Language model fine-tuning
Adapting open models to low-resource languages and narrow domains: assembling the dataset, running the training, then checking the output holds up against real speakers.
Jonathan Immanuel Montolalu: independent AI & data science engineer, Indonesia
I fine-tune language models and write the Python systems around them, so the result is something a person can open on a phone rather than a github repo sitting doing nothing.
Most AI work splits into people who train models and people who ship software. I do both, because the projects I care about are too small to hand off: adapting a model to a dialect with almost no training data, then building the interface that lets someone use it without knowing what a checkpoint is.
The problems tend to start close to home, classmates missing the shuttle, students drilling the same math concept without making progress, a dialect that mainstream models have barely seen. Each one is narrow enough to finish and real enough to be used.
Adapting open models to low-resource languages and narrow domains: assembling the dataset, running the training, then checking the output holds up against real speakers.
The architecture that turns a model or a data source into a product; request flow, scheduling, storage, and the unglamorous parts that have to keep working after launch.
Getting messy, real-world data into a shape a model can learn from and a person can read: collection, cleaning, and the analysis that decides what to build next.
Work that crosses the software–hardware line, where constraints are physical and the margin for a bad assumption is much smaller.
Four projects, each in use by someone other than me.
A fine-tuned large language model that translates formal Bahasa Indonesia into the Manadonese dialect.
Language model Published on Hugging Face
A simple web app that builds fluency in core math concepts through repeated practice.
Web application itermath.jomontolalu.com
An unofficial schedule tracker for the Binus Square shuttle bus, giving students departure and arrival times across the Kemanggisan, Alam Sutera, and JWC/FX routes.
Web application bsqshuttle.jomontolalu.com
A student-led high school organisation I co-founded with some partner, founded to help other students with practical programming skills.
Community Instagram
Have a cool project that aligns with my interests, or just have something to say? Email is the fastest way to reach me.