Datasets:
Add MUDDLE dataset card
Browse files
README.md
ADDED
|
@@ -0,0 +1,35 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: cc-by-4.0
|
| 3 |
+
task_categories:
|
| 4 |
+
- question-answering
|
| 5 |
+
tags:
|
| 6 |
+
- context-degradation
|
| 7 |
+
- long-context
|
| 8 |
+
- multi-document
|
| 9 |
+
- markdown
|
| 10 |
+
size_categories:
|
| 11 |
+
- 1K<n<10K
|
| 12 |
+
---
|
| 13 |
+
|
| 14 |
+
# MUDDLE eval bundle - MARKDOWN modality
|
| 15 |
+
|
| 16 |
+
Each document is rendered once to markdown and referenced by every cell that uses it.
|
| 17 |
+
Same cell ids, same seeded distractor selection, and the same source-first ordering as
|
| 18 |
+
the PDF and page-image bundles; only the input rendering differs.
|
| 19 |
+
|
| 20 |
+
This is the modality the distractor sweep reported in the paper is run in, because a
|
| 21 |
+
source document plus its distractors exceeds current image and PDF input limits.
|
| 22 |
+
|
| 23 |
+
## Related
|
| 24 |
+
|
| 25 |
+
Part of **MUDDLE: Measuring Understanding of Documents under Distractor and Length Effects**
|
| 26 |
+
(COLM 2026 Workshop on Context Beyond the Window).
|
| 27 |
+
|
| 28 |
+
| Piece | Where |
|
| 29 |
+
|---|---|
|
| 30 |
+
| Code | https://github.com/luoojason/muddle |
|
| 31 |
+
| Questions + hard-negative metadata | https://huggingface.co/datasets/luoojason/mmlongbench-text-only |
|
| 32 |
+
| Eval cells, PDF modality | https://huggingface.co/datasets/luoojason/muddle-eval-bundle |
|
| 33 |
+
| Eval cells, markdown modality | https://huggingface.co/datasets/luoojason/muddle-eval-bundle-md |
|
| 34 |
+
| Eval cells, page-image modality | https://huggingface.co/datasets/luoojason/muddle-eval-bundle-img |
|
| 35 |
+
| Paper | https://openreview.net/forum?id=9YS7tSPM9c |
|