---
license: cc-by-nc-4.0
language:
- fa
task_categories:
- audio-classification
pretty_name: PMB
size_categories:
- 10K<n<100K
dataset_info:
  features:
  - name: audio
    dtype:
      audio:
        sampling_rate: 32000
  - name: id
    dtype: string
  - name: artist
    dtype: string
  - name: song
    dtype: string
  - name: duration
    dtype: float64
  - name: genre_primary
    dtype: string
  - name: genre_raw
    dtype: string
  - name: key
    dtype: string
  - name: key_tonic
    dtype: string
  - name: key_mode
    dtype: string
  - name: valence_num
    dtype: float64
  - name: valence_cat
    dtype: string
  - name: arousal
    dtype: string
  - name: tempo
    dtype: string
  - name: popularity
    dtype: float64
  - name: caption_ref
    dtype: string
  - name: instruments_raw
    dtype: string
configs:
- config_name: default
  data_files:
  - split: benchmark
    path: data/benchmark-*
---

# PMB: a zero-shot benchmark for music understanding in Persian music

**13,544 clips (~20 s, 32 kHz mono MP3)** of Persian music with labels for
zero-shot evaluation of audio-language models: **genre** (7 classes),
**musical key** (24 classes; also tonic-only and mode-only granularities),
**emotion** (valence 0–100 + 3-class bins; arousal 3-class), **tempo**
(4 ordered classes), plus reference captions, Spotify popularity, and
artist/song metadata.

Derived from the PMG dataset (supervised split): the persian-pop genre was
downsampled to **1,000 distinct songs (one clip each)**, spread across 113
artists; all clips of the remaining six genres are retained.

| genre | clips |
|---|---|
| afghan pop | 5,449 |
| persian rock | 3,276 |
| classic persian pop | 2,299 |
| persian traditional | 1,277 |
| persian pop | 1,000 |
| persian alternative | 232 |
| persian neo-traditional | 11 |

## Usage

```python
from datasets import load_dataset
ds = load_dataset("keepsolid001/PMB", split="benchmark")
```

## Label provenance & caveats

- Genre/key/emotion/tempo labels are **track-level, Spotify-derived** metadata
  inherited by each clip. A classical key-detection baseline
  (Krumhansl-Schmuckler) agrees with the key labels at 0.37 (24-way; chance
  0.04) and 0.69 (mode), validating them as benchmark gold; tempo categories
  are softer (beat-tracked BPM agrees at only 0.30).
- `valence_cat` bins `valence_num` (0-100) at <40 / 40-60 / >60.
  `arousal` is the dataset's own 3-way energy category.
- `caption_ref` is metadata-templated prose - suitable for attribute-coverage
  metrics, not as human-written caption gold.
- Multiple clips of the same song share labels for the non-persian-pop genres;
  split by `song`/`artist` to avoid leakage when training.

## Benchmark results

Eleven systems (audio-LLMs, contrastive audio-text models, and a classical DSP
baseline) have been evaluated zero-shot on this set across genre, key, emotion,
tempo, captioning (incl. cultural identification), language identification,
instrument recognition, and stem-based hallucination tests. See the paper for
full results and analysis.

## License & provenance

Audio excerpts of commercial Persian music, distributed for
**non-commercial research only** (CC-BY-NC-4.0). If you are a rights holder
and want content removed, open a discussion on this repository.

## Citation

Anonymous — under review. A citation will be added upon publication.