TAGCOS

This is the official implementation of TAGCOS: Task-agnostic Gradient Clustered Coreset Selection for Instruction Tuning Data

1 environment setup

cd release
conda create --name <env> --file requirement.txt

2 data preprocess

bash script/prepare_train_data.sh

3 warmup training

TODO:automatic train 4 epochs while mannual stop at 30 steps needed.

bash script/train/warmup_lora_train.sh ./data meta-llama/Llama-2-7b-hf 0.05 3 llama2-7b-p0.05-lora-seed3

4 gradient data store

CUDA_VISIBLE_DEVICES=0 bash script/prepare_grad_data.sh

5 select data

python3 -m utils.mean_grad --gradient_path grads_full_data/llama2-7b-p0.05-lora-seed3/{}-ckpt{}-adam/dim8192/all_orig.pt --train_file_names hard_coded lima code_alpaca baize cot dolly flan_v2 gpt4_alpaca oasst1 open_orca science self_instruct sharegpt stanford_alpaca super_ni unnatural_instructions wizardlm --ckpts 10 20 30 --checkpoint_weights 2.1569e-06 6.0784e-06 1e-05 --output_path selected_full_data_102030

python3 -m utils.sk_LESS

python3 -m utils.omp out/cluster_LESS_102030_100.pkl out/centers_LESS_102030_100.pkl out/top_omp_LESS102030_100.pkl

python3 -m utils.get_segment_data

We provide our selected data here.

6 train with seleted data

bash ./script/train/lora_train.sh out/top_omp_LESS102030_100_data.jsonl meta-llama/Llama-2-7b-hf llama2-7b-0.05-omp-LESS102030-100-lora

7 eval

Please follow the instructions in the evaluation folder to evaluate the performance of the model trained on the selected data by using open-instruct.

bash script/eval/bbh_test.sh
bash script/eval/mmlu_test.sh
bash script/eval/tydiqa_test.sh

2003pro/TAGCOS