|
|
--- |
|
|
base_model: llm-jp/llm-jp-3-13b |
|
|
tags: |
|
|
- text-generation-inference |
|
|
- transformers |
|
|
- unsloth |
|
|
- llama |
|
|
- trl |
|
|
license: cc-by-nc-4.0 |
|
|
language: |
|
|
- ja |
|
|
datasets: |
|
|
- DeL-TaiseiOzaki/Tengentoppa-sft-v1.0 |
|
|
--- |
|
|
|
|
|
# Uploaded model |
|
|
|
|
|
- **Developed by:** daichira |
|
|
- **License:** cc-by-nc-4.0 |
|
|
- **Finetuned from model :** llm-jp/llm-jp-3-13b |
|
|
|
|
|
This llama model was trained 2x faster with [Unsloth](https://github.com/unslothai/unsloth) and Huggingface's TRL library. |
|
|
|
|
|
[<img src="https://raw.githubusercontent.com/unslothai/unsloth/main/images/unsloth%20made%20with%20love.png" width="200"/>](https://github.com/unslothai/unsloth) |
|
|
|
|
|
### README.md |
|
|
|
|
|
# llm-jp-3-13b-itnew9 |
|
|
|
|
|
## 概要 |
|
|
|
|
|
このプロジェクトは、Hugging Face上で提供される言語モデル`llm-jp/llm-jp-3-13b`を基盤とし、さらなる指示応答タスク向けに微調整（SFT: Supervised Fine-Tuning）を施した'daichira/llm-jp-3-13b-finetune2'から、さらに後述のコードによりSFTを施したモデル`llm-jp-3-13b-itnew9`を公開するものです。このREADMEは、モデルのセットアップ、トレーニング、推論の再現性を確保するための手順を示します。 |
|
|
|
|
|
--- |
|
|
|
|
|
## 前提条件 |
|
|
|
|
|
このプロジェクトを実行するには、以下の環境とツールが必要です： |
|
|
|
|
|
- Python 3.8以上 |
|
|
- Google Colabまたはローカル環境 (GPU推奨) |
|
|
- Hugging Faceアクセストークン (HF\_TOKEN) |
|
|
|
|
|
--- |
|
|
|
|
|
## セットアップ手順 |
|
|
|
|
|
### 1. ライブラリのインストール |
|
|
|
|
|
Google Colabの場合、以下のコマンドを使用して必要なライブラリをインストールします： |
|
|
|
|
|
```bash |
|
|
!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 |
|
|
!pip install ipywidgets --upgrade |
|
|
``` |
|
|
|
|
|
Flash Attention 2をサポートするために、以下をインストールします： |
|
|
|
|
|
```python |
|
|
import torch |
|
|
if torch.cuda.get_device_capability()[0] >= 8: |
|
|
!pip install --no-deps packaging ninja einops "flash-attn>=2.6.3" |
|
|
``` |
|
|
|
|
|
### 2. モデルとトークナイザーのロード |
|
|
|
|
|
以下のコードを使用して、Hugging Faceからベースモデルをロードし、LoRAの設定を適用します： |
|
|
|
|
|
```python |
|
|
from transformers import AutoModelForCausalLM, AutoTokenizer |
|
|
from unsloth import FastLanguageModel |
|
|
|
|
|
max_seq_length = 1024 |
|
|
dtype = None |
|
|
load_in_4bit = True |
|
|
model_id = "daichira/llm-jp-3-13b-finetune2" |
|
|
new_model_id = "llm-jp-3-13b-itnew9" |
|
|
|
|
|
model, tokenizer = FastLanguageModel.from_pretrained( |
|
|
model_name=model_id, |
|
|
dtype=dtype, |
|
|
load_in_4bit=load_in_4bit, |
|
|
trust_remote_code=True, |
|
|
) |
|
|
|
|
|
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, |
|
|
) |
|
|
|
|
|
``` |
|
|
|
|
|
--- |
|
|
|
|
|
## データセットの準備 |
|
|
|
|
|
### データの分割と保存 |
|
|
|
|
|
以下のコードでデータセットをHugging Faceからロードし、分割して保存します： |
|
|
|
|
|
```python |
|
|
HF_TOKEN = "Your_token" # Write権限 |
|
|
|
|
|
!pip install datasets |
|
|
import os |
|
|
from datasets import load_dataset |
|
|
|
|
|
dataset = load_dataset("DeL-TaiseiOzaki/Tengentoppa-sft-v1.0", split="train") |
|
|
chunk_size = 30000 |
|
|
output_dir = "/content/tengentoppa_chunks" |
|
|
os.makedirs(output_dir, exist_ok=True) |
|
|
|
|
|
total_rows = len(dataset) |
|
|
num_chunks = (total_rows + chunk_size - 1) // chunk_size |
|
|
|
|
|
for i in range(num_chunks): |
|
|
start_idx = i * chunk_size |
|
|
end_idx = min(start_idx + chunk_size, total_rows) |
|
|
chunk = dataset.select(range(start_idx, end_idx)) |
|
|
chunk_file = f"{output_dir}/tengentoppa_chunk_{i+1}.json" |
|
|
chunk.to_json(chunk_file) |
|
|
print(f"Saved chunk {i+1}/{num_chunks} to {chunk_file}") |
|
|
|
|
|
print("All chunks have been saved!") |
