---
license: cc-by-sa-4.0
tags:
  - hate-speech
  - toxic-comments
  - classification
  - hatebert
  - jigsaw
  - fine-tuned
base_model: GroNLP/hateBERT
datasets:
  - jigsaw-toxic-comment-classification-challenge
metrics:
  - accuracy
  - f1
---

# HateBERT Fine-Tuned on Jigsaw Toxic Comments (v5)

This model is a fine-tuned version of [GroNLP/hateBERT](https://huggingface.co/GroNLP/hateBERT) on a binary version of the [Jigsaw Toxic Comment Classification Challenge](https://www.kaggle.com/c/jigsaw-toxic-comment-classification-challenge) dataset.

It has been fine-tuned to detect whether a comment is toxic (`1`) or non-toxic (`0`) using class-weighted Focal Loss and evaluation strategies suitable for imbalanced classification tasks.

## 💻 Training Setup

- **Base Model:** GroNLP/hateBERT
- **Dataset:** Jigsaw Toxic Comment Classification Challenge
- **Binary Labeling:** A comment is marked as *toxic* if any of the following labels is `1`: `toxic`, `severe_toxic`, `obscene`, `threat`, `insult`, `identity_hate`
- **Tokenizer Max Length:** 256
- **Loss Function:** Focal Loss with class weights
- **Hardware:** NVIDIA H100 GPU (via SLURM on TU Berlin HPC)
- **Training Time:** ~6 hours
- **Final F1 Score (Validation):** `0.850`

## 📊 Evaluation Metrics

| Metric   | Value  |
|----------|--------|
| F1 Score | 0.850  |
| Accuracy | ~0.84  |
| Confusion Matrix & PR Curves | [Saved and visualized during training] |

## 🧪 How to Use

```python
from transformers import AutoModelForSequenceClassification, AutoTokenizer, pipeline

model = AutoModelForSequenceClassification.from_pretrained("Jensvollends/hatebert-finetuned_v5")
tokenizer = AutoTokenizer.from_pretrained("Jensvollends/hatebert-finetuned_v5")

pipe = pipeline("text-classification", model=model, tokenizer=tokenizer, top_k=None)

text = "You are a kind person"
result = pipe(text)
print(result)