ArGuard-Qwen3-VL-8B-QLoRA-A1

ArGuard-Qwen3-VL-8B-QLoRA-A1

QLoRA adapter for Qwen3-VL-8B-Instruct, fine-tuned for binary Arabic hateful meme detection (Subtask A1 of ArGuard 2026, Track A). The model receives the meme image, the OCR text, and a task instruction, and generates Hateful or Not Hateful.

This was the NAMAA team's official A1 submission and is the best A1 system in our paper: NAMAA at ArGuard Shared Tasks: Multimodal Fusion, Vision–Language Model Adaptation, and In-Context Prompting for Arabic Hateful Meme Detection (ArabicNLP 2026).

Results

Subtask A1, from Table 1 and Table 3 of the paper:

Comparison on the A1 test set (macro-F1):

Repository contents

Training

Generation is deterministic with at most 8 new tokens. The output is mapped onto the official label vocabulary.

Instruction

Detect the classification of hateful Arabic memes from image and extracted text. The classification labels are only ['Hateful', 'Not Hateful'], return only the label.

Usage

import torch
from PIL import Image
from peft import PeftModel
from transformers import AutoProcessor, Qwen3VLForConditionalGeneration

REPO = "NAMAA-Space/ArGuard-Qwen3-VL-8B-QLoRA-A1"
base = Qwen3VLForConditionalGeneration.from_pretrained(
    "unsloth/Qwen3-VL-8B-Instruct-unsloth-bnb-4bit", torch_dtype=torch.bfloat16, device_map="auto")
model = PeftModel.from_pretrained(base, REPO, subfolder="best_adapter").eval()
processor = AutoProcessor.from_pretrained(REPO, subfolder="best_adapter")

INSTRUCTION = ("Detect the classification of hateful Arabic memes from image and extracted text. "
               "The classification labels are only ['Hateful', 'Not Hateful'], return only the label.")

image = Image.open("meme.jpg").convert("RGB")
ocr_text = "..."  # OCR text of the meme
messages = [{"role": "user", "content": [
    {"type": "image"},
    {"type": "text", "text": f"{INSTRUCTION}\n\nOCR text: {ocr_text}"},
]}]
prompt = processor.apply_chat_template(messages, add_generation_prompt=True)
inputs = processor(images=[image], text=[prompt], return_tensors="pt").to(model.device)
out = model.generate(**inputs, max_new_tokens=8, do_sample=False)
print(processor.decode(out[0, inputs["input_ids"].shape[1]:], skip_special_tokens=True).strip())

Limitations and ethical considerations

• The model was trained with a single random seed on about 3.5k memes. It reflects the ArGuard annotation conventions and the dialects and topics covered by the dataset.
• In our error analysis, most errors come from resolving the target and stance (for example self-referential versus group-directed use of the same term), from nonliteral humour, and from noisy OCR. They are not caused by a missing lexicon.
• The model handles hateful content and can reproduce social, cultural or dialectal biases. Use it to support culturally informed human moderation, not to replace it.
• The shared-task data is for research use under its access terms.

NAMAA at ArGuard 2026 models

Citation

@inproceedings{djamai-etal-2026-namaa,
    title     = "{NAMAA} at {ArGuard} Shared Tasks: Multimodal Fusion, Vision{--}Language Model Adaptation, and In-Context Prompting for {A}rabic Hateful Meme Detection",
    author    = "Djamai, Abdelbasset and Al Jallad, Khloud and Emad Eldin, Fatimah and Nacar, Omer",
    booktitle = "Proceedings of the Fourth Arabic Natural Language Processing Conference: Shared Tasks",
    year      = "2026",
    address   = "Budapest, Hungary",
    publisher = "Association for Computational Linguistics",
}

@inproceedings{alam-etal-2026-arguard,
    title     = "{ArGuard} Shared Task: Harmful Content Detection in {A}rabic Memes and {LLM} Prompts",
    author    = "Alam, Firoj and Biswas, Md. Rafiul and Kmainasi, Mohamed Bayan and Shahroor, Ali Ezzat and Mubarak, Hamdy and Mikros, Georgios and Hasnat, Abul and Zaghouani, Wajdi",
    booktitle = "Proceedings of the Fourth Arabic Natural Language Processing Conference: Shared Tasks",
    year      = "2026",
    address   = "Budapest, Hungary",
    publisher = "Association for Computational Linguistics",
}

Authors: Abdelbasset Djamai, Khloud Al Jallad, Fatimah Emad Eldin, Omer Nacar (NAMAA Community). Contact: [email protected]