iclr2027anon/ROOTS

ROOTS

ROOTS is a bioacoustic audio-language training dataset containing generated language tasks paired with references to source audio. This repository contains language/conversation content and public-oriented audio identifiers only; it does not host audio files.

Rows: 43,925,394 Parquet shards: 8,901

Schema

• —id: reassigned stable ROOTS row ID.
• —tier: training tier, from 1 to 4.
• —category: coarse task category.
• —task: semantic task slug.
• —format: broad answer format, such as caption, mcq, binary, open_ended, or multi_label.
• —source_dataset: source dataset slug.
• —template: generation template path included in this repository, or Templated for templated/rule-based rows without a prompt template file.
• —source_id: source row identifier when available.
• —audio_paths: relative audio references. These are never private GCS URIs.
• —source_audio_ids: source audio identifiers aligned to audio_paths.
• —source_urls: public source URLs aligned to audio_paths when available.
• —audio_start_seconds and audio_end_seconds: crop or segment times aligned to audio_paths. See AUDIO.md for null values and SED crop suffixes.
• —num_audios: number of audio references for the row.
• —messages: conversation messages with role and content.

Audio Availability

Audio files are not included in this repository. See AUDIO.md for source mappings, crop conventions, and the resolver.

The companion index provides verified archive locations for the three hosted synthetic audio collections. It also records upstream identities and source-specific paths, with an explicit validation status for each asset. The four call-type/few-shot tasks remain pending an audio mapping; their 300,078 conversation rows are retained.

This cleaned release excludes 864,640 auxiliary AudioSetCaps rows. It does not change the historical model-training mixture. See release_changes.json for the exclusion log and audio_resolution_report.json for validation coverage.

Files

• —data/train-*.parquet: the single train split.
• —templates/: prompt templates referenced by the template column.
• —export_summary.json: row counts by tier, source dataset, task, format, and template.
• —audio_assets.sqlite.gz and roots_audio.py: audio index and resolver.
• —AUDIO.md: source setup instructions and crop conventions.
• —release_changes.json and audio_resolution_report.json: exclusions and validation results.

Loading

from datasets import load_dataset

dataset = load_dataset("iclr2027anon/ROOTS", split="train", streaming=True)

Review release

This is an anonymous review release. Author and hosting identifiers have been removed from documentation and metadata. Underlying source licenses and attribution requirements remain unchanged.