Fix dataset metadata and annotation loading
The dataset viewer currently selects the WebDataset loader for ordinary
audio TAR archives and fails to infer splits. The dataset card also has no YAML
metadata. This change adds explicit per-language configurations backed by typed
annotation Parquet files, plus language, task, license, and paper metadata.
The six added Parquet files contain the same 34,030 annotations as the original
JSONL files; they introduce no new samples or audio. In Parquet,
context_text_list triples become records with text/start/end fields. The original
JSONL files, audio archives, and existing cs_yodas.py are unchanged. The default
subset remains fra, following the legacy loader.
The README provides download and streaming-load examples and explains the cmn/zho
label difference, context-versus-target audio alignment, the six released subsets,
and the lack of an official evaluation split. No new Python scripts are included.
Validation: all six local README configurations load successfully, totaling
34,030 rows. Parquet round trips and streaming first-row comparisons pass for
each subset. The hf download command and remote streaming loads for all six
subsets on the PR branch were tested. Tested with Python 3.14.6, datasets 5.1.0, huggingface_hub 1.33.0, and pyarrow 26.0.0.
This restores annotation loading. The Parquet files do not embed audio, so web
audio playback is outside this change. Full-corpus audio alignment and the Hub
viewer rebuild have not been verified.
Hi @byan , I opened PR #2 (https://huggingface.co/datasets/byan/cs-yodas/discussions/2) to address the missing metadata and dataset loading errors. It updates the dataset card and adds Parquet versions of the existing annotations. All six subsets load successfully from the PR branch. Could you take a look when you have a chance? Thanks!