# Revised (shortcut-controlled) datasets

Datasets for the revised *Emergence of Biological Structural Discovery in General-Purpose
Language Models*. These remove the length and composition shortcuts present in naive constructions.

- `protein_pairs_matched.csv` — standard homology, **length-matched** negatives
  (|len1-len2| ~2.6 for both classes). Columns: sentence1, sentence2, label, split.
- `protein_pairs_remote_matched.csv` — remote homology (<25% identity, SCOP same-superfamily),
  length-matched; **no length/composition/k-mer shortcut** (a length+composition classifier scores
  0.50, chance). Columns: sentence1, sentence2, label, split.

`split` ∈ {train, val, test}. Validation is used only for polarity calibration; test is never used
for model selection. Code: https://github.com/maris205/biopaws