| --- |
| base_model: Qwen/Qwen2.5-3B-Instruct |
| library_name: peft |
| tags: |
| - polymers |
| - molecular-optimization |
| - reinforcement-learning |
| - repo |
| - polyedit |
| --- |
| |
| # PolyEdit RePO-polymer-retrained |
|
|
| This LoRA adapter uses the public RePO `XGRPOTrainer` with PolyEdit polymer references. |
| It is **not** an upstream RePO checkpoint: the official repository publishes training |
| code and recipes but no trained weights. |
|
|
| Training uses all eight properties with equal property mass, 4,096 examples from only |
| PolyEdit train components, 256 optimizer steps, Qwen2.5-3B-Instruct, four sampled |
| generations per prompt, a verifiable reward |
| combining two-anchor validity, structural locality, and a train-only property verifier, |
| the reference-guidance loss, and KL regularization. Exact training and full-test metrics |
| are recorded in the linked repository and `training_meta.json`. |
|
|
| On all 8,176 balanced test requests, this adapter obtains 79.770% RDKit+TDC validity, |
| 51.248% two-anchor polymer validity, 33.745% changed outputs, 1.345% strict |
| MIPS-retrained full-edit hits, and 0.489% observed-DFT strict full-edit hits at 24.352% |
| DFT coverage. It does not outperform the polymer-adapted Molecular Optimization |
| Transformer and should be treated as a reproducible RePO adaptation baseline rather than |
| a claimed state-of-the-art result. |
|
|
| Upstream RePO code: <https://github.com/tmlr-group/RePO> |
|
|
| PolyEdit implementation and record-level evaluation: <https://github.com/promotion-kim/POLYEDIT/tree/tsyou/balanced-polymer-baseline-eval> |
|
|
