---
license: cc-by-nc-sa-4.0
pretty_name: FGMCaps Benchmark
language:
- en
tags:
- music
- music-retrieval
- audio-text-retrieval
- fine-grained
- contrastive-learning
size_categories:
- 10K<n<100K
configs:
- config_name: default
  data_files:
  - split: test
    path: data/test-*
---

# FIGMA: Towards FIne-Grained Music retrievAl (ACL 2026)

**Accepted to ACL 2026.**

**FGMCaps-Test** is a benchmark for **fine-grained music retrieval**: retrieving music from
natural-language descriptions that specify precise musical attributes, introduced in the paper '**FIGMA: Towards FIne-Grained Music retrievAl**'. This
benchmark provides **10,000 music clips**, each paired with a single-sentence caption
that describes fine-grained musical characteristics — **tempo (BPM), musical key,
chord progression, and beat / time signature** — alongside higher-level descriptive
content.

- **Paper:** https://arxiv.org/abs/2606.06615
- **Project page:** https://nishitanand.github.io/figma-website
- **Model:** https://huggingface.co/nishitanand/FIGMA
- **Code:** https://github.com/nishitanand/FIGMA

## Contents

| Column | Type | Description |
|---|---|---|
| `id` | string | Unique clip identifier (UUID) |
| `caption` | string | Fine-grained natural-language description of the clip |
| `audio` | audio | 24 kHz mono audio clip |

Split: `test` (10,000 examples). Audio renders directly in the dataset viewer.

## Usage

```python
from datasets import load_dataset

ds = load_dataset("nishitanand/FGMCaps-benchmark", split="test")
ex = ds[0]
print(ex["caption"])
audio = ex["audio"]          # {'array': np.ndarray, 'sampling_rate': 24000, 'path': ...}
```

Typical evaluation: encode every caption and every clip with a music–text retrieval
model, then measure Recall@K for text→audio and audio→text retrieval.

## Intended use

Intended for **non-commercial academic research** on fine-grained music retrieval and
music–language alignment.

## License

Released under **CC BY-NC 4.0-ShareAlike (CC BY-NC-SA 4.0)** for non-commercial
research use. Redistribution and derivatives must remain non-commercial and be shared
under the same terms, with attribution.

## Citation

```bibtex
@inproceedings{figma2026,
  title     = {FIGMA: Towards FIne-Grained Music retrievAl},
  author    = {Anand, Nishit and Seth, Ashish and Ghosh, Sreyan and Manocha, Dinesh and Duraiswami, Ramani},
  booktitle = {Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics},
  year      = {2026},
  url       = {https://arxiv.org/abs/2606.06615}
}
```