Dataset Preparation

Format your training data for Qwen3-VL fine-tuning

JSON Data Structure

All training data must follow a standardized JSON format with the following components:

Required Fields

• Media field: Either image or video containing the path to media files
• conversations: An array of conversation turns with from and value fields

Media Tags

Use special tokens in prompts to indicate media placement:

• <image> for image understanding tasks
• <video> for video understanding tasks

Important formatting rules:

• One <image> tag must correspond to exactly one image file
• Similarly, <video> tags must correspond to video files
• These special tokens should NOT appear in the answer text

Example Formats

Single Image Example

Basic single-image question-answering format:

{
    "image": "images/001.jpg",
    "conversations": [
        {
            "from": "human",
            "value": "<image>\nWhat's the main object in this picture?"
        },
        {
            "from": "gpt",
            "value": "A red apple on a wooden table"
        }
    ]
}

Multi-Image Example

For tasks requiring comparison or reasoning across multiple images:

{
    "image": ["cats/001.jpg", "cats/002.jpg"],
    "conversations": [
        {
            "from": "human",
            "value": "<image>\n<image>\nWhat are the differences between these two cats?"
        },
        {
            "from": "gpt",
            "value": "The first cat is an orange tabby with short fur and green eyes, while the second is a gray Siamese with blue eyes and pointed coloration. They also appear to be in different environments - the first is indoors on a couch, the second is outdoors in a garden."
        }
    ]
}

The number of <image> tags in the prompt must match the number of images in the image array.

Video Example

For video understanding and temporal reasoning tasks:

{
    "video": "videos/005.mp4",
    "conversations": [
        {
            "from": "human",
            "value": "<video>\nWhat caused the blue object to move?\nOptions:\n(A) Gravity\n(B) Collision\n(C) Magnetic force"
        },
        {
            "from": "gpt",
            "value": "Answer: (B) Collision"
        }
    ]
}

Grounding Example

For object detection and localization tasks:

{
    "image": "demo/COCO_train2014_000000580957.jpg",
    "conversations": [
        {
            "from": "human",
            "value": "<image>\nLocate house in this image and output the bbox coordinates in JSON format."
        },
        {
            "from": "gpt",
            "value": "{\n\"bbox_2d\": [135, 114, 1016, 672]\n}"
        }
    ]
}

If you have grounding data in a different format, use the tools/process_bbox.ipynb notebook to transform your data into the QwenVL format.

Packed Data Example

For improved training efficiency, multiple samples can be packed into a single sequence:

[
    {
        "image": "images/001.jpg",
        "conversations": [
            {
                "from": "human",
                "value": "<image>\nWhat's the main object in this picture?"
            },
            {
                "from": "gpt",
                "value": "A red apple on a wooden table"
            }
        ]
    },
    {
        "image": "images/002.jpg",
        "conversations": [
            {
                "from": "human",
                "value": "<image>\nWhat's the main object in this picture?"
            },
            {
                "from": "gpt",
                "value": "A green orange on a plastic table"
            }
        ]
    }
]

To use packed data, preprocess your dataset with tools/pack_data.py and set data_packing=True in your training configuration.

File Organization

Your dataset can be organized in two ways:

Option 1: Relative Paths with data_path

# In your dataset config
{
    "annotation_path": "/data/annotations.json",
    "data_path": "/data/images/"  # Base path for media files
}

// In your JSON file
{
    "image": "001.jpg",  // Will be resolved to /data/images/001.jpg
    ...
}

Option 2: Absolute Paths

# In your dataset config
{
    "annotation_path": "/data/annotations.json",
    "data_path": ""  # Empty if using absolute paths
}

// In your JSON file
{
    "image": "/absolute/path/to/images/001.jpg",
    ...
}

Data Validation

For open-source datasets that might have missing images or other issues, verify data completeness using:

python tools/check_image.py --annotation_path /path/to/annotations.json

Available Public Datasets

You can use these datasets directly without additional processing:

• nyu-visionx/Cambrian-10M
• lmms-lab/LLaVA-NeXT-Data
• FreedomIntelligence/ALLaVA-4V
• TIGER-Lab/VisualWebInstruct

Example Files

The repository includes example datasets:

• demo/single_images.json - Single image examples
• demo/video.json - Video understanding examples

These files can be used directly for training or as templates for your own datasets.