在私域数据集上微调后做caption with grouding，但是bbox框不到目标，差距很大

[Missfoxsy]

微调数据集大致格式如下：{
"id": "00000026",
"conversations": [
{
"from": "user",
"value": "images/train/ECD0016015_B200A008C0.jpg\nPlease locate the anomalous structure in the image, state which terrain type it belongs to, and summarize its geographic features."
},
{
"from": "assistant",
"value": "Anomalous geological formation in mountainous terrain(647,573),(799,743) The anomalous structure lies within the annotated region, showing topographic irregularities, spectral variations, an increased elevation-to-base ratio, and irregular surface textures."
}
]
}
微调后的模型在同样这张图上做推理的时候输出的box里面坐标是(405,777),(544,942)，和grounding坐标相差很大，完全框不到目标，但是生成的文本又是对的，这是模型本身的能力问题还是我没有找到最合适的参数？