aaravshar/t3_initial_grpo_run

T3 Initial GRPO Run

Private archival dataset for the canonical initial T3 GRPO experiment: full_censored_grpo_20260723_a.

Experiment identity

• Warm start: qwen36-35b-su-t3-practice-step160
• Coding agent: claude-opus-4-6
• Training pool: 192 tasks across 23 apps
• Fixed held-out cohort: 16 tasks
• Batch size: 32
• GRPO group size: 8
• Maximum sequence length: 131,072 tokens
• Completed training length: 32 optimizer steps
• Complete model checkpoints produced at the end of the run: steps 30 and 32

This is the corrected run with enforced censored-rollout filtering, the final thinking-mode contract, bounded policy lag, deterministic curriculum order, and the canonical held-out evaluation cohort.

Files

• nova_rl_rollouts_full_censored_grpo_20260723_a.tar.zst
  • Per-session SU/CA histories, final repositories, traces, and the Bedrock request/cost ledger.
• prime_training_artifacts_full_censored_grpo_20260723_a.tar.zst
  • Prime-RL rollout batches, configs, logs, lightweight control/checkpoint state, maintenance evidence, and offline W&B data.
• canonical_evals_full_censored_grpo_20260723_a.tar.zst
  • Canonical held-out evaluations for the raw base, SFT step 160, and GRPO steps 10, 20, 30, and 32, plus the aligned-thinking gates.
• timeout_canaries_full_censored_grpo_20260723_a.tar.zst
  • Timeout/cancellation canary runs used to validate censoring behavior.
• replays_full_censored_grpo_20260723_a.tar.zst
  • Targeted judge-truncation replay evidence.
• launcher_state_full_censored_grpo_20260723_a.tar.zst
  • The launcher state and run logs from /home/aarav/.nova_full_runs/full_censored_grpo_20260723_a.
• t3_task_dataset.tar.zst
  • The exact /data/aarav/dataset/T3 task data referenced by the run.
• SOURCE_INVENTORY.txt
  • Source paths, raw byte counts, exclusions, and file counts.
• SHA256SUMS
  • Checksums for every archive.

Deliberate exclusions

This dataset intentionally excludes reproducible or separately archived model state:

• Prime full checkpoints at steps 30 and 32
• exported model weights at steps 30 and 32
• Prime broadcast snapshots
• temporary eval-model exports for steps 10 and 20

Model weights belong in Hugging Face model repos, while this dataset contains the experiment evidence needed to reconstruct and audit training and evaluation. No artifacts from the abandoned full_opus_20260717 run are included.

Restore

Verify first:

sha256sum -c SHA256SUMS

Extract an archive:

tar --use-compress-program=unzstd -xf ARCHIVE.tar.zst

The tar members retain the original full_censored_grpo_20260723_a directory name so that provenance remains unambiguous while the dataset repo uses the friendlier name t3_initial_grpo_run.