instruments_more_pretrain.sh

export WANDB_MODE=disabled
export CUDA_LAUNCH_BLOCKING=0

DATASET=Instruments
BASE_MODEL=/datain/v-yinju/llama-7b
CKPT_PATH=/datain/v-yinju/rq-llama/v3-train/Instruments/first
DATA_PATH=/datain/v-yinju/rqvae-zzx/data
OUTPUT_DIR=/datain/v-yinju/rq-llama/v3-train/Instruments/more_pretrain

torchrun --nproc_per_node=8 --master_port=3324 continue_pretrain.py \
    --base_model $BASE_MODEL \
    --rqvae_model $CKPT_PATH \
    --output_dir $OUTPUT_DIR \
    --dataset $DATASET \
    --data_path $DATA_PATH \
    --per_device_batch_size 6 \
    --gradient_accumulation_steps 2 \
    --learning_rate 5e-5 \
    --epochs 4 \
    --weight_decay 0.01 \
    --save_and_eval_strategy epoch \
    --deepspeed ./config/ds_z2_fp16.json \
    --dataloader_num_workers 4 \
    --only_train_response \
    --tasks seqrec,itemsearch,inters2title,inters2description,preferenceobtain,item2index,index2item,intertitles2item,query2item \
    --train_prompt_sample_num 1,1,1,1,1,1,1,1,1 \
    --train_data_sample_num 0,0,0,0,0,0,0,0,0 \
    --fp16 &>>$OUTPUT_DIR/pretrain-log.txt

cd convert
nohup ./convert.sh $OUTPUT_DIR >convert.log 2>&1 &
cd ..