Linly

Why we included this project

Teams working with Chinese-language LLMs will find ready-made base weights and instruction-tuned checkpoints here, instead of having to adapt an English-first model. The project takes LLaMA and Falcon, expands their vocabularies, and continues pretraining on Chinese and mixed corpora, so the results handle Chinese text for generation, extraction, and question answering without a full retrain. Beyond those, Linly-OpenLLaMA was trained from scratch on roughly a terabyte of Chinese and English data with a Chinese-oriented tokenizer, and the ChatFlow dialogue checkpoints were built by following large multilingual instruction sets. The package is unusually open about how the models were made: data preparation, full-parameter training, and evaluation pipelines are documented, and weights come in both TencentPretrain and HuggingFace formats. For anyone who wants Chinese-capable models and the option to inspect or rerun the training, this is a strong reference.