vivianamarquez/adaption-kirundi-sft

An open-source starter repo for improving low-resource Kirundi SFT data with Adaption and evaluating post-training outcomes. # Adaption Kirundi SFT An open-source starter repo for improving low-resource Kirundi SFT data with Adaption and comparing post-training outcomes. > If we improve low-resource SFT data before training, what changes can we observe in downstream model behavior? ## Why This Project Exists In 2023, I taught AI in Burundi and saw the AI gap up close. At the time, many of the most widely used LLM tools were not accessible from the country. But the gap was bigger than platform access alone. It was also a language gap, a data gap, and a design gap. Many AI workflows assume abundant English data, large evaluation sets, and easy access to native-language reviewers. That is not the reality for many low-resource languages. For communities whose languages are underrepresented in training data, AI often does not meet people where they are. This repo explores one practical slice of that problem: can adaptive data improvement make a Kirundi SFT workflow easier to build, test, compare, and improve? The broader question is even more important: what would it look like for AI systems to adapt to the world, instead of expecting the world to adapt to them? The framing is intentionally cautious. Qualitative model outputs are only early signals. Native speaker review matters. This repo does not claim to solve low-resource language AI. Instead, it offers a clear starter workflow that developers, researchers, and product teams can extend responsibly. Teaching AI in Burundi in 2023. At the time, some major LLM tools were not available in the country. ## What This Repo Builds This repo walks through a small SFT experiment in five notebooks: 1. `Notebook 01`: load a small subset of `ptrdvn/kakugo-run`, normalize prompt/response columns, and create the raw SFT file. 2. `Notebook 02`: upload the raw data to Adaption, inspect API-visible metadata, run the adaptation job, and convert the adapted output to SFT format. 3. `Notebook 03`: fine-tune the base model on the raw …