LLM-jp SFT (Supervised Fine-Tuning)

This repository contains the code for supervised fine-tuning of LLM-jp models.

Requirements

• Python: 3.10.12
• torch>=2.0.0 (should meet with cuda version)
• transformers>=4.34.0
• tokenizers>=0.14.0
• accelerate>=0.23.0
• trl>=0.7.2
• peft>=0.5.0

Installation

Install the necessary packages using pip:

pip install -r requirements.txt

To turn on use_flash_attention_2 option:

pip install flash-attn --no-build-isolation

Dataset Preparation

A sample dataset is provided in data/. A training example is structured as follows:

{"text": "以下は、タスクを説明する指示です。要求を適切に満たす応答を書きなさい。\n\n### 指示:\n日本で一番高い山は？\n\n### 応答:\n富士山"}

During training, loss calculation is focused on tokens post the "### 応答:" segment. For the above example, the loss will be based on "富士山".

Training

Here is the command to train a model on the sample dataset.

python train.py \
    --num_train_epochs 1 \
    --per_device_train_batch_size 1 \
    --learning_rate 1e-5 \
    --warmup_ratio 0.1 \
    --lr_scheduler_type cosine \
    --data_files data/example.jsonl \
    --model_name_or_path llm-jp/llm-jp-1.3b-v1.0 \
    --output_dir results/

To Reproduce LLM-jp Models

Datasets

We used the following datasets for fine-tuning.

• Jaster: A collection of automatically transformed data from the existing Japanese NLP datasets.
• Dolly: A Japanese translation of Dolly.
• OpenAssistant: A Japanese translation of OpenAssistant Conversations Dataset.

NOTE: The datasets mentioned above are not public as of now. We're in the process of making them accessible. Stay tuned for updates.

Full Parameter Supervised Fine-tuning

For the 1.3B model (single node; 8 A100 40GB GPUs)

accelerate launch --config_file configs/accelerate_config_zero1.yaml \
    train.py \
    --num_train_epochs 2 \
    --per_device_train_batch_size 8 \
    --gradient_accumulation_steps 8 \
    --learning_rate 1e-5 \
    --warmup_ratio 0.1 \
    --lr_scheduler_type cosine \
    --bf16 \
    --max_seq_length 2048 \
    --logging_steps 1 \
    --data_files jamp.json janli.json jcommonsenseqa.json jemhopqa.json jnli.json jsem.json jsick.json jsquad.json jsts.json niilc.json dolly_deepl.json oasst_deepl.json \
    --model_name_or_path llm-jp/llm-jp-1.3b-v1.0 \
    --output_dir results/llm-jp-1.3b-instruct-full-jaster-dolly-oasst-v1.0

For the 13B model (single node; 8 A100 40GB GPUs)

accelerate launch --config_file configs/accelerate_config_zero3.yaml \
    train.py \
    --num_train_epochs 2 \
    --per_device_train_batch_size 1 \
    --gradient_accumulation_steps 32 \
    --learning_rate 1e-5 \
    --warmup_ratio 0.1 \
    --lr_scheduler_type cosine \
    --bf16 \
    --max_seq_length 2048 \
    --gradient_checkpointing \
    --logging_steps 1 \
    --data_files jamp.json janli.json jcommonsenseqa.json jemhopqa.json jnli.json jsem.json jsick.json jsquad.json jsts.json niilc.json dolly_deepl.json oasst_deepl.json \
    --model_name_or_path llm-jp/llm-jp-13b-v1.0 \
    --output_dir results/llm-jp-13b-instruct-full-jaster-dolly-oasst-v1.0

For the 13B model (8 nodes; 64 A100 40GB GPUs)

Run following lines from all the nodes. ($machine_rank is the sequential number from 0 to 7 assigned to each node, and $main_process_ip is the IP address of the node $machine_rank=0)

accelerate launch --config_file configs/accelerate_config_zero2.8node.yaml \
    --main_process_ip $main_process_ip \
    --main_process_port 29500 \
    --machine_rank $machine_rank \
    train.py \
    --num_train_epochs 2 \
    --per_device_train_batch_size 3 \
    --gradient_accumulation_steps 6 \
    --learning_rate 1e-5 \
    --warmup_ratio 0.1 \
    --lr_scheduler_type cosine \
    --bf16 \
    --max_seq_length 2048 \
    --logging_steps 1 \
    --data_files jamp.json janli.json jcommonsenseqa.json jemhopqa.json jnli.json jsem.json jsick.json jsquad.json jsts.json niilc.json dolly_deepl.json oasst_deepl.json \
    --model_name_or_path llm-jp/llm-jp-13b-v1.0 \
    --output_dir results/llm-jp-13b-instruct-full-jaster-dolly-oasst-v1.0

