README.md

---
library_name: pytorch
tags:
- polymers
- property-prediction
- graph-transformer
- polyedit
---

# PolyEdit MIPS-retrained

This is a leakage-controlled retraining of the public MIPS architecture for the eight
PolyEdit DFT properties: `Egc`, `Egb`, `Eea`, `Ei`, `EPS`, `Nc`, `Xc`, and `Eat`.
It is **not** a checkpoint released by the MIPS authors. The public repository does not
include its required MAE or downstream checkpoints, so weights were trained from scratch
on PolyEdit train components.

The model uses MIPS star-link polymer graphs, complete-graph path features, the 200-D
molecular-descriptor modality, the 512-D 3D atom-pair modality, and the public
`KfuseGraphTransformer` with cross-attention fusion. The split contains 1,300 train,
75 validation, and 953 held-out polymer records with zero connected-component overlap.

The strict target tolerances are much narrower than this retrained evaluator's held-out
errors. When the true reference targets themselves are scored, only 10.996% fall inside
the requested tolerance. MIPS-based generation numbers must therefore be presented as
evaluator hit rates alongside this calibration, not as ground-truth chemical success.

Upstream code: <https://github.com/wjxts/MIPS>

PolyEdit implementation and full metrics: <https://github.com/promotion-kim/POLYEDIT/tree/tsyou/balanced-polymer-baseline-eval>