README.md

---
license: cc-by-nc-sa-4.0
license_link: LICENSE
task_categories:
  - text-to-audio
  - text-retrieval
language:
  - en
tags:
  - music
  - captioning
  - music-information-retrieval
  - clap
pretty_name: AllMusicCaps
size_categories:
  - 100K<n<1M
configs:
  - config_name: default
    data_files: allmusiccaps_v1.jsonl
extra_gated_heading: "Access to AllMusicCaps dataset"
extra_gated_prompt: >-
  AllMusicCaps is released for non-commercial research under CC BY-NC-SA 4.0. 
  
  You may use the dataset for academic and non-commercial research, and redistribute derivative works under the same license, with attribution to the AllMusicCaps paper.
  
  You may not use it for commercial purposes, attempt to identify or contact the authors of the underlying reviews, or redistribute it as if it were your own.
  
  The dataset contains no audio, only track identifiers and LLM-generated captions derived from professional AllMusic album reviews linked to YouTube links. Rights in the original review text remain with their respective owners; this release covers the derived captions and annotations only.
  
  If you recover audio via the youtube_id field, you do so under your own responsibility and subject to the terms of the source platform and the applicable copyright holders. The authors provide no audio and grant no rights to it.
  
  Captions are machine-generated from third-party editorial text and may contain errors or biases. The dataset is provided "as is", without warranty of any kind.
  
  Please cite the paper in any publication that uses this dataset.

extra_gated_fields:
  Company: text
  Country: text
  I agree to use this dataset for non-commercial purposes only: checkbox
---

# AllMusicCaps

Music caption dataset built from professional AllMusic album reviews, cross-referenced with Discogs
releases and YouTube tracks. Captions are generated in two complementary styles by a two-stage LLM
pipeline.

Released with the ISMIR 2026 paper *AllMusicCaps: Album Reviews as Complementary Supervision for
Music CLAP*. Code and models: [github.com/MTG/allmusiccaps](https://github.com/MTG/allmusiccaps).

**This dataset contains no audio**, only identifiers and captions. Recover audio from the
`youtube_id` field, or map the identifiers onto your own collection.

## Versions

| File | Rows | Use it for |
|---|---|---|
| `allmusiccaps_v1.jsonl` | 540,454 | **Default.** Same content, with `generated_quotes_captions` normalized to a consistent type. |
| `allmusiccaps_v0.jsonl` | 540,454 | Exact reproduction of the paper. Raw second-stage LLM output. |

```python
from datasets import load_dataset

ds = load_dataset("mtg-upf/allmusiccaps", split="train")  # v1
```

v0 cannot be read with `load_dataset`: its unstable caption type is exactly what Arrow rejects
(that is why v1 exists). It is still valid JSONL, so read it line by line:

```python
import json
from huggingface_hub import hf_hub_download

path = hf_hub_download("mtg-upf/allmusiccaps", "allmusiccaps_v0.jsonl", repo_type="dataset")
with open(path) as f:
    rows = [json.loads(line) for line in f]
```

### Why there are two versions

v0 is the file the paper's models were trained on, kept unchanged so results stay reproducible.

In v0, `generated_quotes_captions` holds the second-stage LLM output as it came back. For 4,123
rows (0.8% of the dataset, 1.7% of rows that have quote captions) the model wrapped its answer in
something other than a list of strings: a nested list, a dict keyed by attribute name, or an
occasional number or null. The field therefore has an unstable type, and Arrow-backed readers reject
the file outright:

```
JSON parse error: Column(/generated_quotes_captions/[]) changed from string to array in row 25
```

v1 fixes only the container, never the text:

| v0 | v1 |
|---|---|
| `["a", "b"]` | `["a", "b"]` (unchanged) |
| `[["a", "b"]]` | `["a", "b"]` |
| `[{"description": "a"}]` | `["a"]` |
| `[{"mood": "x", "genre": "y"}]` | `["x", "y"]` |
| `[null]`, `[1]`, `[true]` | `null` |

Placeholder values (`"none"`, `"not specified"`, `"n/a"`) are dropped. Where a stray quotation mark
split one caption across a JSON key and its value, the two halves are rejoined. Row order, row
count, all identifier fields, and `generated_structured_captions` are byte-identical to v0; 384 rows
whose only quote entries were non-text end up with `null` instead.

The transformation is reproducible: see
[`scripts/preprocess_am_discotube/normalize_allmusiccaps_v1.py`](https://github.com/MTG/allmusiccaps/blob/main/scripts/preprocess_am_discotube/normalize_allmusiccaps_v1.py).

## Schema

One row per `(youtube_id, discogs_release_id)` pair.

| Field | Type | Description |
|-------|------|-------------|
| `youtube_id` | string | YouTube video ID. |
| `discogs_release_id` | string | Discogs release ID. |
| `allmusic_album_id` | string | AllMusic album ID (`mwXXXXXXXXXX`). |
| `youtube_url` | string | `https://www.youtube.com/watch?v={youtube_id}`. |
| `discogs_release_url` | string | `https://www.discogs.com/release/{discogs_release_id}`. |
| `allmusic_review_link` | string | `https://www.allmusic.com/album/{allmusic_album_id}`. |
| `generated_quotes_captions` | list[string] or null | Quote-style captions extracted from the review by Qwen2.5-32B (chatgpt_v2 prompt, t=0.5). Each string is a self-contained caption sentence. In v0 this field is not consistently typed; see above. |
| `generated_structured_captions` | object or null | Structured per-attribute caption from LLaMA-3.3 (promptv3). Keys: `music_style`, `mood`, `tempo`, `energy`, `instrumentation`, `production_style`. |

A row is emitted whenever at least one of the two caption sources is available; the missing source
is set to `null`.

## Statistics

| | Rows |
|---|---|
| Total | 540,454 |
| Both caption sources (v0) | 245,346 |
| Structured captions only (v0) | 295,096 |
| Quotes only (v0) | 12 |
| Quote captions in v1 | 985,981 across 244,974 rows |

## Two caption styles

**Quotes.** An extractor LLM is restricted to factual descriptions of musical and acoustic
characteristics (genre, production style, instrumentation, mood, rhythm) and explicitly excludes
subjective opinions, biographical information, release dates, commercial performance, and album or
artist names.

**Structured.** A single LLM receives all available metadata (AllMusic album review, YouTube
description and tags, Discogs genre and style tags) and a fixed schema of musical attributes, and
fills in each field, leaving it empty when no evidence is available.

## Licensing

Released for **non-commercial scientific research purposes only**.

Data extracted from the AllMusic database is released for non-commercial scientific research
purposes only. **Any publication of results based on the data extracts of the AllMusic database must
cite AllMusic as the source of the data.**

## Citation

```bibtex
@inproceedings{alonso2026allmusiccaps,
  title = {{AllMusicCaps}: Album Reviews as Complementary Supervision for Music {CLAP}},
  author = {Alonso-Jim{\'e}nez, Pablo and Lizarraga-Seijas, Xavier and Serra, Xavier and Bogdanov, Dmitry},
  booktitle = {International Society for Music Information Retrieval Conference (ISMIR)},
  year = {2026},
}
```