---
pretty_name: MMNeedle
language:
- en
license: cc-by-4.0
homepage: https://mmneedle.github.io/
paper: https://arxiv.org/abs/2406.11230
tagline: Benchmarking long-context multimodal retrieval under extreme visual context lengths.
tags:
- multimodal
- visual-question-answering
- long-context
- evaluation
size_categories:
- 100K<n<1M
---

# MMNeedle

MMNeedle is a stress test for long-context multimodal reasoning. Each example
contains a sequence of haystack images created by stitching MS COCO sub-images
into 1×1, 2×2, 4×4, or 8×8 grids. Given textual needle descriptions (derived
from MS COCO captions), models must predict which haystack image and which
sub-image cell matches the caption—or report that the needle is absent.

This dataset card accompanies the official Hugging Face release so researchers no
longer need to download from Google Drive or regenerate the benchmark from MS
COCO.

## Dataset structure

- **Sequences (`sequence_length`)**: either a single stitched image or a set of 10 stitched images.
- **Grid sizes (`grid_rows`, `grid_cols`)**: {1, 2, 4, 8} with square layouts.
- **Needles per query (`needles_per_query`)**: {1, 2, 5}. Each query provides that many captions.
- **Examples per configuration**: 10,000. Half contain the needle(s); half are negatives.
- **Total examples**: 210,000 (21 configurations × 10k samples).

Every example stores the full list of haystack image paths, the ground-truth
needle locations (`image_index`, `row`, `col`), the MS COCO image IDs for the
needles, the natural-language captions, and a `has_needle` boolean.

## Usage

```python
from datasets import load_dataset

ds = load_dataset("Wang-ML-Lab/MMNeedle", split="test")
example = ds[0]
print(example.keys())
# dict_keys(['id', 'sequence_length', 'grid_rows', 'grid_cols', 'needles_per_query',
#            'haystack_images', 'needle_locations', 'needle_image_ids',
#            'needle_captions', 'has_needle'])
```

Each entry in `haystack_images` is a PIL-compatible image object. `needle_captions`
contains one string per requested needle (even for negative examples, where the
corresponding location is `(-1, -1, -1)`).

## Data fields

| Field | Type | Description |
| --- | --- | --- |
| `id` | string | Unique identifier combining configuration and sample id. |
| `sequence_length` | int | Number of stitched haystack images shown to the model. |
| `grid_rows`, `grid_cols` | int | Dimensions of the stitched grid (each cell is 256×256 px). |
| `needles_per_query` | int | Number of captions provided for the sample (1, 2, or 5). |
| `haystack_images` | list of `Image` | Ordered haystack images for the sequence. |
| `needle_locations` | list of dict | One dict per caption with `image_index`, `row`, and `col` (−1 when absent). |
| `needle_image_ids` | list of string | MS COCO filenames that generated each caption. |
| `needle_captions` | list of string | MS COCO captions used as the needle descriptions. |
| `has_needle` | bool | True if at least one caption corresponds to a haystack cell. |

## Recommended evaluation protocol

1. Feed the ordered haystack images (preserving grid layout) plus the instruction
   template from the MMNeedle paper to your multimodal model.
2. Parse the model output into `(image_index, row, col)` triples.
3. Compare against `needle_locations` to compute accuracy for positives and the
   false-positive rate for negatives.

See the repository’s `needle.py` for a reference implementation.

## Source data

- **Images & Captions**: MS COCO 2014 validation split (CC BY 4.0).
- **Needle Metadata**: Automatically generated by the MMNeedle authors; included
  here as JSON files.

## Licensing

All stitched haystack images inherit the [Creative Commons Attribution 4.0
License](https://creativecommons.org/licenses/by/4.0/) from MS COCO. Attribution
at minimum should cite both MMNeedle and MS COCO.

## Citations

```
@article{wang2024mmneedle,
  title={Multimodal Needle in a Haystack: Benchmarking Long-Context Capability of Multimodal Large Language Models},
  author={Wang, Hengyi and Shi, Haizhou and Tan, Shiwei and Qin, Weiyi and Wang, Wenyuan and Zhang, Tunyu and Nambi, Akshay and Ganu, Tanuja and Wang, Hao},
  journal={arXiv preprint arXiv:2406.11230},
  year={2024}
}

@article{lin2014microsoft,
  title={Microsoft COCO: Common Objects in Context},
  author={Lin, Tsung-Yi and Maire, Michael and Belongie, Serge and Hays, James and Perona, Pietro and Ramanan, Deva and Doll{\'a}r, Piotr and Zitnick, C. Lawrence},
  journal={ECCV},
  year={2014}
}
```