OmegaGenome Distilled Students

Distilled student checkpoints from the OmegaGenome (NeurIPS 2026 revision) study. Two families of students, two teachers.

1. NTv3 single-track KD students (ntv3_kd_students/)

Small NTv3 students (8M / 100M) distilled per single genomic track from the NTv3-650M multitrack teacher (faithful recipe, 19,932 steps, cached teacher logits). Metric = per-track test Pearson r.

track assay student KD base (no-KD) teacher d(KD-base) % teacher
t0 PRO-cap 8M 0.4474 0.3801 0.5344 +0.0673 84%
t3 PRO-cap 8M 0.4485 0.4125 0.5671 +0.0360 79%
t11 eCLIP 8M 0.6142 0.5839 0.6886 +0.0303 89%
t12 ATAC 8M 0.7556 0.7469 0.8326 +0.0087 91%
t14 eCLIP 8M 0.4103 0.3929 0.4782 +0.0174 86%
t18 polyA-RNA 8M 0.6329 0.5933 0.7203 +0.0396 88%
t20 polyA-RNA 8M 0.6366 0.6089 0.6992 +0.0277 91%
t33 Histone 100M 0.8887 0.8828 0.9143 +0.0059 97%

All d>0 (KD beats the matched no-distillation baseline). t33 (Histone) is the 100M capacity-lever result (the 8M ties on t33; 100M opens headroom so KD wins). Files: ntv3_kd_students/{8m,100m}/t<N>_kd.pth (raw state_dict).

2. Carbon-3B BPNet KD students (carbon_bpnet_students/)

Tiny BPNet students distilled from Carbon-3B LoRA teachers on the NT18 classification tasks; best HP config per task (selected by validation MCC from the HP grid). Metric = MCC.

task best_val_mcc best_test_mcc
H2AFZ 0.5233 0.4919
H3K27ac 0.4896 0.4817
H3K27me3 0.6012 0.5899
H3K36me3 0.6176 0.5634
H3K4me1 0.4931 0.4718
H3K4me2 0.5926 0.5481
H3K4me3 0.6842 0.6319
H3K9ac 0.5614 0.4954
H3K9me3 0.4092 0.3972
H4K20me1 0.6360 0.6211
enhancers 0.5151 0.4975
enhancers_types 0.4831 0.4688
promoter_all 0.7541 0.7415
promoter_no_tata 0.7487 0.7305
promoter_tata 0.8544 0.8308
splice_sites_acceptors 0.9060 0.8907
splice_sites_all 0.8853 0.8779
splice_sites_donors 0.9337 0.9025

Files: carbon_bpnet_students/<task>.pt (raw state_dict).

Interim selection from the ongoing HP search; "best" config may shift as the grid completes. NTv3 100M near-miss students (t11/t18/t14/t0/t3) to follow once trained.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support