richardyoung/Phi-3.5-mini-instruct-heretic
richardyoung/Phi-3.5-mini-instruct-heretic is a 3.8 billion parameter instruction-tuned causal language model, a decensored version of Microsoft's Phi-3.5-mini-instruct, built using Heretic v1.4.0. This model supports a 128K token context length and excels in reasoning tasks, particularly in code, math, and logic, while offering enhanced multilingual capabilities. It is optimized for memory/compute constrained environments and latency-bound scenarios.
Loading preview...
Overview
richardyoung/Phi-3.5-mini-instruct-heretic is a 3.8 billion parameter instruction-tuned causal language model, derived from Microsoft's Phi-3.5-mini-instruct and decensored using Heretic v1.4.0. It is part of the Phi-3 model family, trained on 3.4 trillion tokens of high-quality, reasoning-dense data, including synthetic data and filtered public websites. The model supports an extensive 128K token context length, making it suitable for long document summarization and QA tasks.
Key Capabilities
- Decensored Output: Offers a less restrictive response generation compared to the original model, with significantly reduced refusals (12/100 vs 95/100 for the original).
- Strong Reasoning: Excels in tasks requiring reasoning, particularly in code, mathematics, and logic, as evidenced by competitive scores on benchmarks like GSM8K and MATH.
- Multilingual Proficiency: Demonstrates strong performance across multiple languages on benchmarks such as Multilingual MMLU and MGSM, often outperforming larger models in its class.
- Long Context Understanding: Supports a 128K token context window, enabling effective processing of lengthy inputs for tasks like document summarization and information retrieval.
- Code Generation: Achieves solid performance on code generation benchmarks like HumanEval and MBPP, with specific strengths in understanding code across various languages (Python, C++, Rust, Java, TypeScript).
Good For
- Memory/Compute Constrained Environments: Its compact size (3.8B parameters) makes it efficient for deployment where resources are limited.
- Latency-Bound Scenarios: Designed for applications requiring quick response times.
- General Purpose AI Systems: Suitable as a foundational building block for various generative AI features and applications.
- Research on Language Models: Accelerates research, especially concerning decensored model behavior and performance.
- Multilingual Applications: Ideal for use cases requiring strong performance in multiple languages, including Arabic, Chinese, French, German, and Spanish, among others.