Amir Tavanaei’s Post

https://lnkd.in/gNi7rqcG Just described the internals of transformer with a simple example. This could be a nice starting point for anyone interested in learning LLMs

Thrilled to share that our new research article, "An adaptive synthetic sampling and batch generation-oriented hybrid approach for addressing class imbalance problem in software defect prediction," has just been published in Soft Computing (IF 3.1)! This work addresses one of the key challenges in software defect prediction -class imbalance. Our approach aims to enhance predictive performance, providing valuable insights for both researchers and practitioners working in software quality assurance and data-intensive systems. A big thank you to Saif Ur Rehman Khan, PhD for his contributions to this accomplishment! 📖 https://lnkd.in/dTGwxnvV 📬 I’d love to hear your thoughts or discuss potential applications of our approach. Feel free to connect and share your insights! #Research #SoftwareEngineering #ClassImbalance #SoftwareDefectPrediction

This blog post introduces the concept of prompts and prompt engineering for LLMs. If you're unsure about the distinction between user prompts and system prompts, this is a must-read. Check it out here: https://lnkd.in/gZUzNCGQ #LLMs #promptengineering #userprompts #systemprompts

The learning of my phi2 fine tuning. https://lnkd.in/gV4U3TPJ

Really exciting technical report released by Meta on Multi-Modal Llama3 training which surpasses GPT-4o on many benchmarks. The engineering effort is pure genius, and the report is pure gold. No, I didn't read the whole 92 page report, but I read a fantastic summary by Aakash Nain. Thanks to Bharat Shetty B for pointing me to it. Here is the link to the summary: https://lnkd.in/gcr4vZHf In case you want to understand the Llama3 model code, which was released just a few months back. Then you can refer to this blog I published with Hasgeek, where I put my learnings to paper. Here is the link to my blog: https://lnkd.in/gtp-GVqr

🚀Just finished the course “LLM Foundations: Vector Databases for Caching and Retrieval Augmented Generation (RAG)” by Kumaran Ponnambalam!  Here's a glimpse of the key learnings: Explored the world of GenAI with vector databases, understanding the concept of vectors, their role in NLP vectorization, and the significance of vector similarity search. Mastered Milvus DB, including its architecture, collections, partitions, indexes, and data management capabilities. Additionally, gained practical experience by setting up Milvus and working with exercise files. Deepened understanding of Milvus database operations, such as creating connections, databases, and users, inserting data, building indexes, and querying scalar data, along with searching vector fields. Explored the application of Vector DB for LLM query caching, including the prompt caching workflow, setting up Milvus cache, inference process, and cache management. Delved into Retrieval Augmented Generation (RAG), learning about LLMs as a knowledge source, the RAG concept, its question-answering process, and various real-world applications. Implemented RAG with Milvus, covering the setup process, data preparation for the knowledge base, populating the Milvus database, and utilizing RAG for answering questions. Discussed best practices for vector databases, including considerations for choosing a vector database, effectively combining vector and scalar data, understanding distance measure implications, and optimizing performance. Excited to leverage these insights for future AI projects! 🌟 #VectorDatabases #RAG #MilvusDB #AI #KnowledgeCuration

Working as an ML Engineer in this moment of continual technical breakthroughs causes me to think, “is this reality or science fiction?” many times over the course of a given week. The other day I watched an LLM “planner” persona direct other LLM “worker” personas on how to best answer a user’s question, given the faculties available in an OpenAPI REST spec. If a worker persona came up with an API call that resulted in either a mismatch with the schema definition, or the API call failing with a 4xx error, the worker persona would be given feedback that their work was incorrect. The worker would then correct its mistake, and managed to successfully query the necessary API endpoint. It returned the answer, which was combined with other worker answers by a “reporter” LLM to return the correct answer back to the user. The entire conversation among agents played out in the log console of a Docker container. Given that AI agents can perform the sort of planning, coordination, and self-correction that was limited to humans only a few years ago, it’s a good time to look back at the field of cybernetics. Cyberneticist Stafford Beer’s Viable System Model (VSM) has been a major influence to my architectural approach for integrating LLMs into planning and control systems. In his book ‘Brain of the Firm’, Beer explains how VSM can be used to construct systems of people that are both self-correcting and internally stable in the face of a changing environment. His VSM went on to influence business management theory, along with an experiment by the Chilean government in decentralized planning during the early 1970’s. I view this as bridging the gap between traditional software engineering practices, and practices that take advantage of the new capacities available in LLM systems. Brain of the Firm: https://lnkd.in/d4EhZVWf #apokto [Photograph of the Project Cybersyn control + coordination room, courtesy The New Statesman]

Grounding is the process of anchoring LLMs’ responses in real-world knowledge, ensuring relevance to the context. While LLMs possess vast knowledge, it’s often not tailored to specific use-cases. To enhance accuracy and quality, we need to provide LLMs with context-specific information. Retrieval Augmented Generation (RAG) is a powerful technique for grounding, allowing LLMs to incorporate relevant data when generating responses. Grounding has applications in search systems, context-aware suggestions, and more. Remember, LLMs are engines, not databases! Insightful article by Eleanor Berger

After years of interest in the topic of software and the environment, and with my master's thesis giving me the chance to explore it more deeply, I decided to rewrite a part of my theoretical framework as an online article. I believe we're not talking enough about how big the environmental impact of software is, so here are my two cents on making this problem a bit more visible.