Update instruments_pretrain.sh

instruments_pretrain.sh

@@ -2,34 +2,33 @@ export WANDB_MODE=disabled
 export CUDA_LAUNCH_BLOCKING=0
 
 DATASET=Instruments
-BASE_MODEL=/datain/v-yinju/llama-7b
-RQVAE_MODEL=/datain/v-yinju/rqvae-zzx/models/instruments/Apr-01-2024_01-25-11/best_collision_model.pth
-DATA_PATH=/datain/v-yinju/rqvae-zzx/data
-OUTPUT_DIR=/datain/v-yinju/rq-llama/v3-train/Instruments/CoT
+BASE_MODEL=/home/jovyan/workspace/Llama-7b
+RQVAE_MODEL=/home/jovyan/workspace/LC-Rec/index/Ins/Apr-04-2025_07-12-04/best_collision_model.pth
+DATA_PATH=/home/jovyan/workspace
+OUTPUT_DIR=./Ins
 
 mkdir -p $OUTPUT_DIR
 
-torchrun --nproc_per_node=8 pre-train.py \
+torchrun --nproc_per_node=4 pre-train.py \
     --base_model $BASE_MODEL \
     --rqvae_model $RQVAE_MODEL \
     --output_dir $OUTPUT_DIR \
     --dataset $DATASET \
     --data_path $DATA_PATH \
-    --per_device_batch_size 6 \
+    --per_device_batch_size 8 \
     --gradient_accumulation_steps 2 \
     --learning_rate 5e-4 \
     --model_max_length 1024 \
-    --epochs 4 \
+    --epochs 1 \
     --weight_decay 0.01 \
     --save_and_eval_strategy epoch \
-    --fp16 \
-    --deepspeed ./config/ds_z2_fp16.json \
+    --deepspeed ./config/ds_z2_bf16.json \
     --dataloader_num_workers 4 \
     --only_train_response \
     --tasks seqrec,itemsearch,inters2title,inters2description,preferenceobtain,item2index,index2item,intertitles2item,query2item \
     --train_prompt_sample_num 1,1,1,1,1,1,1,1,1 \
     --train_data_sample_num 0,0,0,0,0,0,0,0,0 \
-    --index_file /datain/v-yinju/rqvae-zzx/models/instruments/Apr-01-2024_01-25-11/indices.json
+    --bf16 &>>$OUTPUT_DIR/log.txt
 
 cd convert
 nohup ./convert.sh $OUTPUT_DIR >convert.log 2>&1 &