Records · kv:2610.00001v2

Sebeni morphotactic post-training on MULTI13: SFT, DPO, APO and GRPO arms with SmolLM2-135M

Sebastien Diarra

Submitted by Sebastien Diarra @s7d11

Published 2026-10-10

Official implementation · Research paper · Natural language processing

Code

Changes in version 2: Added the GRPO arm (r4), run with the same commit, settings and K-Veritas capture as the SFT, DPO and APO arms. One difference: GRPO evaluates the held-out split once, at step 500, instead of every 50 steps, because one evaluation takes about 3 hours on this GPU. Like the other arms, sebeni eval reports checkpoint_id baseline, so the test scores match the other arms.

Abstract

Three post-training arms of Sebeni, a morphotactic generation toolkit for extremely low-resource Manding languages, on the MULTI13 scope (13 languages). Each arm distils the grammar and dictionary, fine-tunes SmolLM2-135M with 4-bit LoRA for 500 steps on the packaged 300-sample dataset with a 10% held-out split, and scores the held-out test set (phi, MER, MCS, UWEC). Code: github.com/mlsftwrs/sebeni at commit 8e66775, with K-Veritas capture added to the CLI.

Sealed reports

Cite this record

@misc{kveritas261000001v2,
  title={Sebeni morphotactic post-training on MULTI13: SFT, DPO, APO and GRPO arms with SmolLM2-135M},
  author={Sebastien Diarra},
  year={2026},
  howpublished={K-Veritas Records},
  note={kv:2610.00001v2},
  url={https://kveritas.org/records/2610.00001v2},
}

Versions

• kv:2610.00001v1
• kv:2610.00001v2

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.