernize-scientific-stack", + "torch-tensor-parallelism", + "path-tracing", + "prove-plus-comm", + "feal-linear-cryptanalysis", + "distribution-search", + "pypi-server" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1775733996.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.8736842105263158, + "uncertainty": { + "standard_error": { + "value": 0.067034993594887 + }, + "num_samples": 95 + }, + "details": { + "total_matched_trajectories": "95", + "stopping_reason_count_completed_on_successful_submit": "83", + "stopping_reason_count_token_limit": "11", + "stopping_reason_count_repetition_guard": "1", + "avg_total_tokens_target_model": "1870386.24", + "total_total_tokens_target_model": "177686693.0", + "avg_total_tokens_other_models": "37.01", + "total_total_tokens_other_models": "3516.0", + "avg_total_tokens_all_models": "1870423.25", + "total_total_tokens_all_models": "177690209.0", + "avg_turn_count": "41.79", + "total_turn_count": "3970.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5.2-2025-12-11/cbc8da76-3b4a-4623-9caf-732662f88e81_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "f3b27995c83fa3656e5b9cf7b777eaf7237a1d5482dc10e4f24c884669075210", + "total_rows": 95 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/cbc8da76-3b4a-4623-9caf-732662f88e81_samples.jsonl b/data/terminalbench/openai/gpt-5.2-2025-12-11/cbc8da76-3b4a-4623-9caf-732662f88e81_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..5be2ccc5446323105d6aad3a42d96249983e3a7c --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/cbc8da76-3b4a-4623-9caf-732662f88e81_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f3b27995c83fa3656e5b9cf7b777eaf7237a1d5482dc10e4f24c884669075210 +size 38088306 diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/cda21203-f9fc-4b33-9a41-da7baa3a84e9.json b/data/terminalbench/openai/gpt-5.2-2025-12-11/cda21203-f9fc-4b33-9a41-da7baa3a84e9.json new file mode 100644 index 0000000000000000000000000000000000000000..48db496c7056fd0543a9f01a584ceae67622c747 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/cda21203-f9fc-4b33-9a41-da7baa3a84e9.json @@ -0,0 +1,116 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5.2-2025-12-11/1786165840.465453", + "evaluation_timestamp": "1778229123.0", + "retrieved_timestamp": "1786165840.465453", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive/+1ep/ed31c698 for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 3, + "sample_ids": [ + "mteb-leaderboard", + "rstan-to-pystan", + "dna-assembly" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1778229123.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.5, + "uncertainty": { + "standard_error": { + "value": 0.5 + }, + "num_samples": 2 + }, + "details": { + "total_matched_trajectories": "2", + "stopping_reason_count_token_limit": "1", + "stopping_reason_count_repetition_guard": "1", + "avg_total_tokens_target_model": "6002669.00", + "total_total_tokens_target_model": "12005338.0", + "avg_total_tokens_other_models": "522.00", + "total_total_tokens_other_models": "1044.0", + "avg_total_tokens_all_models": "6003191.00", + "total_total_tokens_all_models": "12006382.0", + "avg_turn_count": "124.00", + "total_turn_count": "248.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5.2-2025-12-11/cda21203-f9fc-4b33-9a41-da7baa3a84e9_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "4d1f77728f3c3e4cf85d2eccbe5296c0d8e96b71572f1710f4e8979f9a3199b4", + "total_rows": 2 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/cda21203-f9fc-4b33-9a41-da7baa3a84e9_samples.jsonl b/data/terminalbench/openai/gpt-5.2-2025-12-11/cda21203-f9fc-4b33-9a41-da7baa3a84e9_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..0d37b6a8c27d91ee4b8f50e47618ea8d41418b53 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/cda21203-f9fc-4b33-9a41-da7baa3a84e9_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4d1f77728f3c3e4cf85d2eccbe5296c0d8e96b71572f1710f4e8979f9a3199b4 +size 1641845 diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/d13db513-3107-4c6a-8fe8-b348b05859e9.json b/data/terminalbench/openai/gpt-5.2-2025-12-11/d13db513-3107-4c6a-8fe8-b348b05859e9.json new file mode 100644 index 0000000000000000000000000000000000000000..0fc9e5926c4153557aeb9eab02f39b5b12b98930 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/d13db513-3107-4c6a-8fe8-b348b05859e9.json @@ -0,0 +1,110 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5.2-2025-12-11/1786165872.178741", + "evaluation_timestamp": "1778269139.0", + "retrieved_timestamp": "1786165872.178741", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive/+3ep/347739d1 for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 1, + "sample_ids": [ + "circuit-fibsqrt" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1778269139.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 1.0, + "uncertainty": { + "num_samples": 2 + }, + "details": { + "total_matched_trajectories": "2", + "stopping_reason_count_repetition_guard": "2", + "avg_total_tokens_target_model": "452258.00", + "total_total_tokens_target_model": "904516.0", + "avg_total_tokens_other_models": "803.50", + "total_total_tokens_other_models": "1607.0", + "avg_total_tokens_all_models": "453061.50", + "total_total_tokens_all_models": "906123.0", + "avg_turn_count": "18.00", + "total_turn_count": "36.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5.2-2025-12-11/d13db513-3107-4c6a-8fe8-b348b05859e9_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "1d5fcd4f1044a58c022a24317caa169eb854078deb5565e2217a0f71d4be4b63", + "total_rows": 2 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/d13db513-3107-4c6a-8fe8-b348b05859e9_samples.jsonl b/data/terminalbench/openai/gpt-5.2-2025-12-11/d13db513-3107-4c6a-8fe8-b348b05859e9_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..d4f6524a2c4baae28fea81e08b2de6b624e46b70 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/d13db513-3107-4c6a-8fe8-b348b05859e9_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1d5fcd4f1044a58c022a24317caa169eb854078deb5565e2217a0f71d4be4b63 +size 497777 diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/d30c48a3-e487-4a4e-a515-33f734acb977.json b/data/terminalbench/openai/gpt-5.2-2025-12-11/d30c48a3-e487-4a4e-a515-33f734acb977.json new file mode 100644 index 0000000000000000000000000000000000000000..65de256a395ce5887b8b4ea57735b2992ff3a83f --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/d30c48a3-e487-4a4e-a515-33f734acb977.json @@ -0,0 +1,119 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5.2-2025-12-11/1786163309.548067", + "evaluation_timestamp": "1776629382.0", + "retrieved_timestamp": "1786163309.548067", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive+C/+3ep/507885b9 for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 10, + "sample_ids": [ + "reshard-c4-data", + "caffe-cifar-10", + "custom-memory-heap-crash", + "adaptive-rejection-sampler", + "multi-source-data-merger", + "configure-git-webserver", + "crack-7z-hash", + "chess-best-move", + "cobol-modernization", + "hf-model-inference" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1776629382.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.8666666666666666, + "uncertainty": { + "standard_error": { + "value": 0.1018350154434631 + }, + "num_samples": 30 + }, + "details": { + "total_matched_trajectories": "30", + "stopping_reason_count_completed_on_successful_submit": "26", + "stopping_reason_count_repetition_guard": "3", + "stopping_reason_count_token_limit": "1", + "avg_total_tokens_target_model": "1293111.80", + "total_total_tokens_target_model": "38793354.0", + "avg_total_tokens_other_models": "253.77", + "total_total_tokens_other_models": "7613.0", + "avg_total_tokens_all_models": "1293365.57", + "total_total_tokens_all_models": "38800967.0", + "avg_turn_count": "38.77", + "total_turn_count": "1163.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5.2-2025-12-11/d30c48a3-e487-4a4e-a515-33f734acb977_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "d24cf89a0a1d7bbc52abc2feb93a31e4792993f82a2d3cb03a1b5d5c53a0673c", + "total_rows": 30 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/d30c48a3-e487-4a4e-a515-33f734acb977_samples.jsonl b/data/terminalbench/openai/gpt-5.2-2025-12-11/d30c48a3-e487-4a4e-a515-33f734acb977_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..6fddd2a85fb86417640bff32a1f0df195acd8e21 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/d30c48a3-e487-4a4e-a515-33f734acb977_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d24cf89a0a1d7bbc52abc2feb93a31e4792993f82a2d3cb03a1b5d5c53a0673c +size 7515161 diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/dc29dfab-f4d2-4686-b703-652d567a8401.json b/data/terminalbench/openai/gpt-5.2-2025-12-11/dc29dfab-f4d2-4686-b703-652d567a8401.json new file mode 100644 index 0000000000000000000000000000000000000000..ecb9d84cb423f3db2560b8df1fdd248a5c02dd41 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/dc29dfab-f4d2-4686-b703-652d567a8401.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5.2-2025-12-11/1786167680.661695", + "evaluation_timestamp": "1775665083.0", + "retrieved_timestamp": "1786167680.661695", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5.2-2025-12-11/dc29dfab-f4d2-4686-b703-652d567a8401_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "0f973e9b88236743370530ab1c9151f5aef25791947881d0cd59bbaf91c7e6c1", + "total_rows": 59 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/dc29dfab-f4d2-4686-b703-652d567a8401_samples.jsonl b/data/terminalbench/openai/gpt-5.2-2025-12-11/dc29dfab-f4d2-4686-b703-652d567a8401_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..7cabc2d8658e99f6d504eaeb0cd81e37ef6e9856 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/dc29dfab-f4d2-4686-b703-652d567a8401_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0f973e9b88236743370530ab1c9151f5aef25791947881d0cd59bbaf91c7e6c1 +size 25849204 diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/debf978c-664c-46de-b01d-76d3a25879b4.json b/data/terminalbench/openai/gpt-5.2-2025-12-11/debf978c-664c-46de-b01d-76d3a25879b4.json new file mode 100644 index 0000000000000000000000000000000000000000..5820512761b9406c2f8d8255b9f4aa24072a3c31 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/debf978c-664c-46de-b01d-76d3a25879b4.json @@ -0,0 +1,117 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5.2-2025-12-11/1786163289.381419", + "evaluation_timestamp": "1777586285.0", + "retrieved_timestamp": "1786163289.381419", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive+C/+1ep/1c622f59 for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 7, + "sample_ids": [ + "sam-cell-seg", + "model-extraction-relu-logits", + "large-scale-text-editing", + "sanitize-git-repo", + "rstan-to-pystan", + "constraints-scheduling", + "vulnerable-secret" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1777586285.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.7142857142857143, + "uncertainty": { + "standard_error": { + "value": 0.18442777839082938 + }, + "num_samples": 7 + }, + "details": { + "total_matched_trajectories": "7", + "stopping_reason_count_completed_on_successful_submit": "5", + "stopping_reason_count_repetition_guard": "2", + "avg_total_tokens_target_model": "1420973.00", + "total_total_tokens_target_model": "9946811.0", + "avg_total_tokens_other_models": "0.00", + "total_total_tokens_other_models": "0.0", + "avg_total_tokens_all_models": "1420973.00", + "total_total_tokens_all_models": "9946811.0", + "avg_turn_count": "37.43", + "total_turn_count": "262.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5.2-2025-12-11/debf978c-664c-46de-b01d-76d3a25879b4_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "365581b8fbd86136d42b2fcd1d1b92bfb7fdb6eb8d6649b1f61dff7d7add82e5", + "total_rows": 7 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/debf978c-664c-46de-b01d-76d3a25879b4_samples.jsonl b/data/terminalbench/openai/gpt-5.2-2025-12-11/debf978c-664c-46de-b01d-76d3a25879b4_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..ce70717379b91fb2d400493abfceb50c7ff76d20 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/debf978c-664c-46de-b01d-76d3a25879b4_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:365581b8fbd86136d42b2fcd1d1b92bfb7fdb6eb8d6649b1f61dff7d7add82e5 +size 3461650 diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/ef5f1761-54b8-4ef2-a65c-1387c143b0cd.json b/data/terminalbench/openai/gpt-5.2-2025-12-11/ef5f1761-54b8-4ef2-a65c-1387c143b0cd.json new file mode 100644 index 0000000000000000000000000000000000000000..74264f3a7112cedbdf0912bcffe06b39ebe75255 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/ef5f1761-54b8-4ef2-a65c-1387c143b0cd.json @@ -0,0 +1,116 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5.2-2025-12-11/1786165869.802668", + "evaluation_timestamp": "1778055257.0", + "retrieved_timestamp": "1786165869.802668", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive/+2ep/04242b81 for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 3, + "sample_ids": [ + "git-multibranch", + "make-doom-for-mips", + "tune-mjcf" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1778055257.