intern-s1-mini-assay-transfer-record-level-v27-ames-general-best

Intern-S1-mini: AMES general record-level assay transfer V27

This is the validation-selected full-model checkpoint from the AMES general V27 record-level assay-transfer run. It is trained to score A/B assay-transfer prompts, not as a general-purpose molecule predictor.

Provenance

• Base model: jiosephlee/Intern-S1-mini-lm
• Base model revision: fcb667c380ae01f57693a45b4b5c2d331052a107
• Training dataset: jiosephlee/assay-transfer-record-level-v27-ames-general-intern
• Dataset revision: 77a918b7aacb32abc9f1fbfae552106e46fb5276
• Training schedule: 1,000 optimizer steps, effective batch size 128, seed 42, soft-target loss
• Selection metric: validation joint ID/OOD level-macro KNN regression MAE@3
• Selected optimizer step: 725
• Weights & Biases training run

Metrics

The selected checkpoint's validation joint ID/OOD level-macro KNN regression MAE@3 was 0.4085 (ID 0.4519, OOD 0.3651; lower is better).

On the 1,000-query held-out test ranking split, the corresponding metric was 0.5576 for this model, 0.9218 for weighted Morgan, and 0.9954 for vanilla Morgan. Test NDCG@3 was 0.7463, versus 0.5218 and 0.5205; Spearman correlation was 0.4763, versus 0.0133 and 0.0143.

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