Microsoft Releases A Bible for Language Models, Calls it Guidance

Microsoft today released a Bible for language models, called Guidance. This new tool allows users to control language models more effectively and efficiently than traditional prompting or chaining. The language model could also include open-source models like Vicuna, besides GPT-4 and others. 

The company said that the Guidance program allows users to interleave generation, prompting, and logical control into a single continuous flow, matching how the language model actually processes the text.  Looks like the tech giant is scripting a new religion in the name of LLM. It even quoted GPT Proverb 11:14 – which says, “Where there is no guidance, a model fails, but in an abundance of instructions, there is safety.”

Check out the GitHub repository here. 

Microsoft said that simple output structures like the chain of thought (COT) and its many variants, including ART, Auto-COT, etc., have been shown to improve the performance of the language model. It said that the advent of more powerful LLMs like GPT-4, allows for an even richer structure, and Guidance makes that structure easier and cheaper. 

Guidance Features 

• The Guidance offers simple and intuitive syntax based on Handlebars templating. 
• Provides rich output structure with multiple generations, selections, conditionals, tool use, etc. 
• Users can experience Playground-like streaming in Jupyter/VSCode Notebooks 
• Offers smart seed-based generation caching. 
• Supports tole-based chat models (ChatGPT) 
• This new feature can be easily integrated with HuggingFace models. This includes guidance acceleration, token healing and regex pattern guides. 

Earlier this month, researchers from Microsoft unveiled automatic prompt optimisation (APO), a simple and general-purpose framework for the automatic optimisation of LLM prompts. Check out the research paper here. 

Last year, at Build 2022, Microsoft announced the launch of open-source tools and datasets designed to audit AI-powered content moderation systems, alongside automatically writing tests that highlight potential bugs in AI models. At the time, the company claimed that the projects, namely AdaTest and (De)ToxiGen, could lead to more reliable language models.