TeichAI/Qwen3.6-27B-Fable-5-Experimental
TeichAI/Qwen3.6-27B-Fable-5-Experimental is a 27 billion parameter Qwen3.6-based language model, fine-tuned by TeichAI with a context length of 32768 tokens. This experimental model is specifically optimized for planning, 3D modeling in three.js, small game development, and ML engineering, having been over-trained on a small dataset of "fable 5 traces." While showing strong planning capabilities and a distinct "fable-like" style, it may exhibit regression on other tasks due to the aggressive fine-tuning on limited data.
Loading preview...
Overview
TeichAI/Qwen3.6-27B-Fable-5-Experimental is a 27 billion parameter model based on Qwen3.6, fine-tuned by TeichAI. This experimental version was developed by over-training on a small dataset of "fable 5 traces" using aggressive style transfer settings. The model aims to replicate the planning capabilities and conversational style of "fable 5," while maintaining its original reasoning abilities.
Key Capabilities & Characteristics
- Enhanced Planning: Demonstrates improved planning capabilities, making it suitable for complex task orchestration.
- Specialized for 3D & ML: Shows proficiency in 3D modeling using three.js, developing small games, and ML engineering tasks.
- Distinct Style: Exhibits a unique "fable-like" conversational and planning style.
- Reasoning Untouched: The base model's reasoning capabilities were preserved during fine-tuning.
Performance & Limitations
Benchmarking against the base Qwen3.6-27B model shows comparable performance on ARC and BoolQ, with slight variations:
- ARC: 0.650 (vs 0.637 for base)
- ARC-E: 0.813 (vs 0.798 for base)
- BoolQ: 0.909 (vs 0.911 for base)
Due to the small dataset and aggressive fine-tuning, the model may regress on tasks outside its specialized domains. It was trained using Teich for data extraction and formatting, and Unsloth for accelerated training.