Wolof-Adapted Industry-Ready Open-Weight Language Models
Quick Answer
The paper introduces 0.6B-14B open-weight language models tailored for Wolof, demonstrating strong performance in understanding and generation across 63 tasks.
Quick Take
The 8B and 14B models compete effectively with larger commercial systems, addressing the need for local deployment without the constraints of API models.
Key Points
- Models adapted through instruction tuning and merging for better Wolof support.
- Performance evaluated on 63 public and in-house tasks, showing competitive results.
- 8B and 14B models outperform larger commercial systems in specific benchmarks.
- Models available for public use and currently in internal customer-care pilots.
- Addresses governance and cost issues associated with API-based language models.
DeepSignal Analysis
What happened
The paper presents open-weight language models ranging from 0.6B to 14B parameters specifically adapted for the Wolof language. These models demonstrate effective performance across 63 tasks, including both understanding and generation. The 8B and 14B models show competitive results against larger commercial systems.
Key evidence
- The Wolof language is spoken by approximately 23 million people, primarily in Senegal, yet it has limited support from existing open-weight language models.
- The models were developed using techniques such as instruction tuning, preference alignment, and model merging, which are detailed in the paper.
- The 8B and 14B models are reported to compete effectively with larger commercial systems, addressing local deployment needs without API constraints.
Why it matters
This development is significant as it addresses the gap in language model support for Wolof, a language spoken by millions. By providing open-weight models, the authors aim to enhance local deployment capabilities, reducing reliance on potentially costly API-based solutions. This could empower local developers and organizations to create applications that better serve Wolof speakers.
Paper Resources
📖 Reader Mode
~2 min readAuthors:Gwénolé Lecorvé, Delphine Charlet, Pierre Adam, Floriane Behanzin, Lionel Delphin-Poulat, Johannes Heinecke, Géraldine Damnati, Pelin Kaytmaz, Nourredine Khaous, Charbel Kindji, Meriam Maadi, Abdoulaye Mbaye, Claire Perroux, Ismaël Rousseau, Papa Séga Wadé, Gilles Xiberras, Olivier Zhang, Malek Zitouni
Abstract:Wolof is spoken by 23M people, mostly in Senegal, but remains weakly supported by open-weight language models, while API models raise cost, governance, and sovereignty constraints for local deployments. In this paper, we present 0.6B-14B open-weight models adapted to Wolof through instruction tuning, preference alignment, and model merging. After presenting the pipeline and data, extensive experiments on 63~public and in-house tasks show good performance on Wolof understanding and generation while remaining effective on other languages and cross-lingual skills, with the 8B and 14B models being particularly competitive with substantially larger commercial systems. The models are currently being evaluated in internal customer-care pilots and are publicly available at this https URL .
| Comments: | 6 + 47 pages (main + appendices), 27 tables, 1 figure |
| Subjects: | Computation and Language (cs.CL); Artificial Intelligence (cs.AI) |
| Cite as: | arXiv:2610.20824 [cs.CL] |
| (or arXiv:2610.20824v1 [cs.CL] for this version) | |
| https://doi.org/10.48550/arXiv.2610.20824 arXiv-issued DOI via DataCite (pending registration) |
Submission history
From: Gwénolé Lecorvé [view email]
[v1]
Wed, 14 Oct 2026 19:10:23 UTC (215 KB)
— Originally published at arxiv.org
Want this in your inbox every morning?
Daily brief at your local 8am — bilingual EN/中文, free.
More from arXiv cs.CL
See more →The "10th Juror": Open-Set Standpoint Screening for Bureaucratic Bias Detection
MARS-Gov introduces a framework for detecting bureaucratic bias in Dutch government documents, achieving a new state-of-the-art F1 score of 0.880. This model outperforms existing zero-shot detectors by 20.2 points and reduces unnecessary interventions to just 2.5%. The framework's dynamic '10th juror' adapts to emerging biases, enhancing legal language processing.