JinchuanTian's picture
Clarify pretraining language metadata
12d8905 verified
|
Raw
History Blame Contribute Delete
8.04 kB
---
pretty_name: Bagpiper Pretraining Data
language:
- en
task_categories:
- automatic-speech-recognition
- text-to-audio
tags:
- audio
- speech
- music
- sound
- multimodal
- rich-caption
- speech-language-model
- parquet
size_categories:
- 100M<n<1B
---
# Bagpiper Pretraining Data
Bagpiper Pretraining Data is the public rich-captioned audio snapshot associated
with [Bagpiper](https://bagpiper-cmu.github.io/), an open-ended audio language
model that learns bidirectional mappings between audio and comprehensive text
descriptions across speech, music, environmental sound, and mixtures.
The `en` metadata describes the primary rich-caption language. Source audio can
contain speech or singing in other languages; it is not an English-only audio
guarantee.
The repository contains 155,151,789 rows in 7,780 valid Parquet shards across
18 source-family directories, using 5.229 TB. It is intended for
large-scale research workflows; select only the source families needed for your
experiment rather than downloading the entire repository by default.
## Repository snapshot
| Source-family directory | Rows | Valid Parquet shards | GB |
|---|---:|---:|---:|
| `audiocaps` | 35,206 | 2 | 1.217 |
| `audioset` | 1,346,522 | 68 | 47.973 |
| `clotho_aqa` | 4,193 | 1 | 0.130 |
| `clotho_train` | 29,434 | 15 | 0.905 |
| `emilia_en` | 15,664,702 | 784 | 528.702 |
| `fma` | 2,317,679 | 116 | 84.010 |
| `laion_audio_300m_part1` | 22,439,013 | 1,122 | 469.972 |
| `laion_audio_300m_part2` | 23,840,302 | 1,193 | 527.090 |
| `laion_audio_300m_part3` | 24,424,316 | 1,222 | 541.761 |
| `laion_audio_300m_part4` | 18,983,198 | 950 | 352.998 |
| `laion_captioned_ai_music_snippets` | 2,448,073 | 123 | 150.037 |
| `laion_in_the_wild_sound_events` | 294,199 | 15 | 7.071 |
| `owsm_v4_caption` | 20,225,474 | 1,012 | 1,570.800 |
| `wavcaps` | 2,051,781 | 103 | 67.056 |
| `yodas_auto` | 14,251,524 | 713 | 516.842 |
| `yodas_manual` | 3,495,090 | 175 | 128.507 |
| `youtube_8m_arkive` | 2,327,341 | 117 | 165.080 |
| `yt8m` | 973,742 | 49 | 68.900 |
| **Total** | **155,151,789** | **7,780** | **5,229.050** |
The `clotho_train` directory also contains 43 zero-row, schema-only
`.parquet.tmp` files; they are temporary artifacts, are excluded from all counts above,
and must not be loaded. These are physical repository statistics for revision
`1997390aef538950729203ed4886b78b077d1a71`. Directory names identify packing
families and do not themselves establish ownership or a uniform license.
## Loading selected Parquet families
Use explicit Parquet globs to keep data acquisition intentional:
```python
from datasets import load_dataset
dataset = load_dataset(
"parquet",
data_files={
"train": [
"hf://datasets/espnet/Bagpiper_PreTrain_Data/audiocaps/*.parquet",
"hf://datasets/espnet/Bagpiper_PreTrain_Data/clotho_train/*.parquet",
]
},
split="train",
streaming=True,
)
print(dataset.features)
example = next(iter(dataset))
```
All audited valid shards use one shared schema:
| Column | Type | Meaning |
|---|---|---|
| `audio` | struct (`bytes`, `path`) | Embedded encoded audio; `path` may be null |
| `rich_caption` | string | Machine-generated comprehensive caption |
| `direction` | string | Training eligibility: `und`, `gen`, or `und,gen` |
| `category` | string | Speech, music, or sound taxonomy |
| `example_id` | string | Stable packed example identifier |
| `dataset` | string | Source-family identifier |
| `source_utt_id` | string | Source utterance identifier |
| `sample_rate` | int32 | Audio sample rate |
| `channels` | int32 | Audio channel count |
| `audio_format` | string | Encoded audio format |
| `start_time` | float64 | Optional source-segment start offset |
| `end_time` | float64 | Optional source-segment end offset |
`direction` denotes whether the pair is eligible for audio-to-text
understanding, text-to-audio generation, or both; it does not imply duplicated
physical rows. Inspect `dataset.features` before writing a consumer and ignore
`.parquet.tmp` files.
