Datasets:
You need to agree to share your Hugging Face information to access this dataset
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
This audio dataset requires manual access approval. Explain your intended research or development use.
Log in or Sign Up to review the conditions and access this dataset content.
Yoruba MOSS-Audio laughter sample over two seconds
This public, manually gated dataset contains 1,000 randomly
selected Yoruba audio rows from the completed portion of the
Yoruba_audio_chunked_10 MOSS-Audio annotation run.
Selection requirements:
duration_seconds > 2has_chuckle_or_laughter == "yes"
The frozen source snapshot contained 53,500 finalized rows across
535 Parquet shards, of which 1,962 met
both requirements. The sample was drawn without replacement using Python
random.Random(20260930). If fewer than 1,000 rows had qualified, every eligible
row would have been included.
Each row preserves its embedded audio and all 69 columns from the MOSS-annotated source. MOSS-Audio-4B-Thinking received only the row audio during inference; transcript text, tags, and other metadata were not provided to the model. Private thinking traces are not stored in result fields. The laughter labels and speech captions are model-generated and should be reviewed before being treated as ground truth.
The full MOSS fleet run was still processing when this snapshot was frozen; this repository is a random sample of the finalized rows available at that time, not a sample from the eventual completed corpus.
selection_manifest.json records the exact seed, frozen source shard list,
candidate count, selected source row positions, validations, and SHA-256 hashes
of every published Parquet shard.
Audio playback
The embedded audio column is declared as a Hugging Face Audio feature and
can be played directly from the dataset viewer after access is approved.
- Downloads last month
- 40