Views
No views yet
| Language | Dataset | description |
|---|---|---|
| Japanese | ichikara-instruction-003-001-1.json | A manually constructed instruction dataset |
1!pip install -U bitsandbytes
2!pip install -U transformers
3!pip install -U accelerate
4!pip install -U datasets1from transformers import (
2 AutoModelForCausalLM,
3 AutoTokenizer,
4 BitsAndBytesConfig,
5)
6import torch
7from tqdm import tqdm
8import jsonHF_TOKEN = "YOUR-HF-TOKEN"model_name = "rambling1228/llm-jp-3-13b-finetune"1bnb_config = BitsAndBytesConfig(
2 load_in_4bit=True,
3 bnb_4bit_quant_type="nf4",
4 bnb_4bit_compute_dtype=torch.bfloat16,
5 bnb_4bit_use_double_quant=False,
6)1model = AutoModelForCausalLM.from_pretrained(
2 model_name,
3 quantization_config=bnb_config,
4 device_map="auto",
5 token = HF_TOKEN
6)
7
8tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True, token = HF_TOKEN)1datasets = []
2with open("./YOUR-DATA.jsonl", "r") as f:
3 item = ""
4 for line in f:
5 line = line.strip()
6 item += line
7 if item.endswith("}"):
8 datasets.append(json.loads(item))
9 item = ""1results = []
2for data in tqdm(datasets):
3
4 input = data["input"]
5
6 prompt = f"""### 指示
7 {input}
8 ### 回答：
9 """
10
11 tokenized_input = tokenizer.encode(prompt, add_special_tokens=False, return_tensors="pt").to(model.device)
12 with torch.no_grad():
13 outputs = model.generate(
14 tokenized_input,
15 max_new_tokens=512,
16 do_sample=False,
17 repetition_penalty=1.2
18 )[0]
19 output = tokenizer.decode(outputs[tokenized_input.size(1):], skip_special_tokens=True)
20
21 results.append({"task_id": data["task_id"], "input": input, "output": output})1import 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')