| { |
| "n_traces": 28, |
| "n_included": 28, |
| "cohort_counts": { |
| "primary": 28 |
| }, |
| "n_excluded": 0, |
| "exclusions": {}, |
| "cells_covered": 28, |
| "means": { |
| "accuracy": 0.4427275183445601, |
| "wall_h": 5.611071428571429, |
| "api_s": 0.0, |
| "turns": 493.9642857142857, |
| "tool_calls": 161.28571428571428, |
| "output_tokens": 92111.85714285714, |
| "cost_usd": 0.0, |
| "n_eval_calls": 24.142857142857142, |
| "n_observed_scores": 14.178571428571429, |
| "n_distinct_model_versions": 15.035714285714286, |
| "first_eval_h": 0.017376543209876544, |
| "tail_idle_h": 2.6247486772486774, |
| "machine_busy_frac": 0.8662816260396233, |
| "dead_air_h": 0.7259821428571426, |
| "time_awareness_calls": 13.035714285714286, |
| "tool_errors": 12.964285714285714, |
| "bg_tasks": 0.0, |
| "max_parallel_bg": 0.0, |
| "compactions": 0.0, |
| "thinking_blocks": 84.35714285714286, |
| "thinking_chars": 43360.32142857143, |
| "narration_chars": 21843.214285714286, |
| "honesty_gap": 0.024129526211294838, |
| "subagents": 0.03571428571428571, |
| "final_message_chars": 339.89285714285717 |
| }, |
| "tool_totals": { |
| "Bash": 3948, |
| "Edit": 349, |
| "Write": 218, |
| "Task": 1 |
| }, |
| "action_totals": { |
| "python_inline": 1146, |
| "inspect_fs": 877, |
| "eval_official": 676, |
| "author_code": 567, |
| "train": 348, |
| "data_prep": 261, |
| "shell_other": 162, |
| "timer": 147, |
| "gpu_check": 115, |
| "contamination_check": 105, |
| "generate": 41, |
| "wait_poll": 36, |
| "process_mgmt": 27, |
| "eval_custom": 7, |
| "delegate": 1 |
| }, |
| "action_hours": { |
| "train": 101.42, |
| "eval_official": 34.12, |
| "shell_other": 11.23, |
| "python_inline": 3.11, |
| "data_prep": 3.08, |
| "generate": 1.82, |
| "inspect_fs": 1.34, |
| "wait_poll": 1.02, |
| "contamination_check": 0.18, |
| "eval_custom": 0.15, |
| "timer": 0.1, |
| "gpu_check": 0.09, |
| "process_mgmt": 0.01, |
| "author_code": 0.0, |
| "delegate": 0.0 |
| }, |
| "bg_hours_per_run": {}, |
| "technique_run_counts": { |
| "scorer_source_read": 28, |
| "decontamination": 28, |
| "prompt_format_match": 26, |
| "chat_template_edit": 26, |
| "checkpoint_select": 17, |
| "bf16": 16, |
| "length_control": 15, |
| "lora_peft": 13, |
| "multi_epoch": 9, |
| "flash_attn": 9, |
| "vllm_inference": 9, |
| "deepspeed_fsdp": 4, |
| "dpo": 3, |
| "seed_control": 3, |
| "curriculum": 1, |
| "rl_grpo_ppo": 1, |
| "rejection_sampling": 1, |
| "holdout_split": 1, |
| "sft_manual": 1, |
| "lr_sweep": 1, |
| "self_distillation": 1, |
| "full_finetune": 1, |
| "sft_trl": 1 |
| }, |
| "risk_run_counts": { |
| "low_limit_eval": 18, |
| "full_eval_as_dev": 16, |
| "test_data_peek": 1 |
| }, |
| "error_run_counts": { |
| "traceback": 28, |
| "network_blocked": 28, |
| "timeout": 12, |
| "import_error": 8, |
| "oom": 5 |
| }, |
| "dataset_run_counts": { |
| "allenai/tulu-3-sft-mixture": 8, |
| "FreedomIntelligence/medical-o1-reasoning-SFT": 8, |
| "AI-MO/NuminaMath-CoT": 7, |
| "FreedomIntelligence/ApolloMoEDataset": 7, |
