A Bolu: A Structured Dataset for the Computational Analysis of Sardinian Improvisational Poetry

The significance here isn't just preservation: improvisational poetry like cantada logudorese is structurally constrained by meter, rhyme, and real-time composition rules, which makes it a stress test for models trained almost entirely on written, edited text. A corpus that captures those constraints in structured form is a different kind of resource than a raw text dump.

This is largely disconnected from the recent activity covered on Modelwire, which has focused on commercial tooling, code generation benchmarks, and expressive speech synthesis (see Google DeepMind's Gemini 3.1 Flash TTS release from mid-April). The A Bolu dataset belongs to a quieter but persistent thread in NLP: the infrastructure problem for low-resource languages, where the bottleneck isn't model architecture but the absence of clean, annotated training data in the first place.

Watch whether any research group publishes a fine-tuned model or automatic scansion tool trained on A Bolu within the next 12 months. That would confirm the corpus is structured well enough to be practically useful, not just archivally complete.