Fineweb-Edu-Chinese V2.2: 10B-100B Tokens for Educational LLMs

OpenCSG released Fineweb-Edu-Chinese V2.2 in February 2026, providing a massive corpus of 10 billion to 100 billion tokens for the Chinese education sector. This collection supports the full development lifecycle of Large Language Models by including both pre-training data and Supervised Fine-Tuning (SFT) instruction pairs.

Use Cases

• Pre-training foundational models using the 10B-100B token educational text corpus
• Supervised Fine-Tuning (SFT) for educational Question Answering using the instruction-tuning features
• Generating synthetic educational content based on the synthetic data patterns provided

Strengths

• Massive scale of 10B to 100B tokens
• Apache 2.0 license for open commercial and research use
• Integrated support for both pre-training and SFT workflows

Limitations

• Presence of synthetic data may introduce specific artifacts or hallucinations
• Metadata indicates a US region tag which may imply geographic biases in source selection despite the Chinese language focus

Provenance

Source

OpenCSG Community

Collection Method

Scraped and synthetic

Freshness

Last updated February 2026.

Geography

China (language focus), US (metadata region)