# Same reasoning as gold/.gitignore, and it survives the licence check.
#
# The modelbiome corpus itself is CC-BY-4.0 with a dataset card and a paper
# (A17.3, checked 2026-08-11), so the CORPUS is redistributable with attribution.
# That is not the question. Each card inside it carries its own model licence on
# top, and the sheet embeds card bodies verbatim to be self-contained -- so
# committing it would redistribute thousands of third-party texts on thousands of
# separate terms. The permissive wrapper does not launder what it wraps.
#
# Nothing is lost. search.py + make_sheet.py regenerate the sheet from the pinned
# corpus (repo revision and LFS sha256 in manifest.json) with a fixed seed, and
# row_hash ties every label to exact text. What gets committed is the COMPLETED
# sheet's labels, which are the author's own work -- see README.md.
candidates.jsonl
enriched_set.csv
# enriched_labels.csv IS committed -- labels only, card bodies stripped,
# the same shape gold_labels.csv travels in.
__pycache__/
