promotion/polyedit-real-hard-task-augmentation

PolyEdit DFT-consistent task augmentation

This repository contains the SFT policies used for the 369, 1,000, and 2,000-request PolyEdit task-augmentation comparison. Each condition has five seed directories containing the policy, a self-contained policy bundle, the corresponding fine-tuned polyBERT Egc verifier, and the run configuration.

Training setup

• Property: DFT-labeled chain band gap, Egc
• Requests: 369 original, then 1,000 and 2,000 after train-only augmentation
• Decision steps: 1,263, 5,145, and 11,133
• Edit budget: at most three matched-pair transitions
• Policy loss: set-valued negative log-likelihood over valid next edits
• Seeds: 1004, 1005, 1006, 1007, 1008
• Evaluation: fixed held-out sources, 112 single-step and 36 multi-step requests

Results

Five-seed means and 95% t confidence intervals:

Requests	Single-target success	Multi-step success
369	0.889 ± 0.033	0.811 ± 0.089
2,000	0.941 ± 0.022	0.928 ± 0.019

Paired improvements were +0.052 for single-target success (p=0.040) and +0.117 for multi-step success (p=0.025).

Limitations

The policies were trained and evaluated only for Egc within the frozen matched-pair graph. Success therefore measures navigation among polymers with available DFT labels. It does not establish performance on arbitrary polymer structures or other properties.