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.3333333333333333, + "uncertainty": { + "standard_error": { + "value": 0.33333333333333337 + }, + "num_samples": 6 + }, + "details": { + "total_matched_trajectories": "6", + "stopping_reason_count_repetition_guard": "4", + "stopping_reason_count_token_limit": "2", + "avg_total_tokens_target_model": "4174282.00", + "total_total_tokens_target_model": "25045692.0", + "avg_total_tokens_other_models": "555.50", + "total_total_tokens_other_models": "3333.0", + "avg_total_tokens_all_models": "4174837.50", + "total_total_tokens_all_models": "25049025.0", + "avg_turn_count": "103.33", + "total_turn_count": "620.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "5", + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "16000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {}, + "max_attempts": 5 + }, + "additional_details": { + "max_retries": "5", + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5.2-2025-12-11/ef5f1761-54b8-4ef2-a65c-1387c143b0cd_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "2b026a7e98532efdfec563a101b5fd0965694f337ea6a1bc052bc37528f4fbfe", + "total_rows": 6 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/ef5f1761-54b8-4ef2-a65c-1387c143b0cd_samples.jsonl b/data/terminalbench/openai/gpt-5.2-2025-12-11/ef5f1761-54b8-4ef2-a65c-1387c143b0cd_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..1938c719760cb740437829f89174c6f27266f73a --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/ef5f1761-54b8-4ef2-a65c-1387c143b0cd_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2b026a7e98532efdfec563a101b5fd0965694f337ea6a1bc052bc37528f4fbfe +size 2999120 diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/f1beb253-fbaa-43d6-b325-91277d8e84e3.json b/data/terminalbench/openai/gpt-5.2-2025-12-11/f1beb253-fbaa-43d6-b325-91277d8e84e3.json new file mode 100644 index 0000000000000000000000000000000000000000..70bc017722dae6689727451ee84371c952dfc859 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/f1beb253-fbaa-43d6-b325-91277d8e84e3.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5.2-2025-12-11/1786169599.921473", + "evaluation_timestamp": "1775078458.0", + "retrieved_timestamp": "1786169599.921473", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.202,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5.2-2025-12-11/f1beb253-fbaa-43d6-b325-91277d8e84e3_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "e44d56dec07674c311a0796aa9be44ba19d50e3d968e02670db16abe0b6f0105", + "total_rows": 30 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/f1beb253-fbaa-43d6-b325-91277d8e84e3_samples.jsonl b/data/terminalbench/openai/gpt-5.2-2025-12-11/f1beb253-fbaa-43d6-b325-91277d8e84e3_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..993d92c3101aef47585db65225da753e027d3443 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/f1beb253-fbaa-43d6-b325-91277d8e84e3_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e44d56dec07674c311a0796aa9be44ba19d50e3d968e02670db16abe0b6f0105 +size 11775126 diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/f7886d15-3ed0-4be0-a0b3-4aed8af63dc1.json b/data/terminalbench/openai/gpt-5.2-2025-12-11/f7886d15-3ed0-4be0-a0b3-4aed8af63dc1.json new file mode 100644 index 0000000000000000000000000000000000000000..22814beef9c6a79efdfaca1a8825d372a81e1ce9 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/f7886d15-3ed0-4be0-a0b3-4aed8af63dc1.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5.2-2025-12-11/1786167581.405014", + "evaluation_timestamp": "1776711189.0", + "retrieved_timestamp": "1786167581.405014", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.2-2025-12-11", + "id": "openai/gpt-5.2-2025-12-11", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5.2-2025-12-11/f7886d15-3ed0-4be0-a0b3-4aed8af63dc1_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "32d56711a14943c007638a6dedde07e93546a5dd5c7ee57ed8a3a338b1b41f62", + "total_rows": 39 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5.2-2025-12-11/f7886d15-3ed0-4be0-a0b3-4aed8af63dc1_samples.jsonl b/data/terminalbench/openai/gpt-5.2-2025-12-11/f7886d15-3ed0-4be0-a0b3-4aed8af63dc1_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..ce1e76e9be92cbdef797899fc8ccdb25124cf1f6 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.2-2025-12-11/f7886d15-3ed0-4be0-a0b3-4aed8af63dc1_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:32d56711a14943c007638a6dedde07e93546a5dd5c7ee57ed8a3a338b1b41f62 +size 22857877 diff --git a/data/terminalbench/openai/gpt-5.4-2026-03-05/013f876a-0a0d-4fb1-88e4-9a5541eeb66e.json b/data/terminalbench/openai/gpt-5.4-2026-03-05/013f876a-0a0d-4fb1-88e4-9a5541eeb66e.json new file mode 100644 index 0000000000000000000000000000000000000000..47f10ac3da0111c46a5ea027138a0a05609d9c48 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.4-2026-03-05/013f876a-0a0d-4fb1-88e4-9a5541eeb66e.json @@ -0,0 +1,112 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5.4-2026-03-05/1786182969.061481", + "evaluation_timestamp": "1774902342.0", + "retrieved_timestamp": "1786182969.061481", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.191.dev111+g7c0c4345,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 89, + "sample_ids": [ + "gpt2-codegolf", + "llm-inference-batching-scheduler", + "break-filter-js-from-html", + "reshard-c4-data", + "write-compressor" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1774902342.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.8, + "uncertainty": { + "standard_error": { + "value": 0.20000000000000004 + }, + "num_samples": 10 + }, + "details": { + "total_matched_trajectories": "10", + "stopping_reason_count_repetition_guard": "10", + "avg_total_tokens_target_model": "1186360.40", + "total_total_tokens_target_model": "11863604.0", + "avg_total_tokens_other_models": "2363.80", + "total_total_tokens_other_models": "23638.0", + "avg_total_tokens_all_models": "1188724.20", + "total_total_tokens_all_models": "11887242.0", + "avg_turn_count": "31.30", + "total_turn_count": "313.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\"}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "3", + "attempt_timeout": "1200", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 6000000 + }, + "sandbox": {}, + "max_attempts": 3 + }, + "additional_details": { + "max_retries": "3" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5.4-2026-03-05/013f876a-0a0d-4fb1-88e4-9a5541eeb66e_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "9c7aab6d3c3b696b5bd33445fe885ead95b3141d6c702d94d6effe21ffa54c63", + "total_rows": 10 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5.4-2026-03-05/013f876a-0a0d-4fb1-88e4-9a5541eeb66e_samples.jsonl b/data/terminalbench/openai/gpt-5.4-2026-03-05/013f876a-0a0d-4fb1-88e4-9a5541eeb66e_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..94f14cdd479eb62be3fff9438dfe3abd40aeee15 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.4-2026-03-05/013f876a-0a0d-4fb1-88e4-9a5541eeb66e_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9c7aab6d3c3b696b5bd33445fe885ead95b3141d6c702d94d6effe21ffa54c63 +size 2961219 diff --git a/data/terminalbench/openai/gpt-5.4-2026-03-05/06c72550-0ab6-43e9-b5c2-f9151cb165d1.json b/data/terminalbench/openai/gpt-5.4-2026-03-05/06c72550-0ab6-43e9-b5c2-f9151cb165d1.json new file mode 100644 index 0000000000000000000000000000000000000000..08a0801196d0e89d6884315b89293790e955f436 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.4-2026-03-05/06c72550-0ab6-43e9-b5c2-f9151cb165d1.json @@ -0,0 +1,112 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5.4-2026-03-05/1786163352.596808", + "evaluation_timestamp": "1777478370.0", + "retrieved_timestamp": "1786163352.596808", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive+C/+2ep/300bd7bc for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 2, + "sample_ids": [ + "raman-fitting", + "dna-assembly" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1777478370.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.5, + "uncertainty": { + "standard_error": { + "value": 0.5 + }, + "num_samples": 4 + }, + "details": { + "total_matched_trajectories": "4", + "stopping_reason_count_completed_on_successful_submit": "2", + "stopping_reason_count_repetition_guard": "2", + "avg_total_tokens_target_model": "3480682.50", + "total_total_tokens_target_model": "13922730.0", + "avg_total_tokens_other_models": "0.00", + "total_total_tokens_other_models": "0.0", + "avg_total_tokens_all_models": "3480682.50", + "total_total_tokens_all_models": "13922730.0", + "avg_turn_count": "62.50", + "total_turn_count": "250.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "32", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "32", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5.4-2026-03-05/06c72550-0ab6-43e9-b5c2-f9151cb165d1_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "57412333dc674323c72d3d71a7344422a083fc210cfcccf2c043f828656348b6", + "total_rows": 4 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5.4-2026-03-05/06c72550-0ab6-43e9-b5c2-f9151cb165d1_samples.jsonl b/data/terminalbench/openai/gpt-5.4-2026-03-05/06c72550-0ab6-43e9-b5c2-f9151cb165d1_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..8c82c90cef44cfcb37ca1d5b52b4fed12ad34c11 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.4-2026-03-05/06c72550-0ab6-43e9-b5c2-f9151cb165d1_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:57412333dc674323c72d3d71a7344422a083fc210cfcccf2c043f828656348b6 +size 3073774 diff --git a/data/terminalbench/openai/gpt-5.4-2026-03-05/08f1f1fc-e078-4f8a-bfdc-ca2c1d6fb82d.json b/data/terminalbench/openai/gpt-5.4-2026-03-05/08f1f1fc-e078-4f8a-bfdc-ca2c1d6fb82d.json new file mode 100644 index 0000000000000000000000000000000000000000..8debc1eee6fa1361379ba5d08d41f532eb401f75 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.4-2026-03-05/08f1f1fc-e078-4f8a-bfdc-ca2c1d6fb82d.json @@ -0,0 +1,109 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5.4-2026-03-05/1786165947.118832", + "evaluation_timestamp": "1776433475.0", + "retrieved_timestamp": "1786165947.118832", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive/+1ep/c9130808 for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 2, + "sample_ids": [ + "password-recovery", + "path-tracing-reverse" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1776433475.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.5, + "uncertainty": { + "standard_error": { + "value": 0.5 + }, + "num_samples": 2 + }, + "details": { + "total_matched_trajectories": "2", + "stopping_reason_count_repetition_guard": "2", + "avg_total_tokens_target_model": "1545614.00", + "total_total_tokens_target_model": "3091228.0", + "avg_total_tokens_other_models": "280.50", + "total_total_tokens_other_models": "561.0", + "avg_total_tokens_all_models": "1545894.50", + "total_total_tokens_all_models": "3091789.0", + "avg_turn_count": "33.00", + "total_turn_count": "66.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5.4-2026-03-05/08f1f1fc-e078-4f8a-bfdc-ca2c1d6fb82d_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "a9b04342ff710d8ec08fbd28fca70bd929dd76f1cad0c1a65359b76131f65ac8", + "total_rows": 2 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5.4-2026-03-05/08f1f1fc-e078-4f8a-bfdc-ca2c1d6fb82d_samples.jsonl b/data/terminalbench/openai/gpt-5.4-2026-03-05/08f1f1fc-e078-4f8a-bfdc-ca2c1d6fb82d_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..c4538205178ebf9746ef4306f84f59709db78a62 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.4-2026-03-05/08f1f1fc-e078-4f8a-bfdc-ca2c1d6fb82d_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a9b04342ff710d8ec08fbd28fca70bd929dd76f1cad0c1a65359b76131f65ac8 +size 553994 diff --git a/data/terminalbench/openai/gpt-5.4-2026-03-05/0c929847-aa17-4921-8d77-09abcc13fc4e.json b/data/terminalbench/openai/gpt-5.4-2026-03-05/0c929847-aa17-4921-8d77-09abcc13fc4e.json new file mode 100644 index 0000000000000000000000000000000000000000..2342975891d667e874d37ee8623835923278f890 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.4-2026-03-05/0c929847-aa17-4921-8d77-09abcc13fc4e.json @@ -0,0 +1,112 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5.4-2026-03-05/1786166000.084478", + "evaluation_timestamp": "1776680886.0", + "retrieved_timestamp": "1786166000.084478", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive/+2ep/878b8bc9 for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 5, + "sample_ids": [ + "overfull-hbox", + "query-optimize", + "large-scale-text-editing", + "sqlite-db-truncate", + "build-pmars" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1776680886.