Views
No views yet
model_idadapter_idload_in_4bit=Truebnb_4bit_quant_type="nf4"bnb_4bit_compute_dtype=torch.bfloat16model.generate()を用いた出力生成
max_new_tokens=100repetition_penalty=1.2pad_token_id=tokenizer.eos_token_idresultsに保存resultsをJSON Lines形式で保存1!pip install -U bitsandbytes
2!pip install -U transformers
3!pip install -U accelerate
4!pip install -U datasets
5!pip install -U peft
6
7from transformers import (
8 AutoModelForCausalLM,
9 AutoTokenizer,
10 BitsAndBytesConfig,
11)
12from peft import PeftModel
13import torch
14from tqdm import tqdm
15import json
16
17
18# Hugging Faceで取得したToken
19HF_TOKEN = "Hugging Face Token"
20
21# ベースとなるモデルと学習したLoRAのアダプタ。
22model_id = "llm-jp/llm-jp-3-13b"
23adapter_id = "" # Hugging FaceのID
24
25# QLoRA config
26bnb_config = BitsAndBytesConfig(
27 load_in_4bit=True,
28 bnb_4bit_quant_type="nf4",
29 bnb_4bit_compute_dtype=torch.bfloat16,
30)
31
32# Load model
33model = AutoModelForCausalLM.from_pretrained(
34 model_id,
35 quantization_config=bnb_config,
36 device_map="auto",
37 token = HF_TOKEN
38)
39
40# Load tokenizer
41tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True, token = HF_TOKEN)
42
43# 元のモデルにLoRAのアダプタを統合。
44model = PeftModel.from_pretrained(model, adapter_id, token = HF_TOKEN)
45
46# データセットの読み込み。
47datasets = []
48with open("./elyza-tasks-100-TV_0.jsonl", "r") as f:
49 item = ""
50 for line in f:
51 line = line.strip()
52 item += line
53 if item.endswith("}"):
54 datasets.append(json.loads(item))
55 item = ""
56
57results = []
58for data in tqdm(datasets):
59
60 input = data["input"]
61
62 prompt = f"""### 指示
63 {input}
64 ### 回答
65 """
66
67 tokenized_input = tokenizer.encode(prompt, add_special_tokens=False, return_tensors="pt").to(model.device)
68 attention_mask = torch.ones_like(tokenized_input)
69 with torch.no_grad():
70 outputs = model.generate(
71 tokenized_input,
72 attention_mask=attention_mask,
73 max_new_tokens=100,
74 do_sample=False,
75 repetition_penalty=1.2,
76 pad_token_id=tokenizer.eos_token_id
77 )[0]
78 output = tokenizer.decode(outputs[tokenized_input.size(1):], skip_special_tokens=True)
79
80 results.append({"task_id": data["task_id"], "input": input, "output": output})
81
82import re
83jsonl_id = re.sub(".*/", "", adapter_id)
84with open(f"./{jsonl_id}-outputs.jsonl", 'w', encoding='utf-8') as f:
85 for result in results:
86 json.dump(result, f, ensure_ascii=False) # ensure_ascii=False for handling non-ASCII characters
87 f.write('\n')