AlphaEdge AI releases French Sparseup-fr-v0 retrieval weights after 55-hour training run

AlphaEdge AI has published Sparseup-fr-v0, a downloadable French-language sparse retrieval model built by adapting Linkup's Sparseup approach. The October 12 author article describes the release as an experiment in reproducing an English sparse-embedding recipe for French search. Instead of generating chat responses, the checkpoint maps queries and documents to sparse lexical-semantic vectors that can be compared for retrieval. This makes the release relevant to French-language search and retrieval-augmented generation pipelines.

The authors changed the multilingual backbone to LightOn's mLateOn-unsupervised and used a trimmed mmBERT masked-language-model head. The resulting vocabulary contains 50,240 tokens, which also defines the sparse output dimensionality in the official model card. AlphaEdge AI reports about 55 hours of training on a single RTX 6000 Pro GPU using LightOn's French fine-tuning dataset. The model card lists a 256-token query limit and a 1,024-token document limit, and the Hugging Face page displays Apache-2.0 license metadata.

The publisher compared the model with other sparse French retrievers using a cleaned NanoBEIR-fr benchmark. It reports average nDCG@10 of 56.37 when documents are truncated and 56.50 under sliding-window processing. Those figures are AlphaEdge AI's own evaluations on its selected benchmark version, not independent Open Model Weights measurements. The author also notes possible remaining annotation issues and describes the result as an initial v0 replication rather than a comprehensive multilingual release.

For developers, the model card provides a Sentence Transformers SparseEncoder example and notes trust_remote_code=True in the published loading snippet. Teams should inspect that code before deployment, confirm their intended search index accepts sparse outputs, and measure latency and relevance on their own French corpora. The public checkpoint is a credible registry intake candidate, but the exact artifact revision, license file, dependencies and operational compatibility require separate field-level verification. A downloadable checkpoint alone is not evidence of a tested production integration.