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.875, + "uncertainty": { + "standard_error": { + "value": 0.125 + }, + "num_samples": 8 + }, + "details": { + "total_matched_trajectories": "8", + "stopping_reason_count_repetition_guard": "8", + "avg_total_tokens_target_model": "489426.00", + "total_total_tokens_target_model": "3915408.0", + "avg_total_tokens_other_models": "529.50", + "total_total_tokens_other_models": "4236.0", + "avg_total_tokens_all_models": "489955.50", + "total_total_tokens_all_models": "3919644.0", + "avg_turn_count": "22.62", + "total_turn_count": "181.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5.4-2026-03-05/0c929847-aa17-4921-8d77-09abcc13fc4e_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "9aa3265b462b0b2bb4618200f37f1869cc2e536af1ef6eeaffa75934f7c0ee9b", + "total_rows": 8 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5.4-2026-03-05/0c929847-aa17-4921-8d77-09abcc13fc4e_samples.jsonl b/data/terminalbench/openai/gpt-5.4-2026-03-05/0c929847-aa17-4921-8d77-09abcc13fc4e_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..c8b7c8926761167ff216c855da3fdf168eb415d0 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.4-2026-03-05/0c929847-aa17-4921-8d77-09abcc13fc4e_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9aa3265b462b0b2bb4618200f37f1869cc2e536af1ef6eeaffa75934f7c0ee9b +size 1255149 diff --git a/data/terminalbench/openai/gpt-5.4-2026-03-05/0dc5402a-38b5-4705-8729-ef6ed0a21209.json b/data/terminalbench/openai/gpt-5.4-2026-03-05/0dc5402a-38b5-4705-8729-ef6ed0a21209.json new file mode 100644 index 0000000000000000000000000000000000000000..f7385f66ae81dbdff9ef51f720a6177dd7e93eff --- /dev/null +++ b/data/terminalbench/openai/gpt-5.4-2026-03-05/0dc5402a-38b5-4705-8729-ef6ed0a21209.json @@ -0,0 +1,108 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5.4-2026-03-05/1786165960.1784", + "evaluation_timestamp": "1776696546.0", + "retrieved_timestamp": "1786165960.1784", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive/+1ep/dabc1965 for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 2, + "sample_ids": [ + "filter-js-from-html", + "dna-assembly" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1776696546.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.0, + "uncertainty": { + "num_samples": 2 + }, + "details": { + "total_matched_trajectories": "1", + "stopping_reason_count_repetition_guard": "1", + "avg_total_tokens_target_model": "432646.00", + "total_total_tokens_target_model": "432646.0", + "avg_total_tokens_other_models": "741.00", + "total_total_tokens_other_models": "741.0", + "avg_total_tokens_all_models": "433387.00", + "total_total_tokens_all_models": "433387.0", + "avg_turn_count": "17.00", + "total_turn_count": "17.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "3600", + "attempt_timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5.4-2026-03-05/0dc5402a-38b5-4705-8729-ef6ed0a21209_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "c74fe00a05b00c388f8990031323ddea7477b14e30f50d5261014f0ced460d2c", + "total_rows": 2 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5.4-2026-03-05/0dc5402a-38b5-4705-8729-ef6ed0a21209_samples.jsonl b/data/terminalbench/openai/gpt-5.4-2026-03-05/0dc5402a-38b5-4705-8729-ef6ed0a21209_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..e62959b959b314f5fbd5d143c334555e61ecaa6e --- /dev/null +++ b/data/terminalbench/openai/gpt-5.4-2026-03-05/0dc5402a-38b5-4705-8729-ef6ed0a21209_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c74fe00a05b00c388f8990031323ddea7477b14e30f50d5261014f0ced460d2c +size 592655 diff --git a/data/terminalbench/openai/gpt-5.4-2026-03-05/187a4bad-11bf-4a46-908d-64a51c95c965.json b/data/terminalbench/openai/gpt-5.4-2026-03-05/187a4bad-11bf-4a46-908d-64a51c95c965.json new file mode 100644 index 0000000000000000000000000000000000000000..3f5e679d5abf33d94167ab03d64678b62ffa8ce0 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.4-2026-03-05/187a4bad-11bf-4a46-908d-64a51c95c965.json @@ -0,0 +1,114 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5.4-2026-03-05/1786182964.203848", + "evaluation_timestamp": "1774895514.0", + "retrieved_timestamp": "1786182964.203848", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.191.dev111+g7c0c4345,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 89, + "sample_ids": [ + "gpt2-codegolf", + "llm-inference-batching-scheduler", + "break-filter-js-from-html", + "reshard-c4-data", + "write-compressor" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1774895514.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.8, + "uncertainty": { + "standard_error": { + "value": 0.20000000000000004 + }, + "num_samples": 10 + }, + "details": { + "total_matched_trajectories": "10", + "stopping_reason_count_completed_on_successful_submit": "8", + "stopping_reason_count_repetition_guard": "1", + "stopping_reason_count_token_limit": "1", + "avg_total_tokens_target_model": "1347733.50", + "total_total_tokens_target_model": "13477335.0", + "avg_total_tokens_other_models": "0.00", + "total_total_tokens_other_models": "0.0", + "avg_total_tokens_all_models": "1347733.50", + "total_total_tokens_all_models": "13477335.0", + "avg_turn_count": "32.40", + "total_turn_count": "324.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": false, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\"}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "max_retries": "3", + "attempt_timeout": "1200", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "64000" + } + }, + "eval_limits": { + "token_limit": 6000000 + }, + "sandbox": {}, + "max_attempts": 3 + }, + "additional_details": { + "max_retries": "3" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5.4-2026-03-05/187a4bad-11bf-4a46-908d-64a51c95c965_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "7744b0b8afe675693963086b6c89222ee9ab8cc5ac16750ba240af96a60188f5", + "total_rows": 10 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5.4-2026-03-05/187a4bad-11bf-4a46-908d-64a51c95c965_samples.jsonl b/data/terminalbench/openai/gpt-5.4-2026-03-05/187a4bad-11bf-4a46-908d-64a51c95c965_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..6104c8e203aa5cb04fc036c11dfda0436bb6d5fd --- /dev/null +++ b/data/terminalbench/openai/gpt-5.4-2026-03-05/187a4bad-11bf-4a46-908d-64a51c95c965_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7744b0b8afe675693963086b6c89222ee9ab8cc5ac16750ba240af96a60188f5 +size 3338875 diff --git a/data/terminalbench/openai/gpt-5.4-2026-03-05/1eae4c0f-23ce-4271-89a8-cd53473852ba.json b/data/terminalbench/openai/gpt-5.4-2026-03-05/1eae4c0f-23ce-4271-89a8-cd53473852ba.json new file mode 100644 index 0000000000000000000000000000000000000000..061e9b00c089ca14f662db6d20054010b25d3b71 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.4-2026-03-05/1eae4c0f-23ce-4271-89a8-cd53473852ba.json @@ -0,0 +1,107 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5.4-2026-03-05/1786163337.831691", + "evaluation_timestamp": "1777298324.0", + "retrieved_timestamp": "1786163337.831691", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive+C/+1ep/6ad0425d for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 1, + "sample_ids": [ + "compile-compcert" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1777298324.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 1.0, + "uncertainty": { + "num_samples": 1 + }, + "details": { + "total_matched_trajectories": "1", + "stopping_reason_count_completed_on_successful_submit": "1", + "avg_total_tokens_target_model": "2600006.00", + "total_total_tokens_target_model": "2600006.0", + "avg_total_tokens_other_models": "0.00", + "total_total_tokens_other_models": "0.0", + "avg_total_tokens_all_models": "2600006.00", + "total_total_tokens_all_models": "2600006.0", + "avg_turn_count": "55.00", + "total_turn_count": "55.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5.4-2026-03-05/1eae4c0f-23ce-4271-89a8-cd53473852ba_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "d30f3ed87e57625c663ff084adfd32be100388d40a651a275ce97c4b0636b35f", + "total_rows": 1 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5.4-2026-03-05/1eae4c0f-23ce-4271-89a8-cd53473852ba_samples.jsonl b/data/terminalbench/openai/gpt-5.4-2026-03-05/1eae4c0f-23ce-4271-89a8-cd53473852ba_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..c03a32f80cef89abe5f42cbb7420a6a5a9ee3951 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.4-2026-03-05/1eae4c0f-23ce-4271-89a8-cd53473852ba_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d30f3ed87e57625c663ff084adfd32be100388d40a651a275ce97c4b0636b35f +size 403353 diff --git a/data/terminalbench/openai/gpt-5.4-2026-03-05/21380eeb-cefb-4642-aaf3-77455603160f.json b/data/terminalbench/openai/gpt-5.4-2026-03-05/21380eeb-cefb-4642-aaf3-77455603160f.json new file mode 100644 index 0000000000000000000000000000000000000000..113cd030beb61f4e706f7ce8cc5ac2772fc9b31c --- /dev/null +++ b/data/terminalbench/openai/gpt-5.4-2026-03-05/21380eeb-cefb-4642-aaf3-77455603160f.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5.4-2026-03-05/1786167728.768656", + "evaluation_timestamp": "1776904610.0", + "retrieved_timestamp": "1786167728.768656", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5.4-2026-03-05/21380eeb-cefb-4642-aaf3-77455603160f_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "422d923397db212b1a15deda4ea24b4856a68ae9bc3bb80cbe635c4a3a371a77", + "total_rows": 15 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5.4-2026-03-05/21380eeb-cefb-4642-aaf3-77455603160f_samples.jsonl b/data/terminalbench/openai/gpt-5.4-2026-03-05/21380eeb-cefb-4642-aaf3-77455603160f_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..b54be69de1e0243b87d6da969caa0074f9c7179f --- /dev/null +++ b/data/terminalbench/openai/gpt-5.4-2026-03-05/21380eeb-cefb-4642-aaf3-77455603160f_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:422d923397db212b1a15deda4ea24b4856a68ae9bc3bb80cbe635c4a3a371a77 +size 3774562 diff --git a/data/terminalbench/openai/gpt-5.4-2026-03-05/229d61ed-7dea-46da-8a82-53ef5d2a1b6e.json b/data/terminalbench/openai/gpt-5.4-2026-03-05/229d61ed-7dea-46da-8a82-53ef5d2a1b6e.json new file mode 100644 index 0000000000000000000000000000000000000000..6069f5c92dbbad157f0738ec7478ca8e918a3ebe --- /dev/null +++ b/data/terminalbench/openai/gpt-5.4-2026-03-05/229d61ed-7dea-46da-8a82-53ef5d2a1b6e.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5.4-2026-03-05/1786169606.719695", + "evaluation_timestamp": "1776601454.0", + "retrieved_timestamp": "1786169606.719695", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5.4-2026-03-05/229d61ed-7dea-46da-8a82-53ef5d2a1b6e_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "b46bc76dd62e7bfd38cb064312357495633e14b698bfda03590122fa142ae9ee", + "total_rows": 10 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5.4-2026-03-05/229d61ed-7dea-46da-8a82-53ef5d2a1b6e_samples.jsonl b/data/terminalbench/openai/gpt-5.4-2026-03-05/229d61ed-7dea-46da-8a82-53ef5d2a1b6e_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..a5150de4e002036e682f1e18e338282f9d97d093 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.4-2026-03-05/229d61ed-7dea-46da-8a82-53ef5d2a1b6e_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b46bc76dd62e7bfd38cb064312357495633e14b698bfda03590122fa142ae9ee +size 2025111 diff --git a/data/terminalbench/openai/gpt-5.4-2026-03-05/2c86add7-fbd7-44be-b472-75e19a44289b.json b/data/terminalbench/openai/gpt-5.4-2026-03-05/2c86add7-fbd7-44be-b472-75e19a44289b.json new file mode 100644 index 0000000000000000000000000000000000000000..5549f0ef6cbb9bf16bcc1496ab65f144c3dee809 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.4-2026-03-05/2c86add7-fbd7-44be-b472-75e19a44289b.json @@ -0,0 +1,128 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5.4-2026-03-05/1786163366.40188", + "evaluation_timestamp": "1777041955.0", + "retrieved_timestamp": "1786163366.40188", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive+C/+3ep/3d693d54 for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 16, + "sample_ids": [ + "winning-avg-corewars", + "path-tracing-reverse", + "regex-chess", + "adaptive-rejection-sampler", + "configure-git-webserver", + "chess-best-move", + "cobol-modernization", + "overfull-hbox", + "polyglot-rust-c", + "hf-model-inference", + "compile-compcert", + "headless-terminal", + "fix-git", + "qemu-startup", + "build-pov-ray", + "kv-store-grpc" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1777041955.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.7083333333333333, + "uncertainty": { + "standard_error": { + "value": 0.11334558757279538 + }, + "num_samples": 45 + }, + "details": { + "total_matched_trajectories": "45", + "stopping_reason_count_completed_on_successful_submit": "31", + "stopping_reason_count_repetition_guard": "3", + "stopping_reason_count_completed_with_submit": "5", + "stopping_reason_count_token_limit": "6", + "avg_total_tokens_target_model": "1780882.64", + "total_total_tokens_target_model": "80139719.0", + "avg_total_tokens_other_models": "120.98", + "total_total_tokens_other_models": "5444.0", + "avg_total_tokens_all_models": "1781003.62", + "total_total_tokens_all_models": "80145163.0", + "avg_turn_count": "34.44", + "total_turn_count": "1550.