JLiangHe commited on
Commit
b7cc937
·
verified ·
1 Parent(s): 4285f0e

Add audited Luna high SciCoding-Agent avg@4 results and complete artifacts

Browse files
.gitattributes CHANGED
@@ -70,3 +70,4 @@ Qwen3.5-9B/frontierscience-research/judge_trajectories.jsonl filter=lfs diff=lfs
70
  Qwen3.5-9B/frontierscience-research/trajectories.jsonl filter=lfs diff=lfs merge=lfs -text
71
  sft-mix/scicode-verified/trajectories.jsonl filter=lfs diff=lfs merge=lfs -text
72
  Qwen3.5-9B/scicode-verified/trajectories.jsonl filter=lfs diff=lfs merge=lfs -text
 
 
70
  Qwen3.5-9B/frontierscience-research/trajectories.jsonl filter=lfs diff=lfs merge=lfs -text
71
  sft-mix/scicode-verified/trajectories.jsonl filter=lfs diff=lfs merge=lfs -text
72
  Qwen3.5-9B/scicode-verified/trajectories.jsonl filter=lfs diff=lfs merge=lfs -text
73
+ GPT-5.6-Luna/scicoding-agent/trajectories.jsonl filter=lfs diff=lfs merge=lfs -text
GPT-5.6-Luna/README.md ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ # GPT-5.6 Luna evaluations
2
+
3
+ [Final SciCoding-Agent avg@4 evaluation](scicoding-agent/README.md): **183/200 (91.5%)**; pass@4 **50/50 (100%)**. High reasoning effort; historical provider-default sampling, not the newer temperature-1 / avg@2 protocol.
GPT-5.6-Luna/manifest.json ADDED
@@ -0,0 +1,38 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "README.md": {
3
+ "size": 250,
4
+ "sha256": "e8fe19f7caa895ce7d0c7feed1041e1376a08ec48ef736d5e1a2e3cbda0958c8"
5
+ },
6
+ "summary.json": {
7
+ "size": 1653,
8
+ "sha256": "46412fd0f472e99595f0439ff5646c4962d2c7c6bc48c1879816bb4305395237"
9
+ },
10
+ "scicoding-agent/README.md": {
11
+ "size": 2943,
12
+ "sha256": "a98f1528e7281ca194d8633a968deb8d1b117570a7ce2a785088357585429b7d"
13
+ },
14
+ "scicoding-agent/judge_trajectories.jsonl": {
15
+ "size": 9749935,
16
+ "sha256": "1691e0226a95b410f319278cbd0024b4f1c91ab99726e6d8fa871065cbb65118"
17
+ },
18
+ "scicoding-agent/native_artifacts.parquet": {
19
+ "size": 15160163,
20
+ "sha256": "810828777b95572734b72c00e3219e7c579429cac30e7187e018f35427c66c9e"
21
+ },
22
+ "scicoding-agent/provenance.json": {
23
+ "size": 1802,
24
+ "sha256": "e025f6d5ce77f9c3f363ebc968ceb442385166bb1d5c405c0f4bc67611b88c3b"
25
+ },
26
+ "scicoding-agent/summary.json": {
27
+ "size": 1548,
28
+ "sha256": "9ea862b9cde2c88f0b21dee426b4e6ee35377db931ae4ee3535a760a9c628d47"
29
+ },
30
+ "scicoding-agent/task_results.json": {
31
+ "size": 142877,
32
+ "sha256": "254fe6ed22b928c272bf9aefb0663e46511af46eeac2c1ec93f8d2152965e06f"
33
+ },
34
+ "scicoding-agent/trajectories.jsonl": {
35
+ "size": 19013790,
36
+ "sha256": "a398a5f4b40b4f128ca59cc63c106e309d503de6351bb4bdd5be596205e8c852"
37
+ }
38
+ }
GPT-5.6-Luna/scicoding-agent/README.md ADDED
@@ -0,0 +1,24 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # SciCoding-Agent — GPT-5.6 Luna (high)
2
+
3
+ Final audited historical evaluation on the frozen 50-task benchmark, 2026-10-22.
4
+ **183/200 successful attempts; avg@4 = 91.5%.** All 50 tasks were solved at least once (pass@4 = 100%). All 200 attempts were scored; no run exceptions. Success requires native verifier reward 1 and no run exception. This is an executable-test benchmark, not an LLM-judged benchmark; Luna is the solver.
5
+
6
+ ## Configuration
7
+
8
+ Terminus-2 / Harbor 0.23.0; model `gpt-5.6-luna`; reasoning effort **high**. Four independent attempts per task. Temperature and top-p were not explicitly set (provider defaults). Context/segment budget 131,072; whole-trajectory budget 262,144; per-call completion limit 16,384 including hidden reasoning and visible output. Maximum 256 turns and 7,200 seconds per attempt. Proactive summarization headroom 32,768 tokens, not a 32K history-length trigger. No attempt triggered summarization or output-length recovery.
9
+
10
+ Mean provider output: 8,898.785 tokens/attempt (including API-reported hidden reasoning). Mean charged trajectory length: 16,203.09 tokens including new prompt/tool text; repeated context is not repeatedly charged. Mean runtime: 127.682 seconds/attempt. Hidden chain-of-thought is not exposed by the API and is not present in these exports.
11
+
12
+ ## Files
13
+
14
+ - `trajectories.jsonl`: 200 uniquely keyed attempts with visible agent messages/actions/tool observations, native trajectories, usage/budget records and exact configurations. `trial` is zero-based; `attempt` is one-based.
15
+ - `judge_trajectories.jsonl`: 200 aligned executable-verifier outcomes and logs; there are no LLM judge calls.
16
+ - `native_artifacts.parquet`: compact, losslessly consolidated original artifacts, one row per attempt, including nested command-output audit records. Already verified by reconstruction and SHA-256, rather than thousands of separate files.
17
+ - `summary.json`, `task_results.json`: aggregate, domain and per-task results.
18
+ - `provenance.json`: frozen source revision, policy and artifact verification.
19
+
20
+ ## Comparison and curation caveats
21
+
22
+ This is not a rerun at the new temperature-1 / avg@2 defaults. The current Qwen-IT/SFT runs use the identical task files but two attempts, explicit temperature 1/top-p .95/top-k 20, 16,384-token compaction headroom, and model-specific tokenization/reasoning accounting. The major context, trajectory, turn and time budgets match. Do not compare pass@4 directly with pass@2; avg@N is the per-attempt mean.
23
+
24
+ The benchmark was curated to exclude easy 4/4 Qwen tasks; it is not a representative random sample. Task repairs were audited: 160 unchanged Luna attempts were reused after task-hash verification, and 40 fresh attempts evaluated ten amended tasks. Only final-version outcomes are included here, never superseded attempts. Publication does not regenerate or regrade answers. Task sources and full curation records remain in `JLiangHe/SciCoding-Agent`.
GPT-5.6-Luna/scicoding-agent/judge_trajectories.jsonl ADDED
The diff for this file is too large to render. See raw diff
 
