# C4 Evaluation Dataset

A multilingual subset of [allenai/c4](https://huggingface.co/datasets/allenai/c4) for measuring information-theoretic properties (perplexity, KL divergence) of LLMs.

## Dataset

- **File:** `c4_evaluation.parquet`
- **Total samples:** 2,994
- **Total tokens:** 2,007,152 (tokenized with `Qwen/Qwen3.5-35B-A3B`)
- **Columns:** `text`, `language`, `source`

## Language Composition

| Language | Code | Family | Samples | Tokens | % of tokens |
|----------|------|--------|--------:|-------:|------------:|
| English | `en` | Germanic | 1,244 | 695,908 | 34.7% |
| Chinese | `zh` | Sino-Tibetan | 316 | 245,202 | 12.2% |
| Japanese | `ja` | Japonic | 183 | 163,306 | 8.1% |
| Spanish | `es` | Romance | 236 | 160,108 | 8.0% |
| French | `fr` | Romance | 141 | 102,462 | 5.1% |
| German | `de` | Germanic | 143 | 101,514 | 5.1% |
| Korean | `ko` | Koreanic | 114 | 101,298 | 5.0% |
| Russian | `ru` | Slavic | 152 | 99,622 | 5.0% |
| Arabic | `ar` | Semitic | 141 | 99,938 | 5.0% |
| Portuguese | `pt` | Romance | 128 | 80,766 | 4.0% |
| Hindi | `hi` | Indo-Aryan | 95 | 78,244 | 3.9% |
| Tamil | `ta` | Dravidian | 101 | 78,784 | 3.9% |