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5.4-2026-03-05/2c86add7-fbd7-44be-b472-75e19a44289b_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "63bbc80d66c079233d2520a99a6771a11af89ac915673f7eee6c7f82b6302ffd", + "total_rows": 45 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5.4-2026-03-05/2c86add7-fbd7-44be-b472-75e19a44289b_samples.jsonl b/data/terminalbench/openai/gpt-5.4-2026-03-05/2c86add7-fbd7-44be-b472-75e19a44289b_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..f00db0718c8af6905fad3af1de03ad7ce68a57c6 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.4-2026-03-05/2c86add7-fbd7-44be-b472-75e19a44289b_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:63bbc80d66c079233d2520a99a6771a11af89ac915673f7eee6c7f82b6302ffd +size 15434820 diff --git a/data/terminalbench/openai/gpt-5.4-2026-03-05/36f0a933-b053-4091-8c8a-6bc1d4c6b0fc.json b/data/terminalbench/openai/gpt-5.4-2026-03-05/36f0a933-b053-4091-8c8a-6bc1d4c6b0fc.json new file mode 100644 index 0000000000000000000000000000000000000000..2f84c97590fb717b9a94fdab55e5b14bd3d01004 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.4-2026-03-05/36f0a933-b053-4091-8c8a-6bc1d4c6b0fc.json @@ -0,0 +1,118 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5.4-2026-03-05/1786165974.76159", + "evaluation_timestamp": "1776439703.0", + "retrieved_timestamp": "1786165974.76159", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive/+2ep/f2ebf8d5 for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 11, + "sample_ids": [ + "portfolio-optimization", + "modernize-scientific-stack", + "path-tracing", + "prove-plus-comm", + "feal-linear-cryptanalysis", + "caffe-cifar-10", + "distribution-search", + "custom-memory-heap-crash", + "adaptive-rejection-sampler", + "multi-source-data-merger", + "configure-git-webserver" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1776439703.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.7727272727272727, + "uncertainty": { + "standard_error": { + "value": 0.12364973189759493 + }, + "num_samples": 22 + }, + "details": { + "total_matched_trajectories": "22", + "stopping_reason_count_repetition_guard": "22", + "avg_total_tokens_target_model": "654498.23", + "total_total_tokens_target_model": "14398961.0", + "avg_total_tokens_other_models": "949.91", + "total_total_tokens_other_models": "20898.0", + "avg_total_tokens_all_models": "655448.14", + "total_total_tokens_all_models": "14419859.0", + "avg_turn_count": "19.73", + "total_turn_count": "434.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "1200", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5.4-2026-03-05/36f0a933-b053-4091-8c8a-6bc1d4c6b0fc_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "1421226fba4a938e518e450eaba246b204e6d67c913f46fac50602a6be79137d", + "total_rows": 22 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5.4-2026-03-05/36f0a933-b053-4091-8c8a-6bc1d4c6b0fc_samples.jsonl b/data/terminalbench/openai/gpt-5.4-2026-03-05/36f0a933-b053-4091-8c8a-6bc1d4c6b0fc_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..19bbab5b7071573520aef20580637bc3eee2dd4f --- /dev/null +++ b/data/terminalbench/openai/gpt-5.4-2026-03-05/36f0a933-b053-4091-8c8a-6bc1d4c6b0fc_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1421226fba4a938e518e450eaba246b204e6d67c913f46fac50602a6be79137d +size 3184591 diff --git a/data/terminalbench/openai/gpt-5.4-2026-03-05/38808746-c37b-49ad-af0f-7c4eb9bb445e.json b/data/terminalbench/openai/gpt-5.4-2026-03-05/38808746-c37b-49ad-af0f-7c4eb9bb445e.json new file mode 100644 index 0000000000000000000000000000000000000000..f9158199e41928db06ff79f9122f857623a268f9 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.4-2026-03-05/38808746-c37b-49ad-af0f-7c4eb9bb445e.json @@ -0,0 +1,115 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5.4-2026-03-05/1786163348.378266", + "evaluation_timestamp": "1777478370.0", + "retrieved_timestamp": "1786163348.378266", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive+C/+1ep/ad00e1f0 for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 5, + "sample_ids": [ + "gpt2-codegolf", + "regex-chess", + "train-fasttext", + "pytorch-model-recovery", + "dna-insert" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1777478370.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.6666666666666666, + "uncertainty": { + "standard_error": { + "value": 0.33333333333333337 + }, + "num_samples": 3 + }, + "details": { + "total_matched_trajectories": "3", + "stopping_reason_count_repetition_guard": "2", + "stopping_reason_count_completed_on_successful_submit": "1", + "avg_total_tokens_target_model": "1590897.67", + "total_total_tokens_target_model": "4772693.0", + "avg_total_tokens_other_models": "0.00", + "total_total_tokens_other_models": "0.0", + "avg_total_tokens_all_models": "1590897.67", + "total_total_tokens_all_models": "4772693.0", + "avg_turn_count": "33.00", + "total_turn_count": "99.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "32", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "32", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5.4-2026-03-05/38808746-c37b-49ad-af0f-7c4eb9bb445e_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "a887fff30ef891a4a33cc461be9af0d0044953333a41a720c8d0051edf540161", + "total_rows": 3 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5.4-2026-03-05/38808746-c37b-49ad-af0f-7c4eb9bb445e_samples.jsonl b/data/terminalbench/openai/gpt-5.4-2026-03-05/38808746-c37b-49ad-af0f-7c4eb9bb445e_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..187197c2820ab2b06f4cc5dd037b1524b34fcaec --- /dev/null +++ b/data/terminalbench/openai/gpt-5.4-2026-03-05/38808746-c37b-49ad-af0f-7c4eb9bb445e_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a887fff30ef891a4a33cc461be9af0d0044953333a41a720c8d0051edf540161 +size 2829356 diff --git a/data/terminalbench/openai/gpt-5.4-2026-03-05/445626bb-f342-4e8e-8ec3-b19aba4fa7cc.json b/data/terminalbench/openai/gpt-5.4-2026-03-05/445626bb-f342-4e8e-8ec3-b19aba4fa7cc.json new file mode 100644 index 0000000000000000000000000000000000000000..ea72ed9cef28b60faf8940f3f1ad120814ec22d0 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.4-2026-03-05/445626bb-f342-4e8e-8ec3-b19aba4fa7cc.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5.4-2026-03-05/1786169664.123357", + "evaluation_timestamp": "1776680886.0", + "retrieved_timestamp": "1786169664.123357", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5.4-2026-03-05/445626bb-f342-4e8e-8ec3-b19aba4fa7cc_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "0b3f5c8bd08096256a32f802ae0effaffe46232e85768834976eeaabb32ae20d", + "total_rows": 23 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5.4-2026-03-05/445626bb-f342-4e8e-8ec3-b19aba4fa7cc_samples.jsonl b/data/terminalbench/openai/gpt-5.4-2026-03-05/445626bb-f342-4e8e-8ec3-b19aba4fa7cc_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..714967ae5b8baded46504aa512d83a42ebcaf72d --- /dev/null +++ b/data/terminalbench/openai/gpt-5.4-2026-03-05/445626bb-f342-4e8e-8ec3-b19aba4fa7cc_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0b3f5c8bd08096256a32f802ae0effaffe46232e85768834976eeaabb32ae20d +size 3747552 diff --git a/data/terminalbench/openai/gpt-5.4-2026-03-05/4d7b79e2-6400-4154-8a10-ac3663dd9327.json b/data/terminalbench/openai/gpt-5.4-2026-03-05/4d7b79e2-6400-4154-8a10-ac3663dd9327.json new file mode 100644 index 0000000000000000000000000000000000000000..714a0909da633312f2ba0f18ebe6a01cb858d2ee --- /dev/null +++ b/data/terminalbench/openai/gpt-5.4-2026-03-05/4d7b79e2-6400-4154-8a10-ac3663dd9327.json @@ -0,0 +1,107 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5.4-2026-03-05/1786165970.89075", + "evaluation_timestamp": "1776696546.0", + "retrieved_timestamp": "1786165970.89075", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive/+2ep/dc014c9a for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 1, + "sample_ids": [ + "schemelike-metacircular-eval" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1776696546.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.0, + "uncertainty": { + "num_samples": 2 + }, + "details": { + "total_matched_trajectories": "2", + "stopping_reason_count_repetition_guard": "2", + "avg_total_tokens_target_model": "824613.00", + "total_total_tokens_target_model": "1649226.0", + "avg_total_tokens_other_models": "9890.50", + "total_total_tokens_other_models": "19781.0", + "avg_total_tokens_all_models": "834503.50", + "total_total_tokens_all_models": "1669007.0", + "avg_turn_count": "20.00", + "total_turn_count": "40.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "3600", + "attempt_timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5.4-2026-03-05/4d7b79e2-6400-4154-8a10-ac3663dd9327_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "430c1897685aa60ad9c3587956604c50e67df5525ab023c9f342c7a507885cdf", + "total_rows": 2 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5.4-2026-03-05/4d7b79e2-6400-4154-8a10-ac3663dd9327_samples.jsonl b/data/terminalbench/openai/gpt-5.4-2026-03-05/4d7b79e2-6400-4154-8a10-ac3663dd9327_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..c63ae495602ab01ac998be6c46bd6147c863cd08 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.4-2026-03-05/4d7b79e2-6400-4154-8a10-ac3663dd9327_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:430c1897685aa60ad9c3587956604c50e67df5525ab023c9f342c7a507885cdf +size 658185 diff --git a/data/terminalbench/openai/gpt-5.4-2026-03-05/5cfc7af6-64bd-4dfa-875e-b03c66c9ec66.json b/data/terminalbench/openai/gpt-5.4-2026-03-05/5cfc7af6-64bd-4dfa-875e-b03c66c9ec66.json new file mode 100644 index 0000000000000000000000000000000000000000..7b9950143e5cd16c3db8019b832a39903d94cda0 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.4-2026-03-05/5cfc7af6-64bd-4dfa-875e-b03c66c9ec66.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5.4-2026-03-05/1786167740.956419", + "evaluation_timestamp": "1776704388.0", + "retrieved_timestamp": "1786167740.956419", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5.4-2026-03-05/5cfc7af6-64bd-4dfa-875e-b03c66c9ec66_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "c82f7ffd53074c2658be6597b9b371a9b595f0e6e75402c0ab8dc1dbebc9f130", + "total_rows": 56 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5.4-2026-03-05/5cfc7af6-64bd-4dfa-875e-b03c66c9ec66_samples.jsonl b/data/terminalbench/openai/gpt-5.4-2026-03-05/5cfc7af6-64bd-4dfa-875e-b03c66c9ec66_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..c747968908f518cd66569004fc2cd536deb1b895 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.4-2026-03-05/5cfc7af6-64bd-4dfa-875e-b03c66c9ec66_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c82f7ffd53074c2658be6597b9b371a9b595f0e6e75402c0ab8dc1dbebc9f130 +size 20726158 diff --git a/data/terminalbench/openai/gpt-5.4-2026-03-05/5f0410ba-83d9-42bb-9e98-18aa17888f30.json b/data/terminalbench/openai/gpt-5.4-2026-03-05/5f0410ba-83d9-42bb-9e98-18aa17888f30.json new file mode 100644 index 0000000000000000000000000000000000000000..8a68de6c3fe6dc8f1cf08b665e5d058d9856c532 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.4-2026-03-05/5f0410ba-83d9-42bb-9e98-18aa17888f30.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5.4-2026-03-05/1786169611.523839", + "evaluation_timestamp": "1776346958.0", + "retrieved_timestamp": "1786169611.523839", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5.4-2026-03-05/5f0410ba-83d9-42bb-9e98-18aa17888f30_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "392f8d8e53973e3a5c8ecb85addfadd26c1217509b629f56dc3fdf5a00797552", + "total_rows": 23 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5.4-2026-03-05/5f0410ba-83d9-42bb-9e98-18aa17888f30_samples.jsonl b/data/terminalbench/openai/gpt-5.4-2026-03-05/5f0410ba-83d9-42bb-9e98-18aa17888f30_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..0df674da5800277aca09ff13ad6ff94fe8b5d891 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.4-2026-03-05/5f0410ba-83d9-42bb-9e98-18aa17888f30_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:392f8d8e53973e3a5c8ecb85addfadd26c1217509b629f56dc3fdf5a00797552 +size 4277246 diff --git a/data/terminalbench/openai/gpt-5.4-2026-03-05/66814689-659f-4db6-9eb8-d8159707f9bc.json b/data/terminalbench/openai/gpt-5.4-2026-03-05/66814689-659f-4db6-9eb8-d8159707f9bc.json new file mode 100644 index 0000000000000000000000000000000000000000..9d67e1ff71b947a84002b54360c40cebda0e283b --- /dev/null +++ b/data/terminalbench/openai/gpt-5.4-2026-03-05/66814689-659f-4db6-9eb8-d8159707f9bc.json @@ -0,0 +1,137 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5.4-2026-03-05/1786165990.397546", + "evaluation_timestamp": "1776629382.0", + "retrieved_timestamp": "1786165990.397546", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive/+2ep/07a40ddf for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 29, + "sample_ids": [ + "train-fasttext", + "install-windows-3.11", + "torch-pipeline-parallelism", + "tune-mjcf", + "extract-moves-from-video", + "gcode-to-text", + "make-mips-interpreter", + "count-dataset-tokens", + "circuit-fibsqrt", + "mteb-leaderboard", + "mailman", + "raman-fitting", + "extract-elf", + "protein-assembly", + "code-from-image", + "fix-code-vulnerability", + "financial-document-processor", + "regex-log", + "feal-differential-cryptanalysis", + "polyglot-c-py", + "cancel-async-tasks", + "bn-fit-modify", + "fix-ocaml-gc", + "git-leak-recovery", + "build-cython-ext", + "pytorch-model-recovery", + "sam-cell-seg", + "model-extraction-relu-logits", + "nginx-request-logging" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1776629382.