EVAL_GRPO20

This dataset contains the evaluation evidence for the Sept. 1 GRPO20 iter_0000019 checkpoint. It intentionally excludes training artifacts, checkpoint weights, training data, W&B metadata, and comparison runs.

Results

Result	Pass@1	Multi turn with feedback (turn=2)	Trials	Samples
Sept. 1 GRPO20, iter 19	10/26 mean	14/26 mean	4	104

Pass@1 SD, range, 95% CI (out of 26)	Turn-2 SD, range, 95% CI (out of 26)	Conditional turn-2 recovery
1.83; 8-12; 6.25-13.75	1.15; 13-15; 9.75-18.25	16/64 (25.0%; CI 13.2-40.4%)

Trials

Trial	Evaluation run ID	Pass@1	Turn 2	Status
1	generalized-cpp-sept1-iter19-fixed26-20260902T103608Z-t04	11/26	15/26	complete
2	generalized-cpp-sept1-iter19-fixed26-20260902T103608Z-t01	12/26	13/26	complete
3	generalized-cpp-sept1-iter19-fixed26-20260902T103608Z-t02	8/26	15/26	complete
4	generalized-cpp-sept1-iter19-fixed26-20260902T103608Z-t03	9/26	13/26	complete

All four evaluations ran sequentially on one GCP Spot a3-highgpu-8g host with 8×H100 80 GB GPUs. Each job completed and synced its evidence before the next trial began. The host was terminated after the fourth sync.

Evaluation identity and integrity

• Contract: fixed26-contract-v2
• Evaluated checkpoint: iter_0000019
• Adapter SHA-256: 0c0da7a1c3553dfda66af622c8eb5002f3d7418321cfda261414ef020353240c
• Serving adapter SHA-256: 6a7b3247245ab652fca0dbf4e628de05e5e636c17770fcb34f5f3ffb9961d4d5
• Evaluation-driver SHA-256: 39044988bb9a73ff62aee78fbfa1051c8c5a3b165dcb02c0b2d4c162d9f015c4
• Aider commit: 5dc9490bb35f9729ef2c95d00a19ccd30c26339c
• Polyglot commit: 7e0611e77b54e2dea774cdc0aa00cf9f7ed6144f
• Runtime image: us-central1-docker.pkg.dev/lifeandhalf-24122025/w8-biayn/glm47-full-v5-unadmitted-grpo@sha256:5df1c41efb6062a292ac68a49d68ad955d9f8209c182d3fc2f7e8c634ca076a2

Every selected receipt has status=complete, terminal_tasks=26, unique_testcases=26, well_formed_tasks=26, tries=2, and verified LoRA activation on both serving shards.

Confidence intervals use 100,000 deterministic task-clustered percentile bootstrap replicates with seed 20260805. Each replicate resamples the 26 task IDs and retains all four outcomes for each sampled task. The 104 task-trial records are not treated as independent Bernoulli observations.

These are assisted regression evaluations, not a pristine held-out benchmark claim.

Contents

• evaluations/<run-id>/run_receipt.json: aggregate receipt and per-task outcomes
• evaluations/<run-id>/receipts/: shard and task-level evidence
• evaluations/<run-id>/shards/: exact evaluated Fixed26 task trees
• evaluations/<run-id>/*.log: benchmark and serving logs
• evaluations/<run-id>/audit.md and audit.json: receipt audit
• statistics.json and statistics.md: cohort statistics
• per_task_success.csv: per-task frequencies across the four trials
• compute_statistics.py: deterministic recomputation
• EVAL_MANIFEST.json and SHA256SUMS: release inventory and checksums

No API keys, tokens, W&B data, training artifacts, model weights, or generated virtual environments are included.