A cleaned version of CCI3 (first 900 files), where:
All files with encoding errors (can't be read with pandas.read_json) are removed.
Samples with fewer than 100 characters are removed.
The jsonl files are converted to parquet for disk space efficiency.
