fusionbase/fusion-guide-12b-0.1

Model Overview

fusionbase/fusion-guide-12b-0.1 is an advanced AI reasoning system developed by fusionbase, leveraging the Mistral-Nemo 12 billion parameter architecture. Its core innovation lies in a two-model approach:

• A "Guide" model first generates a structured, step-by-step plan for a given task.
• A "Response" model then uses this detailed guidance to formulate an accurate and comprehensive answer.

This methodology significantly enhances the model's problem-solving capabilities, particularly for complex reasoning tasks.

Training and Data

The model is fine-tuned on a custom dataset comprising task-based prompts, with a distribution of 90% English and 10% German. The training data includes scenarios designed to be challenging or even unsolvable, which helps the model develop robustness in handling ambiguous situations. Each training sample follows a prompt => guidance structure, explicitly teaching the model to systematically break down complex tasks.

Key Features & Usage

• Enhanced Reasoning: The two-model system provides a unique approach to complex problem-solving by first generating a plan.
• Multilingual Support: Trained on both English and German data, offering capabilities in both languages.
• Structured Prompting: Requires prompts to be enclosed within <guidance_prompt>{PROMPT}</guidance_prompt> tags to activate its unique reasoning mechanism.
• Compatibility: Can be used with vLLM and other Mistral-Nemo-compatible inference engines, with examples provided for unsloth.

Limitations

The model may occasionally struggle to generate complete guidance, especially when prompts include specific instructions on response structure, a limitation stemming from its training methodology. For a detailed description and evaluation, refer to the fusionbase blog post.