francesco-zatto/hateBERT-freeze-all-weighted-L-sexism-detector

HateBERT Sexism Classifier (Linear Probing / Freeze All)

This model is a fine-tuned version of GroNLP/hateBERT, trained for multi-class sexism detection on the EXIST 2023 Task 2 dataset.

Experiment Details: freeze_all

This repository contains the Linear Probing variant of our ablation study.

• —Frozen Backbone: All parameters in the base BERT model (model.bert.parameters()) were frozen during training. Only the final classification head was trained. This approach protects the pre-trained weights from catastrophic forgetting and speeds up training.
• —Weighted Loss: Because the EXIST 2023 dataset contains class imbalances, training was conducted using a weighted Cross-Entropy loss function. This ensures the model does not become heavily biased toward the majority class (e.g., Non-sexist) and adequately penalizes errors on the minority classes.

Intended Use

Categorizes English tweets into one of four sexist intentions:

1. 1.- (Non-sexist)
2. 2.DIRECT (Directly sexist messages)
3. 3.JUDGEMENTAL (Messages condemning sexist behaviors)
4. 4.REPORTED (Messages reporting a sexist situation)

Preprocessing

Because HateBERT is built on bert-base-uncased, it automatically lowercases text. If you are keeping preprocessing consistent across your ablation study, ensure your inputs are cleaned accordingly:

• —Replace user mentions (@user) with the token @user
• —Replace URLs with the token http
• —(Handled by tokenizer) Lowercase all text

Evaluation Results (Test Set)

• —Macro F1: 0.3198
• —Precision: 0.3430
• —Recall: 0.3173

How to Use

from transformers import AutoTokenizer, AutoModelForSequenceClassification

repo_id = "francesco-zatto/hateBERT-freeze-all-weighted-L-sexism-detector"
tokenizer = AutoTokenizer.from_pretrained(repo_id)
model = AutoModelForSequenceClassification.from_pretrained(repo_id)

inputs = tokenizer("Your cleaned tweet text here", return_tensors="pt")
outputs = model(**inputs)