Readable
The system should remain small enough to inspect, change, and explain.
01 / About
Learning the whole system, not just the last API call.
Jarvisn't is an independent language-model project built and trained by Temuujin.
The project
A practical route through modern language modeling.
The project began with a simple question: what changes when you stop treating language models as black boxes and build the pipeline yourself?
Jarvisn't follows that question from raw text and tokenization through model architecture, pretraining, supervised fine-tuning, evaluation, and inference. The goal is technical understanding, expressed in code that can still fit in one person's head.
The model has now been trained, and its checkpoints are available. It is a small experimental model built to beat GPT-2, so its responses can still be unreliable and prone to hallucinations.
Creator
Temuujin
I'm Temuujin, a university student focused on artificial intelligence and software engineering. I built Jarvisn't to learn what it actually takes to build and train a language model rather than treating modern AI as a black box.
Contact and project links →Grounded
Status and claims come from the repository, not aspirational marketing.
Curious
The point is to understand each stage by building and testing it directly.