Prompt2MusicBench: A Structured Prompt Dataset for Controllable Music Generation

Prompt2MusicBench is a large-scale dataset of 24800 structured text prompts designed for studying controllability in text-to-music models such as MusicGen. Prompts vary systematically across genre, tempo (BPM), instrument, and mood, and include 8 structural variants × 2 paraphrase forms for each combination.

This dataset contains only prompts (CSV) — no audio files. A companion dataset (Prompt2MusicLibrary) has all the 24800 audio clips generated from these prompts using facebook/musicgen-small.

Direct Use:

Training or evaluating audio classification models for genre, instrument, mood, or tempo.

Benchmarking controllability of text-to-music models (does the generated audio follow the prompt attributes?).

Studying prompt engineering effects in music generation.

Serving as a resource for multimodal models (text ↔ audio alignment).

Out-of-Scope Use:

The dataset itself contains only text prompts, not raw audio.

It should not be used as a source of real music recordings or as ground-truth for music transcription or speech recognition.

Dataset Structure:

Each row in the CSV corresponds to one prompt instance.

Columns:

id: Unique integer identifier.

prompt_text: The full structured prompt (e.g., “A peaceful ambient track at 80 BPM featuring synthesizer.”).

genre: The musical genre (11 genres: ambient, lo-fi hip hop, jazz, rock, pop, house, techno, trap, drum and bass, classical, folk).

tempo_bpm: Target tempo in beats per minute (BPM).

instrument: The lead instrument requested (genre-specific).

mood: Mood or affective descriptor (peaceful, melancholic, romantic, suspenseful, etc.).

structure_id: Indicates the prompt structure template used. Eight structure types are included, varying how the same musical attributes are phrased (e.g., canonical_simple, genre_first, instrument_first, mood_emphasized, verbose_descriptive, minimal_terse, instructional_command, creative_poetic).

paraphrase_id: Within each structure, two paraphrase variants are provided (0, 1) to increase linguistic diversity.

Example:

0: “A melancholic jazz track at 120 BPM featuring trumpet.”

1: “This is a melancholic jazz song with trumpet, played at 120 BPM.”