Views
No views yet
1# Fine-Tuning (LoRa)
2
3pip install torch==2.5.0 torchvision==0.20.0 torchaudio==2.5.0 --index-url https://download.pytorch.org/whl/cu121
4
5pip install -U pip
6#pip install -U transformers==4.43.0
7pip install -U transformers
8#pip install -U tokenizers==0.19.1
9pip install -U tokenizers
10pip install -U bitsandbytes
11pip install -U accelerate
12pip install -U datasets
13pip install -U peft
14pip install -U trl
15
16from transformers import (
17 AutoModelForCausalLM,
18 AutoTokenizer,
19 BitsAndBytesConfig,
20)
21from peft import PeftModel
22import torch
23from tqdm import tqdm
24import json
25
26HF_TOKEN = ""
27
28model_id = "llm-jp/llm-jp-3-13b"
29adapter_id = "shigedon/llm-jp-3-13b-finetune-16-Dec-num-02-2024"
30
31# QLoRA config
32bnb_config = BitsAndBytesConfig(
33 load_in_4bit=True,
34 bnb_4bit_quant_type="nf4",
35 bnb_4bit_compute_dtype=torch.bfloat16,
36)
37
38# Load model
39model = AutoModelForCausalLM.from_pretrained(
40 model_id,
41 quantization_config=bnb_config,
42 device_map="auto",
43 token = HF_TOKEN
44)
45
46# Load tokenizer
47tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True, token = HF_TOKEN)
48
49
50# 元のモデルにLoRAのアダプタを統合。
51model = PeftModel.from_pretrained(model, adapter_id, token = HF_TOKEN)
52
53# データセットの読み込み。
54datasets = []
55with open("./elyza-tasks-100-TV_0.jsonl", "r") as f:
56 item = ""
57 for line in f:
58 line = line.strip()
59 item += line
60 if item.endswith("}"):
61 datasets.append(json.loads(item))
62 item = ""
63
64
65results = []
66for data in tqdm(datasets):
67
68 input = data["input"]
69 prompt = f"""### 指示
70 {input}
71 ### 回答
72 """
73
74 input_ids = tokenizer(prompt, return_tensors="pt",return_token_type_ids=False,).to(model.device)
75 outputs = model.generate(**input_ids, max_new_tokens=512, do_sample=False, repetition_penalty=1.2,)
76 output = tokenizer.decode(outputs[0][input_ids.input_ids.size(1):], skip_special_tokens=True)
77
78 results.append({"task_id": data["task_id"], "input": input, "output": output})
79
80
81# llmjp
82results = []
83for data in tqdm(datasets):
84
85 input = data["input"]
86
87 prompt = f"""### 指示
88 {input}
89 ### 回答
90 """
91
92 tokenized_input = tokenizer.encode(prompt, add_special_tokens=False, return_tensors="pt").to(model.device)
93 attention_mask = torch.ones_like(tokenized_input)
94 with torch.no_grad():
95 outputs = model.generate(
96 tokenized_input,
97 attention_mask=attention_mask,
98 max_new_tokens=100,
99 do_sample=False,
100 repetition_penalty=1.2,
101 pad_token_id=tokenizer.eos_token_id
102 )[0]
103 output = tokenizer.decode(outputs[tokenized_input.size(1):], skip_special_tokens=True)
104
105 results.append({"task_id": data["task_id"], "input": input, "output": output})
106
107
108# jsonl
109import re
110jsonl_id = re.sub(".*/", "", adapter_id)
111with open(f"./{jsonl_id}-outputs.jsonl", 'w', encoding='utf-8') as f:
112 for result in results:
113 json.dump(result, f, ensure_ascii=False) # ensure_ascii=False for handling non-ASCII characters
114 f.write('\n')
115
116