Intern-S1-mini molecule transfer V9.0.2 TDC V2 — Skin Reaction
Validation-selected checkpoint trained on the Skin Reaction V9.0.2 TDC V2 combined dataset. The training data combines 13,536 TDC and 85,321 Starling direct records with deterministic 1:1 indirect augmentation.
Provenance
- Base model:
jiosephlee/Intern-S1-mini-lm - Base model revision:
fcb667c380ae01f57693a45b4b5c2d331052a107 - Training dataset:
jiosephlee/context-conditioned-molecule-transfer-v9.0.2-tdc-v2-skin-reaction-mixed-continuous-intern - Dataset revision:
4d97c745c735aa8d6dc07b81305e93f4e3bea496 - Training configuration: 10 scheduled epochs, seed 42, 4 GPUs, device batch size 2, gradient accumulation 16, effective batch size 128, and ranking validation every 20 optimizer steps
- Selection metric: validation KNN binary macro F1@5 (higher is better)
- Selected optimizer step: 80
- Selected validation F1@5:
0.7500 - The run was deliberately stopped at step 222; this repository contains its complete best-observed checkpoint.
- Training run
- Held-out model/Morgan comparison
Held-out test comparison
| Ranker | Binary macro-F1@3 | Binary macro-F1@5 | Spearman@3 | Top-1 hit@3 |
|---|---|---|---|---|
| Model | 0.6077 | 0.6192 | -0.2474 | 0.6585 |
| Morgan vanilla | 0.5661 | 0.5495 | -0.1612 | 0.8659 |
| Morgan weighted | 0.6265 | 0.5661 | -0.2136 | 0.8780 |
The combined model was selected over the indirect-only model because its primary test F1@5 was 0.6192 versus 0.5732, and it beats both Morgan variants on F1@5. The top-1 and F1@3 results are included to make the tradeoff explicit.
The repository contains the full Transformers checkpoint and tokenizer files. metric.json is the complete validation-selection record stored with the checkpoint. Loading requires trust_remote_code=True for the custom tokenizer.