Home

The work runs in three strands: curating multilingual datasets for domains where data is sensitive or scarce, designing the model architecture and the tokenizer that feeds it, and measuring what comes out against much larger models. Everything it produces — models, datasets and training code — is published openly with documentation as the work proceeds. More about the project.

What the project publishes

Everything the project produces is public as it is made, not at the end.

• Publications

  Papers, preprints, posters and talks, each with its DOI, PDF and BibTeX where one exists.

• Models and datasets

  The trained models and the curated corpora, released as they become usable.

• Code and experiments

  The data pipeline, the tokenizer sweep, and the training and evaluation code, with every experiment recorded as it runs.