pierretokns/screenspot-mlx-ccmcp
ScreenSpot MLX-VLM Dataset Multi-format VLM dataset for GUI grounding / pixel coordinate prediction. Formatted for training with mlx-vlm. Formats Each model family has its own data directory with the correct format: Model Family Directory Description SmolVLM/IDEFICS3 data/smolvlm/ Structured content blocks Qwen2-VL data/qwen2-vl/ Inline vision tokens PaliGemma data/paligemma/ Prefix token format LLaVA data/llava/ Newline separator format… See the full description on the dataset page: https://huggingface.co/datasets/pierretokns/screenspot-mlx-ccmcp.
ScreenSpot MLX-VLM Dataset
Multi-format VLM dataset for GUI grounding / pixel coordinate prediction. Formatted for training with mlx-vlm.
Formats
Each model family has its own data directory with the correct format:
Statistics
- Total Examples: 1,272
- Train: 1,017
- Validation: 127
- Test: 128
Source Datasets
Usage
With HuggingFace Datasets
from datasets import load_dataset
# Load SmolVLM format
ds = load_dataset("pierretokns/screenspot-mlx-ccmcp", data_dir="data/smolvlm")
# Load Qwen2-VL format
ds = load_dataset("pierretokns/screenspot-mlx-ccmcp", data_dir="data/qwen2-vl")With mlx-vlm
# Train SmolVLM
uv run python -m mlx_vlm.lora \
--model HuggingFaceTB/SmolVLM-256M-Instruct \
--data pierretokns/screenspot-mlx-ccmcp \
--data-dir data/smolvlm \
--iters 500Output Format
All formats output JSON with action coordinates:
{"action": "left_click", "coordinate": [847, 523]}SmolVLM Format
{"messages": [{"role": "user", "content": [{"type": "image"}, {"type": "text", "text": "Click Submit"}]}, {"role": "assistant", "content": [{"type": "text", "text": "{...}"}]}], "images": ["images/001.png"]}Qwen2-VL Format
{"messages": [{"role": "user", "content": "<|vision_start|><|image_pad|><|vision_end|>Click Submit"}, {"role": "assistant", "content": "{...}"}], "images": ["images/001.png"]}License
Apache 2.0
Citation
If you use this dataset, please cite the original sources:
- ScreenSpot: SeeClick Paper
- OS-Atlas: OS-Atlas Paper
