alicenag/llm2024_final_competition

# ##################################################################################
# インストール
# ##################################################################################

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

!pip install --upgrade torch torchvision torchaudio
!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"

# ##################################################################################
# パラメータ設定
# ##################################################################################

device = "cuda" if torch.cuda.is_available() else "cpu"

# データファイルをカレントディレクトリ`.`に配置してから実行してください
input_dir  = "./ichikara-instruction-003-001-1.json"
eval_dir   = "./elyza-tasks-100-TV_0.jsonl"
result_dir = "./"

# ##################################################################################
# 学習パラメータを設定
# ##################################################################################

model_id = "llm-jp/llm-jp-3-13b"
new_model_id = "llm-jp-3-13b-it"      # Fine-Tuningしたモデルにつけたい名前
dtype = None                          # Noneにしておけば自動で設定                                                   +
load_in_4bit = True                   # 今回は13Bモデルを扱うためTrue
max_seq_length = 512                  # unslothではRoPEをサポートしているのでコンテキスト長は自由に設定可能

# パラメータをPack
config={
    "model_id": model_id,
    "learning_rate": 2e-5,
    "per_device_train_batch_size": 4,
    "gradient_accumulation_steps": 4,
    "num_train_epochs":3,
    "warmup_steps": 10,
    "max_steps": -1,

    "model_max_seq_length": max_seq_length,
    "model_dtype": dtype,
    "model_load_in_4bit": load_in_4bit,

    "lora_r": 32,
    "lora_alpha": 32,
    "lora_dropout": 0.05,
    "lora_bias": "none",
    "lora_use_rslora": False,
    "lora_loftq_config": None,

    "seed": 3407,
    "max_seq_length": max_seq_length
}

# ##################################################################################
# llm-jp/llm-jp-3-13bを4bit量子化のqLoRA設定でロード
# ##################################################################################

from unsloth import FastLanguageModel
model, tokenizer = FastLanguageModel.from_pretrained( 
    model_name   = cfg.model_id,
    dtype        = cfg.model_dtype,
    load_in_4bit = cfg.model_load_in_4bit,
    trust_remote_code= True,
)

model = FastLanguageModel.get_peft_model(   
    model,
    r              = cfg.lora_r,
    target_modules = ["q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj",],
    lora_alpha     = cfg.lora_alpha,
    lora_dropout   = cfg.lora_dropout,
    bias           = cfg.lora_bias,
    random_state   = cfg.seed,
    use_rslora     = cfg.lora_use_rslora,
    loftq_config   = cfg.lora_loftq_config,
    max_seq_length = cfg.max_seq_length,
    use_gradient_checkpointing = "unsloth",
)

# ##################################################################################
# データセットをロード
# ##################################################################################

from datasets import load_dataset
train_dataset = load_dataset("json", data_files= input_dir, split="train[:80%]" )
test_dataset = load_dataset("json", data_files= input_dir, split="train[80%:]")


#  プロンプトに合わせた形式に合わせる
EOS_TOKEN = tokenizer.eos_token                  # トークナイザーのEOSトークン（文末トークン）
prompt = f"""### 指示\n{input}\n### 回答\n"""    # 学習時のプロンプトフォーマットの定義

def formatting_prompts_func(examples):
    input = examples["text"]                          # 入力データ
    output = examples["output"]                       # 出力データ
    text = prompt.format(input, output) + EOS_TOKEN   # プロンプトの作成
    return { "formatted_text" : text, }               # 新しいフィールド "formatted_text" を返す

# Promptフォーマットを適用
train_dataset = train_dataset.map( formatting_prompts_func, num_proc= 4 )
test_dataset = test_dataset.map( formatting_prompts_func, num_proc= 4 )

# ##################################################################################
# 学習する
# ##################################################################################

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

# EarlyStoppingコールバック
early_stopping_callback = EarlyStoppingCallback(
    early_stopping_patience  = 3,   
)

# Trainerを準備する
trainer = SFTTrainer(
    model              = model,
    tokenizer          = tokenizer,
    train_dataset      = train_dataset,
    eval_dataset       = test_dataset,
    max_seq_length     = cfg.max_seq_length,
    dataset_text_field = "formatted_text",
    packing            = False,

    callbacks=[early_stopping_callback],

    args = TrainingArguments(
        per_device_train_batch_size = cfg.per_device_train_batch_size,
        gradient_accumulation_steps = cfg.gradient_accumulation_steps,
        num_train_epochs = cfg.num_train_epochs,
        warmup_steps     = cfg.warmup_steps,
        max_steps        = cfg.max_steps,
        learning_rate    = cfg.learning_rate,
        seed             = cfg.seed,
        output_dir       = "outputs",
        report_to        = "no",
        fp16 = not is_bfloat16_supported(),
        bf16 = is_bfloat16_supported(),
        group_by_length     = True,
        logging_steps       = 10,

        evaluation_strategy = "steps",
        eval_steps          = 20,
        save_strategy       = "steps",
        save_steps          = 60,
        save_total_limit    = 3,
        load_best_model_at_end = True,
        metric_for_best_model  = "eval_loss",
        greater_is_better      = False,
    ),
)

# ##################################################################################
# 学習実行
# ##################################################################################
trainer_stats = trainer.train()

# ##################################################################################
# チューニングモデルで、質問タスクを実行
# ##################################################################################

import json
eval_datasets = []
elyza_tasks_dir = eval_dir
with open(elyza_tasks_dir, "r") as f:
    item = ""
    for line in f:
      line = line.strip()
      item += line
      if item.endswith("}"):
        eval_datasets.append(json.loads(item))
        item = ""


# チューニングモデルで、質問タスクを実行
from tqdm import tqdm
FastLanguageModel.for_inference(model)
model.eval()

results = []
for dt in tqdm(eval_datasets):
  input = dt["input"]
  prompt = f"""### 指示\n{input}\n### 回答\n"""
  inputs = tokenizer([prompt], return_tensors = "pt")

  outputs = model.generate(**inputs, max_new_tokens = 512, use_cache = True, do_sample=False, repetition_penalty=1.2)
  prediction = tokenizer.decode(outputs[0], skip_special_tokens=True).split('\n### 回答')[-1]

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



# ##################################################################################
# jsonlで保存
# ##################################################################################

jsonl_result_dir = result_dir + f"/{new_model_id}_output.jsonl",
with open( jsonl_result_dir, 'w', encoding='utf-8') as f:
    for result in results:
        json.dump(result, f, ensure_ascii=False)
        f.write('\n')

# アダプタだけ保存
HF_TOKEN = "ここにhuggingfaceのアクセストークンを設定してください"
model.push_to_hub_merged(
    new_model_id+"_lora",
    tokenizer=tokenizer,
    save_method="lora",
    token=HF_TOKEN,
    private=False
)