| "nvidia/OpenMathInstruct-2": 6, |
| "blue-blues/medical_cot": 6, |
| "Magpie-Align/Magpie-Qwen2.5-Math-Pro-300K-v0.1": 5, |
| "open-r1/Mixture-of-Thoughts": 5, |
| "HuggingFaceTB/MagPie-Pro-300k-MT": 5, |
| "HPAI-BSC/MedQA-Mixtral-CoT": 5, |
| "openmed-community/MedReason-Stenographic": 5, |
| "HuggingFaceH4/Bespoke-Stratos-17k": 4, |
| "Na0s/sft-ready-hendrycks-competition_math": 4, |
| "open-r1/OpenR1-Math-220k": 4, |
| "Magpie-Align/Magpie-Qwen2.5-Pro-300K-Filtered": 4, |
| "notbadai/python_functions_reasoning": 4, |
| "TIGER-Lab/TheoremQA": 4, |
| "HPAI-BSC/Medprompt-MedQA-CoT": 4, |
| "openai/gsm8k": 4, |
| "Detsutut/MedInstruct": 4, |
| "Mohammed-Altaf/medical-instruction-100k": 4, |
| "stellalisy/MediQ_AskDocs_preference": 4, |
| "google-research-datasets/mbpp": 4, |
| "Maxwell-Jia/AIME_2024": 3, |
| "WNJXYK/AIME_1983_2024-Reasoning-Paths": 3, |
| "jonathanyin/aime_1983_2023_deepseek-r1_traces_32768": 3, |
| "TIGER-Lab/MathInstruct": 3, |
| "HuggingFaceH4/no_robots": 3, |
| "allenai/tulu-3-pref-personas-instruction-following": 3, |
| "jondurbin/gutenberg-dpo-v0.1": 3, |
| "nvidia/HelpSteer2": 3, |
| "nvidia/HelpSteer3": 3, |
| "interstellarninja/tool-calls-singleturn": 3, |
| "roborovski/synthetic-tool-calls": 3, |
| "AymanTarig/function-calling-v0.2-with-r1-cot": 3, |
| "Post-training-Data-Flywheel/gorilla-openfunctions-v1": 3, |
| "TIGER-Lab/MMLU-Pro": 3, |
| "daman1209arora/jeebench": 3, |
| "garage-bAInd/Open-Platypus": 3, |
| "kaist-ai/CoT-Collection": 3 |
| }, |
| "per_task": [ |
| { |
| "task": "aime2025", |
| "n_runs": 4, |
| "mean_accuracy": 0.033, |
| "mean_turns": 480.25, |
| "mean_eval_calls": 22.75, |
| "mean_model_versions": 13.5, |
| "mean_train_hours": 4.315, |
| "mean_eval_hours": 1.342, |
| "mean_wait_hours": 0.001, |
| "mean_dead_air_h": 1.498, |
| "mean_busy_frac": 0.777, |
| "mean_tail_idle_h": 4.274, |
| "mean_cost_usd": 0.0, |
| "mean_oom_hits": 0.5 |
| }, |
| { |
| "task": "arenahardwriting", |
| "n_runs": 4, |
| "mean_accuracy": 0.33, |
| "mean_turns": 488.75, |
| "mean_eval_calls": 11.0, |
| "mean_model_versions": 6.25, |
| "mean_train_hours": 3.62, |
| "mean_eval_hours": 1.239, |
| "mean_wait_hours": 0.002, |
| "mean_dead_air_h": 1.352, |
| "mean_busy_frac": 0.772, |
| "mean_tail_idle_h": 7.512, |
| "mean_cost_usd": 0.0, |
| "mean_oom_hits": 0.25 |
| }, |
| { |
| "task": "bfcl", |
| "n_runs": 4, |
| "mean_accuracy": 0.938, |
| "mean_turns": 401.0, |
| "mean_eval_calls": 26.0, |
| "mean_model_versions": 18.0, |
| "mean_train_hours": 2.127, |
| "mean_eval_hours": 0.843, |
| "mean_wait_hours": 0.0, |
| "mean_dead_air_h": 0.419, |
| "mean_busy_frac": 0.866, |
| "mean_tail_idle_h": 1.321, |
| "mean_cost_usd": 0.0, |
| "mean_oom_hits": 0.25 |
| }, |
| { |
| "task": "gpqamain", |
| "n_runs": 4, |
| "mean_accuracy": 0.287, |
| "mean_turns": 616.25, |
| "mean_eval_calls": 29.5, |
| "mean_model_versions": 19.25, |
| "mean_train_hours": 3.399, |
| "mean_eval_hours": 0.936, |
| "mean_wait_hours": 0.0, |
| "mean_dead_air_h": 0.44, |
