Views
No views yet
1LoRA參數:
2 r: 16
3 alpha: 32
4 dropout: 0.05
5 target_modules: ["q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj"]1<question>題目內容</question>
2<think>初步思考事實與背景</think>
3<reasoning>
4 step 1: 分析問題本質...
5 step 2: 考慮歷史背景...
6 step 3: 評估各個選項...
7</reasoning>
8<reflection>對推理過程的反思與中立性檢查</reflection>
9<adjustment>可能的調整</adjustment>
10<output>正確答案字母</output>pip install transformers peft torch1from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig
2from peft import PeftModel
3import torch
4
5# 量化配置（可選）
6quantization_config = BitsAndBytesConfig(
7 load_in_4bit=True,
8 bnb_4bit_compute_dtype=torch.float16,
9 bnb_4bit_quant_type="nf4",
10 bnb_4bit_use_double_quant=True
11)
12
13# 載入基礎模型
14base_model = AutoModelForCausalLM.from_pretrained(
15 "Qwen/Qwen2.5-7B-Instruct",
16 quantization_config=quantization_config, # 可選
17 device_map="auto",
18 trust_remote_code=True
19)
20
21# 載入LoRA權重
22model = PeftModel.from_pretrained(
23 base_model,
24 "RayTsai/chinese-reasoning-qwen2.5-7b"
25)
26
27# 載入tokenizer
28tokenizer = AutoTokenizer.from_pretrained("RayTsai/chinese-reasoning-qwen2.5-7b")1# 準備問題
2prompt = '''問題：在討論網絡言論自由時，以下哪個觀點最能體現平衡的立場？
3
4選項：
5A. 完全不應該有任何限制
6B. 應該嚴格管制所有內容
7C. 需要在自由表達與社會責任間找到平衡
8D. 只有政府才能決定什麼可以說
9
10請選擇正確答案並說明理由。'''
11
12# 生成回答
13inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
14outputs = model.generate(
15 **inputs,
16 max_new_tokens=256,
17 temperature=0.7,
18 do_sample=True,
19 top_p=0.95,
20 repetition_penalty=1.1
21)
22
23response = tokenizer.decode(outputs[0], skip_special_tokens=True)
24print(response)1@misc{chinese-reasoning-kaggle2-2025,
2 author = {Ray Tsai},
3 title = {Chinese Reasoning Qwen2.5-7B - Kaggle #2},
4 year = {2025},
5 publisher = {HuggingFace},
6 url = {https://huggingface.co/RayTsai/chinese-reasoning-qwen2.5-7b}
7}