RESEARCH AT FORTUNAML
Our current research starts with a concrete question: how can focused training make an open model a better Python developer? We’re studying specialization alongside the cost of running useful models.
01QUESTIONS THAT GUIDE US
How can training help a model understand a Python issue, choose a focused repair, and verify the result?
Current work: LoRA training, tool use, and test-based evaluation of repository repairs using Gemma 4.
What does a Python specialist need to learn beyond producing code that looks plausible?
Current work: repository navigation, debugging, Python behavior, and training examples with reproducible checks.
How can specialized models become useful to people with limited memory, compute, and budgets?
Research direction: lower hardware requirements, practical inference, and measuring memory, latency, and repair quality together.
02HOW WE THINK
We want the work to be useful to people who build with models, as well as the people who train them.
A clear hypothesis makes an experiment worth running and its results easier to interpret.
Stronger performance on one task can come with limitations elsewhere. Both belong in the conversation.
Our work is intended for open release: code, model artifacts, methods, and findings that others can inspect and build on.
THE CURRENT EXPERIMENT
The Gemma 4 Python specialist connects our training research to a concrete software-engineering task. The development workflow covers adapter training, inference, patch generation, and independent test checks. Public research notes and reproducible results will accompany future releases.
See the project’s current progressGOOD THINGS BEGIN WITH A CONVERSATION
Research partnerships, funding,
and a shared belief in open intelligence.