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.6551724137931034, + "uncertainty": { + "standard_error": { + "value": 0.08268675040194343 + }, + "num_samples": 57 + }, + "details": { + "total_matched_trajectories": "57", + "stopping_reason_count_repetition_guard": "55", + "stopping_reason_count_token_limit": "2", + "avg_total_tokens_target_model": "1721590.70", + "total_total_tokens_target_model": "98130670.0", + "avg_total_tokens_other_models": "826.40", + "total_total_tokens_other_models": "47105.0", + "avg_total_tokens_all_models": "1722417.11", + "total_total_tokens_all_models": "98177775.0", + "avg_turn_count": "39.49", + "total_turn_count": "2251.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5.4-2026-03-05/66814689-659f-4db6-9eb8-d8159707f9bc_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "b5c712fbd6b5d152744ead9976e9c7747095311c4c24deb7240f22856b255402", + "total_rows": 57 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5.4-2026-03-05/66814689-659f-4db6-9eb8-d8159707f9bc_samples.jsonl b/data/terminalbench/openai/gpt-5.4-2026-03-05/66814689-659f-4db6-9eb8-d8159707f9bc_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..29b1640f5ceeb243d1d686c61382123b98dd72e0 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.4-2026-03-05/66814689-659f-4db6-9eb8-d8159707f9bc_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b5c712fbd6b5d152744ead9976e9c7747095311c4c24deb7240f22856b255402 +size 19818275 diff --git a/data/terminalbench/openai/gpt-5.4-2026-03-05/6d7fbbe3-81cb-4375-b27d-73f05bfc1c8c.json b/data/terminalbench/openai/gpt-5.4-2026-03-05/6d7fbbe3-81cb-4375-b27d-73f05bfc1c8c.json new file mode 100644 index 0000000000000000000000000000000000000000..dbd05b47733651e4ef98de235edc24da676791f3 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.4-2026-03-05/6d7fbbe3-81cb-4375-b27d-73f05bfc1c8c.json @@ -0,0 +1,120 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5.4-2026-03-05/1786163356.75545", + "evaluation_timestamp": "1777298418.0", + "retrieved_timestamp": "1786163356.75545", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive+C/+2ep/19e9a262 for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 9, + "sample_ids": [ + "regex-chess", + "crack-7z-hash", + "git-multibranch", + "train-fasttext", + "video-processing", + "qemu-alpine-ssh", + "torch-pipeline-parallelism", + "tune-mjcf", + "count-dataset-tokens" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1777298418.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.8888888888888888, + "uncertainty": { + "standard_error": { + "value": 0.11111111111111112 + }, + "num_samples": 15 + }, + "details": { + "total_matched_trajectories": "15", + "stopping_reason_count_completed_on_successful_submit": "8", + "stopping_reason_count_repetition_guard": "6", + "stopping_reason_count_token_limit": "1", + "avg_total_tokens_target_model": "1502963.73", + "total_total_tokens_target_model": "22544456.0", + "avg_total_tokens_other_models": "0.00", + "total_total_tokens_other_models": "0.0", + "avg_total_tokens_all_models": "1502963.73", + "total_total_tokens_all_models": "22544456.0", + "avg_turn_count": "34.67", + "total_turn_count": "520.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5.4-2026-03-05/6d7fbbe3-81cb-4375-b27d-73f05bfc1c8c_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "cb57bd1cd0bb28b13aa172a434ff653d26908aedd923d7787b30aef75b3c2568", + "total_rows": 15 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5.4-2026-03-05/6d7fbbe3-81cb-4375-b27d-73f05bfc1c8c_samples.jsonl b/data/terminalbench/openai/gpt-5.4-2026-03-05/6d7fbbe3-81cb-4375-b27d-73f05bfc1c8c_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..9c380f73e44b6118169688bd6f052c9f80d7c945 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.4-2026-03-05/6d7fbbe3-81cb-4375-b27d-73f05bfc1c8c_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cb57bd1cd0bb28b13aa172a434ff653d26908aedd923d7787b30aef75b3c2568 +size 7574008 diff --git a/data/terminalbench/openai/gpt-5.4-2026-03-05/72c2c51c-8977-4241-ba09-58fb24bfb5a5.json b/data/terminalbench/openai/gpt-5.4-2026-03-05/72c2c51c-8977-4241-ba09-58fb24bfb5a5.json new file mode 100644 index 0000000000000000000000000000000000000000..80ee64a0bdfa70033ce1207a3e1255de024477ae --- /dev/null +++ b/data/terminalbench/openai/gpt-5.4-2026-03-05/72c2c51c-8977-4241-ba09-58fb24bfb5a5.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5.4-2026-03-05/1786169630.507742", + "evaluation_timestamp": "1776520455.0", + "retrieved_timestamp": "1786169630.507742", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5.4-2026-03-05/72c2c51c-8977-4241-ba09-58fb24bfb5a5_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "f38a666f80367f06ba06e19622830c7d90c9ab8d8ca5e7ea550e054c0ff28fe7", + "total_rows": 40 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5.4-2026-03-05/72c2c51c-8977-4241-ba09-58fb24bfb5a5_samples.jsonl b/data/terminalbench/openai/gpt-5.4-2026-03-05/72c2c51c-8977-4241-ba09-58fb24bfb5a5_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..ee69fcfdaa075d53c6a5db5d11773ed7557061b6 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.4-2026-03-05/72c2c51c-8977-4241-ba09-58fb24bfb5a5_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f38a666f80367f06ba06e19622830c7d90c9ab8d8ca5e7ea550e054c0ff28fe7 +size 9788006 diff --git a/data/terminalbench/openai/gpt-5.4-2026-03-05/78613643-7f04-4267-a40c-59f84b41f254.json b/data/terminalbench/openai/gpt-5.4-2026-03-05/78613643-7f04-4267-a40c-59f84b41f254.json new file mode 100644 index 0000000000000000000000000000000000000000..390b396f9057186cd3e1e8c194dec13eb6bf407a --- /dev/null +++ b/data/terminalbench/openai/gpt-5.4-2026-03-05/78613643-7f04-4267-a40c-59f84b41f254.json @@ -0,0 +1,110 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5.4-2026-03-05/1786163346.009698", + "evaluation_timestamp": "1776810309.0", + "retrieved_timestamp": "1786163346.009698", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive+C/+1ep/686ca904 for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 4, + "sample_ids": [ + "mteb-retrieve", + "pypi-server", + "custom-memory-heap-crash", + "multi-source-data-merger" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1776810309.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 1.0, + "uncertainty": { + "num_samples": 4 + }, + "details": { + "total_matched_trajectories": "4", + "stopping_reason_count_completed_on_successful_submit": "4", + "avg_total_tokens_target_model": "123037.50", + "total_total_tokens_target_model": "492150.0", + "avg_total_tokens_other_models": "0.00", + "total_total_tokens_other_models": "0.0", + "avg_total_tokens_all_models": "123037.50", + "total_total_tokens_all_models": "492150.0", + "avg_turn_count": "13.00", + "total_turn_count": "52.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5.4-2026-03-05/78613643-7f04-4267-a40c-59f84b41f254_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "9a61dafdfa49fe8e720b70b91a5797767203459d39687c4073dd95d445f934be", + "total_rows": 4 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5.4-2026-03-05/78613643-7f04-4267-a40c-59f84b41f254_samples.jsonl b/data/terminalbench/openai/gpt-5.4-2026-03-05/78613643-7f04-4267-a40c-59f84b41f254_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..dba877de801077c60a3dd9310c8411bf99285a0f --- /dev/null +++ b/data/terminalbench/openai/gpt-5.4-2026-03-05/78613643-7f04-4267-a40c-59f84b41f254_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9a61dafdfa49fe8e720b70b91a5797767203459d39687c4073dd95d445f934be +size 308427 diff --git a/data/terminalbench/openai/gpt-5.4-2026-03-05/7a387d41-9ebf-420d-871d-3ff815fb1c94.json b/data/terminalbench/openai/gpt-5.4-2026-03-05/7a387d41-9ebf-420d-871d-3ff815fb1c94.json new file mode 100644 index 0000000000000000000000000000000000000000..9c79fac04d802c5d48ce37a168a637888efa1699 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.4-2026-03-05/7a387d41-9ebf-420d-871d-3ff815fb1c94.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5.4-2026-03-05/1786169649.541169", + "evaluation_timestamp": "1776653347.0", + "retrieved_timestamp": "1786169649.541169", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5.4-2026-03-05/7a387d41-9ebf-420d-871d-3ff815fb1c94_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "be3781b71c220563bbe2da5fe3faf1564b813f33d56894c6e18e5ab21df88321", + "total_rows": 30 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5.4-2026-03-05/7a387d41-9ebf-420d-871d-3ff815fb1c94_samples.jsonl b/data/terminalbench/openai/gpt-5.4-2026-03-05/7a387d41-9ebf-420d-871d-3ff815fb1c94_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..0d6b8c2bcd789f16538798af0139e322c14bc817 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.4-2026-03-05/7a387d41-9ebf-420d-871d-3ff815fb1c94_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:be3781b71c220563bbe2da5fe3faf1564b813f33d56894c6e18e5ab21df88321 +size 5800169 diff --git a/data/terminalbench/openai/gpt-5.4-2026-03-05/80580ddb-7698-4c98-8105-ddc60282cd5a.json b/data/terminalbench/openai/gpt-5.4-2026-03-05/80580ddb-7698-4c98-8105-ddc60282cd5a.json new file mode 100644 index 0000000000000000000000000000000000000000..56001d8ba9a73133ab4f44e16db67eef02a6d900 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.4-2026-03-05/80580ddb-7698-4c98-8105-ddc60282cd5a.json @@ -0,0 +1,112 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5.4-2026-03-05/1786163339.908598", + "evaluation_timestamp": "1777586285.0", + "retrieved_timestamp": "1786163339.908598", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive+C/+1ep/03d08750 for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 2, + "sample_ids": [ + "regex-chess", + "train-fasttext" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1777586285.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.5, + "uncertainty": { + "standard_error": { + "value": 0.5 + }, + "num_samples": 2 + }, + "details": { + "total_matched_trajectories": "2", + "stopping_reason_count_completed_on_successful_submit": "1", + "stopping_reason_count_token_limit": "1", + "avg_total_tokens_target_model": "5759306.00", + "total_total_tokens_target_model": "11518612.0", + "avg_total_tokens_other_models": "0.00", + "total_total_tokens_other_models": "0.0", + "avg_total_tokens_all_models": "5759306.00", + "total_total_tokens_all_models": "11518612.0", + "avg_turn_count": "54.50", + "total_turn_count": "109.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "16", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5.4-2026-03-05/80580ddb-7698-4c98-8105-ddc60282cd5a_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "1b0e7c4483569be772c5472b58022b7b1500b54e9d80d978afc078b0fd150c53", + "total_rows": 2 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5.4-2026-03-05/80580ddb-7698-4c98-8105-ddc60282cd5a_samples.jsonl b/data/terminalbench/openai/gpt-5.4-2026-03-05/80580ddb-7698-4c98-8105-ddc60282cd5a_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..fc9440bb01f83964705f18f5e92746fde6e7abe5 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.4-2026-03-05/80580ddb-7698-4c98-8105-ddc60282cd5a_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1b0e7c4483569be772c5472b58022b7b1500b54e9d80d978afc078b0fd150c53 +size 2082586 diff --git a/data/terminalbench/openai/gpt-5.4-2026-03-05/84f59525-6682-4c8b-90d6-8302b588a541.json b/data/terminalbench/openai/gpt-5.4-2026-03-05/84f59525-6682-4c8b-90d6-8302b588a541.json new file mode 100644 index 0000000000000000000000000000000000000000..fd403133f3abbe702df67cbc2ec95b3fb475b4ca --- /dev/null +++ b/data/terminalbench/openai/gpt-5.4-2026-03-05/84f59525-6682-4c8b-90d6-8302b588a541.json @@ -0,0 +1,113 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5.4-2026-03-05/1786163375.005341", + "evaluation_timestamp": "1777335415.0", + "retrieved_timestamp": "1786163375.005341", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive+C/+4ep/763118a5 for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 3, + "sample_ids": [ + "schemelike-metacircular-eval", + "constraints-scheduling", + "vulnerable-secret" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1777335415.