|
|
``` |
|
|
|
|
|
### JSON形式でのデータセット読み込み |
|
|
|
|
|
以下のコードでJSONデータセットをロードします： |
|
|
|
|
|
```python |
|
|
json_path = "/content/tengentoppa_chunks/tengentoppa_chunk_3.json" |
|
|
dataset = load_dataset("json", data_files=json_path) |
|
|
print(dataset) |
|
|
``` |
|
|
|
|
|
--- |
|
|
|
|
|
## トレーニングの設定 |
|
|
|
|
|
以下の手順でトレーニングを設定します： |
|
|
|
|
|
```python |
|
|
prompt = """### 指示 |
|
|
{} |
|
|
### 回答 |
|
|
{}""" |
|
|
EOS_TOKEN = tokenizer.eos_token |
|
|
|
|
|
def formatting_prompts_func(examples): |
|
|
input_text = examples["instruction"] |
|
|
output_text = examples["output"] |
|
|
return {"formatted_text": prompt.format(input_text, output_text) + EOS_TOKEN} |
|
|
|
|
|
dataset = dataset.map(formatting_prompts_func, num_proc=4) |
|
|
|
|
|
from trl import SFTTrainer |
|
|
from transformers import TrainingArguments |
|
|
from unsloth import is_bfloat16_supported |
|
|
|
|
|
trainer = SFTTrainer( |
|
|
model=model, |
|
|
tokenizer=tokenizer, |
|
|
train_dataset=dataset["train"], |
|
|
max_seq_length=max_seq_length, |
|
|
dataset_text_field="formatted_text", |
|
|
args=TrainingArguments( |
|
|
per_device_train_batch_size=6, |
|
|
gradient_accumulation_steps=4, |
|
|
num_train_epochs=1, |
|
|
logging_steps=50, |
|
|
warmup_steps=500, |
|
|
save_steps=500, |
|
|
save_total_limit=2, |
|
|
learning_rate=3e-4, |
|
|
fp16=not is_bfloat16_supported(), |
|
|
bf16=is_bfloat16_supported(), |
|
|
group_by_length=True, |
|
|
seed=3407, |
|
|
output_dir="outputs", |
|
|
), |
|
|
) |
|
|
|
|
|
# 学習実行 |
|
|
torch.cuda.empty_cache() |
|
|
trainer.train() |
|
|
``` |
|
|
|
|
|
--- |
|
|
|
|
|
## 推論 |
|
|
|
|
|
以下のコードでトレーニング済みモデルを使用して推論を行います： |
|
|
|
|
|
```python |
|
|
import json |
|
|
from tqdm import tqdm |
|
|
|
|
|
with open("/content/elyza-tasks-100-TV_0.jsonl", "r") as f: |
|
|
datasets = [json.loads(line) for line in f if line.strip().endswith("}")] |
|
|
|
|
|
FastLanguageModel.for_inference(model) |
|
|
|
|
|
results = [] |
|
|
for dt in tqdm(datasets): |
|
|
input_text = dt["input"] |
|
|
prompt = f"### 指示\n{input_text}\n### 回答\n" |
|
|
|
|
|
inputs = tokenizer([prompt], return_tensors="pt").to(model.device) |
|
|
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_text, "output": prediction}) |
|
|
|
|
|
with open(f"{new_model_id}_output.jsonl", 'w', encoding='utf-8') as f: |
|
|
for result in results: |
|
|
json.dump(result, f, ensure_ascii=False) |
|
|
f.write('\n') |
|
|
``` |
|
|
|
|
|
--- |
|
|
|
|
|
## 注意事項 |
|
|
|
|
|
- 本モデルは日本語専用で設計されています。 |
|
|
- 再現性を確保するため、ランダムシードを固定しています (`seed=3407`)。 |
|
|
- モデルのパラメータ量が大きいため、十分なGPUメモリを確保してください (推奨: 16GB以上)。 |
|
|
|
|
|
--- |
|
|
|
|
|
## 実行コード全体 |
|
|
```python |
|
|
# 必要なライブラリのインストール |
|
|
!pip uninstall unsloth -y && pip install --upgrade --no-cache-dir "unsloth[colab-new] @ git+https://github.com/unslothai/unsloth.git" |
|
|
!pip install --upgrade torch |
|
|
!pip install --upgrade xformers |
|
|
!pip install ipywidgets --upgrade |
|
|
|
|
|
# Flash Attention 2のインストール |
|
|
import torch |
|
|
if torch.cuda.get_device_capability()[0] >= 8: |
|
|
!pip install --no-deps packaging ninja einops "flash-attn>=2.6.3" |
|
|
|
|
|
# モデルとトークナイザーのロード |
|
|
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig |
|
|
from unsloth import FastLanguageModel |
|
|
|
|
|
# モデル設定 |
|
|
max_seq_length = 1024 |
|
|
dtype = None |
|
|
load_in_4bit = True |
|
|
model_id = "daichira/llm-jp-3-13b-finetune2" |
|
|
