Build-LLM-Using-LLM-Data

Build LLM/MLLM From Scratch Using Only LLM/MLLM Data Distillation

Chat Test

start vLLM server

python3 -m vllm.entrypoints.openai.api_server \
  --model /the_model_path/ \
  --host 0.0.0.0 \
  --port 8000

do request

curl http://localhost:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "/the_model_path/",
    "messages": [
      {"role": "system", "content": "You are a helpful assistant."},
      {"role": "user", "content": "你好"}
    ],
    "temperature": 0.7,
    "max_tokens": 4096
  }'

Eval Results

Base Model	Train Dataset	ceval	cmmlu	mmlu	math_500
Qwen3-32B-Base	Chinese, Non-thinking, 3 Epoch	0.7077	0.791	0.7789	0.78
Qwen3-32B-Base	Chinese + English, Non-thinking, 2 Epoch	0.7194	0.8075	0.7871	0.8133