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.9166666666666666, + "uncertainty": { + "standard_error": { + "value": 0.08333333333333334 + }, + "num_samples": 12 + }, + "details": { + "total_matched_trajectories": "12", + "stopping_reason_count_completed_on_successful_submit": "10", + "stopping_reason_count_repetition_guard": "2", + "avg_total_tokens_target_model": "938828.17", + "total_total_tokens_target_model": "11265938.0", + "avg_total_tokens_other_models": "0.00", + "total_total_tokens_other_models": "0.0", + "avg_total_tokens_all_models": "938828.17", + "total_total_tokens_all_models": "11265938.0", + "avg_turn_count": "19.58", + "total_turn_count": "235.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5.4-2026-03-05/84f59525-6682-4c8b-90d6-8302b588a541_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "fd1cff7a944f6daa75182a5f0e0b3f02c8c8532fed692c5c0312f870f93526a3", + "total_rows": 12 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5.4-2026-03-05/84f59525-6682-4c8b-90d6-8302b588a541_samples.jsonl b/data/terminalbench/openai/gpt-5.4-2026-03-05/84f59525-6682-4c8b-90d6-8302b588a541_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..5382ba1650cb24ec09e295f72ccdf4d376da2f0c --- /dev/null +++ b/data/terminalbench/openai/gpt-5.4-2026-03-05/84f59525-6682-4c8b-90d6-8302b588a541_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fd1cff7a944f6daa75182a5f0e0b3f02c8c8532fed692c5c0312f870f93526a3 +size 3270312 diff --git a/data/terminalbench/openai/gpt-5.4-2026-03-05/89431b6c-fc52-43e8-a1eb-da84c9c61186.json b/data/terminalbench/openai/gpt-5.4-2026-03-05/89431b6c-fc52-43e8-a1eb-da84c9c61186.json new file mode 100644 index 0000000000000000000000000000000000000000..c08e335e4484b5678831b08809aa0b7c03b3c35b --- /dev/null +++ b/data/terminalbench/openai/gpt-5.4-2026-03-05/89431b6c-fc52-43e8-a1eb-da84c9c61186.json @@ -0,0 +1,109 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5.4-2026-03-05/1786166003.844706", + "evaluation_timestamp": "1776646372.0", + "retrieved_timestamp": "1786166003.844706", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive/+3ep/8dc27caa for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 2, + "sample_ids": [ + "mcmc-sampling-stan", + "sparql-university" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1776646372.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.75, + "uncertainty": { + "standard_error": { + "value": 0.25 + }, + "num_samples": 5 + }, + "details": { + "total_matched_trajectories": "3", + "stopping_reason_count_repetition_guard": "3", + "avg_total_tokens_target_model": "450155.67", + "total_total_tokens_target_model": "1350467.0", + "avg_total_tokens_other_models": "849.00", + "total_total_tokens_other_models": "2547.0", + "avg_total_tokens_all_models": "451004.67", + "total_total_tokens_all_models": "1353014.0", + "avg_turn_count": "23.33", + "total_turn_count": "70.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5.4-2026-03-05/89431b6c-fc52-43e8-a1eb-da84c9c61186_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "75e55fc6f6a75c79e9e48511d098f6b32b39ddfe4fd09fad388f709518197e77", + "total_rows": 5 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5.4-2026-03-05/89431b6c-fc52-43e8-a1eb-da84c9c61186_samples.jsonl b/data/terminalbench/openai/gpt-5.4-2026-03-05/89431b6c-fc52-43e8-a1eb-da84c9c61186_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..3cc6fbd11c5c320ecc73379f7e11fa0712f3777c --- /dev/null +++ b/data/terminalbench/openai/gpt-5.4-2026-03-05/89431b6c-fc52-43e8-a1eb-da84c9c61186_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:75e55fc6f6a75c79e9e48511d098f6b32b39ddfe4fd09fad388f709518197e77 +size 1050177 diff --git a/data/terminalbench/openai/gpt-5.4-2026-03-05/8ba572a6-bd5d-4525-b930-419766646bf7.json b/data/terminalbench/openai/gpt-5.4-2026-03-05/8ba572a6-bd5d-4525-b930-419766646bf7.json new file mode 100644 index 0000000000000000000000000000000000000000..3a94d3bf6e2785eb5aa4cf38a22f212754561820 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.4-2026-03-05/8ba572a6-bd5d-4525-b930-419766646bf7.json @@ -0,0 +1,128 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5.4-2026-03-05/1786165956.6284", + "evaluation_timestamp": "1776629382.0", + "retrieved_timestamp": "1786165956.6284", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive/+1ep/856a14eb for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 20, + "sample_ids": [ + "regex-chess", + "torch-tensor-parallelism", + "mteb-retrieve", + "pypi-server", + "crack-7z-hash", + "chess-best-move", + "cobol-modernization", + "polyglot-rust-c", + "hf-model-inference", + "compile-compcert", + "db-wal-recovery", + "headless-terminal", + "fix-git", + "qemu-startup", + "build-pov-ray", + "git-multibranch", + "video-processing", + "qemu-alpine-ssh", + "kv-store-grpc", + "make-doom-for-mips" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1776629382.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.65, + "uncertainty": { + "standard_error": { + "value": 0.1094243309804831 + }, + "num_samples": 20 + }, + "details": { + "total_matched_trajectories": "20", + "stopping_reason_count_repetition_guard": "19", + "stopping_reason_count_token_limit": "1", + "avg_total_tokens_target_model": "1213150.55", + "total_total_tokens_target_model": "24263011.0", + "avg_total_tokens_other_models": "492.85", + "total_total_tokens_other_models": "9857.0", + "avg_total_tokens_all_models": "1213643.40", + "total_total_tokens_all_models": "24272868.0", + "avg_turn_count": "35.25", + "total_turn_count": "705.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "attempt_timeout": "3600", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5.4-2026-03-05/8ba572a6-bd5d-4525-b930-419766646bf7_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "c350942bb90c4178d536e65580fe5f530933791bd96b65d8c9bee9359673efed", + "total_rows": 20 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5.4-2026-03-05/8ba572a6-bd5d-4525-b930-419766646bf7_samples.jsonl b/data/terminalbench/openai/gpt-5.4-2026-03-05/8ba572a6-bd5d-4525-b930-419766646bf7_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..fe885030a105c5f31fe87a4989621de802ecf7cf --- /dev/null +++ b/data/terminalbench/openai/gpt-5.4-2026-03-05/8ba572a6-bd5d-4525-b930-419766646bf7_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c350942bb90c4178d536e65580fe5f530933791bd96b65d8c9bee9359673efed +size 5820713 diff --git a/data/terminalbench/openai/gpt-5.4-2026-03-05/8be2e495-c6f2-454f-bb5a-8ba6ee5b73d8.json b/data/terminalbench/openai/gpt-5.4-2026-03-05/8be2e495-c6f2-454f-bb5a-8ba6ee5b73d8.json new file mode 100644 index 0000000000000000000000000000000000000000..1ca739e7159e95741434f82bde73c6ea0d3e1e6c --- /dev/null +++ b/data/terminalbench/openai/gpt-5.4-2026-03-05/8be2e495-c6f2-454f-bb5a-8ba6ee5b73d8.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5.4-2026-03-05/1786169666.804156", + "evaluation_timestamp": "1776683465.0", + "retrieved_timestamp": "1786169666.804156", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5.4-2026-03-05/8be2e495-c6f2-454f-bb5a-8ba6ee5b73d8_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "e9ee7e8adf1f2f56673a6fb77148a1fd9f9ae7109196f6e673e83e5ea24d6ade", + "total_rows": 1 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5.4-2026-03-05/8be2e495-c6f2-454f-bb5a-8ba6ee5b73d8_samples.jsonl b/data/terminalbench/openai/gpt-5.4-2026-03-05/8be2e495-c6f2-454f-bb5a-8ba6ee5b73d8_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..2489518452c219507480538b1c1fb243b246d40c --- /dev/null +++ b/data/terminalbench/openai/gpt-5.4-2026-03-05/8be2e495-c6f2-454f-bb5a-8ba6ee5b73d8_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e9ee7e8adf1f2f56673a6fb77148a1fd9f9ae7109196f6e673e83e5ea24d6ade +size 295085 diff --git a/data/terminalbench/openai/gpt-5.4-2026-03-05/922969c3-26e3-4b15-8220-1a2ee8221bb5.json b/data/terminalbench/openai/gpt-5.4-2026-03-05/922969c3-26e3-4b15-8220-1a2ee8221bb5.json new file mode 100644 index 0000000000000000000000000000000000000000..607243261f7874eafcb9c29fc3c062ba8cc761a5 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.4-2026-03-05/922969c3-26e3-4b15-8220-1a2ee8221bb5.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5.4-2026-03-05/1786169658.404657", + "evaluation_timestamp": "1776359447.0", + "retrieved_timestamp": "1786169658.404657", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5.4-2026-03-05/922969c3-26e3-4b15-8220-1a2ee8221bb5_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "cdfe060a22af2dee908765c79cf3f51b7bb153d641b459d4fdd06260b7f9db1e", + "total_rows": 58 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5.4-2026-03-05/922969c3-26e3-4b15-8220-1a2ee8221bb5_samples.jsonl b/data/terminalbench/openai/gpt-5.4-2026-03-05/922969c3-26e3-4b15-8220-1a2ee8221bb5_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..9c153960a81852927a3378e370b6a712f3bbd250 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.4-2026-03-05/922969c3-26e3-4b15-8220-1a2ee8221bb5_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cdfe060a22af2dee908765c79cf3f51b7bb153d641b459d4fdd06260b7f9db1e +size 12997186 diff --git a/data/terminalbench/openai/gpt-5.4-2026-03-05/9a7ebb2b-160d-4d72-ac10-8761266c9162.json b/data/terminalbench/openai/gpt-5.4-2026-03-05/9a7ebb2b-160d-4d72-ac10-8761266c9162.json new file mode 100644 index 0000000000000000000000000000000000000000..829fea914cd07913414c6720b945cbbaa3d84632 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.4-2026-03-05/9a7ebb2b-160d-4d72-ac10-8761266c9162.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5.4-2026-03-05/1786167775.73097", + "evaluation_timestamp": "1777160954.0", + "retrieved_timestamp": "1786167775.73097", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5.4-2026-03-05/9a7ebb2b-160d-4d72-ac10-8761266c9162_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "99946dc114f9140502ddabe59f3bcf86805ea49e373c3c91dc879122df648bcb", + "total_rows": 3 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5.4-2026-03-05/9a7ebb2b-160d-4d72-ac10-8761266c9162_samples.jsonl b/data/terminalbench/openai/gpt-5.4-2026-03-05/9a7ebb2b-160d-4d72-ac10-8761266c9162_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..a7bd6d3f25cbbebe8d7af1218c4b9081f6e7ed6c --- /dev/null +++ b/data/terminalbench/openai/gpt-5.4-2026-03-05/9a7ebb2b-160d-4d72-ac10-8761266c9162_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:99946dc114f9140502ddabe59f3bcf86805ea49e373c3c91dc879122df648bcb +size 318619 diff --git a/data/terminalbench/openai/gpt-5.4-2026-03-05/a7f0f636-15a1-48a4-a49d-e82984d5b166.json b/data/terminalbench/openai/gpt-5.4-2026-03-05/a7f0f636-15a1-48a4-a49d-e82984d5b166.json new file mode 100644 index 0000000000000000000000000000000000000000..7f83f9dc54b8100e050744a83603ac2af6019dfd --- /dev/null +++ b/data/terminalbench/openai/gpt-5.4-2026-03-05/a7f0f636-15a1-48a4-a49d-e82984d5b166.json @@ -0,0 +1,108 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5.4-2026-03-05/1786163372.186576", + "evaluation_timestamp": "1777484230.0", + "retrieved_timestamp": "1786163372.186576", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive+C/+3ep/64ae8c02 for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 1, + "sample_ids": [ + "sam-cell-seg" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1777484230.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.0, + "uncertainty": { + "num_samples": 3 + }, + "details": { + "total_matched_trajectories": "3", + "stopping_reason_count_repetition_guard": "1", + "stopping_reason_count_token_limit": "2", + "avg_total_tokens_target_model": "7211433.67", + "total_total_tokens_target_model": "21634301.0", + "avg_total_tokens_other_models": "1629.33", + "total_total_tokens_other_models": "4888.0", + "avg_total_tokens_all_models": "7213063.00", + "total_total_tokens_all_models": "21639189.0", + "avg_turn_count": "105.67", + "total_turn_count": "317.