Skip to content

Add dataset: FSD50K #5222

Description

@yaswanth169

dataset link on Hugging dataset

https://huggingface.co/datasets/Fhrozen/FSD50k

Arxiv link

https://arxiv.org/abs/2010.00475

Description of the dataset

Part of the MOEB tracker (#4842) — audio→audio retrieval open slot.

FSD50K (Fonseca et al.) is a large-scale dataset of 51,213 real audio clips (Freesound recordings) spanning 200 sound-event categories organized under the AudioSet ontology, originally built for sound event classification/tagging. CC-BY-4.0, not gated — real audio files confirmed present in the HF mirror (not just metadata).

Plan: build an audio→audio retrieval task using category co-membership as the relevance signal (query clip → other clips sharing its sound-event label), sampling a benchmark-sized subset across a good spread of categories. Note: FSD50K's audio is already used elsewhere in mteb for classification (#2056) — this reuses the same permissively-licensed source for a genuinely new task direction (audio→audio retrieval), the same pattern already established for NSynth (used for both clustering and retrieval) elsewhere in this tracker.

Metadata

Metadata

Assignees

No one assigned

    Labels

    new datasetIssues related to adding a new task or dataset

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions