preethamBelweave/macos-ax-agent-sft
macOS AX-Agent — SFT dataset 1,887 next-action examples for training a text-only accessibility-tree → GUI action model. Source Curated from markov-ai/computer-use (Apache-2.0): successful computer-use agent trajectories on OSWorld tasks, produced by Gemini 3 Flash with linearized accessibility-tree grounding. Only fully-successful trajectories (score = 1.0) are included. Format Each row is {"messages": [...]} (chat template of… See the full description on the dataset page: https://huggingface.co/datasets/preethamBelweave/macos-ax-agent-sft.
macOS AX-Agent — SFT dataset
1,887 next-action examples for training a text-only accessibility-tree → GUI action model.
Source
Curated from `markov-ai/computer-use` (Apache-2.0): successful computer-use agent trajectories on OSWorld tasks, produced by Gemini 3 Flash with linearized accessibility-tree grounding. Only fully-successful trajectories (score = 1.0) are included.
Format
Each row is {"messages": [...]} (chat template of Qwen/Qwen3-4B-Instruct-2507):
- system: computer-use agent prompt + JSON action schema (coordinates normalized to 0..1)
- user:
TASK: ...+ optionalLAST ACTION:+CURRENT SCREEN (accessibility tree):(linearized AX tree:tag\ttext\tposition (center x & y)\tsize (w & h)) - assistant: canonical next action, e.g.
{"action": "click", "x": 0.62, "y": 0.41}
Action types: click (1030), exit (275), type (274), hotkey (164), openapp (63), scroll (38), wait (24), draganddrop (15), switchwindow (4).
Splits
train.jsonl— 1,604 examplesvalid.jsonl— 283 examples (split by trajectory, no leakage)
Reproducibility
scripts/prep_data.py regenerates the JSONL from the raw source parquet.
