The Epsilon Trap: When Adam Stops Being Adam
Beyond numerical stability, we investigate an often overlooked hyperparameter in the Adam optimizer: epsilon.
Sifal Klioui
Entropic Instruction Following: Does Semantic Coherence Help LLMs Follow Instructions?
Testing whether semantic relatedness of instructions affects a model's ability to follow them under cognitive load
Elements Of Mechanistic Interpretability: From Observation to Causation
We strip down mechanistic interpretability to three key experiments: watching a model 'think', finding where it stores concepts, and performing 'causal surgery' to change its 'thought process'
SFT vs. DPO (/ RLHF)- A Visual Guide to What Your LLM Actually Learns
A visual guide and toy experiment to build intuition for the practical differences between Supervised Fine-Tuning (SFT) and Direct Preference Optimization (DPO).
Do You Need A Matryoshka Model?
An analysis conducted on the informations hotspots on embeddings.
Doing MORE To consume LESS – Flash Attention V1
Flash Attention played a major role in making LLMs more accessible to consumers. This algorithm embodies how a set of what one might consider "trivial ideas" can come together and form a powerful s...
KV cache – The how not to waste your FLOPS starter
You've probably heard of the Transformers by now, they're everywhere, so much so that new born babies are gonna start saying Transformers as their first word, this blog will explore an important co...
Attention scores, Scaling and Softmax
If you're familiar with the Attention Mechansim, then you know that before applying a softmax to the attention scores, we need to rescale them by a factor of $\frac{1}{\sqrt{D_k}}$ where $D_k$ is t...
The Hidden Beauty of Sinusoidal Positional Encodings in Transformers
In this blog we will shed the light into a crucial component of the Transformers architecture that hasn't been given the attention it deserves, and you'll also get to see some pretty vizualizations!
Chunking Strategies for RAG - Breaking Down Documents for Better Retrieval
A comprehensive guide to chunking strategies for Retrieval-Augmented Generation, from basic splitting to advanced semantic and agentic approaches.