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "32", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "32", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5.4-2026-03-05/a7f0f636-15a1-48a4-a49d-e82984d5b166_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "011c5e5543570e3919cc3ccdb44d2e126d984735a9236c1891f56197e78d0c09", + "total_rows": 3 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5.4-2026-03-05/a7f0f636-15a1-48a4-a49d-e82984d5b166_samples.jsonl b/data/terminalbench/openai/gpt-5.4-2026-03-05/a7f0f636-15a1-48a4-a49d-e82984d5b166_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..0279bb28bb1e2fb8d1123de96ea95cee2df890a3 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.4-2026-03-05/a7f0f636-15a1-48a4-a49d-e82984d5b166_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:011c5e5543570e3919cc3ccdb44d2e126d984735a9236c1891f56197e78d0c09 +size 4107897 diff --git a/data/terminalbench/openai/gpt-5.4-2026-03-05/af98a8f9-dd26-42a1-8cac-d245b49bdb79.json b/data/terminalbench/openai/gpt-5.4-2026-03-05/af98a8f9-dd26-42a1-8cac-d245b49bdb79.json new file mode 100644 index 0000000000000000000000000000000000000000..6143db0c7d1955441ceb80b9e147871770f2aa71 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.4-2026-03-05/af98a8f9-dd26-42a1-8cac-d245b49bdb79.json @@ -0,0 +1,136 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5.4-2026-03-05/1786163343.900086", + "evaluation_timestamp": "1777369892.0", + "retrieved_timestamp": "1786163343.900086", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive+C/+1ep/55f5b1f7 for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 25, + "sample_ids": [ + "regex-chess", + "train-fasttext", + "install-windows-3.11", + "make-doom-for-mips", + "tune-mjcf", + "extract-moves-from-video", + "gcode-to-text", + "make-mips-interpreter", + "circuit-fibsqrt", + "mteb-leaderboard", + "mailman", + "extract-elf", + "protein-assembly", + "code-from-image", + "fix-code-vulnerability", + "financial-document-processor", + "regex-log", + "feal-differential-cryptanalysis", + "cancel-async-tasks", + "bn-fit-modify", + "git-leak-recovery", + "model-extraction-relu-logits", + "nginx-request-logging", + "sparql-university", + "large-scale-text-editing" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1777369892.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.8260869565217391, + "uncertainty": { + "standard_error": { + "value": 0.08081046758996392 + }, + "num_samples": 23 + }, + "details": { + "total_matched_trajectories": "23", + "stopping_reason_count_completed_on_successful_submit": "19", + "stopping_reason_count_repetition_guard": "3", + "stopping_reason_count_completed_with_submit": "1", + "avg_total_tokens_target_model": "1344625.96", + "total_total_tokens_target_model": "30926397.0", + "avg_total_tokens_other_models": "0.00", + "total_total_tokens_other_models": "0.0", + "avg_total_tokens_all_models": "1344625.96", + "total_total_tokens_all_models": "30926397.0", + "avg_turn_count": "28.09", + "total_turn_count": "646.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5.4-2026-03-05/af98a8f9-dd26-42a1-8cac-d245b49bdb79_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "1e7ead0b55fa6c32415deacf49e86efde4d2e718c7622331d090f91ebf3d411d", + "total_rows": 23 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5.4-2026-03-05/af98a8f9-dd26-42a1-8cac-d245b49bdb79_samples.jsonl b/data/terminalbench/openai/gpt-5.4-2026-03-05/af98a8f9-dd26-42a1-8cac-d245b49bdb79_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..e9e64ccd30a43940709fed174ed8d1dc97fa5f61 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.4-2026-03-05/af98a8f9-dd26-42a1-8cac-d245b49bdb79_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1e7ead0b55fa6c32415deacf49e86efde4d2e718c7622331d090f91ebf3d411d +size 7769217 diff --git a/data/terminalbench/openai/gpt-5.4-2026-03-05/afbc530e-a251-4960-96e8-4dd196e1787e.json b/data/terminalbench/openai/gpt-5.4-2026-03-05/afbc530e-a251-4960-96e8-4dd196e1787e.json new file mode 100644 index 0000000000000000000000000000000000000000..48b1e932d4806c6581db0cf51f3636a3f6856b5a --- /dev/null +++ b/data/terminalbench/openai/gpt-5.4-2026-03-05/afbc530e-a251-4960-96e8-4dd196e1787e.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5.4-2026-03-05/1786167793.368191", + "evaluation_timestamp": "1777384608.0", + "retrieved_timestamp": "1786167793.368191", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5.4-2026-03-05/afbc530e-a251-4960-96e8-4dd196e1787e_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "f8c98cc2c423b27578294a29537a2821c94866c620de34441cd936c91c32c423", + "total_rows": 26 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5.4-2026-03-05/afbc530e-a251-4960-96e8-4dd196e1787e_samples.jsonl b/data/terminalbench/openai/gpt-5.4-2026-03-05/afbc530e-a251-4960-96e8-4dd196e1787e_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..ab40b3698edaf28ab26878ac7eb1494abbcf3a97 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.4-2026-03-05/afbc530e-a251-4960-96e8-4dd196e1787e_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f8c98cc2c423b27578294a29537a2821c94866c620de34441cd936c91c32c423 +size 21624511 diff --git a/data/terminalbench/openai/gpt-5.4-2026-03-05/b2f92bfe-70a1-4050-a96e-79378a2b3c9b.json b/data/terminalbench/openai/gpt-5.4-2026-03-05/b2f92bfe-70a1-4050-a96e-79378a2b3c9b.json new file mode 100644 index 0000000000000000000000000000000000000000..1477e04e8b0fff5bbc35a69f2ecb09b287b4b0a6 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.4-2026-03-05/b2f92bfe-70a1-4050-a96e-79378a2b3c9b.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5.4-2026-03-05/1786169603.96434", + "evaluation_timestamp": "1776417565.0", + "retrieved_timestamp": "1786169603.96434", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5.4-2026-03-05/b2f92bfe-70a1-4050-a96e-79378a2b3c9b_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "e45a4088af30e6f7870154183f7bac7f463c2830381bceeb831f818de4fb6c56", + "total_rows": 10 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5.4-2026-03-05/b2f92bfe-70a1-4050-a96e-79378a2b3c9b_samples.jsonl b/data/terminalbench/openai/gpt-5.4-2026-03-05/b2f92bfe-70a1-4050-a96e-79378a2b3c9b_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..e0d3f1e45698fb3ac622f685754711333203a046 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.4-2026-03-05/b2f92bfe-70a1-4050-a96e-79378a2b3c9b_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e45a4088af30e6f7870154183f7bac7f463c2830381bceeb831f818de4fb6c56 +size 1969226 diff --git a/data/terminalbench/openai/gpt-5.4-2026-03-05/bd89ac6f-862e-429f-88a2-6541fbd8fc88.json b/data/terminalbench/openai/gpt-5.4-2026-03-05/bd89ac6f-862e-429f-88a2-6541fbd8fc88.json new file mode 100644 index 0000000000000000000000000000000000000000..be07583c977169eafbeb1ecc3db11fcb24855082 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.4-2026-03-05/bd89ac6f-862e-429f-88a2-6541fbd8fc88.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5.4-2026-03-05/1786167780.395723", + "evaluation_timestamp": "1776704175.0", + "retrieved_timestamp": "1786167780.395723", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5.4-2026-03-05/bd89ac6f-862e-429f-88a2-6541fbd8fc88_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "285a570d98ace0c2c8f09705aa3b9d0d6eb2f363ac72c52dfb5ec5e4fdb04349", + "total_rows": 75 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5.4-2026-03-05/bd89ac6f-862e-429f-88a2-6541fbd8fc88_samples.jsonl b/data/terminalbench/openai/gpt-5.4-2026-03-05/bd89ac6f-862e-429f-88a2-6541fbd8fc88_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..05b064248b4cd7c1721b3bf57e5414ae24a2bf6e --- /dev/null +++ b/data/terminalbench/openai/gpt-5.4-2026-03-05/bd89ac6f-862e-429f-88a2-6541fbd8fc88_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:285a570d98ace0c2c8f09705aa3b9d0d6eb2f363ac72c52dfb5ec5e4fdb04349 +size 9575702 diff --git a/data/terminalbench/openai/gpt-5.4-2026-03-05/c2cf4832-8205-4d87-a6be-279fcfd399ed.json b/data/terminalbench/openai/gpt-5.4-2026-03-05/c2cf4832-8205-4d87-a6be-279fcfd399ed.json new file mode 100644 index 0000000000000000000000000000000000000000..2ebcfb88e71b4a0f257c41784dd844055565d05a --- /dev/null +++ b/data/terminalbench/openai/gpt-5.4-2026-03-05/c2cf4832-8205-4d87-a6be-279fcfd399ed.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5.4-2026-03-05/1786167756.360727", + "evaluation_timestamp": "1777311171.0", + "retrieved_timestamp": "1786167756.360727", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5.4-2026-03-05/c2cf4832-8205-4d87-a6be-279fcfd399ed_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "aab3723f22087ddc7a0acc5b3be8b75a24fa6ea037eea9472ef300c700bd631b", + "total_rows": 53 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5.4-2026-03-05/c2cf4832-8205-4d87-a6be-279fcfd399ed_samples.jsonl b/data/terminalbench/openai/gpt-5.4-2026-03-05/c2cf4832-8205-4d87-a6be-279fcfd399ed_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..e4096eda7fe61c614ea134bdb5f4232c46a4c0fb --- /dev/null +++ b/data/terminalbench/openai/gpt-5.4-2026-03-05/c2cf4832-8205-4d87-a6be-279fcfd399ed_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:aab3723f22087ddc7a0acc5b3be8b75a24fa6ea037eea9472ef300c700bd631b +size 20418699 diff --git a/data/terminalbench/openai/gpt-5.4-2026-03-05/cc66e3cd-857c-4c7a-900f-89f02bbe9a16.json b/data/terminalbench/openai/gpt-5.4-2026-03-05/cc66e3cd-857c-4c7a-900f-89f02bbe9a16.json new file mode 100644 index 0000000000000000000000000000000000000000..8d32570f73e11e8b1e03f14c5c8124f1dc5a943b --- /dev/null +++ b/data/terminalbench/openai/gpt-5.4-2026-03-05/cc66e3cd-857c-4c7a-900f-89f02bbe9a16.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5.4-2026-03-05/1786167770.290398", + "evaluation_timestamp": "1777042398.0", + "retrieved_timestamp": "1786167770.290398", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5.4-2026-03-05/cc66e3cd-857c-4c7a-900f-89f02bbe9a16_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "e8cbaea45654e63da19824ef386944de47874f82ced1c1dd1c8863a800fb3738", + "total_rows": 48 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5.4-2026-03-05/cc66e3cd-857c-4c7a-900f-89f02bbe9a16_samples.jsonl b/data/terminalbench/openai/gpt-5.4-2026-03-05/cc66e3cd-857c-4c7a-900f-89f02bbe9a16_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..345a5248defeae6c238b99e2357914bcc8d1690b --- /dev/null +++ b/data/terminalbench/openai/gpt-5.4-2026-03-05/cc66e3cd-857c-4c7a-900f-89f02bbe9a16_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e8cbaea45654e63da19824ef386944de47874f82ced1c1dd1c8863a800fb3738 +size 23296516 diff --git a/data/terminalbench/openai/gpt-5.4-2026-03-05/d4b97809-736f-4fb5-9075-35c32ed479a1.json b/data/terminalbench/openai/gpt-5.4-2026-03-05/d4b97809-736f-4fb5-9075-35c32ed479a1.json new file mode 100644 index 0000000000000000000000000000000000000000..af27d0f46577e7b01054d36ce96e1b958bfafa23 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.4-2026-03-05/d4b97809-736f-4fb5-9075-35c32ed479a1.json @@ -0,0 +1,107 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5.4-2026-03-05/1786163350.236961", + "evaluation_timestamp": "1776811964.0", + "retrieved_timestamp": "1786163350.236961", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive+C/+2ep/ab56b2a9 for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 1, + "sample_ids": [ + "llm-inference-batching-scheduler" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1776811964.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 1.0, + "uncertainty": { + "num_samples": 2 + }, + "details": { + "total_matched_trajectories": "2", + "stopping_reason_count_completed_on_successful_submit": "2", + "avg_total_tokens_target_model": "513000.00", + "total_total_tokens_target_model": "1026000.0", + "avg_total_tokens_other_models": "0.00", + "total_total_tokens_other_models": "0.0", + "avg_total_tokens_all_models": "513000.00", + "total_total_tokens_all_models": "1026000.0", + "avg_turn_count": "20.00", + "total_turn_count": "40.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5.4-2026-03-05/d4b97809-736f-4fb5-9075-35c32ed479a1_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "f3ee41c2d5b687de524d141eeaba56b233d6cd8fa563bf3d00eb23836d8a463a", + "total_rows": 2 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5.4-2026-03-05/d4b97809-736f-4fb5-9075-35c32ed479a1_samples.jsonl b/data/terminalbench/openai/gpt-5.4-2026-03-05/d4b97809-736f-4fb5-9075-35c32ed479a1_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..403cd3cc9e4fa4e88e614569b6e410aafde1f4c5 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.4-2026-03-05/d4b97809-736f-4fb5-9075-35c32ed479a1_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f3ee41c2d5b687de524d141eeaba56b233d6cd8fa563bf3d00eb23836d8a463a +size 440043 diff --git a/data/terminalbench/openai/gpt-5.4-2026-03-05/d616b534-e88f-4af4-8857-d61f6867d6f9.json b/data/terminalbench/openai/gpt-5.4-2026-03-05/d616b534-e88f-4af4-8857-d61f6867d6f9.json new file mode 100644 index 0000000000000000000000000000000000000000..78c2d83b26230e17d223703f99fb4f4bfe76a173 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.4-2026-03-05/d616b534-e88f-4af4-8857-d61f6867d6f9.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5.4-2026-03-05/1786169617.333786", + "evaluation_timestamp": "1776433475.0", + "retrieved_timestamp": "1786169617.333786", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5.4-2026-03-05/d616b534-e88f-4af4-8857-d61f6867d6f9_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "7a3b74c78908b83e59053aa2e017fe3404acb2b3b1135804af56c66b941f2d7f", + "total_rows": 30 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5.4-2026-03-05/d616b534-e88f-4af4-8857-d61f6867d6f9_samples.jsonl b/data/terminalbench/openai/gpt-5.4-2026-03-05/d616b534-e88f-4af4-8857-d61f6867d6f9_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..5fb3fb7db2663292b87b8ba9a20968e893ae7716 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.4-2026-03-05/d616b534-e88f-4af4-8857-d61f6867d6f9_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7a3b74c78908b83e59053aa2e017fe3404acb2b3b1135804af56c66b941f2d7f +size 9131822 diff --git a/data/terminalbench/openai/gpt-5.4-2026-03-05/d8d512bb-5c44-4343-8f7d-e86b42491f67.json b/data/terminalbench/openai/gpt-5.4-2026-03-05/d8d512bb-5c44-4343-8f7d-e86b42491f67.json new file mode 100644 index 0000000000000000000000000000000000000000..1c949e9a92f51094eeb5b750644ea75c62781eec --- /dev/null +++ b/data/terminalbench/openai/gpt-5.4-2026-03-05/d8d512bb-5c44-4343-8f7d-e86b42491f67.json @@ -0,0 +1,107 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5.4-2026-03-05/1786163369.259788", + "evaluation_timestamp": "1777379967.0", + "retrieved_timestamp": "1786163369.259788", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive+C/+3ep/19b0e2cf for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 1, + "sample_ids": [ + "pytorch-model-recovery" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1777379967.