Records · kv:2610.00041v1

Seben: Morphotactic-Guided Preference Alignment for Extremely Low-Resource Language

Sebastien Diarra, Mamadou K. Keita, Sokona Coulibaly

Submitted by s7d11 (RobotsMali) @s7d11

Published 2026-10-13

Official implementation · Research paper · Machine learning

Abstract

Preference-based post-training improves language models, but the standard methods use surface tokens as the training signal and target high-resource languages. We present Sɛbɛn, a post-training framework that adds morphotactic verification to the alignment loop for small language models in extremely low-resource languages. A rule-based morphological analyzer checks each generated analysis against a per-language grammar and dictionary, and a morpheme-level reward turns the check into a training signal for supervised fine-tuning, DPO, APO, and GRPO. The framework also gates automatic grammar and dictionary updates behind an integrity threshold. We run Sɛbɛn on the 13 national languages of Mali with five small models from 135M to 2B parameters and report morphological integrity, morpheme error rate, consistency, and an uncertainty-weighted cost on a seed-fixed held-out split. GRPO gives the lowest morpheme error rate for SmolLM2-135M, Gemma-4-E2B, and TinyLlama-1.1B, while the four arms tie for LFM2.5-230M. We record and seal every run with K-Veritas, so each reported score is bound to the computation that produced it and can be verified publicly. We report the full configuration and evaluation records so that others can reproduce the matrix.

Sealed reports

Cite this record

@misc{kveritas261000041v1,
  title={Seben: Morphotactic-Guided Preference Alignment for Extremely Low-Resource Language},
  author={Sebastien Diarra and Mamadou K. Keita and Sokona Coulibaly},
  year={2026},
  howpublished={K-Veritas Records},
  note={kv:2610.00041v1},
  url={https://kveritas.org/records/2610.00041v1},
}

References

1. [1] Mamadou K. Keita and Christopher Homan. Computer Science Conferences Should Require Nonrepudiable Experimental Results. NeurIPS 2026 Position Paper Track. arXiv:2605.08586.

Sealed with K-Veritas [1]. Listed, not endorsed. A record shows these results came from the sealed code, unchanged. Whether the work is sound is for the reader to judge.