Intern-S1-mini context-conditioned molecule transfer V10.3 TDC-v2-HP — Oral bioavailability
Validation-selected checkpoint from the direct-inclusive TDC-v2-HP Oral bioavailability run. This was the best held-out macro-F1@5 result among the indirect-only and Starling-direct TDC variants.
Provenance
- Base model:
jiosephlee/Intern-S1-mini-lm - Base model revision:
fcb667c380ae01f57693a45b4b5c2d331052a107 - Training dataset:
jiosephlee/context-conditioned-molecule-transfer-v10.3-tdc-v2-hp-bioavailability-ma-mixed-continuous-intern - Dataset revision:
9db99d12536f5dbb3d906ab90e5acff1b036405c - Training data: 49,224 direct records, including TDC and Starling direct data, plus 49,224 indirect records
- Training configuration: 10 scheduled epochs, seed 42, soft-target loss, 8 GPUs, device batch size 4, gradient accumulation 4, Flash Attention 2, BFD packing, and padding-free training
- Selection metric: validation
knn_binary_macro_f1_at_5 - Selected optimizer step: 80
- Selected validation macro-F1@5:
0.6297 - Training was intentionally stopped at step 202 of 440; this repository contains the saved step-80 validation-selected checkpoint.
- Training run
- Held-out model/Morgan comparison
Held-out test comparison
| Ranker | Macro-F1@3 | Macro-F1@5 | Spearman@3 | Top-1 hit@3 |
|---|---|---|---|---|
| Model | 0.5657 | 0.5657 | 0.5774 | 0.7266 |
| Morgan vanilla | 0.5987 | 0.5060 | -0.0207 | 0.9297 |
| Morgan weighted | 0.6169 | 0.5439 | -0.0142 | 0.9219 |
Checkpoint selection and the TDC-variant comparison use macro-F1@5. All model and Morgan test metrics are available in the linked W&B evaluation run.
The repository contains the full Transformers checkpoint and tokenizer files. metric.json is the complete validation-selection record stored with the checkpoint.