VITA-Group / HardFocusTraining

[Preprint] "Take the Bull by the Horns: Hard Sample-Reweighted Continual Training Improves LLM Generalization" by Xuxi Chen, Zhendong Wang, Daouda Sow, Junjie Yang, Tianlong Chen, Yingbin Liang, Mingyuan Zhou, Zhangyang Wang

HardFocusTraining

Codes for [Preprint] "Take the Bull by the Horns: Hard Sample-Reweighted Continual Training Improves LLM Generalization" by Xuxi Chen, Zhendong Wang, Daouda Sow, Junjie Yang, Tianlong Chen, Yingbin Liang, Mingyuan Zhou, Zhangyang Wang

Requirements

We provide the environment requirement in requirements.txt. Please run the following commands to install necessary dependencies:

pip install -r requirements.txt
git submodule update --init
cd lm-evaluation-harness
pip install -e .
cd ../

Experiments

Training

accelerate launch batch_training.py --data_path </path/to/dataset> --model_path </path/to/models> --batch_size 8 --frac 1 --save_dir opt_125m_baseline_sgd_bs8_n100_dro_2e-6 --num_batch 100 --model facebook/opt-125m --save_freq 100 --mode dro --lr 2e-6

accelerate launch batch_training.py --data_path </path/to/dataset> --model_path </path/to/models> --batch_size 8 --frac 1 --save_dir opt_350m_baseline_sgd_bs8_n100_dro_2e-6 --num_batch 100 --model facebook/opt-350m --save_freq 100 --mode dro --lr 2e-6

accelerate launch batch_training.py --data_path </path/to/dataset> --model_path </path/to/dataset> --batch_size 4 --frac 1 --save_dir sheared_1.3b_baseline_sgd_bs4_n100_dro_2e-6 --mode dro --num_batch 100 --model princeton-nlp/Sheared-LLaMA-1.3B --save_freq 10 --lr 2e-6 

Evaluation

CUDA_VISIBLE_DEVICES=2 lm_eval --model hf \
    --model_args pretrained=facebook/opt-125m \
    --tasks piqa,hellaswag,winogrande,arc_challenge,boolq,mmlu \
    --num_fewshot 0 \
    --batch_size 32 \
    --checkpoint_path <path>/model_99.pt

CUDA_VISIBLE_DEVICES=2 lm_eval --model hf \
    --model_args pretrained=facebook/opt-125m \
    --tasks mmlu \
    --num_fewshot 5 \
    --batch_size 32 \
    --checkpoint_path <path>/model_99.pt

Todos

• Upload codes related to instruction-tuning experiment.