baseten/isomorphic-strings-sgd-qwen38_27b-expert

Isomorphic Strings — Qwen3.8-27B SGD expert

A full-weight reinforcement-learning expert trained by Baseten on Isomorphic Strings using GRPO with SGD. Determine whether two strings admit a consistent one-to-one character mapping. This release targets the final post-trained inference weights at optimizer update 100, with 6,400 consumed completions in the retained training trajectory.

• Base model: Qwen/Qwen3.8-27B, pinned revision 1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0.
• Environment source: Isomorphic Strings from Reasoning Gym, pinned revision 49b07130b3fcd12f2d064bba7c43869543a0e7e7. The experiment used a frozen task bank and completed-answer binary reward adapter derived from this environment; these results are not an official upstream benchmark score.
• Final training segment: training run j40le0t4.
• Settings, native results and complete retained checkpoint lineage: expert-provenance.json.

Evaluation

Checkpoint	Accuracy	Correct / scored completions	Truncation
Pinned base model before environment training	57.25%	229 / 400	0.00%
Final expert, update 100	93.00%	372 / 400	2.25%

Accuracy improved by 35.75 percentage points. Each evaluation uses 50 held-out tasks × 8 sampled completions = 400 scored completions. Accuracy is the mean binary reward across all completions; truncated responses remain in the denominator and count as failures. The endpoint is the final checkpoint, not an earlier peak.

The frozen bank contains 400 training cases and 50 held-out validation cases. A dedicated inference role evaluates every 10 optimizer updates with fixed seeds [17, 43, 101, 137, 211, 307, 401, 503]. All 11 logical evaluations, final inference/full-state archives and raw completion coverage were verified. The same validation bank was inspected throughout training, with no separate final confirmation set. These results do not establish cross-environment transfer, broad reasoning gains, or replicated causal effects of the optimizer.

Training and generation settings

Setting	Value
Training	Full-weight GRPO; SGD
Learning rate	Constant 0.1
SGD momentum / weight decay	0.0 / 0.0
Policy clip epsilon / upper epsilon	0.2 / 0.28
KL / entropy coefficients	0.0 / 0.0
Batch per update	8 questions × 8 completions = 64 responses
Retained optimizer budget	100 updates; 6,400 training responses
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 / context limit	16384 / 32768
Async consumption	Ordered batches; maximum completed-update lag 1
Checkpoint / validation interval	Every 10 optimizer updates

For comparable text generation, use the supplied tokenizer and chat template, keep thinking enabled with reasoning_effort="medium", and retain these sampling limits. The original runtime pinned Transformers 5.12.1. The experiment trained and evaluated text inputs; multimodal performance and retention of unrelated capabilities were not measured.

Retained training lineage

This final expert retains updates 1–90 from run04 and updates 91–100 from run06. Run06 restored complete training state from checkpoint 90; the parent’s later updates 91–93 were rolled back and their 192 responses remain preserved but excluded from logical coverage. The missing checkpoint 90 evaluation completed in run06 before optimizer update 91. Five ready groups retained their original parent captures; three pending groups were regenerated under the child identity. The unsubmitted run05 alternative contributed no training or evaluation data.

The total logical training budget remains 100 updates and 6,400 consumed responses. Native checkpoint receipts preserve model, optimizer and data provenance. Raw responses and detailed derived grader labels remain separate from the unchanged binary rewards.

Checkpoint and reporting provenance

The source inference snapshot is miles-rg-isomorphic-strings-sgd100-20261022-r06-93ae68a87258/rg-isomorphic-strings-sgd100-20261022-r06-93ae68a87258/training/eval-snapshots/step_99. It belongs to training job qjm9xpq in project qvvnk9q and is COMPLETE in durable platform storage. Its original final evaluation completed at 2026-10-22T19:09:31.542879+00:00.

The W&B summary observed at 2026-10-23T04:06:26.978558+00:00 still reports update 90 at 91.75%. Full history was not rescanned, so this is specifically a summary discrepancy. Existing-run correction writes are permission-blocked. The 93.00% at update 100 reported here comes from original native receipts and verified raw coverage.

Only the final post-trained inference weights and required tokenizer/configuration/chat-template assets are selected for publication. The export contract permits file repacking while requiring exact tensor-bit preservation, with separate per-tensor hash and remote-file verification before publication is declared complete. No numerical conversion, quantization, averaging or base-weight substitution is permitted. Resumable full-state checkpoints, including SGD state, FP32 master weights, scheduler, RNG and cursor, and raw traces remain in Baseten archives and owning experiment evidence.

Intended use and limitations

This is a specialized research checkpoint for the named text environment. Performance depends on the frozen task bank, prompt/grading contract, sampling settings and token budget. Broader task transfer, production safety, multimodal behavior and different decoding settings were not evaluated.

License

Apache-2.0, following the pinned base model. See the accompanying license and 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 cccc67c54fa029aced953c2c9fda70f7a0b6e551.