---
library_name: transformers
base_model: sagawa/ReactionT5v2-forward
tags:
  - chemistry
  - polymers
  - reactiont5
  - text2text-generation
  - molecular-editing
---

# PolyEdit ReactionT5 Egc checkpoints

This repository contains two full-fine-tuned checkpoints derived from
`sagawa/ReactionT5v2-forward` for the Egc portion of PolyEdit.

## Repository layout

| Hub folder | Training target | Intended use |
|---|---|---|
| `product-sft/` | Complete target PSMILES | Property-conditioned polymer generation baseline |
| `fragment-sft/` | Single-cut replacement fragment | Experimental source-constrained fragment generation |

These folders are independent Transformers checkpoints and each includes model weights, tokenizer,
generation configuration, training arguments, and training metadata.

## Loading a checkpoint

```python
from transformers import AutoModelForSeq2SeqLM, AutoTokenizer

repo_id = "luca0621/polyedit-reactiont5-egc"
tokenizer = AutoTokenizer.from_pretrained(repo_id, subfolder="product-sft")
model = AutoModelForSeq2SeqLM.from_pretrained(repo_id, subfolder="product-sft")
```

PolyEdit task strings must be processed with the sentinel-token escape map stored in
`training_meta.json`. The project helper `polyedit.reactiont5.encode_for_tokenizer` performs this
conversion and `decode_from_tokenizer` reverses it.

## Product-SFT checkpoint

The model receives a source PSMILES, source Egc, and requested Egc interval and generates a complete
product PSMILES. Evaluation uses five deterministic beams and a train-only fingerprint property
reranker.

On the frozen 31,756-request test split:

| Metric | Result |
|---|---:|
| Valid selected product | 98.29% |
| DFT-evaluable selection | 86.32% |
| DFT success given evaluable | 33.96% |
| Overall DFT-verified yield | 29.32% |
| Exact target selected | 1 / 31,756 |
| Audited source-edit plus property success | 2 / 31,756 |

The 29.32% result is a property-conditioned generation metric. It does **not** mean that the output
is a valid single-step edit of the supplied source.

Weight SHA-256:

```text
2a14f5fce517bbad16ec79667e094096c51e338a7a1f0d75e174719f60eaf6f5
```

## Fragment-SFT checkpoint

The fragment model was trained on 15,016 train examples from source/target pairs sharing a valid
two-anchor single-cut context. It predicts only the target replacement fragment. Generated
fragments must be applied to the source through the PolyEdit executor; they are not complete
polymer products.

- Train fragment classes: 405
- Validation examples: 232
- Best reported validation loss: 0.7923
- Training epochs: 5

Open-ended fragment beam decoding remained weak. The best source-constrained benchmark result in
the project uses a closed-vocabulary fragment ensemble and site ranker, not this checkpoint.

Weight SHA-256:

```text
686a3c99c86ca99df4861631fd8a4c513ac3da23a16049291a77f02784b7e081
```

## DFT metric caveat

No new quantum-chemistry calculation is run during evaluation. A generated product is DFT-evaluable
only when its `PolymerGraphID` has a frozen Egc value in `labeled_mmp_pairs.csv`. Outputs without a
frozen label are unknown, not confirmed failures. Always report DFT coverage together with
conditional success and overall verified yield.

One duplicated polymer has conflicting Egc labels (3.8424 and 4.7875 eV), affecting 174 requests.
The project audit document reports both the raw 31,756-request metrics and a clean 31,582-request
view that excludes this conflict.

## Scope and limitations

- Property: Egc only
- Split: leakage-safe MMP connected components, seed 42
- The product checkpoint is a generator baseline, not a guaranteed editor.
- The fragment checkpoint requires the PolyEdit source executor.
- Rule-SFT checkpoints that failed executable-rule evaluation are intentionally not published here.