Lab-MSP/Long-Listening-Thought

Long Listening Thoughts (LLT) data

Data for Long Listening Thoughts: Eliciting Open Auditory Reasoning with Deliberative Perception and Cognitive Refinement. Code: https://github.com/Lab-MSP/Long-Listening-Thought · Model: Lab-MSP/llt-dpo-qwen2.5-omni-7b · Demo: https://jaeyeonkim99.github.io/llt/

File	Content
afthink_metadata.json	368,860 AF-Think questions (audio ≤ 30 s) with normalized answers
llt_hard_metadata.json	323,400 LLT stage2 questions (<key>_hard; original_key points to the source clip)
llt_full_reasoning_traces.json	all reasoning traces: o (original short), oo / xo (verified long), x (incorrect short)
llt_sft_sharegpt.jsonl	SFT data used in the paper (1,159,275 rows, LLaMA-Factory ShareGPT format)
llt_dpo_60k.jsonl	DPO pairs used in the paper (60,000 rows, ms-swift format)

The metadata files also contain questions that were not used for training (e.g., a validation split and a set kept for further experiments), so they are larger than the counts reported in the paper (338k original / 316k stage2 questions).

Metadata entries:

"YouTube8M/4Bftz5jhHFY_23_10_hard": {
  "wav_path": "YouTube8M/4Bftz5jhHFY_23_10.mp3",
  "duration": 10.0,
  "original_key": "YouTube8M/4Bftz5jhHFY_23_10",
  "question": "...",
  "response": "short reasoning + answer",
  "answer": "D",
  "is_mcq": true,
  "choices": ["(A) ...", "(B) ...", "(C) ...", "(D) ..."]
}

Trace entries: {"key", "wav_path", "question", "reasoning", "answer", "type"} where key is the metadata key.

Audio is not included. Download the clips following the AF-Think instructions (each audio source has its own license). All wav_path / audios fields are relative to your audio root.

Audio layout

Keys and paths come from the AF-Think afthink/<Source>.json files:

• key = <Source>/<id>
• wav_path = <Source>/<id>.<ext>

Here <Source> is the json file name and <id> is the entry's id. Each file holds the audio of the entry's sound field, mixed down to mono at the original sample rate. When sound lists several segments (LibriSpeech, VoxPopuli), they are concatenated in order into one file. For VGGSound the audio track is extracted from the video.

<Source> (AF-Think json)	Clips	Ext	AF-Think sound	Our wav_path
YouTube8M	306,212	mp3	v1yO_FviAZM_74_10.mp3	YouTube8M/v1yO_FviAZM_74_10.mp3
AudioSet	55,872	flac	Y5M587BUfo0g.wav	AudioSet/Y5M587BUfo0g.flac
VGG	2,278	flac	oK2d2JF27Qk_000000.wav	VGG/oK2d2JF27Qk_000000.flac
AudioSet_SL	946	flac	YtF6G0Xryh3c.wav	AudioSet_SL/YtF6G0Xryh3c.flac
Freesound	798	flac	644531.wav	Freesound/644531.flac
DailyTalk	752	flac	1147.wav	DailyTalk/1147.flac
CochlScene	706	flac	Park_user0515_15311135_002.wav	CochlScene/Park_user0515_15311135_002.flac
FMA	396	mp3	047701.wav	FMA/047701.mp3
MELD	393	flac	492.wav	MELD/492.flac
MagnaTagATune	238	flac	<artist>-<album>-<track>-<n>.mp3	MagnaTagATune/<same name>.flac
FSD50k	98	flac	241546.wav	FSD50k/241546.flac
MACS	57	wav	park-milan-1063-41819-a.wav	MACS/park-milan-1063-41819-a.wav
MusicCaps	36	flac	19Pp9QEw17U.wav	MusicCaps/19Pp9QEw17U.flac
Clotho-v2	31	flac	MetroStation.wav	Clotho-v2/MetroStation.flac
LibriSpeech	29	flac	['7505-258958-0000.wav', '7505-258958-0001.wav']	LibriSpeech/7505_258958.flac (concatenated)
VoxPopuli	10	flac	['en/2017/20171114-0900-PLENARY-7-en_..._0.wav', ...]	VoxPopuli/20171114-0900-PLENARY-7.flac (concatenated)
BBC_Sound_Effects	6	flac	NHU05014209.wav	BBC_Sound_Effects/NHU05014209.flac
ESC-50	2	wav	3-151081-B-20.wav	ESC-50/3-151081-B-20.wav

Only clips of at most 30 s are kept. The extension does not matter to the scripts (audio is read with

License

This dataset is derived from AF-Think and is released under the NVIDIA OneWay Noncommercial License (a copy is included in this repository). It may be used for non-commercial research purposes only. The AF-Think questions and answers retain their original terms (see the AF-Think dataset card, including the OpenAI Terms of Use and the Qwen Research License for its synthetic content). New content in this dataset was generated with Qwen3 models.

Citation

@inproceedings{kim2026llt,
  title     = {Long Listening Thoughts: Eliciting Open Auditory Reasoning with Deliberative Perception and Cognitive Refinement},
  author    = {Kim, Jaeyeon and Yang, Chao-Han Huck and Zhang, Luoyi and Hsu, Chan-Jan and Ruiloba Portilla, Fernando and Tian, Jinchuan and Watanabe, Shinji and Busso, Carlos},
  booktitle = {Findings of the Association for Computational Linguistics: EMNLP 2026},
  year      = {2026}
}