Fine-tuning with PEFT

For the 1.3B model (single node; single A100 40GB GPU)

CUDA_VISIBLE_DEVICES=0 python train.py \
    --num_train_epochs 5 \
    --per_device_train_batch_size 8 \
    --gradient_accumulation_steps 4 \
    --learning_rate 1e-4 \
    --warmup_ratio 0.1 \
    --lr_scheduler_type cosine \
    --bf16 \
    --max_seq_length 2048 \
    --data_files jamp.json janli.json jcommonsenseqa.json jemhopqa.json jnli.json jsem.json jsick.json jsquad.json jsts.json niilc.json dolly_deepl.json oasst_deepl.json \
    --use_peft \
    --model_name_or_path llm-jp/llm-jp-1.3b-v1.0 \
    --output_dir results/llm-jp-1.3b-instruct-lora-jaster-dolly-oasst-v1.0

For the 13B model (single node; single A100 40GB GPU)

CUDA_VISIBLE_DEVICES=0 python train.py \
    --num_train_epochs 5 \
    --per_device_train_batch_size 1 \
    --gradient_accumulation_steps 32 \
    --learning_rate 1e-4 \
    --warmup_ratio 0.1 \
    --lr_scheduler_type cosine \
    --bf16 \
    --max_seq_length 2048 \
    --gradient_checkpointing \
    --data_files jamp.json janli.json jcommonsenseqa.json jemhopqa.json jnli.json jsem.json jsick.json jsquad.json jsts.json niilc.json dolly_deepl.json oasst_deepl.json \
    --use_peft \
    --model_name_or_path llm-jp/llm-jp-13b-v1.0 \
    --output_dir results/llm-jp-13b-instruct-lora-jaster-dolly-oasst-v1.0

For the 1.3B model (single node; 8 A100 40GB GPUs)

accelerate launch --config_file configs/accelerate_config_zero1.yaml \
    train.py \
    --num_train_epochs 5 \
    --per_device_train_batch_size 8 \
    --gradient_accumulation_steps 8 \
    --learning_rate 1e-4 \
    --warmup_ratio 0.1 \
    --lr_scheduler_type cosine \
    --bf16 \
    --max_seq_length 2048 \
    --data_files jamp.json janli.json jcommonsenseqa.json jemhopqa.json jnli.json jsem.json jsick.json jsquad.json jsts.json niilc.json dolly_deepl.json oasst_deepl.json \
    --use_peft \
    --model_name_or_path llm-jp/llm-jp-1.3b-v1.0 \
    --output_dir results/llm-jp-1.3b-instruct-lora-jaster-dolly-oasst-v1.0

For the 13B model (single node; 8 A100 40GB GPUs)

accelerate launch --config_file configs/accelerate_config_zero1.yaml \
    train.py \
    --num_train_epochs 5 \
    --per_device_train_batch_size 1 \
    --gradient_accumulation_steps 16 \
    --learning_rate 1e-4 \
    --warmup_ratio 0.1 \
    --lr_scheduler_type cosine \
    --bf16 \
    --max_seq_length 2048 \
    --data_files jamp.json janli.json jcommonsenseqa.json jemhopqa.json jnli.json jsem.json jsick.json jsquad.json jsts.json niilc.json dolly_deepl.json oasst_deepl.json \
    --use_peft \
    --model_name_or_path llm-jp/llm-jp-13b-v1.0 \
    --output_dir results/llm-jp-13b-instruct-lora-jaster-dolly-oasst-v1.0

Using flash-attn

The use_flash_attention_2 option in transformers v4.36 only supports for the models based on Llama and Falcon.

For the 7B model (single node; single A100 40GB GPU)

CUDA_VISIBLE_DEVICES=0 python train.py \
    --num_train_epochs 5 \
    --per_device_train_batch_size 4 \
    --gradient_accumulation_steps 16 \
    --learning_rate 1e-4 \
    --warmup_ratio 0.1 \
    --lr_scheduler_type cosine \
    --bf16 \
    --max_seq_length 2048 \
    --gradient_checkpointing \
    --data_files jamp.json janli.json jcommonsenseqa.json jemhopqa.json jnli.json jsem.json jsick.json jsquad.json jsts.json niilc.json dolly_deepl.json oasst_deepl.json \
    --use_flash_attention_2 True \
    --use_peft \
    --model_name_or_path llm-jp/llm-jp-7b \
    --output_dir results/llm-jp-7b-instruct-lora-jaster-dolly-oasst-v1.0

GPTQ Converter

python converter/gptq_converter.py \
    --model_name_or_path llm-jp/llm-jp-13b-v1.0 \
    --dataset ptb \
    --output_dir results/llm-jp-13b-v1.0-gptq