Jarbas/music_queries_templates

🎵 MusicQueries Dataset

MusicQueries is a synthetic natural language dataset focused on music-related utterances designed for media playback scenarios. Every sentence in this dataset is a playback request — a query that should result in a media search followed by music playback.

This dataset is ideal for training and evaluating models in intent classification, named entity recognition (NER), and retrieval-based music assistants.

This repository contains the sentence templates only, the generated datasets can be found in this collection

---

✨ Key Features

• Template-driven generation: Sentence templates covering a variety of music contexts (artist, track, album, genre, etc.) serve as the backbone for generating realistic utterances.
• Human-authored templates: All templates were carefully crafted by humans to reflect natural language usage across different music-related scenarios.
• LLM-augmented values: The sentence templates are populated with artist, track, genre, and theme values extracted and enhanced using language models and real-world metadata.
• Grounded synthetic sentences: Though synthetic, all generated queries are grounded in real-world data. For example:
  • Tracks actually belong to the specified albums.
  • Artists are correctly matched with countries, locations, and genres.
  • Themes and styles are derived from real band bios and discographies.
• Extensive artist database: Artist metadata has been compiled from publicly available information into the following datasets:
  • Jarbas/metal-archives-bands
  • Jarbas/metal-archives-tracks
  • Jarbas/jazz-music-archives
  • Jarbas/prog-archives
  • Jarbas/classic-composers
  • Jarbas/trance_tracks

---

🗂 Sentence Categories

Templates are organized by type, reflecting different combinations of available metadata:

• Artist-only: “Play something by {artist_name}”
• Track or Album queries: “Play the track {track_name}” or “Play the album {album_name}”
• Location-grounded: “Play artists from {location}”
• Genre and theme: “Play {music_genre} music” or “Play songs about {theme}”
• Combined contexts: “Play a {music_genre} track by {artist_name} from {country}”

---

📦 Dataset Contents

Each sample in the final synthetic_music_queries.csv includes:

• synthetic_query: the full playback sentence
• sentence_template: the original template used
• Music metadata fields (may be null if not used in that query):
  • track_name
  • artist_name
  • album_name
  • music_genre
  • country
  • location
  • theme

Example entry:

{
  "synthetic_query": "Play a progressive rock track by Rush from Canada",
  "sentence_template": "Play a {music_genre} track by {artist_name} from {country}",
  "track_name": null,
  "artist_name": "Rush",
  "album_name": null,
  "music_genre": "progressive rock",
  "country": "Canada",
  "location": null,
  "theme": null
}

---

🔧 Generation Method

The dataset was generated using a Python pipeline that:

1. Loads structured templates grouped by semantic category.
2. Samples real-world music metadata from curated datasets.
3. Fills templates with contextually consistent values.

The sentence templates and values were selected to maximize linguistic variety and domain coverage.

---

🧪 Use Cases

This dataset can be used for:

• 🔍 Intent classification (e.g., media playback vs. information queries)
• 🧠 NER training for music-specific entities (artists, albums, genres)
• 🗂 Query understanding for music assistant frontends
• 📈 Retrieval evaluation: matching user utterances to real-world content
• 🎧 Data augmentation: boost domain-specific examples for your ASR/NLU stack

---

📜 License

This dataset is released under MIT License. Template writing is original work, and all metadata is derived from openly available datasets.