评测方法：25 道专业题（5 课程 × 5 题），与基座模型 Qwen3.5-2B 进行对比
评分规则：关键词命中率得分（0–6 分）+ 回答完整性得分（0–4 分）= 综合得分（0–10 分）
1import torch
2from transformers import AutoTokenizer, AutoModelForCausalLM
3
4model_path = "XuehangCang/MicroMajor-2B-AIAppTech"
5
6tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
7model = AutoModelForCausalLM.from_pretrained(
8 model_path,
9 torch_dtype=torch.bfloat16,
10 device_map="auto",
11 trust_remote_code=True,
12)
13model.eval()
14
15question = "请简述 Transformer 架构中 Self-Attention 的计算过程。"
16messages = [{"role": "user", "content": question}]
17
18text = tokenizer.apply_chat_template(
19 messages,
20 tokenize=False,
21 add_generation_prompt=True,
22)
23inputs = tokenizer(text, return_tensors="pt").to(model.device)
24
25with torch.no_grad():
26 output_ids = model.generate(
27 **inputs,
28 max_new_tokens=512,
29 temperature=0.7,
30 top_p=0.9,
31 do_sample=True,
32 pad_token_id=tokenizer.eos_token_id,
33 )
34
35new_ids = output_ids[0][inputs["input_ids"].shape[-1]:]
36print(tokenizer.decode(new_ids, skip_special_tokens=True))
1from transformers import pipeline
2
3pipe = pipeline(
4 "text-generation",
5 model="XuehangCang/MicroMajor-2B-AIAppTech",
6 torch_dtype="bfloat16",
7 device_map="auto",
8)
9
10messages = [{"role": "user", "content": "什么是 RAG？它在 AI 应用中有什么作用？"}]
11result = pipe(messages, max_new_tokens=512, temperature=0.7)
12print(result[0]["generated_text"][-1]["content"])
1@misc{MicroMajor-2B-AIAppTech,
2 author = {XuehangCang},
3 title = {MicroMajor-2B-AIAppTech: A Domain-Specific LLM for AI Application Technology Education},
4 year = {2026},
5 publisher = {Hugging Face},
6 url = {https://huggingface.co/XuehangCang/MicroMajor-2B-AIAppTech}
7}
