francesco-zatto/twitter-roberta-base-hate-freeze-all-weighted-L-sexism-detector

RoBERTa Sexism Classifier (Linear Probing / Freeze All)

This model is a fine-tuned version of cardiffnlp/twitter-roberta-base-hate, 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 RoBERTa model (model.roberta.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: 0. - (Non-sexist)

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

Preprocessing

Inputs must be preprocessed to match the CardiffNLP base model formatting. Note that this model is case-sensitive, so do not aggressively lowercase your text if you want to preserve capitalization signals:

• Replace user mentions (@user) with the token @user
• Replace URLs with the token http

Evaluation Results (Test Set)

• Macro F1:  0.4752
• Precision: 0.4680
• Recall: 0.5108

How to Use

from transformers import AutoTokenizer, AutoModelForSequenceClassification

repo_id = "francesco-zatto/twitter-roberta-base-hate-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)