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