explcre/nt-2.5b-lora-teachers-r32

NT-2.5B per-task LoRA teacher adapters (r=32)

Per-task LoRA adapters fine-tuned on the Nucleotide Transformer revised 18-task benchmark.

• Base model: InstaDeepAI/nucleotide-transformer-2.5b-multi-species
• PEFT LoRA: r=32, lora_alpha=32, target_modules=[query, value], task_type=SEQ_CLS
• One folder per task under teacher_adapters/<task>/ (best-val checkpoint).

Best validation MCC per task

Task	Best val MCC
H2AFZ	0.5267
H3K27ac	0.5272
H3K27me3	0.5984
H3K36me3	0.6762
H3K4me1	0.5179
H3K4me2	0.5754
H3K4me3	0.6872
H3K9ac	0.5635
H3K9me3	0.4462
H4K20me1	0.6819
enhancers	0.5477
enhancers_types	0.4995
promoter_all	0.7689
promoter_no_tata	0.797
promoter_tata	0.6733
splice_sites_acceptors	0.9694
splice_sites_all	0.9577
splice_sites_donors	0.981

Load: base NT-2.5b-multi-species + PEFT adapter (merge for inference).