Intern-S1-mini assay-transfer model: Carcinogens degree 24
Validation-selected checkpoint for the Carcinogens V27 degree-24 assay-transfer task. The checkpoint was selected at training step 500 using the joint ID/OOD level-macro MAE@3.
Provenance
- Base model:
jiosephlee/Intern-S1-mini-lm(revisionfcb667c380ae01f57693a45b4b5c2d331052a107) - Training data:
jiosephlee/assay-transfer-record-level-v27-carcinogens-general-intern - Dataset revision:
086291e05ee17c87fb3933075d1d57680acb2c3c - Training setup: 10 scheduled epochs, device batch size 4, gradient accumulation 4, 8 GPUs, FlashAttention-2, padding-free BFD packing, soft-target loss
- Selection metric:
Carcinogens/degree24/overall/knn_regression_id_ood_level_macro_mae_at_3 - Best validation metric:
0.8675038371
The run was stopped after the best checkpoint had been preserved; this uploaded directory is that best checkpoint.
Held-out test comparison
Metrics are on the joint ID/OOD test panel. Lower MAE is better; higher Spearman and Top-1@3 are better.
| Method | Macro MAE@3 | Macro MAE@5 | Spearman | Spearman@3 | Top-1@3 |
|---|---|---|---|---|---|
| This model | 0.5587 | 0.5541 | 0.4197 | -0.0286 | 0.1438 |
| Morgan fingerprint (vanilla) | 0.8142 | 0.8046 | -0.0211 | 0.0506 | 0.0771 |
| Morgan fingerprint (weighted) | 0.7630 | 0.7327 | -0.0177 | 0.0095 | 0.0685 |
The test evaluation is logged in Weights & Biases.
Usage
Load with the standard Transformers APIs and use trust_remote_code=True for the bundled tokenizer implementation.