jiosephlee/intern-s1-mini-assay-transfer-record-level-v27-skin-reaction-combined-best

Intern-S1-mini: Skin Reaction record-level assay transfer V27

This is the selected full-model checkpoint from a 10-epoch assay-transfer training run on the Skin Reaction V27 combined L2/L3 dataset, pinned to dataset revision 1fa33cf3a7df7b6bd07a3b959f37502f951fc8ed. It is trained to score A/B assay-transfer prompts, not as a general-purpose molecule predictor.

The checkpoint was selected at optimizer step 380 by the lowest validation ranking joint ID/OOD level-macro KNN regression MAE@3 (0.1531). Ranking validation ran every 20 optimizer steps; the effective batch size was 128 (4 GPUs × device batch 4 × gradient accumulation 8). Training details and curves are in the W&B run.

On the 12-query held-out test ranking split, KNN regression MAE@3 was 0.248 for this model, 0.568 for vanilla Morgan, and 0.666 for weighted Morgan (lower is better). The model's top-1 hit@3 was 0/12, versus 2/12 for each Morgan baseline; these small-panel metrics should be read together.

The repository includes the tokenizer, model weights, and metric.json with checkpoint-selection details. Loading requires trust_remote_code=True for the custom tokenizer.