Build-LLM-Using-LLM-Data

Build LLM/MLLM From Scratch Using LLM/MLLM Data

Chat Test

Start vLLM server

python3 -m vllm.entrypoints.openai.api_server --model /my_model_path/ --host 0.0.0.0 --port 8001 --served-model-name my_model

CHAT

curl http://localhost:8001/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "my_model",
    "messages": [
      {"role": "system", "content": "You are a helpful assistant."},
      {"role": "user", "content": "你好"}
    ],
    "temperature": 0.7,
    "max_tokens": 4096
  }'

EVAL

evalscope eval --model my_model --eval-type openai_api --api-url http://127.0.0.1:8001/v1 --api-key EMPTY --datasets ceval --generation-config '{"max_tokens":6000}' --limit 10

Eval Results

Base Model	SFT 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