🛠️ Core Capabilities & Training Data Elaboration
This model was fine-tuned on a heavily curated, deduplicated mixture of four distinct datasets. The objective was to combine high-level architectural reasoning with native visual-interaction capabilities without suffering from catastrophic forgetting.
1. Agentic Logic & Chain-of-Thought (CoT)
Dataset: Glint-Research/Fable-5-traces
- What it learned: This dataset injected 4,659 cleartext agentic-coding traces distilled directly from Anthropic's Fable 5 preview model.
- The resulting capability: The model learned how to "think" methodically before acting. It forces a rigorous Chain-of-Thought logical buffer before outputting its final response. It also learned how to perfectly format and execute strict
<tool_use> XML schemas for multi-step software engineering cycles, interactive debugging, and autonomous file-system manipulation.
2. Native Desktop GUI Control
Dataset: Writer/omniact
- What it learned: OmniACT provided thousands of image-instruction pairs mapping natural language directly to host-level operating system actions.
- The resulting capability: The model can analyze a visual screenshot of a heavy desktop application and output precise X/Y coordinates for PyAutoGUI execution, allowing the agent to effectively "click" and "type" anywhere on a native OS screen.
3. Web Dashboard Navigation & DOM Parsing
Dataset: McGill-NLP/WebLINX
- What it learned: WebLINX provided multi-turn conversational web navigation data, heavily emphasizing DOM trees and UI component grounding.
- The resulting capability: Instead of just reading text, the model can parse complex browser layouts, enabling it to autonomously navigate web interfaces, visual management dashboards, and router portals by understanding the relationship between visual web elements and underlying HTML/DOM structures.
4. Visual Schema Enforcement
Dataset: ScaleAI/VisualToolBench
- What it learned: A highly specific dataset engineered to map visual contexts directly to structured tool calls.
- The resulting capability: This acts as the connective tissue between the model's vision and its logic. It ensures that when the agent "sees" a specific UI element or a visual error code in a screenshot, it outputs the exact, deterministic JSON/XML syntax required by the backend routing framework without hallucinating conversational filler.
⚙️ Architecture & Fine-Tuning Details
- Framework: Unsloth
FastVisionModel (Multimodal setup).
- Precision: Merged 16-bit
.safetensors.
- Prompt/Chat Template:
gemma-4-thinking
- Modality Formatting Note: When inferencing with multimodal inputs, ensure the image and/or visual content is placed before the text in the system/user prompt for optimal attention mapping.