reelva/phi3-mini-reasoning-beast
reelva/phi3-mini-reasoning-beast is a 3.8 billion parameter fine-tune of the Phi-3-mini-4k-instruct architecture, developed by reelva. Optimized for complex cognitive tasks, it features native Chain-of-Thought (CoT) capabilities through integration of the Opus 4.6 Filtered Thinking dataset. This model excels at logical reasoning, mathematical deduction, and code analysis, utilizing a 4K token context length.
Loading preview...
Overview
reelva/phi3-mini-reasoning-beast is a 3.8 billion parameter model based on the Phi-3-mini-4k-instruct architecture, specifically fine-tuned for enhanced reasoning. Its core differentiator is the integration of the Opus 4.6 Filtered Thinking dataset, which imbues it with native Chain-of-Thought (CoT) capabilities. This allows the model to perform structured internal deliberation, marked by a <think> tag, before generating its final output.
Key Capabilities
- Native Chain-of-Thought (CoT): Designed for step-by-step logical processing, enabling meta-analysis even for undefined tasks.
- Optimized for Complex Cognition: Excels in logical reasoning, mathematical deduction, and code analysis.
- 4K Token Context: Supports a substantial context window for detailed problem-solving.
- RoPE Patch Included: Addresses
RoPE KeyErrorfor modern Transformers compatibility.
Good For
- Logic Puzzles: Solving complex logical problems with detailed, step-by-step reasoning.
- Mathematical Deduction: Tasks requiring structured mathematical problem-solving.
- Code Analysis: Understanding and processing code-related logical structures.
- Zero-Shot Reasoning: Demonstrates robust meta-analysis when encountering ambiguous or undefined tasks, rather than hallucinating.
Limitations
- Experimental Reasoning Stream: Standard safety guardrails may behave differently; professional discretion is advised.
- Prompt Structure: Best performance is achieved using the
<|user|>and<|assistant|>prompt format.