# commonwealth-wiki-mix

FineWeb-style Parquet shards (LLM-training friendly) created by merging multiple Wikipedia language datasets into a single dataset to reduce looping during training.

## What’s inside

- **Format**: `nanochat-parquet-v1`
- **Layout**: `shard_*.parquet` + `metadata.json`
- **Text column**: `text`
- **Parquet settings**: zstd (level 3), `row_group_size=1024`, `use_dictionary=False`, `write_statistics=False`

## Sources included

This mix was built from the already-exported wiki datasets in `~/.cache/nanochat/datasets/monolingual/`:

- South Asia: Urdu, Hindi, Bengali, Tamil, Telugu, Malayalam, Marathi, Kannada, Gujarati, Punjabi, Nepali, Sinhala
- Southeast Asia: Malay
- Africa: Swahili, Yoruba, Afrikaans, Zulu

## Loading

In Python, you can load shards directly with PyArrow / Pandas. For HuggingFace `datasets`, Parquet loading works out of the box:

```python
from datasets import load_dataset

ds = load_dataset("JayJayThrowThrow/commonwealth-wiki-mix", split="train")
print(ds.column_names)  # ["text"]
```