Views
No views yet
1
2!pip install -U bitsandbytes
3!pip install -U transformers
4!pip install -U accelerate
5!pip install -U datasets
6!pip install -U peft
7
8!pip install ipywidgets --upgrade
9
10from transformers import (
11 AutoModelForCausalLM,
12 AutoTokenizer,
13 BitsAndBytesConfig,
14)
15from peft import PeftModel
16import torch
17from tqdm import tqdm
18import json
19import re
20from google.colab import files
21
22HF_TOKEN = "YOUR TOKEN"
23your_path = '/elyza-tasks-100-TV_0.jsonl'
24
25model_id = "llm-jp/llm-jp-3-13b"
26adapter_id = "mss6/f4"
27
28model = PeftModel.from_pretrained(model, adapter_id, token = HF_TOKEN)
29
30datasets = []
31with open(your_path, "r") as f:
32 item = ""
33 for line in f:
34 line = line.strip()
35 item += line
36 if item.endswith("}"):
37 datasets.append(json.loads(item))
38 item = ""
39
40from tqdm import tqdm
41
42results = []
43for i in tqdm(range(100)):
44 data = datasets[i]
45 input = data["input"]
46
47 prompt = f"""### 指示
48 {input}
49 ### 回答
50 """
51
52 tokenized_input = tokenizer.encode(prompt, add_special_tokens=False, return_tensors="pt").to(model.device)
53 attention_mask = torch.ones_like(tokenized_input)
54
55 with torch.no_grad():
56 outputs = model.generate(
57 tokenized_input,
58 attention_mask=attention_mask,
59 max_new_tokens=1000,
60 do_sample=False,
61 repetition_penalty=1.2,
62 pad_token_id=tokenizer.eos_token_id
63 )[0]
64 output = tokenizer.decode(outputs[tokenized_input.size(1):], skip_special_tokens=True)
65
66 results.append({"task_id": data["task_id"], "input": input, "output": output})
67
68jsonl_id = re.sub(".*/", "", 'ans')
69with open(f"/content/drive/MyDrive/{jsonl_id}-outputs.jsonl", 'w', encoding='utf-8') as f:
70 for result in results:
71 json.dump(result, f, ensure_ascii=False)
72 f.write('\n')
73| Language | Dataset | url |
|---|---|---|
| Japanese | ichikara-instruction | ichikara-instruction |
| Synthesized data from Elyza-tasks-100 | Elyza-tasks-100 |