- πββοΈ My name is JoΓ£o Lages
- π·β Deep Learning Engineer @ Revolut
- π± Iβm interested in everything about machine learning, with focus on deep learning applied to text, images, tabular data, video, speech, time-series, anything!
- π« How to reach me: joaop.glages@gmail.com
- βοΈ Blog posts:
- Model Merging: MoE, Frankenmerging, SLERP, and Task Vector Algorithms π§ - Deep dive on how LLM merging methods work (co-authored with Deci AI)
- OpenAI JSON Mode vs Functions - Practical differences between these two ways of using OpenAI API
- Direct Preference Optimization (DPO) - A simplified explanation of the DPO algorithm applied to large language models, like Zephyr
- Reinforcement Learning from Human Feedback (RLHF) πββοΈ - A simplified explanation of the RLHF algorithm applied to large language models, like ChatGPT
- Transformers KV Caching Explained πΎ - A short writing on how Key and Value states are cached in transformers for faster inference
- Transformers Positional Encodings Explained π - Positional encoding and how it limits the input size of language models
- Mahalanobis for outlier detection - A simple demo on how to use mahalanobis distance for outlier detection
- β Main open-source contributions:
- Diffusers-Interpret π€π§¨π΅οΈββοΈ - Own package, a model explainability tool built on top of π€ Diffusers
- Ecco - Major contributions to this package that is used to explain, analyze, and visualize NLP language models
- AI Reading Group - Co-author of an open AI reading group from 2019-2023
- RATransformers π - Own package, used to make transformer models relation-aware
- π€ datasets - implemented the mahalanobis distance metric