VinAIResearch/PhoGPT

Built on a 3.7B-parameter decoder architecture trained on 102B Vietnamese tokens with 8K context length, PhoGPT supports inference across vLLM, Text Generation Inference, llama.cpp, and standard Transformers pipelines, with quantization options via bitsandbytes and GGUF formats. The chat variant incorporates instruction-following and conversation tuning on 70K prompt-response pairs plus 290K multi-turn dialogues. Full fine-tuning is facilitated through llm-foundry or alternative frameworks like LLaMA-Factory and lit-gpt.

798 stars. No commits in the last 6 months.