about multi-image finetune

[univa-JASON]

when i finetune v2_5, can i compose dataset like this?

[
    {
      "id": "0",
      "image": {
        "<image_00>": "path/to/image_0.jpg",
        "<image_01>": "path/to/image_1.jpg",
        "<image_02>": "path/to/image_2.jpg"
      },
      "conversations": [
        {
          "role": "user", 
          "content": "<image_00>"
        }, 
        {
          "role": "assistant", 
          "content": "main_text"
        }
        {
          "role": "user", 
          "content": "<image_01>"
        }, 
        {
          "role": "assistant", 
          "content": "caption"
        }
        {
          "role": "user", 
          "content": "<image_02>"
        }, 
        {
          "role": "assistant", 
          "content": "caption"
        }
      ]
    }
  ]

[LDLINGLINGLING]

hi，2.5 can not support multi-image finetune

[univa-JASON]

oh, thanks for your answer. then only v2_6 support multi image? And when finetune with interleaved dataset, that format can be used for v2_6?

[LDLINGLINGLING]

yes，you can finetune with interleaved dataset

[univa-JASON]

i see. thank you so much!!