AURA-Chat-Edit: Conversational Music Editing Dataset

Overview

AURA-Chat-Edit is a large-scale conversational music editing dataset used to train AURA, a unified multimodal framework for conversational music editing. The dataset contains 66,539 multi-turn dialogues pairing natural-language edit instructions with structured edit-token outputs across 7 edit types.

Each dialogue simulates a user requesting a music edit (e.g., "remove the drums", "add a jazzy saxophone") and an assistant responding conversationally while emitting typed edit-token blocks [EDIT_<KIND>][EDIT_0..7] that condition the audio generation bridge.

Dataset Statistics

Source Distribution

Edit Type Distribution

Conversation Types

Files

Dialogue Format (dialogues.jsonl)

Each line is a JSON object with the following fields:

{
  "id": "suno_xxx_c0_e0",
  "song_id": "suno_xxx",
  "chunk_id": "suno_xxx_c0",
  "conv_type": "text",
  "image_path": null,
  "chunk_path": "data/edit_dataset/chunks/suno_xxx/suno_xxx_c0.wav",
  "edited_path": "data/edit_dataset/edited/suno_xxx_c0/suno_xxx_c0_e0.wav",
  "edit_instruction": "add a subtle sub-bass drone underneath the pads",
  "edit_type": "instrument_add",
  "has_edit": true,
  "messages": [
    {
      "role": "user",
      "content": [
        {"type": "audio", "audio": "data/edit_dataset/chunks/...wav"},
        {"type": "text", "text": "can you add a subtle sub-bass drone..."}
      ]
    },
    {
      "role": "assistant",
      "content": [
        {"type": "text", "text": "I've added the sub-bass drone... [EDIT_ADD][EDIT_0]...[EDIT_7]"}
      ]
    }
  ],
  "split": "train",
  "llm_model": "qwen/qwen3.6-35b-a3b"
}

CoT Variant (dialogues_cot.jsonl)

Same format as dialogues.jsonl, but assistant responses include a <plan> tag before the edit tokens:

<plan> input: unknown | action: add synth | output: copy(mix) + gen(synth) </plan> I've added the sub-bass drone... [EDIT_ADD][EDIT_0]...[EDIT_7]

The plan encodes the stem-hybrid executor routing (Stem-Hybrid-CoT), used for the CoT planner variant of AURA.

Manifest Format (dataset.jsonl)

Each line contains audio chunk metadata including:

• Audio features: BPM, key, spectral features, energy bucket, onset density
• Chunk timing: start/end timestamps, duration, position (intro/verse/chorus/etc.)
• Edit pair references: source chunk path, edited audio path, edit type

Quick Start

from huggingface_hub import snapshot_download

# Download the dataset
repo_dir = snapshot_download(
    "OpenRB-Lab/AURA-Chat-Edit",
    repo_type="dataset"
)

Or download individual files:

from huggingface_hub import hf_hub_download

dialogues = hf_hub_download(
    "OpenRB-Lab/AURA-Chat-Edit",
    "dialogues.jsonl",
    repo_type="dataset"
)

Loading with Python

import json

dialogues = []
with open("dialogues.jsonl") as f:
    for line in f:
        dialogues.append(json.loads(line))

# Filter by split
train = [d for d in dialogues if d["split"] == "train"]
val = [d for d in dialogues if d["split"] == "val"]

# Filter by edit type
add_edits = [d for d in dialogues if d["edit_type"] == "instrument_add"]

Usage with AURA

This dataset is used to train the AURA thinker (Stage 1 — SFT) and joint thinker+bridge system (Stage 3). See the AURA model card and GitHub repository for training instructions.

Note: The audio files referenced in chunk_path and edited_path are not included in this release due to licensing constraints on the source audio. The dialogue annotations and metadata are released under Apache 2.0. To reproduce the full training pipeline, follow the dataset generation instructions in the GitHub repository.

Citation

@misc{trinh2026auraunifiedmultimodalframework,
      title={AURA: Unified Multimodal Framework for Conversational Music Editing}, 
      author={Quoc-Huy Trinh and Minh-Van Nguyen and Debesh Jha},
      year={2026},
      eprint={2609.14344},
      archivePrefix={arXiv},
      primaryClass={cs.SD},
      url={https://arxiv.org/abs/2609.14344}, 
}

License

The dialogue annotations and metadata in this dataset are released under the Apache License 2.0. The source audio referenced in the paths is subject to the licenses of their respective sources (Slakh2100, Suno, etc.) and is not redistributed here.