14.8K ·GitHub
What is this?
Llama-Chinese is a Chinese-language community project built around Meta's Llama models (Llama2, Llama3, Llama4). It provides Chinese pretraining and fine-tuning, a custom Atom model series optimized for Chinese, and resources for training, deploying and using Llama in Chinese for free commercial use.
Llama-Chinese is a Chinese-language community project built around Meta's Llama models (Llama2, Llama3, Llama4). It provides Chinese pretraining and fine-tuning, a custom Atom model series optimized for Chinese, and resources for training, deploying and using Llama in Chinese for free commercial use.Is this a model, a toolkit, or both?It's both: a collection of pretrained/fine-tuned Chinese Llama model weights (the Atom series and Llama2/3/4 Chinese fine-tunes) plus scripts and docs for pretraining, fine-tuning, quantization, deployment and inference acceleration.
It's both: a collection of pretrained/fine-tuned Chinese Llama model weights (the Atom series and Llama2/3/4 Chinese fine-tunes) plus scripts and docs for pretraining, fine-tuning, quantization, deployment and inference acceleration.How do I get a model running quickly?Install Python 3.10+, set up the environment, download a model like Atom-7B-Chat from HuggingFace/ModelScope/WiseModel, then run a quickstart.py script using transformers' AutoTokenizer and AutoModelForCausalLM. Docker, llama.cpp, gradio, API server, and ollama setups are also documented.
Install Python 3.10+, set up the environment, download a model like Atom-7B-Chat from HuggingFace/ModelScope/WiseModel, then run a quickstart.py script using transformers' AutoTokenizer and AutoModelForCausalLM. Docker, llama.cpp, gradio, API server, and ollama setups are also documented.What do I need before trying this?Python 3.10 or higher, the transformers library, and enough GPU resources to load the chosen model size (7B, 13B, 70B, up to 405B for Llama3.1 or 2T-parameter Llama4 Behemoth).
Python 3.10 or higher, the transformers library, and enough GPU resources to load the chosen model size (7B, 13B, 70B, up to 405B for Llama3.1 or 2T-parameter Llama4 Behemoth).Who is this community for?Chinese-speaking developers, NLP researchers, and enthusiasts who want to use, fine-tune, or deploy Llama models with better Chinese-language support, from beginners to experienced practitioners.
Chinese-speaking developers, NLP researchers, and enthusiasts who want to use, fine-tune, or deploy Llama models with better Chinese-language support, from beginners to experienced practitioners.Can I run these models fully offline after downloading?Yes, once downloaded the models can be run locally via Anaconda, Docker, llama.cpp, gradio, or ollama without needing a hosted API.
Yes, once downloaded the models can be run locally via Anaconda, Docker, llama.cpp, gradio, or ollama without needing a hosted API.What is the Atom model series?Atom is a family of Chinese-optimized models (1B, 7B, 13B) built on top of Llama2, trained with large-scale Chinese corpora, an expanded 65,000-word Chinese vocabulary for ~350% faster Chinese encoding/decoding, and adaptive context extension up to 32K tokens.
Atom is a family of Chinese-optimized models (1B, 7B, 13B) built on top of Llama2, trained with large-scale Chinese corpora, an expanded 65,000-word Chinese vocabulary for ~350% faster Chinese encoding/decoding, and adaptive context extension up to 32K tokens.Does the project support fine-tuning my own models?Yes, it provides full LoRA and full-parameter fine-tuning scripts for both Llama and Atom base models, plus guidance on data preparation and loading the fine-tuned weights afterward.
Yes, it provides full LoRA and full-parameter fine-tuning scripts for both Llama and Atom base models, plus guidance on data preparation and loading the fine-tuned weights afterward.Can I try the models without installing anything?Yes, the community hosts free online demo links and a free API for several Atom and Llama model sizes so you can test them in a browser first.
Yes, the community hosts free online demo links and a free API for several Atom and Llama model sizes so you can test them in a browser first.Does it support LangChain integration?Yes, LangChain support was added to let you plug the Chinese Llama/Atom models into LangChain-based applications.
Yes, LangChain support was added to let you plug the Chinese Llama/Atom models into LangChain-based applications.What are similar projects?Hy-MT2, decider, Chinese-LLaMA-Alpaca, airllm, unsloth, jeeves, rizzo-flow, minimind
Hy-MT2, decider, Chinese-LLaMA-Alpaca, airllm, unsloth, jeeves, rizzo-flow, minimindWhich topics is it in?rl, chinese-nlp, pretraining, llama-cpp, gguf, llama, lora, natural-language-processing, langchain, llm-inference, vllm, quantization
rl, chinese-nlp, pretraining, llama-cpp, gguf, llama, lora, natural-language-processing, langchain, llm-inference, vllm, quantization