NAMAA-Space/araseg-2026

AraSeg 2026 — NAMAA system hub

The four segmentation systems NAMAA submitted to the Arabic Segmentation Shared Task 2026 (AraSeg, ArabicNLP 2026), covering every closed-track subtask. Start here: this repo holds what turns the member checkpoints into the submitted systems.

Subtask	Combiner	Thr.	Development F1	Blind F1
PA	logit average of 3 members	0.25	94.49	94.4
NP	OOF-fitted linear stack over 5 members	0.36	92.84	91.3
NoPnx-PA	OOF-fitted MEMM structural decoder over 8 members	0.46	87.82	89.9
NoPnx-NP	OOF-fitted MEMM structural decoder over 7 members	0.34	86.49	87.0

What is in this repo

Path	Contents
heads/decoder_nopnx_pa/	MEMM decoder weights + threshold, 8 members
heads/decoder_nopnx_np/	MEMM decoder weights + threshold, 7 members
heads/oof_stack_np/	NP linear stacker weights + threshold
oof/	frozen out-of-fold TRAIN matrices, 105 member JSONs
oof_sat/	out-of-fold TRAIN matrices for the SaT member

PA needs no fitted head — it is a plain logit average at threshold 0.25.

The oof/ caches cannot be regenerated. The five-fold models they came from no longer exist, so they are the only way to refit the decoders and stacker. They include the e75 folds, which are not part of any lock: the 15-subset gate that selected the final NoPnx-PA membership cannot be reproduced without them.

Members

Nineteen member repos, listed by subtask in the collection. e38 and sat-ft are each members of three of the four systems.

External dependencies

Not redistributed here, and the systems do not run without them:

Model	Needed by
FacebookAI/xlm-roberta-large	9 members
Qwen/Qwen3.5-9B	e32, e40, e41, e42
google/gemma-4-12b	e33 — note Gemma's own terms, not Apache-2.0
MostafaMaroof/Naqta	e69, e70 (base encoder) and e76 at inference time
Segment-any-Text base	sat_ft

e76 is the one to watch: it was trained on text with Naqta-predicted commas inserted, so reproducing NoPnx-PA means running Naqta as a preprocessing step at min_p=0.3 with a Latin , as the comma character.

Reproducing

Code, configs, and the run recipe: https://github.com/NAMAA-ORG/NAMAA-Community-AraSeg-2026

Citation

@inproceedings{namaa2026araseg,
  title     = {NAMAA at Arabic Segmentation Shared Task 2026},
  author    = {NAMAA Community},
  booktitle = {Proceedings of ArabicNLP 2026},
  year      = {2026}
}