In an October 12 article, AlphaEdgeAI describes Sparseup-fr-v0, a French replication of Linkup’s sparse approach. The project adapts the multilingual mLateOn backbone and an mmBERT language-model head with vocabulary trimmed to selected French tokens. It concerns information retrieval, not a general-purpose generative assistant.
The authors release their model and a cleaned NanoBEIR-fr version after inspecting training/evaluation leakage and removing suspicious rows. They compare several retrieval families and two document treatments: truncation and sliding windows. Those choices materially affect a document-search system and can change the relative ordering of models.
The article claims a leading result among the compared French sparse models, not superiority across every family and test. The authors explicitly say they did not investigate possible annotation errors. The useful contribution is therefore an inspectable recipe and artifacts, not universal superiority or proof of quality on legal or medical corpora. Repository and dataset terms require separate review before use.