| --- |
| license: mit |
| task_categories: |
| - text-generation |
| language: |
| - en |
| tags: |
| - ui-grounding |
| - gui-automation |
| - function-calling |
| - agents |
| - accessibility |
| size_categories: |
| - 1K<n<10K |
| --- |
| |
| # GenieLM UI-Grounding |
|
|
| Synthetic supervised fine-tuning data for **text-based UI grounding**: given a list of |
| on-screen elements (label + pixel center) and a natural-language instruction, pick the |
| single element to act on and emit a strict JSON action. |
|
|
| Built for [GenieLM](https://github.com/Barath19/GenieLM), a macOS agent that reads |
| the **accessibility tree as text** (not pixels) and lets a small LLM drive the cursor. |
|
|
| ## Format |
|
|
| Conversational SFT (`messages` column): |
|
|
| ```json |
| {"messages": [ |
| {"role": "system", "content": "You are a UI grounding model ... reply with ONLY JSON ..."}, |
| {"role": "user", "content": "Elements:\n[0] Button \"Go back\" center=(122,40)\n... \n\nInstruction: take me back"}, |
| {"role": "assistant", "content": "{\"action\":\"click\",\"id\":0}"} |
| ]} |
| ``` |
|
|
| Assistant target schema: `{"action":"click"|"type","id":<int>,"text":<string if typing>}`, |
| or `{"action":"none"}` when nothing matches. |
|
|
| ## Composition |
|
|
| - Synthetic element lists across 6 app types: browser, system settings, login forms, |
| code editors, e-commerce, media players. |
| - Direct and paraphrased / indirect instructions ("log in", "take me back", "go fullscreen"). |
| - `type` actions into text/search fields; ~10% no-match negatives. |
| - Reproducible: generated by [`gen_grounding_dataset.py`](https://github.com/Barath19/GenieLM/blob/main/tools/gen_grounding_dataset.py) with a fixed seed. |
|
|
| ## Intended use |
|
|
| Fine-tune a small decoder (e.g. Qwen3-4B, Gemma) to ground UI instructions to elements + |
| coordinates. A model trained on this drives GenieLM's `do`/`auto` agent loop. |
|
|
