Chinese Instruction-Tuning Data for Large Language Models

Alpaca GPT4 Zh provides Chinese-language instruction-response pairs derived from GPT-4 outputs for model fine-tuning. The dataset was curated by llamafactory, who removed over 6,100 mistruncated examples to improve quality. It was last updated in June 2024.

Use Cases

• Fine-tuning language models on Chinese instruction-response pairs for improved task following.
• Training models to generate Chinese text in a style similar to GPT-4 outputs.
• Creating benchmarks for evaluating Chinese question-answering and instruction-following capabilities.
• Augmenting multilingual training data with high-quality Chinese examples.

Strengths

• Dataset was cleaned by removing 6,103 mistruncated examples.
• Specifically formatted for direct use in the LLaMA Factory training framework.
• Focuses on Chinese language, a key resource for non-English model development.

Limitations

• The original total size and final row count after cleaning are unknown.
• Potential bias towards the style and knowledge cutoff of the source GPT-4 model.
• No information on the diversity of topics or potential class imbalance in instruction types.

Provenance

Source

Derived from GPT-4 outputs via the 'Instruction-Tuning-with-GPT-4/GPT-4-LLM' GitHub project.

Collection Method

Instruction-response pairs were generated by GPT-4 and subsequently cleaned.

Freshness

Last updated on 2024-06-07.