Dataset Viewer issue

#2
by Jeronymous - opened
OpenLLM France org

The dataset viewer is not working for "speech.asr" subset

The speech.asr config contains 49 while it should generally contain 3 splits maximum (train/validation/test). If the splits FLEURS_fr, Multilingual_LibriSpeech_fr, CommonVoice_fr, Yodas_fr, PxSLU_fr... are not used to differentiate between training and evaluation, please consider defining configs of this dataset instead.

We could not find a better way to package our dataset. Will the restriction on the number of splits be relaxed in the future ?

cc @lhoestq @cfahlgren1 .

You can have one split with all the language if you only specify data_dir.

  - config_name: speech.ast
    data_dir: "data/speech/ast"

This will fix the viewer

Abd users can then filter by language by specifying which files to load e.g. in datasets

load_dataset("OpenLLM-France/Luciole-Audio-Training-Dataset", "speech.ast", data_files="*_fr_en.jsonl")

if you really wish to keep two levels structure (task + language), I'd recommend having one dataset per task and then one config per language

Sign up or log in to comment