new_model_id = "llm-jp-3-13b-itnew9" |
|
|
|
|
|
model, tokenizer = FastLanguageModel.from_pretrained( |
|
|
model_name=model_id, |
|
|
dtype=dtype, |
|
|
load_in_4bit=load_in_4bit, |
|
|
trust_remote_code=True, |
|
|
) |
|
|
|
|
|
# SFT用のモデル設定 |
|
|
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, |
|
|
) |
|
|
|
|
|
# Hugging Faceのトークン設定 |
|
|
HF_TOKEN = "your_token" |
|
|
|
|
|
# データセットの準備 |
|
|
!pip install datasets |
|
|
import os |
|
|
from datasets import load_dataset |
|
|
|
|
|
dataset = load_dataset("DeL-TaiseiOzaki/Tengentoppa-sft-v1.0", split="train") |
|
|
chunk_size = 30000 |
|
|
output_dir = "/content/tengentoppa_chunks" |
|
|
os.makedirs(output_dir, exist_ok=True) |
|
|
|
|
|
total_rows = len(dataset) |
|
|
num_chunks = (total_rows + chunk_size - 1) // chunk_size |
|
|
|
|
|
for i in range(num_chunks): |
|
|
start_idx = i * chunk_size |
|
|
end_idx = min(start_idx + chunk_size, total_rows) |
|
|
chunk = dataset.select(range(start_idx, end_idx)) |
|
|
chunk_file = f"{output_dir}/tengentoppa_chunk_{i+1}.json" |
|
|
chunk.to_json(chunk_file) |
|
|
print(f"Saved chunk {i+1}/{num_chunks} to {chunk_file}") |
|
|
|
|
|
print("All chunks have been saved!") |
|
|
|
|
|
# JSON形式のデータセットをロード |
|
|
json_path = "/content/tengentoppa_chunks/tengentoppa_chunk_3.json" |
|
|
dataset = load_dataset("json", data_files=json_path) |
|
|
print(dataset) |
|
|
|
|
|
# プロンプトフォーマットの適用 |
|
|
prompt = """### 指示 |
|
|
{} |
|
|
### 回答 |
|
|
{}""" |
|
|
EOS_TOKEN = tokenizer.eos_token |
|
|
|
|
|
def formatting_prompts_func(examples): |
|
|
input_text = examples["instruction"] |
|
|
output_text = examples["output"] |
|
|
return {"formatted_text": prompt.format(input_text, output_text) + EOS_TOKEN} |
|
|
|
|
|
dataset = dataset.map(formatting_prompts_func, num_proc=4) |
|
|
|
|
|
# トレーニングの設定 |
|
|
from trl import SFTTrainer |
|
|
from transformers import TrainingArguments |
|
|
from unsloth import is_bfloat16_supported |
|
|
|
|
|
trainer = SFTTrainer( |
|
|
model=model, |
|
|
tokenizer=tokenizer, |
|
|
train_dataset=dataset["train"], |
|
|
max_seq_length=max_seq_length, |
|
|
dataset_text_field="formatted_text", |
|
|
args=TrainingArguments( |
|
|
per_device_train_batch_size=6, |
|
|
gradient_accumulation_steps=4, |
|
|
num_train_epochs=1, |
|
|
logging_steps=50, |
|
|
warmup_steps=500, |
|
|
save_steps=500, |
|
|
save_total_limit=2, |
|
|
learning_rate=3e-4, |
|
|
fp16=not is_bfloat16_supported(), |
|
|
bf16=is_bfloat16_supported(), |
|
|
group_by_length=True, |
|
|
seed=3407, |
|
|
output_dir="outputs", |
|
|
), |
|
|
) |
|
|
|
|
|
# 学習実行 |
|
|
torch.cuda.empty_cache() |
|
|
trainer.train() |
|
|
|
|
|
# 推論の準備 |
|
|
import json |
|
|
from tqdm import tqdm |
|
|
|
|
|
with open("/content/elyza-tasks-100-TV_0.jsonl", "r") as f: |
|
|
datasets = [json.loads(line) for line in f if line.strip().endswith("}")] |
|
|
|
|
|
FastLanguageModel.for_inference(model) |
|
|
|
|
|
results = [] |
|
|
for dt in tqdm(datasets): |
|
|
input_text = dt["input"] |
|
|
prompt = f"""### 指示\n{input_text}\n### 回答\n""" |
|
|
|
|
|
inputs = tokenizer([prompt], return_tensors="pt").to(model.device) |
|
|
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_text, "output": prediction}) |
|
|
|
|
|
# 推論結果の保存 |
|
|
with open(f"{new_model_id}_output.jsonl", 'w', encoding='utf-8') as f: |
|
|
for result in results: |
|
|
json.dump(result, f, ensure_ascii=False) |
|
|
f.write('\n') |
|
|
``` |
|
|
|
|
|
|
|
|
``` |