GPT-5.6-Luna/scicoding-agent/native_artifacts.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:810828777b95572734b72c00e3219e7c579429cac30e7187e018f35427c66c9e
3
+ size 15160163
GPT-5.6-Luna/scicoding-agent/provenance.json ADDED
@@ -0,0 +1,41 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "source_run": "/workspace/data/sci-llm-training/benchmarks/scidoing-50-20261022/panel-v3",
3
+ "source_benchmark_repo": "JLiangHe/SciCoding-Agent",
4
+ "benchmark_revision": "d9b75e37fe71313a721bcf7a107422c2ef73cdf6",
5
+ "evaluation_policy": {
6
+ "model": "gpt-5.6-luna",
7
+ "agent": "terminus-2",
8
+ "harbor_version": "0.23.0",
9
+ "reasoning_effort": "high",
10
+ "attempts_per_task": 4,
11
+ "context_tokens": 131072,
12
+ "completion_tokens_per_call": 16384,
13
+ "trajectory_budget_tokens": 262144,
14
+ "max_turns": 256,
15
+ "agent_timeout_seconds": 7200,
16
+ "concurrency_per_runner": 4,
17
+ "temperature": null,
18
+ "top_p": null,
19
+ "reasoning_cap_note": "OpenAI hidden reasoning has no separate numeric cap here. The 16384 total completion cap includes reasoning and visible output.",
20
+ "budget_scope": "New prompt text plus API completion usage across all calls, including summaries. Failed length-limited calls charged full requested completion budget.",
21
+ "length_recovery": "Native same-attempt Harbor recovery; truncated text suppressed and no truncated actions executed."
22
+ },
23
+ "proactive_summarization_headroom_tokens": 32768,
24
+ "regenerated_for_publication": false,
25
+ "rejudged_for_publication": false,
26
+ "amendments": {
27
+ "unchanged_reused_attempts": 160,
28
+ "fresh_attempts_on_ten_amended_tasks": 40
29
+ },
30
+ "native_artifact_verification": {
31
+ "path": "evaluations/luna/final.parquet",
32
+ "rows": 200,
33
+ "original_files": 22990,
34
+ "passed": 183,
35
+ "roundtrip_verified": true,
36
+ "bytes": 15160163,
37
+ "sha256": "810828777b95572734b72c00e3219e7c579429cac30e7187e018f35427c66c9e"
38
+ },
39
+ "final_manifest_sha256": "f9d7fb4b250eb2e5bcc44ceb90463f1aca328c3a717b559c785a7b01b0d2a3e5",
40
+ "report_sha256": "a4b2ef29e0da459d545ad6d4ee4797b828e4f444030a6579d0cb28e734c083cf"
41
+ }
GPT-5.6-Luna/scicoding-agent/summary.json ADDED
@@ -0,0 +1,41 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "expected_trials": 200,
3
+ "finished_results": 200,
4
+ "scored_trials": 200,
5
+ "passed_trials": 183,
6
+ "avg_at_4_full_denominator": 0.915,
7
+ "provisional_scored_accuracy": 0.915,
8
+ "questions_at_least_one_pass": 50,
9
+ "fully_finished_questions": 50,
10
+ "exception_counts": {},
11
+ "mean_charged_trajectory_tokens": 16203.09,
12
+ "pass_at_4_full_denominator": 1.0,
13
+ "success_histogram_0_to_4": {
14
+ "0": 0,
15
+ "1": 2,
16
+ "2": 2,
17
+ "3": 7,
18
+ "4": 39
19
+ },
20
+ "mean_elapsed_seconds": 127.681890455,
21
+ "median_elapsed_seconds": 93.775335,
22
+ "elapsed_seconds_available_trials": 200,
23
+ "mean_provider_output_tokens": 8898.785,
24
+ "median_provider_output_tokens": 7317.0,
25
+ "provider_output_tokens_available_trials": 200,
26
+ "mean_agent_episodes": 5.215,
27
+ "median_agent_episodes": 4.0,
28
+ "agent_episodes_available_trials": 200,
29
+ "mean_summarizations": 0,
30
+ "median_summarizations": 0.0,
31
+ "summarizations_available_trials": 200,
32
+ "total_length_recoveries": 0,
33
+ "trials_with_length_recovery": 0,
34
+ "observed_result_span_seconds": 6475.409374,
35
+ "token_note": "Provider output tokens include hidden reasoning where reported by the API; charged trajectory tokens also include new prompt/tool text and conservatively charged capped calls. Repeated input-context tokens are not added to the charged trajectory budget.",
36
+ "avg_at_4_question_bootstrap_95_interval": [
37
+ 0.86,
38
+ 0.96
39
+ ],
40
+ "uncertainty_note": "Resamples the 50 curated tasks (not 200 independent trials); conditional descriptive interval, not a population-representative guarantee."
41
+ }
GPT-5.6-Luna/scicoding-agent/task_results.json ADDED
The diff for this file is too large to render. See raw diff
 
