Current version: s3-v1
- Lineage: Qwen3.8-27B → S1 v2 (chemical-judgment SFT, 819 rows) → S2 v2 (agentic SFT on 167 teacher trajectories, medium reasoning) → S3 (RFT-style training on the student's own harness rollouts selected by route audit; 7 optimizer steps).
- LoRA r=8, alpha=16, dropout 0.05; attention, linear-attention and MLP projections of the language model.
- adapter sha256
e69aa10429dfd7aaa83b3a51201832e7380cd460f720eda52b9a169e62f9b146
Evaluation (test40, medium reasoning, ReTRACE harness; LLM-judge chemistry audit, not expert ground truth)
Table with columns: Model, native success, Audit@1 strict, Audit@1 lenient| Model | native success | Audit@1 strict | Audit@1 lenient |
|---|
| Qwen3.8-27B base | 35/40 | 4/40 | 11/40 |
| ReTRACE-27B s3-v1 | 40/40 | 4/40 | 19/40 |
Small test set (40 targets); the lenient difference is not statistically significant (exact McNemar p=0.077).
Usage
Intended for use inside the ReTRACE agent harness (tool-calling retrosynthesis planning). Load with PEFT on top of Qwen/Qwen3.8-27B, or serve with vLLM --enable-lora.
Status and limitations
- Outputs are planning suggestions; every route must be checked by a chemist before any laboratory use.