---
library_name: pytorch
tags:
- polymer
- molecular-editing
- multi-objective-optimization
- hypervolume
- preference-alignment
- reproducibility
base_model: HAYDERphd/polyBERT
---

# PolyEdit multi-objective Hypervolume and MIP, five seeds

This repository preserves the preference-conditioned policies, Egc and Egb verifiers, and frozen-DFT evaluation records used in the 2026 GPU project interim report.

## Contents

- `checkpoints`: policy and two fine-tuned verifier bundles for seeds 1004 to 1008
- `results`: per-seed task-level results and W&B links
- `summaries`: aggregate metrics, analysis, evaluation manifest, and verification receipt
- `code`: exact metric, training, aggregation, and launch code
- `load_polyedit_bundle.py`: minimal policy and verifier loader

## Protocol

- Data: jointly labeled real DFT Egc and Egb records
- Split: 290 train and 253 evaluation polymers, with zero node overlap between induced edit graphs
- Evaluation: 22 sources, four property directions, and 11 preference weights
- Scale: 88 source-direction tasks and 968 preference decisions per seed
- Candidate budget: 11 outputs per task, each with edit budget 3
- Training: two polyBERT verifiers for 8 epochs and preference-conditioned SFT for 60 epochs
- Seeds: 1004, 1005, 1006, 1007, 1008
- Final evaluator: frozen DFT label lookup
- Base: `HAYDERphd/polyBERT`, revision `7bf9ed32ac54dea5bc163cf90100728b49341750`

## Mean results

Values are five-seed means with 95% t confidence intervals.

| Method | Hypervolume | MIP | Preference regret | Near-optimal |
|---|---:|---:|---:|---:|
| SFT | 0.326 ± 0.009 | 0.554 ± 0.012 | 0.034 ± 0.012 | 0.782 ± 0.091 |
| Random | 0.328 ± 0.005 | 0.500 ± 0.000 | 0.088 ± 0.000 | 0.448 ± 0.019 |
| Greedy verifier | 0.337 ± 0.004 | 0.568 ± 0.003 | 0.019 ± 0.003 | 0.833 ± 0.021 |
| Oracle grid | 0.355 ± 0.000 | 0.588 ± 0.000 | 0.000 ± 0.000 | 1.000 ± 0.000 |

SFT did not improve over the predeclared greedy verifier baseline. Its relative changes were -3.30 ± 3.08% for Hypervolume and -2.50 ± 2.30% for MIP. The negative result is retained to avoid method or seed selection bias.

## Limitations

The 88 tasks come from only 22 independent polymer sources. The four direction combinations and 11 weights discretize the preference space. Oracle grid uses evaluation DFT labels for candidate selection and is only a diagnostic upper bound. The result is restricted to the sparse within-dataset graph with both Egc and Egb labels.

## Code

Main project: https://github.com/promotion-kim/POLYEDIT