GPT-5.6-Luna/scicoding-agent/trajectories.jsonl ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:a398a5f4b40b4f128ca59cc63c106e309d503de6351bb4bdd5be596205e8c852
3
+ size 19013790
GPT-5.6-Luna/summary.json ADDED
@@ -0,0 +1,43 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "scicoding-agent": {
3
+ "expected_trials": 200,
4
+ "finished_results": 200,
5
+ "scored_trials": 200,
6
+ "passed_trials": 183,
7
+ "avg_at_4_full_denominator": 0.915,
8
+ "provisional_scored_accuracy": 0.915,
9
+ "questions_at_least_one_pass": 50,
10
+ "fully_finished_questions": 50,
11
+ "exception_counts": {},
12
+ "mean_charged_trajectory_tokens": 16203.09,
13
+ "pass_at_4_full_denominator": 1.0,
14
+ "success_histogram_0_to_4": {
15
+ "0": 0,
16
+ "1": 2,
17
+ "2": 2,
18
+ "3": 7,
19
+ "4": 39
20
+ },
21
+ "mean_elapsed_seconds": 127.681890455,
22
+ "median_elapsed_seconds": 93.775335,
23
+ "elapsed_seconds_available_trials": 200,
24
+ "mean_provider_output_tokens": 8898.785,
25
+ "median_provider_output_tokens": 7317.0,
26
+ "provider_output_tokens_available_trials": 200,
27
+ "mean_agent_episodes": 5.215,
28
+ "median_agent_episodes": 4.0,
29
+ "agent_episodes_available_trials": 200,
30
+ "mean_summarizations": 0,
31
+ "median_summarizations": 0.0,
32
+ "summarizations_available_trials": 200,
33
+ "total_length_recoveries": 0,
34
+ "trials_with_length_recovery": 0,
35
+ "observed_result_span_seconds": 6475.409374,
36
+ "token_note": "Provider output tokens include hidden reasoning where reported by the API; charged trajectory tokens also include new prompt/tool text and conservatively charged capped calls. Repeated input-context tokens are not added to the charged trajectory budget.",
37
+ "avg_at_4_question_bootstrap_95_interval": [
38
+ 0.86,
39
+ 0.96
40
+ ],
41
+ "uncertainty_note": "Resamples the 50 curated tasks (not 200 independent trials); conditional descriptive interval, not a population-representative guarantee."
42
+ }
43
+ }