Sparseup-fr brings Linkup's open sparse embedding recipe to French on Hugging Face
Loïck Bourdois published Sparseup-fr-v0, a 150-million-parameter French sparse embedding model trained by adapting Linkup's Sparseup method. On a cleaned NanoBEIR-fr benchmark it leads other open French sparse models he compared.
Photo: Benh LIEU SONG / Wikimedia Commons, Public domain. Hugging Face logo via Wikimedia Commons (public domain).
On the Hugging Face blog on October 12, Loïck Bourdois of AlphaEdge AI described Sparseup-fr, a French replication of Linkup's open-source sparse embedding model linkup-sparseup-embed-v1.
Bourdois adapted the method by swapping in LightOn's multilingual mLateOn-unsupervised backbone and a trimmed mmBERT-base masked-language-modeling head kept to 50,240 French tokens, aiming for a model size close to Linkup's English Sparseup. He trained for about 55 hours on a single RTX 6000 Pro for one epoch on LightOn's French training data, doubling query and document lengths relative to Linkup's setup.
On a cleaned NanoBEIR-fr benchmark, Sparseup-fr-v0 averaged 56.37 nDCG@10 with truncation and 56.50 with a sliding window, ahead of the other open French sparse models he compared: Amazon's opensearch-neural-sparse-encoding-multilingual-v1 and splade-camembert-base-v2. Late-interaction mLateOn still led overall, matching the English pattern Linkup reported.
Bourdois released the v0 model and cleaned benchmark in a Hugging Face collection and said further work may include more data, weight merging, a larger French model or a multilingual version.