evolve-instruct

evolve llm training instruction, from english instruction to any language. this sample code is targeting Korean.

evolve code is based on h2o-wizardlm.
base_instruction.json is from wizardlm.
also based on paper WizardLM: Empowering Large Language Models to Follow Complex Instructions

evol_instruct.json is sample generation, about 10,000 q&a pair from base_instruction.jsonl and another 26,000 from alpaca_data.json using ChatGPT and it costs about $80.

Korean llm demo using this dataset. this model is on Huggingface lcw99/polyglot-ko-12.8b-chang-instruct-chat

You can try above Korean model on famouse LLM framework FastChat.