Robust Adaptation of Large Multimodal Models for Retrieval Augmented Hateful Meme Detection

This repository contains a PEFT fine-tuned Large Multimodal Model (LMM) for hateful meme detection, as presented in the paper Robust Adaptation of Large Multimodal Models for Retrieval Augmented Hateful Meme Detection.

Model Details

Model Description

Hateful memes have become a significant concern on the Internet, necessitating robust automated detection systems. While Large Multimodal Models (LMMs) have shown promise in hateful meme detection, they face notable challenges like sub-optimal performance and limited out-of-domain generalization capabilities. Recent studies further reveal the limitations of both supervised fine-tuning (SFT) and in-context learning when applied to LMMs in this setting. To address these issues, this work proposes a robust adaptation framework for hateful meme detection that enhances in-domain accuracy and cross-domain generalization while preserving the general vision-language capabilities of LMMs. Analysis reveals that this approach achieves improved robustness under adversarial attacks compared to SFT models. Experiments on six meme classification datasets show that this approach achieves state-of-the-art performance, outperforming larger agentic systems. Moreover, the method generates higher-quality rationales for explaining hateful content compared to standard SFT, enhancing model interpretability.

Developed by: Jingbiao Mei, Jinghong Chen, Guangyu Yang, Weizhe Lin, Bill Byrne
Model type: Large Multimodal Model (LMM), fine-tuned using PEFT (LoRA) for hateful meme detection.
Language(s) (NLP): English
License: cc-by-4.0
Finetuned from model: QWen/QWen2-VL-7B-Instruct

Model Sources

Paper: https://huggingface.co/papers/2502.13061

Uses

Direct Use

This model is intended for the robust detection of hateful memes. It can be used to classify multimodal content (image and text) for hate speech, offering improved accuracy and cross-domain generalization. It also provides rationales for its classifications, aiding interpretability.

Out-of-Scope Use

This model should not be used for generating hateful content, propagating misinformation, or any other malicious purposes. It is a detection tool and its application should align with ethical AI principles for combating harmful online content.

Bias, Risks, and Limitations

All AI models, especially those dealing with sensitive content like hate speech, may exhibit biases from their training data or limitations in understanding complex nuances, sarcasm, or evolving slang. This could lead to misclassifications or biased explanations.

Recommendations

Users (both direct and downstream) should be made aware of the risks, biases, and limitations of the model. Continuous monitoring, human oversight in critical applications, and further evaluation on diverse and evolving datasets are recommended.

How to Get Started with the Model

This model is a PEFT (Parameter-Efficient Fine-Tuning) adapter built on top of QWen/QWen2-VL-7B-Instruct. To use it, you typically load the base model and its tokenizer using the Hugging Face Transformers library, and then load this model as a PEFT adapter.

Training Details

Training Data

The model was trained and evaluated on six meme classification datasets, as mentioned in the paper.

Training Procedure

The paper proposes a robust adaptation framework that enhances in-domain accuracy and cross-domain generalization while preserving the general vision-language capabilities of LMMs. The training involved fine-tuning using PEFT (LoRA).

Training Hyperparameters

Training regime: Based on the base model, Qwen2-VL-7B-Instruct often uses mixed precision (e.g., bfloat16).
Epochs: 3
Train Batch Size: 16

Evaluation

Testing Data, Factors & Metrics

Testing Data

The model was evaluated on six meme classification datasets.

Metrics

The evaluation metrics recorded during training include:

accuracy
auroc (Area Under the Receiver Operating Characteristic curve)
f1 (F1 Score)
precision
recall

Results

The proposed approach achieved state-of-the-art performance across six meme classification datasets, outperforming larger agentic systems. It also demonstrated improved robustness under adversarial attacks and generated higher-quality rationales compared to standard SFT models.

Citation

If you find this work helpful, please consider citing the following papers:

@inproceedings{RGCL2024Mei,
    title = "Improving Hateful Meme Detection through Retrieval-Guided Contrastive Learning",
    author = "Mei, Jingbiao  and
      Chen, Jinghong  and
      Lin, Weizhe  and
      Byrne, Bill  and
      Tomalin, Marcus",
    editor = "Ku, Lun-Wei  and
      Martins, Andre  and
      Srikumar, Vivek",
    booktitle = "Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)",
    month = aug,
    year = "2024",
    address = "Bangkok, Thailand",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2024.acl-long.291",
    doi = "10.18653/v1/2024.acl-long.291",
    pages = "5333--5347"
}

 @article{RAHMD2025Mei, title={Robust Adaptation of Large Multimodal Models for Retrieval Augmented Hateful Meme Detection},
          url={http://arxiv.org/abs/2502.13061},
          DOI={10.48550/arXiv.2502.13061},
          note={arXiv:2502.13061 [cs]},
          number={arXiv:2502.13061},
          publisher={arXiv},
          author={Mei, Jingbiao and Chen, Jinghong and Yang, Guangyu and Lin, Weizhe and Byrne, Bill},
          year={2025},
          month=may }

Framework versions

PEFT 0.12.0

Downloads last month: -; Downloads are not tracked for this model. How to track