adzcai/babylm-eng-zho-50-50-stratified

---
language:
- eng
- zho
license: cc-by-4.0
---

# BabyLM English–Chinese 50/50 Stratified

A bilingual training corpus for the [Multilingual BabyLM](https://github.com/adzcai/multilingual) project, combining 50% of the [English](https://huggingface.co/datasets/BabyLM-community/babylm-eng) and [Chinese](https://huggingface.co/datasets/BabyLM-community/babylm-zho) BabyBabelLM corpora.

Sampling is **stratified by `category`**: each source category is sampled independently to preserve the original category proportions within each language.

## Token counts

| Language | Tokens | Share |
|----------|-------:|------:|
| English (`eng`) | 49,481,353 | 41.8% |
| Chinese (`zho`) | 68,925,921 | 58.2% |
| **Total** | **118,407,274** | **100%** |

## Category breakdown

| Category | English tokens | English % | Chinese tokens | Chinese % |
|----------|--------------:|---------:|--------------:|---------:|
| `child-available-speech` | 4,551,767 | 9.2% | 3,702,039 | 5.4% |
| `child-books` | 13,406,594 | 27.1% | 7,992,009 | 11.6% |
| `child-directed-speech` | 13,857,314 | 28.0% | 4,821,121 | 7.0% |
| `child-wiki` | 7,304,856 | 14.8% | 12,967 | 0.0% |
| `educational` | 0 | 0.0% | 6,732,733 | 9.8% |
| `padding-opensubtitles` | 9,856,922 | 19.9% | 0 | 0.0% |
| `padding-wikipedia` | 503,900 | 1.0% | 0 | 0.0% |
| `subtitles` | 0 | 0.0% | 45,665,052 | 66.3% |

## Source datasets

- [`BabyLM-community/babylm-eng`](https://huggingface.co/datasets/BabyLM-community/babylm-eng)
- [`BabyLM-community/babylm-zho`](https://huggingface.co/datasets/BabyLM-community/babylm-zho)