The Hub datasets-server currently exposes only a default `clotho_train` view
(29,434 rows), not the complete 18-family snapshot. Use the explicit glob
approach above for the intended families.
## Relationship to the paper
The Bagpiper paper describes a broader pretraining pipeline beginning from
approximately 422 million raw audio-caption pairs, with clips capped at 30
seconds and captions generated by a Qwen3-Omni captioner. Its 600B figure is a
training-token budget—300B text-to-audio, 150B audio-to-text, and 150B
text-only—not the row count or byte size of this Hub snapshot.
This repository is a public artifact associated with that pipeline. Do not
infer that its 18 directories are a complete one-to-one dump of every raw pair
or every text-only source described in the paper.
## Construction overview
Audio from heterogeneous speech, music, and sound collections is paired with
rich captions generated by the paper's Qwen3-Omni-30B-A3B-Captioner. A
Qwen3-32B classifier separates speech, music, and sound; text/audio quality and
alignment filtering uses heuristic and LLM judgments, UTMOS for speech,
AudioBox-Aesthetics for non-speech, CLAP alignment, Gumbel top-k sampling, and
MinHash text deduplication. Rich captions can describe:
- speech transcription, speaker attributes, language/accent, timing, and
acoustic conditions;
- instruments, genre, rhythm, melody, harmony, and musical structure;
- environmental events, temporal ordering, spatial context, ambience, and
recording quality.
Bagpiper uses these descriptions in both directions: audio-to-rich-caption for
understanding and rich-caption-to-audio for generation, alongside text-only
language-model data.
## Limitations, provenance, and responsible use
- Captions are machine-generated and may hallucinate content, timing, speaker
attributes, or acoustic details.
- Source families are heterogeneous and can include web-derived or transformed
media, copyrighted speech/music, identifiable voices, and sensitive content.
- This repository does not declare a blanket license. Users must review the
terms, attribution requirements, privacy/consent constraints, and
redistribution permissions of the underlying source represented by each
row before use or further redistribution.
- Directory-level public availability is not proof that every underlying asset
is cleared for every commercial or biometric use.
- The corpus is not exhaustively moderated for personal information, unsafe
content, bias, or offensive language.
- Generated captions and source media can inherit demographic, linguistic,
geographic, and cultural biases.
For provenance corrections or takedown requests, use the repository community
tab and provide the source-family directory plus a stable row identifier. Do
not repost sensitive media in the report.
## Related resources
- Paper: [Bagpiper: Solving Open-Ended Audio Tasks via Rich Captions](https://openreview.net/forum?id=FuHs64E3X6)
- Project: [bagpiper-cmu.github.io](https://bagpiper-cmu.github.io/)
- Base model: [espnet/bagpiper](https://huggingface.co/espnet/bagpiper)
- Bagpiper SFT data: [espnet/Bagpiper_SFT_Data](https://huggingface.co/datasets/espnet/Bagpiper_SFT_Data)
- Bagpiper-TTS SFT data: [espnet/Bagpiper_TTS_SFT_Data](https://huggingface.co/datasets/espnet/Bagpiper_TTS_SFT_Data)
- ESPnet: [espnet/espnet](https://github.com/espnet/espnet)
## Citation
```bibtex
@inproceedings{tian2026bagpiper,
title={Bagpiper: Solving Open-Ended Audio Tasks via Rich Captions},
author={Tian, Jinchuan and Wang, Haoran and Su, Bo-Hao and Huang, Chien-yu and Wang, Qingzheng and Shi, Jiatong and Chen, William and Gong, Xun and Arora, Siddhant and Li, Chin-Jou and Someki, Masao and Maekaku, Takashi and Goto, Keita and Shinohara, Yusuke and Sakuma, Jin and Yang, Chao-Han Huck and Watanabe, Shinji},
booktitle={Third Conference on Language Modeling},
year={2026},
url={https://openreview.net/forum?id=FuHs64E3X6}
}
```