NAMAA-Space/araseg-sat-ft

AraSeg 2026 · sat_ft — NP, NoPnx-PA and NoPnx-NP ensemble member

One member of NAMAA's NP, NoPnx-PA and NoPnx-NP system for the Arabic Segmentation Shared Task 2026 (AraSeg, ArabicNLP 2026). Fine-tuned Segment-any-Text. Char-level, so genuinely diverse from the XLM-R/LLM pool -- it earns the single largest weight in the NP stack. The unused PA/ checkpoint is not shipped.

This is not a standalone segmenter. It is one voter inside an ensemble, and it produces uncalibrated per-word boundary probabilities. Used alone it does not reproduce any published score. The system that does is NAMAA-Space/araseg-2026, which holds the combiner weights and thresholds.

Subtask	NP, NoPnx-PA and NoPnx-NP
Role	member of a OOF-fitted linear stack over 5 members
System threshold	0.36
Base model	segment-any-text/sat-12l-sm
Training	char-level, full fine-tune
Weights	best_*.pt — a PyTorch state_dict, not an HF-format checkpoint
System score (practice test / blind)	92.84 / 91.3 macro-F1
License	mit, inherited from the base model

Loading

from_pretrained will not work. The file is a bare state_dict; the architecture is built from the experiment's config YAML and the base model, then the weights are loaded in:

import torch
from huggingface_hub import hf_hub_download

path = hf_hub_download("NAMAA-Space/araseg-sat-ft", "best_NP.pt")
state = torch.load(path, map_location="cpu")
# build the architecture first -- see ensemble.py / verify_offcluster.py in the repo

The five LoRA members additionally need transformers==5.12.1 to instantiate their base classes. Full pinned stack: requirements-llm.txt in the code repo.

Reproducing the system

Code, configs and the full member→subtask map: 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}
}