---
pretty_name: CulturalMenuBench
language:
  - zh
  - en
tags:
  - multimodal
  - culinary-reasoning
  - benchmark
  - multiple-choice
size_categories:
  - 10K<n<100K
---

# CulturalMenuBench

**CulturalMenuBench: Probing the Knowledge-Application Gap in Multimodal Culinary Reasoning**

[Paper](https://arxiv.org/abs/2609.03526) · [Code](https://github.com/BobTsang-NLP/CulturalMenuBench)

## Status and permissions

This is the **v3 release candidate** of CulturalMenuBench, hosted as a Hugging
Face dataset repository. The dataset is distributed as three independent ZIP
archives (one metadata archive and two media archives). This layout is intended
for archive download; the Hugging Face Dataset Viewer and
`datasets.load_dataset` are not supported. Extract the archives into one empty
directory and use the included loader.

No dataset license is granted here. Permission to redistribute third-party
recipe images, source text and translations remains subject to the authors'
rights review and any required institutional approval; see `RIGHTS.md`. The
companion code repository's Apache-2.0 license does not extend to third-party
content.

## Dataset scope

| Suite | Questions |
| --- | ---: |
| Binary | 4,870 |
| Chinese multiple-choice | 2,607 |
| Foreign multiple-choice | 3,217 |
| Subregion | 354 |
| **Total** | **11,048** |

The candidate contains 587 main dishes and 12,057 physical JPEG images addressed
by 12,060 image IDs; three IDs share existing files. All questions and images
are included. This is not a sample-only package.

## Download and extraction

Download these **three independent archives** and `SHA256SUMS`:

- `MenuBench_metadata_v3_candidate_20261021.zip`
- `MenuBench_media_v3_candidate_20261021_01-of-02.zip`
- `MenuBench_media_v3_candidate_20261021_02-of-02.zip`

Do not concatenate the archives. Verify the downloaded files, then extract all
three archives into the **same empty directory**:

```bash
sha256sum -c SHA256SUMS
mkdir data
python3 -m zipfile -e MenuBench_metadata_v3_candidate_20261021.zip data
python3 -m zipfile -e MenuBench_media_v3_candidate_20261021_01-of-02.zip data
python3 -m zipfile -e MenuBench_media_v3_candidate_20261021_02-of-02.zip data
cd data
sha256sum -c SHA256SUMS
python3 scripts/load_dataset.py --questions questions/binary.jsonl
python3 scripts/load_dataset.py --questions questions/mc-chinese.jsonl
python3 scripts/load_dataset.py --questions questions/mc-foreign.jsonl
python3 scripts/load_dataset.py --questions questions/subregion.jsonl
```

For the first checksum command, keep this repository's `README.md` and
`RIGHTS.md` alongside the three archives. The second checksum command checks
the extracted dataset rather than the download files.

Python 3.9+ with the standard library is sufficient for the included loader and
scorer. This repository layout is a downloadable archive distribution; it does
not promise automatic Hugging Face Dataset Viewer or `datasets.load_dataset`
support. Use the included loader or the code repository's scripts.

## Schema and evaluation

`questions/*.jsonl` contains the four task suites. `items.jsonl` contains dish
metadata, and `media_manifest.jsonl` resolves `asset://image:...` references to
image files. Preserve ordered question input blocks, including repeated images.
Never supply answer labels, source recipe metadata, or review records as model
inputs. `scripts/score.py` provides local scoring; it makes no model API calls.

## Version boundary and limitations

Dish display names and their task-specific question spans were cleaned, and
non-JPEG parent images were encoded as native-resolution JPEG Q95/4:4:4.
Existing JPEG files retain their original bytes. Question IDs, answers, option
order, image IDs and image order remain unchanged from the v2 candidate.
There is no super-resolution or new image repair in this packaging step.

Earlier image preprocessing includes LaMa/Qwen repair. Neither that processing
nor JPEG encoding guarantees watermark-free images; residual or uncertain
overlays may remain. Source attribution and preprocessing records are retained
inside the metadata archive, not discarded to make this bundle smaller.

This revised candidate has **not been reevaluated**. Historical paper scores
must not be presented as results measured on these revised inputs.
`KNOWN_ISSUES.jsonl` retains eight answer-change proposals and 21 scope proposals;
they have not been applied. See the extracted `README.md`, `RELEASE.json`, and
preprocessing/name-review records for detailed version information.

## Distribution

The dataset is hosted at
https://huggingface.co/datasets/brainz/CulturalMenuBench. Keep the three ZIP
archives, `README.md`, `RIGHTS.md` and `SHA256SUMS` together at the repository
root; the outer transport ZIP is not part of the dataset and should not be
uploaded as a fourth archive.

For updates, preserve the archive layout and refresh `SHA256SUMS`, then re-run
the extraction and loader checks. Any change to visibility or redistribution
scope must be consistent with `RIGHTS.md`, including attribution, public
contact and takedown information.