This repository contains the reference implementation for the paper on Efficient Detection of Hate Speech and Sexism.

The proposed method uses prefix tuning (PEFT) to adapt a frozen Large Language Model (LLM) backbone for abusive language detection with minimal trainable parameters.
It supports both encoder-style backbones (BERT, RoBERTa) and decoder-style backbones (LLaMA, Llama 2).

• Prefix Tuning (PEFT): lightweight adaptation of frozen LLMs for classification tasks.
• Supports multiple backbones: BERT-base, RoBERTa-large, LLaMA, and Llama 2 with LlamaForSequenceClassification.
• Binary and multi-class classification: covers both offensive/hate speech detection and nuanced sexism taxonomy.
• Evaluation metrics: Accuracy, Macro-F1, confusion matrix.
• Fairness: subgroup evaluation (per-group F1, worst-group F1, F1/FPR gaps) when a group column is present.
• Robustness: adversarial obfuscation (homoglyphs, leetspeak) and Macro-F1 evaluation.
• Calibration: Expected Calibration Error (ECE) and optional temperature scaling for reliable probabilities.
• Optimization: separate learning rates for prefix parameters and classifier head, early stopping on Macro-F1.

Clone the repo and install dependencies:

git clone https://github.com/<your-username>/prefixguard-llms.git cd prefixguard-llms python -m venv .venv && source .venv/bin/activate # Linux/Mac .venv\Scripts\activate # Windows pip install -r requirements.txt

Install requirements.txt:

transformers>=4.41.0 datasets>=2.18.0 peft>=0.10.0 accelerate>=0.30.0 scikit-learn>=1.3.0 scipy>=1.10.0 pandas>=2.0.0 numpy>=1.23.0 torch>=2.1.0 bitsandbytes>=0.43.0 

The code expects datasets in CSV format.

Required columns:

• text: input sentence/post
• label: integer label (0/1 for binary; 0..K-1 for multi-class)

Optional column:

• group: subgroup identifier (e.g., women, immigrants) for fairness evaluation

Example:

text,label,group "She's pretty good for a girl",1,women "She is a great engineer",0,women

python -m src.prefixguard.train_prefixguard \ --train_csv data/edos_train.csv \ --dev_csv data/edos_dev.csv \ --output_dir outputs/edos_roberta_prefix \ --model_name_or_path roberta-base \ --num_labels 2 \ --prefix_tokens 10 \ --per_device_train_batch_size 16 \ --per_device_eval_batch_size 32 \ --learning_rate 5e-5 \ --lr_head 1e-4 \ --lr_prefix 5e-5 \ --max_seq_length 256 \ --num_train_epochs 3 \ --early_stopping_patience 2 \ --seed 42

python -m src.prefixguard.train_prefixguard \ --train_csv data/edos_train.csv \ --dev_csv data/edos_dev.csv \ --output_dir outputs/edos_llama2_7b_prefix \ --model_name_or_path meta-llama/Llama-2-7b-hf \ --num_labels 2 \ --prefix_tokens 10 \ --per_device_train_batch_size 2 \ --per_device_eval_batch_size 4 \ --learning_rate 5e-5 \ --lr_head 1e-4 \ --lr_prefix 5e-5 \ --max_seq_length 256 \ --num_train_epochs 3 \ --early_stopping_patience 2 \ --seed 42

After training, the adapted model is saved in the chosen output_dir.

Evaluate on the test set, including fairness, robustness, and calibration:

python -m src.prefixguard.evaluate_prefixguard \ --test_csv data/edos_test.csv \ --model_name_or_path outputs/edos_roberta_prefix \ --per_device_eval_batch_size 32 \ --max_seq_length 256 \ --calibrate_on_dev_csv data/edos_dev.csv \ --evaluate_obfuscation \ --report_path outputs/edos_roberta_prefix/report.json

Metrics (Accuracy, Macro-F1, subgroup fairness gaps, robustness, calibration) are printed and stored in report.json.

prefixguard-llms/ ├── README.md ├── requirements.txt ├── data/ # place datasets here │ ├── edos_train.csv │ ├── edos_dev.csv │ └── edos_test.csv └── src/ └── prefixguard/ ├── __init__.py ├── utils.py ├── train_prefixguard.py └── evaluate_prefixguard.py 

The paper experiments use:

• EDOS (sexism) – SemEval-2023 Task 10
• OLID (offense) – SemEval-2019 Task 6
• HatEval (hate targeting women or immigrants) – SemEval-2019 Task 5

Prepare CSVs with the above schema and place them in the data/ folder.

If you use this code, please cite our paper:

@inproceedings{.., title={Efficient Detection of Hate Speech and Sexism}, author={...}, booktitle={}, year={2025} }

MIT