fajoie/llmjp3_lora

Uploaded model

• Developed by: fajoie
• License: apache-2.0
• Finetuned from model : llm-jp/llm-jp-3-13b

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

使用方法

!pip install -U bitsandbytes
!pip install -U transformers
!pip install -U accelerate
!pip install -U datasets
!pip install -U peft
!pip install ipywidgets --upgrade
from transformers import (
    AutoModelForCausalLM,
    AutoTokenizer,
    BitsAndBytesConfig,
)
from peft import PeftModel
import torch
from tqdm import tqdm
import json

# Hugging Faceで取得したTokenをこちらに貼る。
HF_TOKEN = "xxx"
# ベースとなるモデルと学習したLoRAのアダプタ。
model_id = "llm-jp/llm-jp-3-13b"
adapter_id = "fajoie/llmjp3_lora"
# QLoRA config
bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.bfloat16,
)
# Load model
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    quantization_config=bnb_config,
    device_map="auto",
    token = HF_TOKEN
)

# Load tokenizer
tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True, token = HF_TOKEN)

# 元のモデルにLoRAのアダプタを統合。
model = PeftModel.from_pretrained(model, adapter_id, token = HF_TOKEN)

# データセットの読み込み。
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 = ""

# llmjp
results = []
for data in tqdm(datasets):

  input = data["input"]

  prompt = f"""### 指示
  {input}
  ### 回答
  """

  tokenized_input = tokenizer.encode(prompt, add_special_tokens=False, return_tensors="pt").to(model.device)
  attention_mask = torch.ones_like(tokenized_input)
  with torch.no_grad():
      outputs = model.generate(
          tokenized_input,
          attention_mask=attention_mask,
          max_new_tokens=100,
          do_sample=False,
          repetition_penalty=1.2,
          pad_token_id=tokenizer.eos_token_id
      )[0]
  output = tokenizer.decode(outputs[tokenized_input.size(1):], skip_special_tokens=True)

  results.append({"task_id": data["task_id"], "input": input, "output": output})

# ファイル保存
import re
jsonl_id = re.sub(".*/", "", adapter_id)
with open(f"./{jsonl_id}-outputs.jsonl", 'w', encoding='utf-8') as f:
    for result in results:
        json.dump(result, f, ensure_ascii=False)  # ensure_ascii=False for handling non-ASCII characters
        f.write('\n')

学習手法

!pip uninstall unsloth -y
!pip install --upgrade --no-cache-dir "unsloth[colab-new] @ git+https://github.com/unslothai/unsloth.git"
!pip install --upgrade torch
!pip install --upgrade xformers

# Install Flash Attention 2 for softcapping support
import torch
if torch.cuda.get_device_capability()[0] >= 8:
    !pip install --no-deps packaging ninja einops "flash-attn>=2.6.3"

# Hugging Face Token を指定
HF_TOKEN = "xxx"

# llm-jp/llm-jp-3-13bを4bit量子化のqLoRA設定でロード。
from unsloth import FastLanguageModel
import torch
max_seq_length = 512
dtype = None 
load_in_4bit = True 

model_id = "llm-jp/llm-jp-3-13b"
new_model_id = "llm-jp-3-13b-it" 
# FastLanguageModel 
model, tokenizer = FastLanguageModel.from_pretrained(
    model_name=model_id,
    dtype=dtype,
    load_in_4bit=load_in_4bit,
    trust_remote_code=True,
)

# SFT用のモデルを用意
model = FastLanguageModel.get_peft_model(
    model,
    r = 32,
    target_modules = ["q_proj", "k_proj", "v_proj", "o_proj",
                      "gate_proj", "up_proj", "down_proj",],
    lora_alpha = 32,
    lora_dropout = 0.05,
    bias = "none",
    use_gradient_checkpointing = "unsloth",
    random_state = 3407,
    use_rslora = False,
    loftq_config = None,
    max_seq_length = max_seq_length,
)

from datasets import Dataset, load_dataset, concatenate_datasets

# 使用したいデータセットのパス　すべてのichikaraのデータセットを利用
data_dir = "/content/"
data_files = [
    "ichikara-instruction-003-001-1.json",
    "ichikara-instruction-003-001-2.1.json",
    "ichikara-instruction-003-001-2.2.json",
    "ichikara-instruction-003-001-5.1.json",
    "ichikara-instruction-003-001-5.2.json",
    "ichikara-instruction-003-003-1.json"
    ]

dataset = Dataset.from_dict({"ID": [], "text": [], "output":[]})
for data_file in data_files:
    tmp = load_dataset("json", data_files=f"{data_dir}{data_file}", split="train", streaming=False)
    if len(dataset) == 0:
        dataset = tmp
    else:
        dataset = concatenate_datasets([dataset,tmp])

# 学習時のプロンプトフォーマットの定義
prompt = """### 指示
{}
### 回答
{}"""

EOS_TOKEN = tokenizer.eos_token 
def formatting_prompts_func(examples):
    input = examples["text"] 
    output = examples["output"] 
    text = prompt.format(input, output) + EOS_TOKEN 
    return { "formatted_text" : text, } 
pass

# # 各データにフォーマットを適用
dataset = dataset.map(
    formatting_prompts_func,
    num_proc= 4, 
)

from trl import SFTTrainer
from transformers import TrainingArguments
from unsloth import is_bfloat16_supported

# 学習の設定
trainer = SFTTrainer(
    model = model,
    tokenizer = tokenizer,
    train_dataset=dataset,
    max_seq_length = max_seq_length,
    dataset_text_field="formatted_text",
    packing = False,
    args = TrainingArguments(
        per_device_train_batch_size = 16, #Google Colab Pro+を使ったのでバッチサイズを上げた
        gradient_accumulation_steps = 1, #蓄積は逆になしに
        num_train_epochs = 1, #上げたら過学習してさがったので、最終的に1回にした
        logging_steps = 10,
        warmup_steps = 100,
        save_steps=100,
        save_total_limit=2,
        max_steps=-1,
        learning_rate = 2e-4,
        fp16 = not is_bfloat16_supported(),
        bf16 = is_bfloat16_supported(),
        group_by_length=True,
        seed = 3407,
        output_dir = "outputs",
        report_to = "none",
    ),
)

# 学習実行
trainer_stats = trainer.train()

# LoRAアダプタだけ保存
model.push_to_hub_merged(
    new_model_id+"_lora_4",
    tokenizer=tokenizer,
    save_method="lora",
    token=HF_TOKEN,
    private=True
)