Views
No views yet
1import json
2import re
3
4import torch
5from peft import PeftModel
6from tqdm import tqdm
7from transformers import (
8 AutoModelForCausalLM,
9 AutoTokenizer,
10 BitsAndBytesConfig,
11)
12
13
14model_id = "hiroakihara/llm-jp-3-13b-finetune"
15hf_token = "YOUR HUGGINGFACE TOKEN"
16test_jsonl_data = "elyza-tasks-100-TV_0.jsonl"
17
18bnb_config = BitsAndBytesConfig(
19 load_in_4bit=True,
20 bnb_4bit_quant_type="nf4",
21 bnb_4bit_compute_dtype=torch.bfloat16)
22
23model = AutoModelForCausalLM.from_pretrained(
24 model_id,
25 quantization_config=bnb_config,
26 device_map="auto",
27 token = hf_token)
28
29tokenizer = AutoTokenizer.from_pretrained(
30 model_id,
31 trust_remote_code=True,
32 token=hf_token)
33
34datasets = []
35with open(test_jsonl_data) as f:
36 item = ""
37 for line in f:
38 line = line.strip()
39 item += line
40 if item.endswith("}"):
41 datasets.append(json.loads(item))
42 item = ""
43
44results = []
45for data in tqdm(datasets):
46 input = data["input"]
47 prompt = f"""### 指示
48 {input}
49 ### 回答
50 """
51 tokenized_input = tokenizer.encode(prompt, add_special_tokens=False, return_tensors="pt").to(model.device)
52 attention_mask = torch.ones_like(tokenized_input)
53
54 with torch.no_grad():
55 outputs = model.generate(
56 tokenized_input,
57 attention_mask=attention_mask,
58 max_new_tokens=100,
59 do_sample=False,
60 repetition_penalty=1.2,
61 pad_token_id=tokenizer.eos_token_id
62 )[0]
63 output = tokenizer.decode(outputs[tokenized_input.size(1):], skip_special_tokens=True)
64 results.append({"task_id": data["task_id"], "input": input, "output": output})
65
66jsonl_id = re.sub(".*/", "", model_id)
67with open(f"./{jsonl_id}-outputs.jsonl", 'w', encoding='utf-8') as f:
68 for result in results:
69 json.dump(result, f, ensure_ascii=False)
70 f.write('\n')