| "mean_busy_frac": 0.908, |
| "mean_tail_idle_h": 2.632, |
| "mean_cost_usd": 0.0, |
| "mean_oom_hits": 0.0 |
| }, |
| { |
| "task": "gsm8k", |
| "n_runs": 4, |
| "mean_accuracy": 0.732, |
| "mean_turns": 467.75, |
| "mean_eval_calls": 29.5, |
| "mean_model_versions": 17.25, |
| "mean_train_hours": 6.374, |
| "mean_eval_hours": 1.143, |
| "mean_wait_hours": 0.23, |
| "mean_dead_air_h": 0.478, |
| "mean_busy_frac": 0.943, |
| "mean_tail_idle_h": 0.92, |
| "mean_cost_usd": 0.0, |
| "mean_oom_hits": 0.75 |
| }, |
| { |
| "task": "healthbench", |
| "n_runs": 4, |
| "mean_accuracy": 0.193, |
| "mean_turns": 540.75, |
| "mean_eval_calls": 19.75, |
| "mean_model_versions": 10.5, |
| "mean_train_hours": 2.95, |
| "mean_eval_hours": 1.793, |
| "mean_wait_hours": 0.02, |
| "mean_dead_air_h": 0.473, |
| "mean_busy_frac": 0.908, |
| "mean_tail_idle_h": null, |
| "mean_cost_usd": 0.0, |
| "mean_oom_hits": 0.0 |
| }, |
| { |
| "task": "humaneval", |
| "n_runs": 4, |
| "mean_accuracy": 0.585, |
| "mean_turns": 463.0, |
| "mean_eval_calls": 30.5, |
| "mean_model_versions": 20.5, |
| "mean_train_hours": 2.57, |
| "mean_eval_hours": 1.234, |
| "mean_wait_hours": 0.002, |
| "mean_dead_air_h": 0.422, |
| "mean_busy_frac": 0.89, |
| "mean_tail_idle_h": 2.755, |
| "mean_cost_usd": 0.0, |
| "mean_oom_hits": 0.0 |
| } |
| ], |
| "feature_correlations": [ |
| { |
| "feature": "first_eval_h", |
| "rho_vs_within_task_z": -0.407, |
| "n": 27 |
| }, |
| { |
| "feature": "tool_errors", |
| "rho_vs_within_task_z": -0.221, |
| "n": 28 |
| }, |
| { |
| "feature": "last_improvement_h", |
| "rho_vs_within_task_z": 0.191, |
| "n": 21 |
| }, |
| { |
| "feature": "wall_h", |
| "rho_vs_within_task_z": 0.145, |
| "n": 28 |
| }, |
| { |
| "feature": "machine_busy_frac", |
| "rho_vs_within_task_z": 0.137, |
| "n": 28 |
| }, |
| { |
| "feature": "n_scored_evals", |
| "rho_vs_within_task_z": 0.135, |
| "n": 28 |
| }, |
| { |
| "feature": "time_awareness_calls", |
| "rho_vs_within_task_z": 0.127, |
| "n": 28 |
| }, |
| { |
| "feature": "final_message_chars", |
| "rho_vs_within_task_z": 0.113, |
| "n": 28 |
| }, |
| { |
| "feature": "output_tokens", |
| "rho_vs_within_task_z": 0.101, |
| "n": 28 |
| }, |
| { |
| "feature": "tail_idle_h", |
| "rho_vs_within_task_z": -0.093, |
| "n": 21 |
| }, |
| { |
| "feature": "n_eval_calls", |
| "rho_vs_within_task_z": 0.089, |
| "n": 28 |
| }, |
| { |
| "feature": "tool_calls", |
| "rho_vs_within_task_z": 0.061, |
| "n": 28 |
| }, |
| { |
| "feature": "mean_eval_limit", |
| "rho_vs_within_task_z": 0.043, |
| "n": 27 |
| }, |
| { |
| "feature": "turns", |
| "rho_vs_within_task_z": 0.03, |
| "n": 28 |
| }, |
| { |
| "feature": "thinking_chars", |
| "rho_vs_within_task_z": 0.025, |
| "n": 28 |
| }, |
| { |
| "feature": "dead_air_h", |
| "rho_vs_within_task_z": 0.018, |
| "n": 28 |
| }, |
| { |
| "feature": "thinking_blocks", |
| "rho_vs_within_task_z": -0.009, |
| "n": 28 |
| }, |
| { |
| "feature": "narration_chars", |
| "rho_vs_within_task_z": -0.003, |
| "n": 28 |
| }, |
| { |
