Add portable adapter metadata and model card
Browse files
README.md
ADDED
|
@@ -0,0 +1,26 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
base_model: Qwen/Qwen2.5-3B-Instruct
|
| 3 |
+
library_name: peft
|
| 4 |
+
tags:
|
| 5 |
+
- polymers
|
| 6 |
+
- molecular-optimization
|
| 7 |
+
- reinforcement-learning
|
| 8 |
+
- repo
|
| 9 |
+
- polyedit
|
| 10 |
+
---
|
| 11 |
+
|
| 12 |
+
# PolyEdit RePO-polymer-retrained
|
| 13 |
+
|
| 14 |
+
This LoRA adapter uses the public RePO `XGRPOTrainer` with PolyEdit polymer references.
|
| 15 |
+
It is **not** an upstream RePO checkpoint: the official repository publishes training
|
| 16 |
+
code and recipes but no trained weights.
|
| 17 |
+
|
| 18 |
+
Training uses all eight properties with equal property mass, only PolyEdit train
|
| 19 |
+
components, Qwen2.5-3B-Instruct, four sampled generations per prompt, a verifiable reward
|
| 20 |
+
combining two-anchor validity, structural locality, and a train-only property verifier,
|
| 21 |
+
the reference-guidance loss, and KL regularization. Exact training and full-test metrics
|
| 22 |
+
are recorded in the linked repository and `training_meta.json`.
|
| 23 |
+
|
| 24 |
+
Upstream RePO code: <https://github.com/tmlr-group/RePO>
|
| 25 |
+
|
| 26 |
+
PolyEdit implementation and record-level evaluation: <https://github.com/promotion-kim/POLYEDIT/tree/tsyou/balanced-polymer-baseline-eval>
|