Upload folder using huggingface_hub

Upload hard_negatives.json with huggingface_hub

Upload folder using huggingface_hub

Upload hard_negatives.json with huggingface_hub

Upload folder using huggingface_hub

Upload hard_negatives.json with huggingface_hub

Upload folder using huggingface_hub

Upload hard_negatives.json with huggingface_hub

Upload folder using huggingface_hub

Upload hard_negatives.json with huggingface_hub

Upload folder using huggingface_hub

Upload hard_negatives.json with huggingface_hub

Filter page-locator questions (300 -> 286 rows). Drops questions referencing specific pages (page N, on page X, first/cover page, page ranges) since markdown strips page boundaries.

Replace dataset with cleaned v2: 371 → 300 rows. Dropped unanswerable, page-reference, page-count, and missing-md questions. Source-of-truth for the 0/2/2 + random K-sweep context-degradation experiment.

Add 7 gap-fill HN PDFs + updated hard_negatives_validated.json (all 15 in-plan docs now ≥4 valid HNs)

Add 37 web-sourced HN PDFs + updated hard_negatives_validated.json (script 16)

Upload hard_negatives_validated.json with huggingface_hub

Upload hard_negative_pdfs/arxiv_2405_20834.pdf with huggingface_hub

Upload hard_negative_pdfs/arxiv_2404_15406.pdf with huggingface_hub

Add 38 more hard negative PDFs via Unpaywall (total 140)

Add 102 hard negative PDFs (resumed)

Add hard negatives (validated, 0 contaminated out of 570 checks)

Upload folder using huggingface_hub

Upload dataset

initial commit