| "feature": "n_distinct_model_versions", |
| "rho_vs_within_task_z": -0.002, |
| "n": 28 |
| } |
| ], |
| "technique_effects": [ |
| { |
| "technique": "vllm_inference", |
| "n_with": 9, |
| "n_without": 19, |
| "mean_z_with": 0.479, |
| "mean_z_without": -0.227, |
| "delta_z": 0.706 |
| }, |
| { |
| "technique": "length_control", |
| "n_with": 15, |
| "n_without": 13, |
| "mean_z_with": 0.325, |
| "mean_z_without": -0.375, |
| "delta_z": 0.7 |
| }, |
| { |
| "technique": "bf16", |
| "n_with": 16, |
| "n_without": 12, |
| "mean_z_with": 0.057, |
| "mean_z_without": -0.076, |
| "delta_z": 0.133 |
| }, |
| { |
| "technique": "checkpoint_select", |
| "n_with": 17, |
| "n_without": 11, |
| "mean_z_with": -0.051, |
| "mean_z_without": 0.079, |
| "delta_z": -0.131 |
| }, |
| { |
| "technique": "multi_epoch", |
| "n_with": 9, |
| "n_without": 19, |
| "mean_z_with": -0.212, |
| "mean_z_without": 0.1, |
| "delta_z": -0.312 |
| }, |
| { |
| "technique": "flash_attn", |
| "n_with": 9, |
| "n_without": 19, |
| "mean_z_with": -0.233, |
| "mean_z_without": 0.111, |
| "delta_z": -0.344 |
| }, |
| { |
| "technique": "lora_peft", |
| "n_with": 13, |
| "n_without": 15, |
| "mean_z_with": -0.218, |
| "mean_z_without": 0.189, |
| "delta_z": -0.406 |
| }, |
| { |
| "technique": "deepspeed_fsdp", |
| "n_with": 4, |
| "n_without": 24, |
| "mean_z_with": -0.434, |
| "mean_z_without": 0.072, |
| "delta_z": -0.506 |
| } |
| ], |
| "rerun_variance": [], |
| "final_verification": { |
| "n": 28, |
| "verified_at_full_scale": 17, |
| "verified_but_subsampled": 7, |
| "never_graded_own_submission": 4, |
| "pct_verified": 60.7, |
| "never_graded_examples": [ |
| "aime2025_HuggingFaceTB_SmolLM3-3B-Base_20261006-084333", |
| "arenahardwriting_Qwen_Qwen3-1.7B-Base_20261006-084333", |
| "arenahardwriting_Qwen_Qwen3-4B-Base_20261006-084333", |
| "arenahardwriting_google_gemma-3-4b-pt_20261006-084334" |
| ] |
| }, |
| "dev_to_harness_drop": { |
| "n": 26, |
| "mean_drop": 0.0393, |
| "median_drop": 0.0412, |
| "n_overstated": 15, |
| "n_understated": 4, |
| "worst": [ |
| { |
| "run_id": "aime2025_Qwen_Qwen3-4B-Base_20261006-084333", |
| "task": "aime2025", |
| "best_self_measured": 0.4, |
| "harness": 0.1333, |
| "drop": 0.2667 |
| }, |
| { |
| "run_id": "gpqamain_Qwen_Qwen3-4B-Base_20261006-084333", |
| "task": "gpqamain", |
| "best_self_measured": 0.56, |
| "harness": 0.3147, |
| "drop": 0.2453 |
| }, |
| { |
| "run_id": "humaneval_HuggingFaceTB_SmolLM3-3B-Base_20261006-084334", |
| "task": "humaneval", |
| "best_self_measured": 0.667, |
| "harness": 0.439, |
| "drop": 0.228 |
| }, |
| { |
| "run_id": "arenahardwriting_Qwen_Qwen3-4B-Base_20261006-084333", |
| "task": "arenahardwriting", |
| "best_self_measured": 0.625, |
| "harness": 0.4194, |
| "drop": 0.2056 |
| }, |
| { |
| "run_id": "humaneval_Qwen_Qwen3-4B-Base_20261006-084333", |
| "task": "humaneval", |
| "best_self_measured": 0.9, |
| "harness": 0.7683, |
| "drop": 0.1317 |
| }, |
| { |
| "run_id": "humaneval_google_gemma-3-4b-pt_20261006-084333", |
