LLM Live Book - Project Update - 2023-08-11
2Summary
-------
Amir Feizpour presented an open source project that aims to revolutionize knowledge-ops by leveraging LLMs. The project focuses on operationalizing knowledge, enabling machines to read and provide real-time information. The goal is to enhance communication, collaboration, and project management by utilizing machine learning models with language and reasoning skills.
Topics:
-------
Project Overview
* Machine learning agents collaborating to find and provide necessary information for knowledge-ops
* Integration of AI into workflows to streamline processes and improve productivity
* Emphasis on language models with larger language skills and reasoning abilities
Live Book and Sherpa
* Development of a 'live book' called Sherpa, an interactive book that writes itself
* Sherpa can be accessed through a companion app and provides on-demand information
* Sherpa is currently in the Alpha stage and user contributions are encouraged
Community Involvement
* Collaborative effort involving numerous individuals from the community and company
* New contributors are welcome to join and have their picture added to the project's progress slide
Sherpa's Functionality
* Sherpa serves as a question-answering agent built on top of a large language model-based pipeline
* Provides answers based on the content of the open book and indexed documentation
Project Updates
* Significant progress has been made since the last update
* Features such as resource citation and improved error handling have been added
* Ongoing work to enhance multiplayer behavior in the Slack app and improve user experience
Future Plans
* Focus on expanding the project by adding more tools and resources
* Indexing up-to-date documentation and creating specialized tools for specific topics
* Community involvement and contributions are encouraged to enhance project capabilities
Improvements in Sherpa
* Addressed the issue of taxonomy by fine-tuning a smaller open source model
* Improved resource citation by providing an ID to the document instead of the actual link
* Implemented a method to handle links posted in the chat by summarizing the linked page
Availability and Future Plans of Sherpa
* Sherpa is currently available in public slack channels
* Plans to make it available through dms for more private inquiries
* Considerations to implement a rate limit for the deployed version
* Planning to install Sherpa in other communities and make it available on the marketplace
Finetuning Llama 2:
* Utilization of Collab for finetuning the model
* Fine-tuning the model involves an auto-regressive approach
* Options for hosting the model include Hugging Face and Sagemaker Jumpstart