DavidrPatton/Gemma-4-31B-Fable-Coder
DavidrPatton/Gemma-4-31B-Fable-Coder is a 31-billion parameter Vision-Language-Action (VLA) engine, fine-tuned from unsloth/gemma-4-31b-it-unsloth-bnb-4bit, designed for autonomous orchestration in self-hosted AI frameworks. It specializes in multi-step agentic reasoning, strict tool-calling schema adherence, and multimodal graphical user interface (GUI) interaction. This model excels at processing visual inputs to generate precise actions, making it suitable for automating complex software engineering tasks and desktop/web navigation. It features a 32768 token context length and was trained using Unsloth for faster fine-tuning.
Loading preview...
Model Overview
DavidrPatton/Gemma-4-31B-Fable-Coder is a highly specialized 31-billion parameter Vision-Language-Action (VLA) engine, developed by DavidrPatton. Fine-tuned from unsloth/gemma-4-31b-it-unsloth-bnb-4bit, this model is engineered for autonomous orchestration within self-hosted AI frameworks, focusing on agentic reasoning and multimodal interaction rather than general language tasks.
Core Capabilities
This model's unique capabilities stem from its fine-tuning on a curated mixture of four distinct datasets:
- Agentic Logic & Chain-of-Thought (CoT): Learned from
Glint-Research/Fable-5-traces, it develops methodical reasoning before acting, and adheres to strict<tool_use>XML schemas for multi-step software engineering, debugging, and file-system manipulation. - Native Desktop GUI Control: Trained on
Writer/omniact, it can analyze desktop screenshots and output precise X/Y coordinates for PyAutoGUI execution, enabling interaction with native operating systems. - Web Dashboard Navigation & DOM Parsing: Utilizing
McGill-NLP/WebLINX, the model parses complex browser layouts and understands the relationship between visual web elements and underlying HTML/DOM structures for autonomous web navigation. - Visual Schema Enforcement: With
ScaleAI/VisualToolBench, it connects visual contexts directly to structured tool calls, ensuring deterministic JSON/XML syntax output based on visual cues like UI elements or error codes.
Architecture & Fine-Tuning
The model uses Unsloth's FastVisionModel for its multimodal setup and was fine-tuned 2x faster using Unsloth and Huggingface's TRL library. It operates with a gemma-4-thinking prompt/chat template and requires image/visual content to be placed before text in multimodal inputs for optimal attention mapping.