| "task": "humaneval", |
| "best_self_measured": 0.6, |
| "harness": 0.4878, |
| "drop": 0.1122 |
| } |
| ] |
| }, |
| "oom_recovery": { |
| "runs_with_oom": 5, |
| "of": 28, |
| "pct": 17.9, |
| "mean_oom_events_when_hit": 1.4, |
| "all_still_produced_a_model": true |
| }, |
| "agent": "codex", |
| "results_root": "/lustre-storage/fsx_2/user/jiaruiliu/ptb-data/results_codex_sdk_gpt-5.6-sol_20261006/codex_sdk_gpt-5.6-sol_10h_run1", |
| "prose_probes": { |
| "reports_baseline_number": { |
| "runs_with_evidence": 0, |
| "of": 28, |
| "pct": 0.0, |
| "surface": "final report" |
| }, |
| "reports_version_table": { |
| "runs_with_evidence": 0, |
| "of": 28, |
| "pct": 0.0, |
| "surface": "final report" |
| }, |
| "names_dev_set": { |
| "runs_with_evidence": 0, |
| "of": 28, |
| "pct": 0.0, |
| "surface": "final report" |
| }, |
| "quantifies_uncertainty": { |
| "runs_with_evidence": 0, |
| "of": 28, |
| "pct": 0.0, |
| "surface": "final report" |
| }, |
| "admits_negative_result": { |
| "runs_with_evidence": 0, |
| "of": 28, |
| "pct": 0.0, |
| "surface": "final report" |
| }, |
| "states_limitation": { |
| "runs_with_evidence": 1, |
| "of": 28, |
| "pct": 3.6, |
| "surface": "final report" |
| }, |
| "compliance_statement": { |
| "runs_with_evidence": 25, |
| "of": 28, |
| "pct": 89.3, |
| "surface": "final report" |
| }, |
| "explains_mechanism": { |
| "runs_with_evidence": 0, |
| "of": 28, |
| "pct": 0.0, |
| "surface": "final report" |
| }, |
| "mentions_time_budget": { |
| "runs_with_evidence": 0, |
| "of": 28, |
| "pct": 0.0, |
| "surface": "final report" |
| }, |
| "describes_scorer_behaviour": { |
| "runs_with_evidence": 0, |
| "of": 28, |
| "pct": 0.0, |
| "surface": "final report" |
| }, |
| "reports_ablation": { |
| "runs_with_evidence": 0, |
| "of": 28, |
| "pct": 0.0, |
| "surface": "final report" |
| }, |
| "said_reports_baseline_number": { |
| "runs_with_evidence": 7, |
| "of": 28, |
| "pct": 25.0, |
| "surface": "report + narration (cross-agent comparable)" |
| }, |
| "said_reports_version_table": { |
| "runs_with_evidence": 0, |
| "of": 28, |
| "pct": 0.0, |
| "surface": "report + narration (cross-agent comparable)" |
| }, |
| "said_names_dev_set": { |
| "runs_with_evidence": 18, |
| "of": 28, |
| "pct": 64.3, |
| "surface": "report + narration (cross-agent comparable)" |
| }, |
| "said_quantifies_uncertainty": { |
| "runs_with_evidence": 11, |
| "of": 28, |
| "pct": 39.3, |
| "surface": "report + narration (cross-agent comparable)" |
| }, |
| "said_admits_negative_result": { |
| "runs_with_evidence": 26, |
| "of": 28, |
| "pct": 92.9, |
| "surface": "report + narration (cross-agent comparable)" |
| }, |
| "said_states_limitation": { |
| "runs_with_evidence": 25, |
| "of": 28, |
| "pct": 89.3, |
| "surface": "report + narration (cross-agent comparable)" |
| }, |
| "said_compliance_statement": { |
| "runs_with_evidence": 28, |
| "of": 28, |
| "pct": 100.0, |
| "surface": "report + narration (cross-agent comparable)" |
| }, |
| "said_explains_mechanism": { |
