Frozen-FM embedding cache for 'Three Routes to Failure' (HyenaDNA, DNABERT-2)
收藏官方服务:
资源简介:
Frozen-embedding cache (last-hidden-state features from HyenaDNA-tiny and DNABERT-2-117M, mean-pooled and via the sequence-summary token, plus k-mer feature caches) that reproduces the cross-architecture frozen-FM existence-proof result in "Three Routes to Failure: An Interpretable Diagnosis of Task Difficulty and Negative-Set Contamination in Genomic Sequence-Classification Benchmarks" (PSB 2027) with zero PyTorch. Produced with a single fixed seed (42); 95% confidence intervals use a 1000-resample percentile bootstrap. Code: https://github.com/rikhinkavuru/gb_kmer_benchmark . See cache_manifest.txt for a per-file description.
提供机构:
Zenodo创建时间:
2026-06-01



