AlphaEdge AI releases French sparse text-embedding model
AlphaEdge AI released Sparseup-fr, a French adaptation of Linkup's open-source method, together with a cleaned NanoBEIR-fr benchmark.
Event / data period:
Published: · Updated:
AlphaEdge AI released Sparseup-fr on Hugging Face, a French-language reproduction of Linkup's sparse text-embedding approach. The team said it used a multilingual mLateOn-unsupervised backbone and an mmBERT-base masked-language-model head trimmed to a French vocabulary; training took about 55 hours on one RTX 6000 Pro GPU.
On the authors' cleaned 13-dataset NanoBEIR-fr benchmark, Sparseup-fr led eight datasets, the compared OpenSearch multilingual sparse model led three and SPLADE CamemBERT led two. The team said it checked for possible leakage between training data and the benchmark and removed suspicious rows, but did not examine potential annotation problems. The release includes the model files and cleaned benchmark; results are the model authors' own evaluation.
Data
Leading sparse models on NanoBEIR-fr
Number of datasets led by each model on the cleaned 13-dataset French NanoBEIR benchmark.
- Datasets led by Sparseup-fr
- 8 count
- 2026-10-12 · Observed
- Datasets led by the OpenSearch multilingual sparse model
- 3 count
- 2026-10-12 · Observed
- Datasets led by the SPLADE CamemBERT model
- 2 count
- 2026-10-12 · Observed
Methodology
Counted from the cleaned NanoBEIR-fr results table in AlphaEdge AI's 12 October 2026 post. The authors said they removed data leakage and suspicious rows but did not investigate possible annotation problems. Counts are not an independent evaluation.
Updated:
Subjects and places
Subject proposals
- Artificial intelligenceSuggested
- Subject suggested from the title or summary: AIhuggingface.co ↗