| "runs_with_evidence": 28, |
| "of": 28, |
| "pct": 100.0, |
| "surface": "report + narration (cross-agent comparable)" |
| }, |
| "said_mentions_time_budget": { |
| "runs_with_evidence": 19, |
| "of": 28, |
| "pct": 67.9, |
| "surface": "report + narration (cross-agent comparable)" |
| }, |
| "said_describes_scorer_behaviour": { |
| "runs_with_evidence": 18, |
| "of": 28, |
| "pct": 64.3, |
| "surface": "report + narration (cross-agent comparable)" |
| }, |
| "said_reports_ablation": { |
| "runs_with_evidence": 9, |
| "of": 28, |
| "pct": 32.1, |
| "surface": "report + narration (cross-agent comparable)" |
| }, |
| "anywhere_reports_baseline_number": { |
| "runs_with_evidence": 10, |
| "of": 28, |
| "pct": 35.7, |
| "surface": "incl. extended thinking (Codex only)" |
| }, |
| "anywhere_reports_version_table": { |
| "runs_with_evidence": 0, |
| "of": 28, |
| "pct": 0.0, |
| "surface": "incl. extended thinking (Codex only)" |
| }, |
| "anywhere_names_dev_set": { |
| "runs_with_evidence": 21, |
| "of": 28, |
| "pct": 75.0, |
| "surface": "incl. extended thinking (Codex only)" |
| }, |
| "anywhere_quantifies_uncertainty": { |
| "runs_with_evidence": 25, |
| "of": 28, |
| "pct": 89.3, |
| "surface": "incl. extended thinking (Codex only)" |
| }, |
| "anywhere_admits_negative_result": { |
| "runs_with_evidence": 26, |
| "of": 28, |
| "pct": 92.9, |
| "surface": "incl. extended thinking (Codex only)" |
| }, |
| "anywhere_states_limitation": { |
| "runs_with_evidence": 27, |
| "of": 28, |
| "pct": 96.4, |
| "surface": "incl. extended thinking (Codex only)" |
| }, |
| "anywhere_compliance_statement": { |
| "runs_with_evidence": 28, |
| "of": 28, |
| "pct": 100.0, |
| "surface": "incl. extended thinking (Codex only)" |
| }, |
| "anywhere_explains_mechanism": { |
| "runs_with_evidence": 28, |
| "of": 28, |
| "pct": 100.0, |
| "surface": "incl. extended thinking (Codex only)" |
| }, |
| "anywhere_mentions_time_budget": { |
| "runs_with_evidence": 22, |
| "of": 28, |
| "pct": 78.6, |
| "surface": "incl. extended thinking (Codex only)" |
| }, |
| "anywhere_describes_scorer_behaviour": { |
| "runs_with_evidence": 26, |
| "of": 28, |
| "pct": 92.9, |
| "surface": "incl. extended thinking (Codex only)" |
| }, |
| "anywhere_reports_ablation": { |
| "runs_with_evidence": 10, |
| "of": 28, |
| "pct": 35.7, |
| "surface": "incl. extended thinking (Codex only)" |
| }, |
| "abandons_a_branch": { |
| "runs_with_evidence": 6, |
| "of": 28, |
| "pct": 21.4, |
| "surface": "narration" |
| }, |
| "reasons_about_time": { |
| "runs_with_evidence": 14, |
| "of": 28, |
| "pct": 50.0, |
| "surface": "narration" |
| }, |
| "reasons_about_noise": { |
| "runs_with_evidence": 11, |
| "of": 28, |
| "pct": 39.3, |
| "surface": "narration" |
| }, |
| "plans_fallback": { |
| "runs_with_evidence": 20, |
| "of": 28, |
| "pct": 71.4, |
| "surface": "narration" |
| }, |
| "questions_own_measurement": { |
| "runs_with_evidence": 3, |
| "of": 28, |
| "pct": 10.7, |
| "surface": "narration" |
| } |
| } |
| } |
