NAMAA-Space/araseg-e33-gemma4-12b-pa

AraSeg 2026 · e33 — PA ensemble member

One member of NAMAA's PA system for the Arabic Segmentation Shared Task 2026 (AraSeg, ArabicNLP 2026). Second LLM member of the PA logit average, for architecture diversity.

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	PA
Role	member of a logit average of 3 members
System threshold	0.25
Base model	google/gemma-4-12b
Training	LoRA r=16, alpha=32, bf16
Weights	best_*.pt — a PyTorch state_dict, not an HF-format checkpoint
System score (practice test / blind)	94.49 / 94.4 macro-F1
License	gemma, 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-e33-gemma4-12b-pa", "best_PA.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}
}