Machine-Learning

What’s included Multi-Agent RL for Adaptive UIs Is ‘Programming by Example’ (PBE) solved by LLM’s Learning Iterative Reasoning through Energy Diffusion LORA: Low-Rank Adaptation of LLMs Automating the Enterprise with Foundational Models MARLUI - Multi-Agent RL for Adaptive UI ACM Link: https://dl.acm.org/doi/10.1145/3661147 Paper: MARLUI Adaptive UIs Adaptive UIs - as opposed to regular UI’s, are UI’s that can actually adapt to the users needs. In this paper, the UI is adapting to optimize the number of clicks needed to reach the outcome....

Introduction I previously wrote about writing react code with Deepseek-coder 33b model, and whether we could improve some of these shortcomings with the latest research in the LLM space But to really measure and mark progress, it would require the build of a benchmark to test various hypothesis around it. So in this post, I’m going to evaluate existing benchmarks that specifically measures LLM capabilities on coding capabilities. My goal is to be able to build a benchmark that can test their React/Typescript coding capabilities....

Introduction In my previous post, I tested a coding LLM on its ability to write React code. Specifically, I tried the currently leading open source model in the HumanEval+ benchmark leaderboard - DeepseekCoder:33b-instruct. I used this model in development for a few weeks, and published a subset of examples in the post. Even though I tried this on a relatively small problem size, there were some obvious issues that were recognisable to me, namely:-...

Introduction The goal of this post is to deep-dive into LLMs that are specialized in code generation tasks and see if we can use them to write code. Note: Unlike copilot, we’ll focus on locally running LLM’s. This should be appealing to any developers working in enterprises that have data privacy and sharing concerns, but still want to improve their developer productivity with locally running models. To test our understanding, we’ll perform a few simple coding tasks, compare the various methods in achieving the desired results, and also show the shortcomings....

Part 1/3 - Evaluating LLM’s that are specialised in code generation tasks, and evaluating their performance on writing code. This post starts with concepts and theory, while the next 2 parts evaluate specific code models.

Introduction The goal of this series of posts, is to form foundational knowledge that helps us understanding modern state-of-the-art LLM models, and gain a comprehensive understanding of GPT via reading the seminal papers themselves. In my previous post, I covered transformers via the original paper “Attention is all you need” that brought the innovation that made all this progress possible. This post will focus on GPT-3 and its predecessors GPT-1 and 2....

Part 2/3 - Understanding how modern LLMS work. From RNNs, to transformers, towards modern scaling laws.

Introduction ChatGPT has took the world by storm, and has possibly started the 6th wave. Given its importance, the rush to build new products and research on top is understandable. But, I’ve always liked to ground myself with foundational knowledge on how things work, before exploring anything additive. To gain such foundational knowledge, I believe understanding the progression of techniques and models is important to comprehend and appreciate how these LLM models work under the hood....

Introduction Loss functions tell the algorithm how far we are from actual truth, and their gradients/derivates help understand how to reduce the overall loss (by changing the parameters being trained on) All losses in keras defined here But why is the loss function expressed as a negative loss? Plot: As probabilities only lie between [0-1], the plot is only relevant between X from 0-1 This means, that it penalises a low probability of success exponentially more....