Datasets:
Fix `languages=` usage example — it raises ValueError since the parquet conversion
The README's fourth usage example doesn't run. languages= was a loading-script
parameter, and the script was removed when this dataset moved to parquet, so the kwarg
is no longer accepted.
load_dataset("facebook/voxpopuli", "multilang", languages=["hr", "sk", "sl", "cs", "pl"])
# ValueError: BuilderConfig ParquetConfig(name='multilang', ...) doesn't have a 'languages' key
Tested
datasets==5.0.1, all four examples from the README, exactly as written:
| example | result |
|---|---|
load_dataset("facebook/voxpopuli", "hr") |
works |
load_dataset("facebook/voxpopuli", "multilang") |
works |
load_dataset("facebook/voxpopuli", "multilang", languages=[...]) |
ValueError |
load_dataset("facebook/voxpopuli", "en_accented") |
works |
The failing case was run through load_dataset() itself — it raises during config
resolution, before any download. The three passing cases were checked withload_dataset_builder() to avoid pulling the audio; that's the same config-resolution
path, so it confirms the configs resolve but not that the data downloads cleanly.
Fix
Replace the example with the parquet equivalent — load each language config and
concatenate:
from datasets import concatenate_datasets
languages = ["hr", "sk", "sl", "cs", "pl"]
voxpopuli_slavic = concatenate_datasets(
[load_dataset("facebook/voxpopuli", language, split="train") for language in languages]
)
All five language configs (hr, sk, sl, cs, pl) were verified to resolve withtrain/validation/test splits. README only; no data or config files touched.
Possibly related
Issues #11,
#12,
#13 and
#14 all report load
failures using trust_remote_code=True, i.e. the old script path. Those should be
resolved by the parquet conversion — worth closing if so, though I haven't reproduced
each one.
Also minor: en_accented has only a test split, which the README doesn't mention.
@lhoestq @polinaeterna — tagging you as this is a small follow-up to #16.
The parquet conversion dropped the loading script, so languages= is no longer an accepted kwarg, but the README still documents it as the way to load a subset of languages:
load_dataset("facebook/voxpopuli", "multilang", languages=["hr", "sk", "sl", "cs", "pl"])
# ValueError: BuilderConfig ParquetConfig(name='multilang', ...) doesn't have a 'languages' key
The other three examples in that section still work — I tested all four on datasets 5.0.1. This PR swaps the broken one for the parquet equivalent (load each language config and concatenate). README only.
Also possibly worth a look while you're here: #11–#14 all report load failures using trust_remote_code=True, which is the old script path — those look like the conversion already fixed them.
No urgency, and happy to redirect if someone else is better placed.