ArchiveStudio/Phi-4-reasoning

Overview

ArchiveStudio/Phi-4-reasoning is a 14.7 billion parameter dense decoder-only Transformer model developed by Microsoft Research, fine-tuned from Phi-4. It is specifically designed for advanced reasoning tasks, leveraging supervised fine-tuning on chain-of-thought traces and reinforcement learning. The training data includes synthetic prompts and high-quality filtered public domain data focused on math, science, and coding, alongside alignment data for safety.

Key Capabilities

• Advanced Reasoning: Excels in complex problem-solving across math, science, and coding domains.
• Optimized for Efficiency: Designed for memory/compute-constrained and latency-bound environments.
• Chain-of-Thought Generation: Produces detailed reasoning steps followed by a concise solution, requiring a specific ChatML system prompt.
• High Context Length: Supports a 32k token context window, beneficial for complex queries requiring extensive reasoning.
• Strong Performance: Outperforms significantly larger open-weight models like DeepSeek-R1-Distill-70B on various reasoning benchmarks, including AIME, OmniMath, GPQA-Diamond, and LiveCodeBench.

Good For

• Accelerating Research: Serves as a building block for generative AI features and research in language models.
• General Purpose AI Systems: Suitable for applications requiring strong reasoning and logic, primarily in English.
• Educational Tools: Can be used for generating explanations and solutions for complex math and science problems.
• Code Generation and Analysis: Demonstrates strong functional code generation capabilities, particularly in Python.
• Benchmarking and Evaluation: Useful for evaluating reasoning capabilities against established benchmarks like Eureka and internal metrics.