llama/tune.sh

# BASE_MODEL="facebook/opt-125m" #"meta-llama/Llama-2-7b-hf   facebook/opt-125m
BASE_MODEL="meta-llama/Llama-2-7b-hf"
# DATA_PATH="./data/MetaMathQA.json"
DATA_PATH="./data/MetaMathQA-40K.json"
OUTPUT="output/cp3e5"

export WANDB_PROJECT="HRA_MetaMath395"

# python finetune_32.py \
#     --model_name_or_path $BASE_MODEL \
#     --output_dir $OUTPUT \
#     --hrft_r 32 \
#     --init_a 1e-4 \
#     --eps 1e-4 \
#     --add_orth "none" \
#     --lamda 1e-4 \
#     --data_path $DATA_PATH \
#     --dataset_split "train[:100000]"\
#     --dataset_field query response \
#     --num_train_epochs 2 \
#     --per_device_train_batch_size 8 \
#     --gradient_accumulation_steps 4 \
#     --save_strategy "steps" \
#     --save_steps 0 \
#     --save_total_limit 1 \
#     --learning_rate 3e-5 \
#     --weight_decay 0. \
#     --warmup_ratio 0.005 \
#     --lr_scheduler_type "cosine" \
#     --logging_steps 1000 \
#     --bf16 True \ 
#     --tf32 True \
#     --report_to "none" \
# wandb sync wandb/latest-run

# OUTPUT="output/cp1e5N"
# python finetune_32.py \
#     --model_name_or_path $BASE_MODEL \
#     --output_dir $OUTPUT \
#     --hrft_r 32 \
#     --init_a 1e-4 \
#     --eps 1e-4 \
#     --add_orth "none" \
#     --lamda 1e-4 \
#     --data_path $DATA_PATH \
#     --dataset_split "train[:100000]"\
#     --dataset_field query response \
#     --num_train_epochs 2 \
#     --per_device_train_batch_size 8 \
#     --gradient_accumulation_steps 4 \
#     --save_strategy "steps" \
#     --save_steps 0 \
#     --save_total_limit 1 \
#     --learning_rate 1e-5 \
#     --weight_decay 0. \
#     --warmup_ratio 0.005 \
#     --lr_scheduler_type "cosine" \
#     --logging_steps 1000 \
#     --bf16 True \
#     --tf32 True \
#     --report_to "wandb"
# wandb sync wandb/latest-run

# OUTPUT="output/cpr1"
# python finetune_32.py \
#     --model_name_or_path $BASE_MODEL \
#     --output_dir $OUTPUT \
#     --hrft_r 1 \
#     --init_a 1e-4 \
#     --eps 1e-4 \
#     --add_orth "none" \
#     --lamda 1e-4 \
#     --data_path $DATA_PATH \
#     --dataset_split "train"\
#     --dataset_field query response \
#     --num_train_epochs 2 \
#     --per_device_train_batch_size 32 \
#     --gradient_accumulation_steps 1 \
#     --save_strategy "steps" \
#     --save_steps 0 \
#     --save_total_limit 1 \
#     --learning_rate 3e-5 \
#     --weight_decay 0. \
#     --warmup_ratio 0.005 \
#     --lr_scheduler_type "cosine" \
#     --logging_steps 1000 \
#     --bf16 True \
#     --tf32 True \
#     --report_to "wandb"
# wandb sync wandb/latest-run

# OUTPUT="output/cpr2"
# python finetune_32.py \
#     --model_name_or_path $BASE_MODEL \
#     --output_dir $OUTPUT \
#     --hrft_r 1 \
#     --init_a 1e-4 \
#     --eps 1e-4 \
#     --add_orth "none" \
#     --lamda 1e-4 \
#     --data_path $DATA_PATH \
#     --dataset_split "train"\
#     --dataset_field query response \
#     --num_train_epochs 3 \
#     --per_device_train_batch_size 32 \
#     --gradient_accumulation_steps 1 \
#     --save_strategy "steps" \
#     --save_steps 0 \
#     --save_total_limit 1 \
#     --learning_rate 3e-5 \
#     --weight_decay 0. \
#     --warmup_ratio 0.005 \
#     --lr_scheduler_type "cosine" \
#     --logging_steps 200 \
#     --bf16 True \
#     --tf32 True \
#     --report_to "wandb"
# wandb sync wandb/latest-run

OUTPUT="output/cms3"
python finetune_32.py \
    --model_name_or_path $BASE_MODEL \
    --output_dir $OUTPUT \
    --hrft_r 32 \
    --init_a 1e-4 \
    --eps 1e-4 \
    --add_orth "none" \
    --lamda 1e-4 \
    --data_path $DATA_PATH \
    --dataset_split "train"\
    --dataset_field query response \
    --num_train_epochs 2 \
    --per_device_train_batch_size 8 \
    --gradient_accumulation_steps 4 \
    --save_strategy "steps" \
    --save_steps 0 \
    --save_total_limit 1 \
    --learning_rate 1e-5 \
    --weight_decay 0. \
    --warmup_ratio 0.005 \
    --lr_scheduler_type "cosine" \
    --logging_steps 200 \
    --bf16 True \
    --tf32 True \
    --report_to "wandb"
date +"%F %T"
# wandb sync wandb/latest-run