williamjeong2 / nox

Efficient fine-tuning for ko-llm models

nox

Efficient fine-tuning for ko-llm models.

News or Update

2024.03.15

• upstage/open-ko-llm-leaderboard 1위 (2024/03/15) https://huggingface.co/spaces/upstage/open-ko-llm-leaderboard

Hardware and Software

• nvidia driver : 535.86.10
• CUDA Version: 12.2

GPU Memory

Method	gpu memory
SFT LoRA 16bits	23GB
SFT QLoRA 8bits	20GB
SFT QLoRA 4bits	9GB
DPO LoRA 16bits	25GB
DPO QLoRA 8bits	21GB
DPO QLoRA 4bits	10GB

Released Model Checkpoints

davidkim205/nox-solar-10.7b-v4

Model	Average	Ko-ARC	Ko-HellaSwag	Ko-MMLU	Ko-TruthfulQA	Ko-CommonGen V2
davidkim205/nox-solar-10.7b-v4	67.77	73.55	72.07	57.93	79.32	55.96

davidkim205/nox-solar-10.7b-v2

Model	Average	Ko-ARC	Ko-HellaSwag	Ko-MMLU	Ko-TruthfulQA	Ko-CommonGen V2
davidkim205/nox-solar-10.7b-v2	65.38	73.46	67.32	58.7	71.94	55.49

Datasets

davidkim205/kollm-converations

sft학습시 사용되는 데이터셋으로 conversations 형식으로 되어 있습니다.

• 총 1.12M 대화로 구성

davidkim205/kollm-comparision

dpo학습시 시용되는 데이터셋으로 comparsision 형식 되어 있습니다.

• 총 116K 질문과 답변(chosen,rejected)으로 구성

installation

conda create -n nox python=3.10
conda activate nox
pip install -r requirements.txt

Train

SFT : Supervised Fine-Tuning

LoRA

python train_sft.py \
    --model_name_or_path davidkim205/nox-solar-10.7b-v2 \
    --dataset davidkim205/kollm-converations \
    --output_dir outputs/sft \
    --stage sft\
    --do_train\
    --finetuning_type lora\
    --overwrite_cache\
    --per_device_train_batch_size 16\
    --gradient_accumulation_steps 1\
    --lr_scheduler_type cosine\
    --logging_steps 10\
    --save_steps 100\
    --learning_rate 1e-7\
    --num_train_epochs 3\
    --plot_loss\
    --overwrite_output_dir\
    --bf16\
    --use_fast_tokenizer
  

QLoRA

python train_sft.py \
    --stage sft \
    --do_train \
    --quantization_bit 4 \
    --model_name_or_path davidkim205/nox-solar-10.7b-v2 \
    --dataset davidkim205/kollm-converations \
    --template solar \
    --finetuning_type lora \
    --lora_target all \
    --output_dir outputs/-comp-35k-test-dpo-lora-lr1e-7-e3-b1 \
    --overwrite_cache \
    --per_device_train_batch_size 1 \
    --gradient_accumulation_steps 1 \
    --lr_scheduler_type cosine \
    --logging_steps 10 \
    --save_steps 1000 \
    --learning_rate 1e-7 \
    --num_train_epochs 3 \
    --plot_loss \
    --overwrite_output_dir \
    --use_fast_tokenizer \
    --bf16

DPO : Direct Preference Optimization Trainer

LoRA

python train_dpo.py \
   --stage dpo \
   --do_train \
   --model_name_or_path davidkim205/nox-solar-10.7b-v2 \
   --dataset davidkim205/kollm-comparision \
   --template solar \
   --finetuning_type lora \
   --lora_target all \
   --output_dir outputs/test \
   --overwrite_cache \
   --per_device_train_batch_size 1 \
   --gradient_accumulation_steps 1 \
   --lr_scheduler_type cosine \
   --logging_steps 10 \
   --save_steps 100 \
   --learning_rate 1e-5 \
   --num_train_epochs 1 \
   --plot_loss \
   --overwrite_output_dir \
   --use_fast_tokenizer \
   --bf16

QLoRA

python train_dpo.py \
   --stage dpo \
   --do_train \
   --quantization_bit 4 \
   --model_name_or_path davidkim205/nox-solar-10.7b-v2 \
   --dataset davidkim205/kollm-comparision \
   --template solar \
   --finetuning_type lora \
   --lora_target all \
   --output_dir outputs/test \
   --overwrite_cache \
   --per_device_train_batch_size 1 \
   --gradient_accumulation_steps 1 \
   --lr_scheduler_type cosine \
   --logging_steps 10 \
   --save_steps 100 \
   --learning_rate 1e-5 \
   --num_train_epochs 1 \
   --plot_loss \
   --overwrite_output_dir \
   --use_fast_tokenizer \
   --bf16

Load adapter

--adapter_name_or_path 을 추가하면 adapter weight를 load할수 있습니다. 아래는 outputs/sft에 저장된 adapter를 불러오는 예입니다.

python train_dpo.py \
   --stage dpo \
   --do_train \
   --adapter_name_or_path outputs/sft/
   --quantization_bit 4 \
   --model_name_or_path davidkim205/nox-solar-10.7b-v2 \
   --dataset davidkim205/kollm-comparision \
   --template solar \
   --finetuning_type lora \
   --lora_target all \
   --output_dir outputs/test \
   --overwrite_cache \
   --per_device_train_batch_size 1 \
   --gradient_accumulation_steps 1 \
   --lr_scheduler_type cosine \
   --logging_steps 10 \
   --save_steps 100 \
   --learning_rate 1e-5 \
   --num_train_epochs 1 \
   --plot_loss \
   --overwrite_output_dir \
   --use_fast_tokenizer \
   --bf16

Evaluation

https://github.com/davidkim205/kollm_evaluation

etc

template

sft와 dpo 학습은 SOLAR 모델에 기준으로 작성되었기때문에 다른 모델을 사용시 template를 변경해야합니다. preprocess.py의 get_template를 참고하세요.

def get_template(user, input='', gpt=''):
    if len(input) >= 1:
        return f"### User:\n{user}\n{input}\n### Assistant:{gpt}\n"
    else:
        return f"### User:\n{user}\n### Assistant:\n"

References

• https://github.com/hiyouga/LLaMA-Factory
• https://github.com/OpenAccess-AI-Collective/axolotl
• https://huggingface.co/blog/dpo-trl