Views
No views yet
1!pip install -U bitsandbytes
2!pip install -U transformers
3!pip install -U accelerate
4!pip install -U datasets!pip install ipywidgets --upgrade1from transformers import (
2 AutoModelForCausalLM,
3 AutoTokenizer,
4 BitsAndBytesConfig,
5)
6import torch
7from tqdm import tqdm
8import jsonHF_TOKEN = "xxx"1# QLoRA config
2bnb_config = BitsAndBytesConfig(
3 load_in_4bit=True,
4 bnb_4bit_quant_type="nf4",
5 bnb_4bit_compute_dtype=torch.bfloat16,
6 bnb_4bit_use_double_quant=False,
7)1# Load model
2model = AutoModelForCausalLM.from_pretrained(
3 model_name,
4 quantization_config=bnb_config,
5 device_map="auto",
6 token = HF_TOKEN
7)
8
9# Load tokenizer
10tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True, token = HF_TOKEN)1# データセットの読み込み。
2datasets = []
3with open("/content/elyza-tasks-100-TV_0.jsonl", "r") as f:
4 item = ""
5 for line in f:
6 line = line.strip()
7 item += line
8 if item.endswith("}"):
9 datasets.append(json.loads(item))
10 item = ""1# llmjp
2results = []
3for data in tqdm(datasets):
4
5 input = data["input"]
6
7 prompt = f"""### 指示
8 {input}
9 ### 回答：
10 """
11
12 tokenized_input = tokenizer.encode(prompt, add_special_tokens=False, return_tensors="pt").to(model.device)
13 with torch.no_grad():
14 outputs = model.generate(
15 tokenized_input,
16 max_new_tokens=100,
17 do_sample=False,
18 repetition_penalty=1.2
19 )[0]
20 output = tokenizer.decode(outputs[tokenized_input.size(1):], skip_special_tokens=True)
21
22 results.append({"task_id": data["task_id"], "input": input, "output": output})
231import re
2model_name = re.sub(".*/", "", model_name)
3with open(f"./{model_name}-outputs.jsonl", 'w', encoding='utf-8') as f:
4 for result in results:
5 json.dump(result, f, ensure_ascii=False) # ensure_ascii=False for handling non-ASCII characters
6 f.write('\n')