After fine-tuning qwen2.5 vl on my own dataset and using the fine-tuned weights for inference, most of the outputs were normal, but some became abnormal.
The fine-tuning data format I used was based on https://github.com/modelscope/ms-swift/blob/main/examples/notebook/qwen2_5-vl-grounding/zh.ipynb ： {“messages”: [{“role”: ‘system’, “content”: “You are a helpful assistant.”}, {“role”: “user”, “content”: “Find the in the image”}, {“role”: “assistant”, ‘content’: “”}], “images”: [“/xxx/x.jpg”], “objects”: {“ref”: [‘sheep’], “bbox”: [[90.9, 160.8, 135, 212.8], [360.9, 480.8, 495, 532.8]]}}. Here, I am using absolute coordinates.
Examples of abnormal outputs from the fine-tuned model include:

1. “<|box_start|>(764,25),(1659,1)”,
2. “<|box_start|><|box_start|>”,
3. “ <|box_start|>“,
4. ”<|box_start|>(987,369),(10825,370),(1141,760)<|box_end|>"

The fine-tuning script I used is:

MAX_PIXELS=1003520 \
swift sft \
    --model_type qwen2_5_vl \
    --model ../../qwen_25_vl_7B_awq \
    --dataset train_data.jsonl \
    --val_dataset val_data.jsonl \
    --attn_impl flash_attn\
    --train_type lora \
    --torch_dtype bfloat16 \
    --num_train_epochs 1 \
    --per_device_train_batch_size 2 \
    --per_device_eval_batch_size 2 \
    --learning_rate 1e-4 \
    --lora_rank 8 \
    --lora_alpha 32 \
    --target_modules all-linear \
    --freeze_vit true \
    --eval_steps 300 \
    --save_steps 100 \
    --save_total_limit 5 \
    --logging_steps 5 \
    --max_length 2048 \
    --output_dir output \
    --warmup_ratio 0.05 \
    --dataloader_num_workers 4 \
    --dataset_num_proc 4