SVFT-main/MetaMath/run_math.sh

export MODEL_PATH='google/gemma-2b'
export MASTER_ADDR="localhost"
export MASTER_PORT="1231"
export GLOO_SOCKET_IFNAME="lo"
export NCCL_SOCKET_IFNAME="lo"


#SVFT_PLAIN
export SAVE_PATH='./Gemma_2B_metamath40k_svft_plain'
CUDA_VISIBLE_DEVICES=0 python3 -m torch.distributed.launch --master_addr ${MASTER_ADDR} --master_port ${MASTER_PORT} --nproc_per_node=1 --use_env train_math.py \
    --model_name_or_path $MODEL_PATH \
    --data_path "./data/train/MetaMathQA-40K.json" \
    --data_length 10000000 \
    --bf16 True \
    --output_dir $SAVE_PATH \
    --per_device_train_batch_size 4 \
    --per_device_eval_batch_size 4 \
    --gradient_accumulation_steps 16 \
    --evaluation_strategy "no" \
    --save_strategy "steps" \
    --save_steps 1000 \
    --save_total_limit 2 \
    --learning_rate 5e-2\
    --weight_decay 0. \
    --warmup_ratio 0.1 \
    --lr_scheduler_type "cosine" \
    --logging_steps 1 \
    --num_train_epochs 2 \
    --pattern "banded" \
    --off_diag 0 \
    --target_modules q_proj k_proj v_proj up_proj down_proj o_proj gate_proj \
    --adapter_name "svft"

#SVFT_Random_d=16
export SAVE_PATH='./Gemma_2B_metamath40k_svft_16diag_random'
CUDA_VISIBLE_DEVICES=0 python3 -m torch.distributed.launch --master_addr ${MASTER_ADDR} --master_port ${MASTER_PORT} --nproc_per_node=1 --use_env train_math.py \
    --model_name_or_path $MODEL_PATH \
    --data_path "./data/train/MetaMathQA-40K.json" \
    --data_length 10000000 \
    --bf16 True \
    --output_dir $SAVE_PATH \
    --per_device_train_batch_size 4 \
    --per_device_eval_batch_size 4 \
    --gradient_accumulation_steps 16 \
    --evaluation_strategy "no" \
    --save_strategy "steps" \
    --save_steps 1000 \
    --save_total_limit 2 \
    --learning_rate 5e-2\
    --weight_decay 0. \
    --warmup_ratio 0.1 \
    --lr_scheduler_type "cosine" \
    --logging_steps 1 \
    --num_train_epochs 2 \
    --pattern "random" \
    --off_diag 16 \
    --target_modules q_proj k_proj v_proj up_proj down_proj o_proj gate_proj \
    --adapter_name "svft"


#SVFT_Random_d=16
export SAVE_PATH='./Gemma_2B_metamath40k_svft_16diag_banded'
CUDA_VISIBLE_DEVICES=0 python3 -m torch.distributed.launch --master_addr ${MASTER_ADDR} --master_port ${MASTER_PORT} --nproc_per_node=1 --use_env train_math.py \
    --model_name_or_path $MODEL_PATH \
    --data_path "./data/train/MetaMathQA-40K.json" \
    --data_length 10000000 \
    --bf16 True \
    --output_dir $SAVE_PATH \
    --per_device_train_batch_size 4 \
    --per_device_eval_batch_size 4 \
    --gradient_accumulation_steps 16 \
    --evaluation_strategy "no" \
    --save_strategy "steps" \
    --save_steps 1000 \
    --save_total_limit 2 \
    --learning_rate 5e-3\
    --weight_decay 0. \
    --warmup_ratio 0.1 \
    --lr_scheduler_type "cosine" \
    --logging_steps 1 \
    --num_train_epochs 2 \
    --pattern "banded" \
    --off_diag 16 \
    --target_modules q_proj k_proj v_proj up_proj down_proj o_proj gate_proj \
    --adapter_name "svft"

#SVFT_Random_d=16
export SAVE_PATH='./Gemma_2B_metamath40k_svft_16diag_topk'
CUDA_VISIBLE_DEVICES=0 python3 -m torch.distributed.launch --master_addr ${MASTER_ADDR} --master_port ${MASTER_PORT} --nproc_per_node=1 --use_env train_math.py \
    --model_name_or_path $MODEL_PATH \
    --data_path "./data/train/MetaMathQA-40K.json" \
    --data_length 10000000 \
    --bf16 True \
    --output_dir $SAVE_PATH \
    --per_device_train_batch_size 4 \
    --per_device_eval_batch_size 4 \
    --gradient_accumulation_steps 16 \
    --evaluation_strategy "no" \
    --save_strategy "steps" \
    --save_steps 1000 \
    --save_total_limit 2 \
    --learning_rate 5e-3\
    --weight_decay 0. \
    --warmup_ratio 0.1 \
    --lr_scheduler_type "cosine" \
    --logging_steps 1 \
    --num_train_epochs 2 \
    --pattern "top_k" \
    --off_diag 16 \
    --target_modules q_proj k_proj v_proj up_proj down_proj o_proj gate_proj \
    --adapter_name "svft"


#EVAL
#python eval_gsm8k.py --model './Gemma_2B_metamath40k_svft_16diag_random' --data_file ../MetaMath/data/test/GSM8K_test.jsonl
#python eval_math.py --model './Gemma_2B_metamath40k_svft_16diag_random' --data_file ../MetaMath/data/test/MATH_test.jsonl