June 2023 Summaries

Clarifai has integrated various large language models (LLMs) from vendors like Cohere, AI21 Labs, and OpenAI into its platform, enabling a wide range of capabilities such as text summarization, generation, embedding, translation, and more. New features include a Python Regex-Based Classifier for pattern-matching text classification and enhancements to the model viewer interface, like updated navigation bars and model management tools. Users can now perform transfer learning and auto-labeling with GPT models and customize non-user-owned model outputs in workflows. Numerous bug fixes have been applied to improve the functionality of input management, task management, and email verification processes, while the Clarifai-Python-Utils SDK has been expanded with utilities for data uploads and processing. The legacy portal is being deprecated, encouraging users to transition to the new portal for an improved experience.

The tutorial explores the use of Clarifai's General Recognition model to efficiently perform bulk labeling and create a custom model using transfer learning. Through the example of labeling a dataset containing various animal species, the tutorial demonstrates how to speed up the annotation process by allowing the model to identify and label images swiftly, significantly reducing manual effort. The process involves confirming the presence of specific animals like horses, dogs, elephants, butterflies, and chickens, and addressing any unlabeled images to ensure comprehensive training. By leveraging transfer learning, a new 'animal-classifier' model is created, which is trained on the labeled dataset and tested on a fresh set of images, yielding high accuracy in predicting animal types. The tutorial concludes with a successful demonstration of the model's capabilities and encourages users to explore machine learning applications using these techniques.

Large language models, such as GPT-4, LLaMA, and PaLM, are advanced tools capable of generating human-like text and providing insights, but they are limited by the data they were trained on and can struggle with untrained questions, sometimes leading to incorrect or "hallucinated" answers. To enhance their accuracy, particularly when dealing with complex documents like PDFs, a method has been developed to make such static formats more interactive by splitting the documents into manageable sections, creating embeddings, and storing them in a vector database. This approach allows the models to reference specific, relevant sections of a document when responding to queries, thereby improving the accuracy of the answers and mitigating the risk of hallucinations. An example of this process is demonstrated using documents from the International Crisis Group, where the system efficiently identifies and retrieves information relevant to specific queries, such as finding documents about terrorism or identifying individuals like Saefuddin Zuhri, and even plotting geographical locations on a map. This innovation not only overcomes the input limitations of large language models but also enhances their utility in research and accessibility.