Modelpublic

jiosephlee/intern-s1-mini-assay-transfer-record-level-v27-bioavailability-ma-l2-best

sourceHugging Faceupdated 6d agoView on Hugging Face

0likes180downloads

Model Card

Intern-S1-mini Oral Bioavailability Assay Transfer v27 L2

Best validation checkpoint from record-level assay-transfer training on the L2 oral bioavailability dataset. The dataset was pinned to revision fecee3c43fe00d264fd7cbfbbcc82c6eb18379a6.

Training

—Selected checkpoint: step 500
—Training stopped after step 580 (4.5 of 10 planned epochs)
—Device batch size: 4 on 8 GPUs
—Gradient accumulation: 4
—Attention: FlashAttention 2
—Packing: padding-free BFD packing
—Selection metric: joint ID/OOD level-macro KNN regression MAE@3

Validation at the selected checkpoint

Metric	Value
Joint ID/OOD MAE@3	0.7386
ID MAE@3	0.9007
OOD MAE@3	0.5765
Spearman@3	0.0816
Top1@3	0.2231

Held-out test comparison

The v27 L2 test panel contains 500 ID queries. Lower is better for MAE; higher is better for the other metrics.

Ranker	MAE@3	MAE@5	Spearman	Spearman@3	Top1@3	NDCG@3
Model	0.7043	0.6953	0.3741	0.0075	0.1880	0.4030
Morgan vanilla	0.8312	0.8003	0.0205	0.0522	0.1140	0.3018
Morgan weighted	0.8302	0.8127	0.0412	0.0353	0.1040	0.3066

The model reduces MAE@3 by 15.3% versus vanilla Morgan and 15.2% versus weighted Morgan. Full per-query, per-bucket, and subgroup metrics are tracked in the associated W&B evaluation run.