Hugging Face Overview
Overview
unsloth/Phi-4-mini-instruct is a 3.8 billion parameter instruction-tuned model from the Phi-4 family, developed by Microsoft and further refined by Unsloth with critical bug fixes. It boasts a substantial 131072-token context length and an expanded 200K vocabulary, supporting broad multilingual capabilities. The model is built on a foundation of 5 trillion tokens, combining filtered public data with newly created synthetic, "textbook-like" data specifically designed to enhance reasoning in math, coding, and common sense.
Key Capabilities
- Strong Reasoning: Excels in mathematical and logical reasoning, as evidenced by high scores on benchmarks like GSM8K (88.6) and MATH (64.0).
- Multilingual Support: Features a larger vocabulary and improved architecture for multilingual understanding, covering 22 languages.
- Efficiency: Designed for memory/compute-constrained and latency-bound scenarios, making it suitable for edge deployments.
- Instruction Adherence: Underwent supervised fine-tuning and direct preference optimization for precise instruction following and robust safety.
- Function Calling: Supports tool-enabled function calling with a specific input format.
Good For
- Resource-constrained environments: Ideal for applications where memory or computational power is limited.
- Latency-sensitive applications: Provides fast inference for scenarios requiring quick responses.
- Reasoning-heavy tasks: Particularly strong in math and logic problems.
- Multilingual applications: Benefits from an expanded vocabulary and training for diverse languages.
- Research and development: Serves as a building block for generative AI features and accelerating research in language models.