baseten/caesar-cipher-adam-qwen38_27b-expert

Caesar Cipher — Qwen3.8-27B Adam expert

A full-weight reinforcement-learning expert trained by Baseten on Caesar Cipher using GRPO with Adam. Decode Caesar-cipher text. This repository contains the final inference checkpoint after 100 optimizer updates.

• Base model: Qwen/Qwen3.8-27B, pinned revision 1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0.
• Environment source: Caesar Cipher from Reasoning Gym, pinned revision 49b07130b3fcd12f2d064bba7c43869543a0e7e7. The experiment used a frozen task bank and a completed-answer binary reward adapter derived from this environment; these results are not an official upstream benchmark score.
• Training record: training run lo5clpo9.
• Machine-readable settings, results, and checkpoint lineage: expert-provenance.json.

Evaluation

Checkpoint	Accuracy	Correct / scored completions	Truncation
Pinned base model, before environment training	39.00%	156 / 400	0.25%
Final expert, update 100	95.25%	381 / 400	3.25%

Accuracy improved by 56.25 percentage points from the pinned base model. Each evaluation uses 50 held-out tasks × 8 sampled completions = 400 scored completions. Accuracy is the mean binary correctness across all completions, not best-of-eight or pass-at-eight. Truncated responses remain in the denominator and count as failures. The reported endpoint is the final checkpoint, not an earlier peak.

The bank contains 114 training plaintexts from the finite approved corpus, with 50 separate validation plaintexts and no train/validation plaintext overlap. Validation uses a dedicated inference role every 10 optimizer updates with fixed seeds [17, 43, 101, 137, 211, 307, 401, 503]. All scheduled evaluations and final inference/full-state archival were verified against native receipts. The same validation bank was inspected during training; there was no separate final confirmation set. Results establish performance on held-out cases within this environment and do not establish cross-environment transfer or broad general reasoning improvement.

Training and generation settings

Setting	Value
Training	Full-weight GRPO; Adam
Learning rate	Constant 1e-6
Adam betas / weight decay	(0.9, 0.98) / 0.1
Policy clip epsilon / upper epsilon	0.2 / 0.28
KL / entropy coefficients	0.0 / 0.0
Batch per optimizer update	8 questions × 8 completions = 64 responses
Total consumed training responses	6,400
Precision	BF16 model training and sampling
Thinking	Enabled; reasoning_effort="medium"
Sampling	Temperature 1.0, top-p 0.95, top-k 20
Maximum generated tokens	16384
Context limit used in this experiment	32768
Checkpoint / validation interval	Every 10 optimizer updates
Async consumption	Ordered batches; maximum completed-update lag of 1

For comparable text generation, use the supplied tokenizer and chat template, keep thinking enabled with reasoning_effort="medium", and retain the sampling and output limits above. The original runtime pinned Transformers 5.12.1; detailed runtime revisions are recorded in the provenance JSON. This expert was trained and evaluated on text inputs. Vision or video performance and retention of unrelated base-model capabilities were not evaluated.

Provenance and reporting

The source inference snapshot is miles-qwen3.8-27b-06448b429f21/run-06448b429f21/training/eval-snapshots/step_99. It was emitted at optimizer update 100 and verified complete in durable platform storage. The native final evaluation completed at 2026-10-19T13:25:51.973586+00:00. This release contains inference weights and required inference assets; resumable optimizer state and raw training/evaluation traces remain in the experiment archives.

W&B's final validation point is missing: its recorded validation history stops at update 90 (95.00%). An existing-run update permission restriction prevents reconciliation. The final 95.25% at update 100 reported here comes from verified native evaluation receipts and archived rollout evidence; the W&B gap does not change that measured endpoint.

Intended use and limitations

This is a specialized research checkpoint for text reasoning in the named environment. Its results depend on the frozen bank, prompt and grading contract, decoding settings, and token budget. Broader task transfer, production safety, multimodal behavior, and performance under different decoding settings have not been established by these experiments.

License

Apache-2.0, following the pinned base model. See the accompanying license and the pinned base-model card.

Training and validation curves

The plots below show the recorded training trajectory and verified validation results for this environment. Training uses a running average of 10 with the original series visible; validation is unsmoothed. No training, evaluation, or interpolation was performed to create these images.

Training metric: rollout/raw_reward against rollout/step. Validation metric: eval/validation against completed optimizer updates.

Both published experts contain 100 optimizer updates. Their training curves use the full native axis 0–99; validation includes every scheduled point from 0 through 100. Restored runs retain their original checkpoint ancestry and exclude abandoned branches.

The overlay compares these recorded trajectories; it does not establish an optimizer-only causal effect. See each model’s settings and limitations above.

ADAM training and validation

Final ADAM weights

SGD training and validation

Final SGD weights

Adam and SGD comparison

Adam is blue; SGD is pink.

Chart provenance and settings. The original weight commit is e4119e628b198e7da0bc188f047a4015e37b3579.