After fine-tuning, the model outputs repetitive phrases

[Jackyzjz]

Thanks for your good job。

I am trying to fine-tune the videollama2 model with my own data. However, after fine-tuning, the model starts to repeatedly output the same content. Could you help me solve this issue?

[thisurawz1]

Can you share the inference script that you used to do the inference with fine-tuned LoRA weights?

[Jackyzjz]

I am performing LoRA fine-tuning based on videollama2-7b, and the script is as follows:

#!/bin/bash
export NCCL_P2P_DISABLE="1"
export NCCL_IB_DISABLE="1"

Environment Variables

ARG_WORLD_SIZE=${1:-1}
ARG_NPROC_PER_NODE=${2:-8}
ARG_MASTER_ADDR="127.0.0.1"
ARG_MASTER_PORT=16666
ARG_RANK=0

Multiple conditions

if [ ! -n "$WORLD_SIZE" ] || [ ! -n "$NPROC_PER_NODE" ]; then
WORLD_SIZE=$ARG_WORLD_SIZE
NPROC_PER_NODE=$ARG_NPROC_PER_NODE
fi
if [ ! -n "$MASTER_ADDR" ] || [ ! -n "$MASTER_PORT" ] || [ ! -n "$RANK" ]; then
MASTER_ADDR=$ARG_MASTER_ADDR
MASTER_PORT=$ARG_MASTER_PORT
RANK=$ARG_RANK
fi

echo "WORLD_SIZE: $WORLD_SIZE"
echo "NPROC_PER_NODE: $NPROC_PER_NODE"

Training Arguments

GLOBAL_BATCH_SIZE=8
LOCAL_BATCH_SIZE=1
GRADIENT_ACCUMULATION_STEPS=$[$GLOBAL_BATCH_SIZE/($WORLD_SIZE*$NPROC_PER_NODE*$LOCAL_BATCH_SIZE)]

Log Arguments

export TRANSFORMERS_OFFLINE=1
export WANDB_PROJECT=videollama2
RUN_NAME=new_dataset_lora
DATA_DIR=datasets
OUTP_DIR=/ssd/jacky
torchrun --nnodes $WORLD_SIZE
--nproc_per_node $NPROC_PER_NODE
--master_addr=$MASTER_ADDR
--master_port=$MASTER_PORT
--node_rank $RANK
videollama2/train_flash_attn.py
--lora_enable True --lora_r 128 --lora_alpha 256 --mm_projector_lr 2e-5
--deepspeed scripts/zero2.json
--model_type videollama2
--model_path /ssd/jacky/VideoLLaMA2-7B
--vision_tower /ssd/jacky/clip-vit-large-patch14-336
--mm_projector_type stc_connector
--data_path ${DATA_DIR}/videollava_sft/image_train.json
--data_folder ${DATA_DIR}/videollava_sft/
--mm_vision_select_layer -2
--num_frames 8
--bf16 True
--tf32 True
--fp16 False
--output_dir ${OUTP_DIR}/finetune_${RUN_NAME}
--num_train_epochs 5
--per_device_train_batch_size $LOCAL_BATCH_SIZE
--per_device_eval_batch_size 4
--gradient_accumulation_steps $GRADIENT_ACCUMULATION_STEPS
--evaluation_strategy "no"
--save_strategy "steps"
--save_steps 375
--save_total_limit 99
--learning_rate 2e-5
--weight_decay 0.
--warmup_ratio 0.03
--lr_scheduler_type "cosine"
--logging_steps 1
--model_max_length 2048
--gradient_checkpointing True
--dataloader_num_workers 4
--report_to tensorboard
--run_name $RUN_NAME \

[LiangMeng89]

Thanks for your good job。

I am trying to fine-tune the videollama2 model with my own data. However, after fine-tuning, the model starts to repeatedly output the same content. Could you help me solve this issue?

I also have this problem,do you solve it?

[LiangMeng89]

Thanks for your good job。

I am trying to fine-tune the videollama2 model with my own data. However, after fine-tuning, the model starts to repeatedly output the same content. Could you help me solve this issue?

Hello,I'm a phD student from ZJU, I also use videollama2 to do my own research,we create a WeChat group to discuss some issues of videollama2 and help each other,could you join us? Please contact me: WeChat number == LiangMeng19357260600, phone number == +86 19357260600,e-mail == [email protected].