StanceEval-2026 — Track 2 (Unseen Targets): ALLaM-7B-Instruct-preview (LoRA)

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 2: Unseen Targets

Trained on the Track 2 training pool, evaluated on target topics not seen during training (cross-target generalization). Train 2,721 / Dev 1,400 tweets (Mawqif-v2).

Base model

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

Labels

Dev set result

85.33 Favg2 (dev set), the shared task's primary metric (macro-F1 over Favor + Against, None excluded).

Usage

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

base_model_id = "humain-ai/ALLaM-7B-Instruct-preview"
adapter_id = "HassanB4/t2_s8_allam7b_lora_prompt"

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 2). A system description paper is in preparation; citation details will be added once available.