HF_TOKEN = "your token"
import torch
max_new_tokens = 1024

new_model_id = "MikenekoDyn/llm-jp-13b-061215b" #'_lora'は読み込むときに自動で付加される
load_in_4bit = False 
load_in_8bit = not load_in_4bit 
do_sample=True
repetition_penalty=1.05
temperature=0.7
top_p=0.95

## ELYZA-tasks-100-TVの読み込み
import json
datasets = []
with open("./elyza-tasks-100-TV_0.jsonl", "r") as f:
    item = ""
    for line in f:
      line = line.strip()
      item += line
      if item.endswith("}"):
        datasets.append(json.loads(item))
        item = ""

## Config設定
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
# QLoRA config
bnb_config = BitsAndBytesConfig(
    load_in_4bit=load_in_4bit,
    load_in_8bit=load_in_8bit,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.bfloat16,
    bnb_4bit_use_double_quant=False,
)
# Load model
model = AutoModelForCausalLM.from_pretrained(
    new_model_id+"_lora",
    quantization_config=bnb_config,
    device_map="auto",
    token = HF_TOKEN
)
# Load tokenizer
tokenizer = AutoTokenizer.from_pretrained(new_model_id+"_lora", trust_remote_code=True, token = HF_TOKEN)

## 推論実施
from tqdm import tqdm
results = []
ii=0
for data in tqdm(datasets):
  input = data["input"]
  prompt = f"""### User
  {input}
  ### Assistant
  """
  tokenized_input = tokenizer.encode(prompt, add_special_tokens=False, return_tensors="pt").to(model.device)
  with torch.no_grad():
      outputs = model.generate(
          tokenized_input,
          max_new_tokens=max_new_tokens,
          do_sample=do_sample,
          repetition_penalty=repetition_penalty,
          temperature=temperature,
          top_p=top_p
      )[0]
  output = tokenizer.decode(outputs[tokenized_input.size(1):], skip_special_tokens=True)
  if ii<3:
    print(output)
  results.append({"task_id": data["task_id"], "input": input, "output": output})
  ii=ii+1

# jsonlで保存
import re
save_model_name = re.sub(".*/", "", new_model_id)
with open(f"{save_model_name}_output.jsonl", 'w', encoding='utf-8') as f:
    for result in results:
        json.dump(result, f, ensure_ascii=False)
        f.write('\n')

Instruction Tuning

使用データセット

ichikara-instruction dataset
magpie-sft
ELYZA-tasks-100

This llama model was trained 2x faster with Unsloth and Huggingface's TRL library.

MikenekoDyn/llm-jp-13b-061215b_lora

Uploaded model

Sample Use

Instruction Tuning

MikenekoDyn/
llm-jp-13b-061215b_lora