The problem / use case
Context: make-gold already turns an authoring spec (query, source_id, answer)
into verified source-span gold (it checks quoted_text matches the offsets). A CSV
front-door would lower the barrier for real corpora.
What: Add import-csv --corpus docs.jsonl --spec spec.csv --out queries.jsonl that
reads a CSV with query,source_id,answer[,answer_all] columns and produces verified gold.
Acceptance:
Pointers: reuse make-gold / gold.py's verification — don't re-implement span checking.
Proposed idea
Add a CSV importer that takes query, source_id, answer rows and runs them through the existing make-gold authoring path, so people can bring their own data without hand-writing JSONL or computing offsets.
The problem / use case
Context:
make-goldalready turns an authoring spec (query,source_id,answer)into verified source-span gold (it checks
quoted_textmatches the offsets). A CSVfront-door would lower the barrier for real corpora.
What: Add
import-csv --corpus docs.jsonl --spec spec.csv --out queries.jsonlthatreads a CSV with
query,source_id,answer[,answer_all]columns and produces verified gold.Acceptance:
queries.jsonl; malformed offsets still fail closedtests/test_authoring.py;CHANGELOG.mdupdatedPointers: reuse
make-gold/gold.py's verification — don't re-implement span checking.Proposed idea
Add a CSV importer that takes query, source_id, answer rows and runs them through the existing make-gold authoring path, so people can bring their own data without hand-writing JSONL or computing offsets.