Arabic stance-detection classifier for StanceEval-2026, an ArabicNLP 2026 (@ EMNLP 2026, Budapest) shared task. Given an Arabic tweet and a target topic, the model predicts whether the tweet's author is in Favor of, Against, or has None (neutral/irrelevant) stance toward the target.

Track 1: Seen Targets

Trained and evaluated on targets that appear in the training data: COVID-19 Vaccine, Digital Transformation, Women Empowerment. Train 3,502 / Dev 619 tweets (Mawqif-v2).

Base model

humain-ai/ALLaM-7B-Instruct-preview

Labels

ID	Label	Meaning
0	`Favor`	Tweet supports the target
1	`Against`	Tweet opposes the target
2	`None`	Neutral, irrelevant, or unclear stance

Dev set result

No held-out dev score for this checkpoint, by design: this is a final refit of the t1_s10_allam7b_lora_prompt recipe, retrained on the combined train+dev data immediately before the official test submission (dev Favg2 for the pre-refit checkpoint: 86.06). Trading the held-out check for more training data is standard practice right before a competition test submission.

Usage

python

from transformers import AutoTokenizer, AutoModelForSequenceClassification
from peft import PeftModel
import torch

base_model_id = "humain-ai/ALLaM-7B-Instruct-preview"
adapter_id = "HassanB4/t1_s11_allam7b_lora_prompt_final"

id2label = {0: "Favor", 1: "Against", 2: "None"}

tokenizer = AutoTokenizer.from_pretrained(adapter_id)
base_model = AutoModelForSequenceClassification.from_pretrained(base_model_id, num_labels=3)
model = PeftModel.from_pretrained(base_model, adapter_id)
model.eval()

text = "..."
target = "..."
prompt = f"{text} [SEP] {target}"
inputs = tokenizer(prompt, return_tensors="pt", truncation=True, max_length=512)

with torch.no_grad():
    logits = model(**inputs).logits

predicted_label = id2label[int(torch.argmax(logits, dim=-1)[0])]
print(predicted_label)

Status

Part of the NAMAA Community StanceEval-2026 submission (Track 1). A system description paper is in preparation; citation details will be added once available.