Cleaned negative-set splits for human enhancer classification benchmarks (Three Routes to Failure)

Cleaned train/test splits for three enhancer-classification benchmarks (human_enhancers_cohn from Genomic Benchmarks, nt_enhancers from the Nucleotide Transformer downstream tasks, and drosophila_enhancers_stark as a clean control). The cleaning removes a seeded random subset of negative-class sequences carrying high-confidence TATA-box hits (TBP PSSM ≥ 1.3 bits/pos, JASPAR2024 CORE vertebrate profile), with the subset size equal to the negative-minus-positive TATA-box rate gap (seed 42). Positives are unchanged. Drosophila negatives have no TATA excess and are unchanged bit-for-bit (control). Files are CSV (sequence, label). The cleaning criterion and rationale are described in the associated paper.