hironzeAgadir, Morocco
I work somewhere between data, language models, and software. Lately I've been interested in small models, languages with less data, and figuring out what actually works outside a notebook.
see what I'm working onI like taking things apart, figuring out why they work, and building my own version.
My background is in computer science and data science. Most of the things I build end up somewhere around machine learning, language models, data, or software engineering.
I'm particularly interested in smaller models and low-resource languages: the part where an idea leaves the paper and has to actually run.
Right now: efficient inference, model routing, and Amazigh language data.
An ongoing set of experiments in parameter-efficient fine-tuning and smaller language models, testing how much performance survives under 3 billion parameters. Current playground: agriculture and hydrology question answering.
Python, PyTorch, QLoRA, PEFT, and vLLM.
A clean English-Tachlhit parallel dataset, built by hand, meant to make experimenting with this underrepresented language a little easier. Collecting, cleaning, aligning, and checking everything myself.
Dataset design, Tachlhit, multilingual NLP, and data cleaning.
I'm less interested in making something look impressive and more interested in finding out whether it actually works.
Not everything has to be about machine learning.
Reading: ML papers, language research, and things I find interesting.
Building: small model experiments and the Tachlhit dataset.
There's always another experiment to run, another dataset to clean, or another idea that probably shouldn't work but might.