luoojason/mmlongbench-text-only

Dataset

MUDDLE - source questions and hard-negative metadata

0

26 commits

1 linked in READMEs

updated Aug 4, 2026

See the code

README

MUDDLE - source questions and hard-negative metadata

The 270 human-annotated questions behind MUDDLE, derived from MMLongBench-Doc, plus the curated hard-negative pool (hard_negatives.json) and the source / hard-negative PDFs.

Each question is tied to a single source document and is instantiated in five context conditions in the eval bundles: source alone, source + 2 or 4 topically similar hard negatives, and source + 2 or 4 length-matched random distractors.

Part of MUDDLE: Measuring Understanding of Documents under Distractor and Length Effects (COLM 2026 Workshop on Context Beyond the Window).

Contributors

luoojason

26 commits

luoojason/mmlongbench-text-only

Dataset

MUDDLE - source questions and hard-negative metadata

0

26 commits

1 linked in READMEs

updated Aug 4, 2026

See the code

README

MUDDLE - source questions and hard-negative metadata

The 270 human-annotated questions behind MUDDLE, derived from MMLongBench-Doc, plus the curated hard-negative pool (hard_negatives.json) and the source / hard-negative PDFs.

Each question is tied to a single source document and is instantiated in five context conditions in the eval bundles: source alone, source + 2 or 4 topically similar hard negatives, and source + 2 or 4 length-matched random distractors.

Part of MUDDLE: Measuring Understanding of Documents under Distractor and Length Effects (COLM 2026 Workshop on Context Beyond the Window).

Contributors

luoojason

26 commits