Phi2-mini-Chinese

Why we included this project

This 0.2B-parameter Chinese model follows the Phi-2 architecture, but the real draw is that the repo documents the whole training journey rather than just shipping a checkpoint. You get the data cleaning steps, BPE tokenizer training, causal language model pretraining, supervised instruction tuning, and DPO preference optimization, each laid out in sequence. The Hugging Face weights load directly for Chinese text generation, and there is a notebook that shows hooking the model into a local knowledge base through LangChain for retrieval-augmented question answering. The authors are upfront that this is an experimental project with limited pretraining data, so treat it as a hands-on starting point for training your own small model rather than a production assistant.