heretic-org/Phi-4-mini-reasoning-heretic: Decensored Math Reasoning

This model is a 3.8 billion parameter, decensored version of Microsoft's Phi-4-mini-reasoning, created using the Heretic v1.2.0 tool. It maintains the original model's architecture, including a 32768-token context length, and is specifically designed for advanced mathematical reasoning.

Key Capabilities

• Enhanced Mathematical Reasoning: Excels at multi-step, logic-intensive mathematical problem-solving, including formal proof generation and symbolic computation.
• Decensored Output: Modified to reduce refusals compared to the original Microsoft model (9/100 vs. 84/100 refusals in testing).
• Efficiency: Optimized for deployment in memory/compute-constrained environments and latency-bound scenarios.
• Synthetic Data Training: Fine-tuned on 150 billion tokens of synthetic mathematical content generated by a more capable model (Deepseek-R1), comprising over one million diverse math problems with verified correct solutions.
• Competitive Performance: Achieves strong results on reasoning benchmarks like AIME (57.5), MATH-500 (94.6), and GPQA Diamond (52.0), performing comparably to much larger models in its domain.

Good for

• Applications requiring robust mathematical problem-solving on edge or mobile systems.
• Educational tools for advanced math tutoring and problem generation.
• Scenarios where a compact model with strong reasoning capabilities and reduced content moderation is desired.
• Research and development in formal proof generation and symbolic computation.