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 1.0, + "uncertainty": { + "num_samples": 2 + }, + "details": { + "total_matched_trajectories": "2", + "stopping_reason_count_completed_on_successful_submit": "2", + "avg_total_tokens_target_model": "184751.50", + "total_total_tokens_target_model": "369503.0", + "avg_total_tokens_other_models": "0.00", + "total_total_tokens_other_models": "0.0", + "avg_total_tokens_all_models": "184751.50", + "total_total_tokens_all_models": "369503.0", + "avg_turn_count": "13.50", + "total_turn_count": "27.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5.4-2026-03-05/d8d512bb-5c44-4343-8f7d-e86b42491f67_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "aff50562ba423ab5f0b20b6299558b44d8dc07e40fde27290122dfeec0fa85d2", + "total_rows": 2 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5.4-2026-03-05/d8d512bb-5c44-4343-8f7d-e86b42491f67_samples.jsonl b/data/terminalbench/openai/gpt-5.4-2026-03-05/d8d512bb-5c44-4343-8f7d-e86b42491f67_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..098ca3a65f4b95ed4cddb6024b6b9e6129a31ef4 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.4-2026-03-05/d8d512bb-5c44-4343-8f7d-e86b42491f67_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:aff50562ba423ab5f0b20b6299558b44d8dc07e40fde27290122dfeec0fa85d2 +size 278309 diff --git a/data/terminalbench/openai/gpt-5.4-2026-03-05/ebc8f9ed-7491-4ace-950d-33dbdde29820.json b/data/terminalbench/openai/gpt-5.4-2026-03-05/ebc8f9ed-7491-4ace-950d-33dbdde29820.json new file mode 100644 index 0000000000000000000000000000000000000000..1348531cdbdd10325311ce01d9d9b299472e33da --- /dev/null +++ b/data/terminalbench/openai/gpt-5.4-2026-03-05/ebc8f9ed-7491-4ace-950d-33dbdde29820.json @@ -0,0 +1,112 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5.4-2026-03-05/1786165968.391534", + "evaluation_timestamp": "1776805272.0", + "retrieved_timestamp": "1786165968.391534", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive/+1ep/8b47492d for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 3, + "sample_ids": [ + "gcode-to-text", + "rstan-to-pystan", + "dna-assembly" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1776805272.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.3333333333333333, + "uncertainty": { + "standard_error": { + "value": 0.33333333333333337 + }, + "num_samples": 3 + }, + "details": { + "total_matched_trajectories": "3", + "stopping_reason_count_repetition_guard": "3", + "avg_total_tokens_target_model": "3096513.00", + "total_total_tokens_target_model": "9289539.0", + "avg_total_tokens_other_models": "797.00", + "total_total_tokens_other_models": "2391.0", + "avg_total_tokens_all_models": "3097310.00", + "total_total_tokens_all_models": "9291930.0", + "avg_turn_count": "63.33", + "total_turn_count": "190.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5.4-2026-03-05/ebc8f9ed-7491-4ace-950d-33dbdde29820_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "6bb8cdd60ad6ac219f230c47718c620bf221b02d0d77d444910d5864b02bf5ad", + "total_rows": 3 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5.4-2026-03-05/ebc8f9ed-7491-4ace-950d-33dbdde29820_samples.jsonl b/data/terminalbench/openai/gpt-5.4-2026-03-05/ebc8f9ed-7491-4ace-950d-33dbdde29820_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..73af96c1b88332ce1823e2b36fda872acfc85607 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.4-2026-03-05/ebc8f9ed-7491-4ace-950d-33dbdde29820_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6bb8cdd60ad6ac219f230c47718c620bf221b02d0d77d444910d5864b02bf5ad +size 2357981 diff --git a/data/terminalbench/openai/gpt-5.4-2026-03-05/f0f4e041-5683-49ac-985f-ecb216892e71.json b/data/terminalbench/openai/gpt-5.4-2026-03-05/f0f4e041-5683-49ac-985f-ecb216892e71.json new file mode 100644 index 0000000000000000000000000000000000000000..f3f0464666c522db1a558c0a22cccf9d4e3b3b63 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.4-2026-03-05/f0f4e041-5683-49ac-985f-ecb216892e71.json @@ -0,0 +1,120 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5.4-2026-03-05/1786163360.37773", + "evaluation_timestamp": "1777369892.0", + "retrieved_timestamp": "1786163360.37773", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [ + { + "evaluation_name": "accuracy on terminalbench/S-adaptive+C/+2ep/48adf858 for scorer harbor_scorer", + "source_data": { + "dataset_name": "terminalbench", + "source_type": "hf_dataset", + "samples_number": 9, + "sample_ids": [ + "polyglot-c-py", + "fix-ocaml-gc", + "build-cython-ext", + "sqlite-db-truncate", + "sanitize-git-repo", + "build-pmars", + "rstan-to-pystan", + "sqlite-with-gcov", + "openssl-selfsigned-cert" + ], + "additional_details": { + "shuffled": "False" + } + }, + "evaluation_timestamp": "1777369892.0", + "metric_config": { + "evaluation_description": "accuracy", + "lower_is_better": false, + "score_type": "continuous", + "min_score": 0.0, + "max_score": 1.0 + }, + "score_details": { + "score": 0.8888888888888888, + "uncertainty": { + "standard_error": { + "value": 0.11111111111111112 + }, + "num_samples": 18 + }, + "details": { + "total_matched_trajectories": "18", + "stopping_reason_count_completed_on_successful_submit": "16", + "stopping_reason_count_token_limit": "1", + "stopping_reason_count_repetition_guard": "1", + "avg_total_tokens_target_model": "1326743.67", + "total_total_tokens_target_model": "23881386.0", + "avg_total_tokens_other_models": "353.83", + "total_total_tokens_other_models": "6369.0", + "avg_total_tokens_all_models": "1327097.50", + "total_total_tokens_all_models": "23887755.0", + "avg_turn_count": "27.61", + "total_turn_count": "497.0" + } + }, + "generation_config": { + "generation_args": { + "reasoning": true, + "agentic_eval_config": { + "available_tools": [] + }, + "eval_plan": { + "name": "plan", + "steps": [ + "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"bash\", \"params\": {\"timeout\": 180}}, {\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}" + ], + "config": { + "timeout": "5400", + "attempt_timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"", + "reasoning_tokens": "32000" + } + }, + "eval_limits": { + "token_limit": 10000000 + }, + "sandbox": {} + }, + "additional_details": { + "timeout": "5400", + "max_connections": "10", + "reasoning_effort": "\"high\"" + } + } + } + ], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5.4-2026-03-05/f0f4e041-5683-49ac-985f-ecb216892e71_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "12478055671189c53f28b5bf4ae1c9c92b17df96672c27a9e06433ede224dd0f", + "total_rows": 18 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5.4-2026-03-05/f0f4e041-5683-49ac-985f-ecb216892e71_samples.jsonl b/data/terminalbench/openai/gpt-5.4-2026-03-05/f0f4e041-5683-49ac-985f-ecb216892e71_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..dd59bab8983dbb032579878f76bb89b06cd690a9 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.4-2026-03-05/f0f4e041-5683-49ac-985f-ecb216892e71_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:12478055671189c53f28b5bf4ae1c9c92b17df96672c27a9e06433ede224dd0f +size 4866816 diff --git a/data/terminalbench/openai/gpt-5.4-2026-03-05/f592eef2-02ee-44e8-b647-43afecc9eb05.json b/data/terminalbench/openai/gpt-5.4-2026-03-05/f592eef2-02ee-44e8-b647-43afecc9eb05.json new file mode 100644 index 0000000000000000000000000000000000000000..0472e2eb416054656b7fcf285945cab392dc3d45 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.4-2026-03-05/f592eef2-02ee-44e8-b647-43afecc9eb05.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5.4-2026-03-05/1786167732.211296", + "evaluation_timestamp": "1776660335.0", + "retrieved_timestamp": "1786167732.211296", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5.4-2026-03-05/f592eef2-02ee-44e8-b647-43afecc9eb05_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "cb7907a270d864b9bc34f03d83536a8d48ebad348de52db7ce8835a3347d3f9d", + "total_rows": 21 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5.4-2026-03-05/f592eef2-02ee-44e8-b647-43afecc9eb05_samples.jsonl b/data/terminalbench/openai/gpt-5.4-2026-03-05/f592eef2-02ee-44e8-b647-43afecc9eb05_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..486bd2fae742c9a8df4a1c629ce0580e2787ac82 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.4-2026-03-05/f592eef2-02ee-44e8-b647-43afecc9eb05_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cb7907a270d864b9bc34f03d83536a8d48ebad348de52db7ce8835a3347d3f9d +size 2575929 diff --git a/data/terminalbench/openai/gpt-5.4-2026-03-05/fd2fa93a-fb98-4749-8123-7cc639e8771e.json b/data/terminalbench/openai/gpt-5.4-2026-03-05/fd2fa93a-fb98-4749-8123-7cc639e8771e.json new file mode 100644 index 0000000000000000000000000000000000000000..6cce340a733f8ea5b33a53b6b300e0e154347fc6 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.4-2026-03-05/fd2fa93a-fb98-4749-8123-7cc639e8771e.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5.4-2026-03-05/1786169641.152913", + "evaluation_timestamp": "1776520455.0", + "retrieved_timestamp": "1786169641.152913", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5.4-2026-03-05/fd2fa93a-fb98-4749-8123-7cc639e8771e_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "b9b998fdb21e6c1606cba078853be76aed5d63afad54c505f7dea634d8a2b1f0", + "total_rows": 50 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5.4-2026-03-05/fd2fa93a-fb98-4749-8123-7cc639e8771e_samples.jsonl b/data/terminalbench/openai/gpt-5.4-2026-03-05/fd2fa93a-fb98-4749-8123-7cc639e8771e_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..85c1bca13ffde90c16d58d43368d7c090a98b84f --- /dev/null +++ b/data/terminalbench/openai/gpt-5.4-2026-03-05/fd2fa93a-fb98-4749-8123-7cc639e8771e_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b9b998fdb21e6c1606cba078853be76aed5d63afad54c505f7dea634d8a2b1f0 +size 16618000 diff --git a/data/terminalbench/openai/gpt-5.4-2026-03-05/fd904b7b-1489-4591-906e-7c4c2a7de816.json b/data/terminalbench/openai/gpt-5.4-2026-03-05/fd904b7b-1489-4591-906e-7c4c2a7de816.json new file mode 100644 index 0000000000000000000000000000000000000000..ead965e534f56b29b15efeff4f3d87e6e649880c --- /dev/null +++ b/data/terminalbench/openai/gpt-5.4-2026-03-05/fd904b7b-1489-4591-906e-7c4c2a7de816.json @@ -0,0 +1,36 @@ +{ + "schema_version": "0.3.0", + "evaluation_id": "terminalbench/openai_gpt-5.4-2026-03-05/1786169624.880402", + "evaluation_timestamp": "1776359447.0", + "retrieved_timestamp": "1786169624.880402", + "source_metadata": { + "source_name": "How Inference Compute Shapes Frontier LLM Evaluation", + "source_type": "evaluation_run", + "source_organization_name": "UK AI Security Initiative", + "evaluator_relationship": "third_party", + "source_organization_url": "https://www.aisi.gov.uk/", + "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg" + }, + "eval_library": { + "name": "inspect_ai", + "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0" + }, + "model_info": { + "name": "openai/gpt-5.4-2026-03-05", + "id": "openai/gpt-5.4-2026-03-05", + "developer": "openai", + "inference_platform": "openai", + "additional_details": { + "deployment_type": "externally_managed", + "model_availability": "closed_weights" + } + }, + "evaluation_results": [], + "detailed_evaluation_results": { + "format": "jsonl", + "file_path": "data/terminalbench/openai/gpt-5.4-2026-03-05/fd904b7b-1489-4591-906e-7c4c2a7de816_samples.jsonl", + "hash_algorithm": "sha256", + "checksum": "a7002e94156d3175e16f010cd001206bcc5e14cb65380bc9eec85ec9db782f20", + "total_rows": 51 + } +} \ No newline at end of file diff --git a/data/terminalbench/openai/gpt-5.4-2026-03-05/fd904b7b-1489-4591-906e-7c4c2a7de816_samples.jsonl b/data/terminalbench/openai/gpt-5.4-2026-03-05/fd904b7b-1489-4591-906e-7c4c2a7de816_samples.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..6f5a5be11d35e4905eec5fa26352601985061329 --- /dev/null +++ b/data/terminalbench/openai/gpt-5.4-2026-03-05/fd904b7b-1489-4591-906e-7c4c2a7de816_samples.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a7002e94156d3175e16f010cd001206bcc5e14cb65380bc9eec85ec9db782f20 +size 16555266