Frozen Embeddings: Only the base embedding layer (model.bert.embeddings.parameters()) was frozen during training. All subsequent transformer encoder layers and the classification head were fully fine-tuned. This allows the model to learn task-specific attention patterns while anchoring the fundamental word representations.
Standard Loss: Training was conducted using a standard (unweighted) Cross-Entropy loss function to establish a baseline for how the architecture handles the natural class imbalances present in the EXIST dataset.

Intended Use

Categorizes English tweets into one of four sexist intentions: 0. - (Non-sexist)

DIRECT (Directly sexist messages)
JUDGEMENTAL (Messages condemning sexist behaviors)
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: 0.4391
Precision: 0.4461
Recall: 0.5500

How to Use

python

from transformers import AutoTokenizer, AutoModelForSequenceClassification

repo_id = "francesco-zatto/hateBERT-freeze-embeddings-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)

francesco-zatto/hateBERT-freeze-embeddings-sexism-detector

HateBERT Sexism Classifier (Freeze Embeddings / Unweighted Loss)

Experiment Details: freeze_embeddings

Intended Use

Preprocessing

Evaluation Results (Test Set)

How to Use

francesco-zatto/
hateBERT-freeze-embeddings-sexism-detector

Experiment Details: `freeze_embeddings`