---
license: apache-2.0
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
  - split: validation
    path: data/validation-*
  - split: test
    path: data/test-*
dataset_info:
  source_dataset: jigsaw-toxic-comment-classification-challenge
  processed_by: Koushik (https://huggingface.co/datasets/Koushim)
  tokenizer: bert-base-uncased
  label_format: float multi-label binary vector
  label_columns:
    - toxicity
    - severe_toxicity
    - obscene
    - threat
    - insult
    - identity_attack
    - sexual_explicit
  features:
  - name: text
    dtype: string
  - name: toxicity
    dtype: float32
  - name: severe_toxicity
    dtype: float32
  - name: obscene
    dtype: float32
  - name: threat
    dtype: float32
  - name: insult
    dtype: float32
  - name: identity_attack
    dtype: float32
  - name: sexual_explicit
    dtype: float32
  - name: labels
    sequence: float64
  - name: input_ids
    sequence: int32
  - name: token_type_ids
    sequence: int8
  - name: attention_mask
    sequence: int8
  splits:
  - name: train
    num_bytes: 2110899324
    num_examples: 1804874
  - name: validation
    num_bytes: 113965680
    num_examples: 97320
  - name: test
    num_bytes: 113712324
    num_examples: 97320
  download_size: 693905946
  dataset_size: 2338577328
annotations_creators:
  - crowdsourced
language_creators:
  - found
language:
  - en
multilinguality:
  - monolingual
pretty_name: Processed Jigsaw Toxic Comment Classification
tags:
  - text classification
  - toxicity
  - multi-label classification
  - NLP
  - BERT
  - hate speech
size_categories:
  - 1M<n<10M
task_categories:
  - text-classification
task_ids:
  - multi-label-classification
---

# Processed Jigsaw Toxic Comments Dataset

This is a **preprocessed and tokenized** version of the original [Jigsaw Toxic Comment Classification Challenge](https://www.kaggle.com/competitions/jigsaw-toxic-comment-classification-challenge) dataset, prepared for **multi-label toxicity classification** using transformer-based models like BERT.

⚠️ **Important Note**: I am **not the original creator** of the dataset. This dataset is a cleaned and restructured version made for quick use in PyTorch deep learning models.

---

## 📦 Dataset Features

Each example contains:

- `text`: The original user comment
- `labels`: A list of 7 binary float values indicating presence of toxicity categories
- `input_ids`, `attention_mask`: Tokenized fields using `bert-base-uncased` (max length 128)

### Toxicity Categories:

1. `toxicity`
2. `severe_toxicity`
3. `obscene`
4. `threat`
5. `insult`
6. `identity_attack`
7. `sexual_explicit`

---

## 🧪 Dataset Splits

| Split       | # Examples  |
|-------------|-------------|
| Train       | ~1.8M       |
| Validation  | ~97K        |
| Test        | ~97K        |

---

## 🔧 Processing Details

1. **Original Source**: Manually downloaded from [Kaggle](https://www.kaggle.com/c/jigsaw-toxic-comment-classification-challenge)
2. **Preprocessing**:
   - Combined multiple toxicity columns into a single `labels` vector
   - Converted label values to floats (0.0 or 1.0)
3. **Tokenization**:
   - Used Hugging Face `bert-base-uncased` tokenizer
   - Applied padding and truncation to max length of 128
4. **Formatting**:
   - Final dataset set to return PyTorch `input_ids`, `attention_mask`, and `labels`

---

## 💡 Usage Example

```python
from datasets import load_dataset

dataset = load_dataset("Koushim/processed-jigsaw-toxic-comments")

from torch.utils.data import DataLoader

train_loader = DataLoader(dataset["train"], batch_size=32, shuffle=True)

batch = next(iter(train_loader))
print(batch['input_ids'].shape)  # torch.Size([32, 128])
print(batch['labels'].shape)     # torch.Size([32, 7])
````

---

## 📚 Citation

If you use this dataset, please cite the original Jigsaw authors:

```bibtex
@misc{jigsawtoxic,
  title={Toxic Comment Classification Challenge},
  author={Jigsaw and Google},
  year={2018},
  url={https://www.kaggle.com/c/jigsaw-toxic-comment-classification-challenge}
}
```

---

## 🙏 Acknowledgements

* Original dataset by **Jigsaw/Google**
* Processing, formatting, and tokenization by [Koushik](https://huggingface.co/koushik)