Ishowbackup/Phi-4-reasoning-heretic

Model Overview

Ishowbackup/Phi-4-reasoning-heretic is a 14.7 billion parameter decoder-only Transformer model, a decensored version of Microsoft Research's Phi-4-reasoning. It was created using Heretic v1.1.0, resulting in a lower refusal rate (46/100) compared to the original model (67/100), while maintaining a low KL divergence of 0.0050. The base Phi-4-reasoning model was fine-tuned using supervised fine-tuning on chain-of-thought traces and reinforcement learning, focusing on high-quality data for math, science, and coding skills, as well as safety alignment.

Key Capabilities

• Advanced Reasoning: Excels in complex reasoning tasks across math, science, and coding, as demonstrated by strong performance on benchmarks like AIME, OmniMath, GPQA-Diamond, and LiveCodeBench.
• Decensored Output: Offers a reduced refusal rate, providing more direct responses compared to its original counterpart.
• Efficient Performance: Designed for memory/compute constrained environments and latency-bound scenarios, making it suitable for applications requiring quick and efficient processing.
• Structured Output: Generates responses with a distinct reasoning chain-of-thought block followed by a summarization block, requiring a specific ChatML template with a detailed system prompt for optimal inference.

Intended Use Cases

• Research Acceleration: A building block for generative AI features and general-purpose AI systems, particularly in English.
• Reasoning and Logic Applications: Ideal for tasks requiring advanced logical deduction and problem-solving.
• Resource-Constrained Environments: Suitable for deployment where computational resources or latency are critical factors.