README.md

---
library_name: transformers
tags: [text-classification, bert, bullying-detection, hate-speech, social-good]
---

# Model Card for Davephoenix/bert-bullying-detector

A BERT-based binary classifier that detects whether a given English text contains bullying content or not. It is fine-tuned for use in moderation tools, education platforms, and social media analysis.

## Model Details

### Model Description

This model is based on `bert-base-uncased` and fine-tuned for binary text classification. The goal is to distinguish between bullying and non-bullying text, providing a tool to support online safety and moderation.

- **Developed by:** Davephoenix
- **Funded by [optional]:** Independent project
- **Shared by [optional]:** Davephoenix
- **Model type:** Text classification (binary)
- **Language(s) (NLP):** English
- **License:** Apache 2.0
- **Finetuned from model [optional]:** bert-base-uncased

### Model Sources [optional]

- **Repository:** [https://huggingface.co/Davephoenix/bert-bullying-detector](https://huggingface.co/Davephoenix/bert-bullying-detector)
- **Demo [optional]:** API in progress

## Uses

### Direct Use

- Used for classifying short- to medium-length English text as "Bullying" or "Not Bullying".
- Can be integrated into moderation tools, educational apps, or awareness platforms.

### Downstream Use [optional]

- As a building block in broader moderation or digital well-being systems.
- Further fine-tuning possible for specific platforms/domains.

### Out-of-Scope Use

- Multilingual or non-English bullying detection.
- Misuse in legal or disciplinary decision-making without human oversight.
- Inference on sarcasm, coded language, or highly contextual text may be unreliable.

## Bias, Risks, and Limitations

The model may exhibit limitations in:

- Cultural or contextual understanding of bullying.
- Identifying subtle or sarcastic forms of harassment.
- False positives in emotionally intense or confrontational but non-abusive language.

### Recommendations

Users (both direct and downstream) should:

- Use the model alongside human review, especially in sensitive domains.
- Avoid deploying in high-stakes environments without thorough testing.
- Consider domain-specific fine-tuning if used outside general English online text.

## How to Get Started with the Model

```python
from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch
import torch.nn.functional as F

model_name = "Davephoenix/bert-bullying-detector"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSequenceClassification.from_pretrained(model_name)

def classify_text(text):
    inputs = tokenizer(text, return_tensors="pt", truncation=True, padding=True)
    with torch.no_grad():
        outputs = model(**inputs)
    probs = F.softmax(outputs.logits, dim=1)
    pred = torch.argmax(probs, dim=1).item()
    return pred, probs[0][pred].item()

label_map = {0: "Not Bullying", 1: "Bullying"}
text = "You are so dumb and nobody likes you."
pred, confidence = classify_text(text)
print(f"Prediction: {label_map[pred]} (Confidence: {confidence:.2f})")
````

## Training Details

### Training Data

* Approximately 20,000 English text samples labeled as "bullying" or "not bullying"
* Balanced dataset curated from public moderation datasets and synthetic augmentation

### Training Procedure

#### Preprocessing \[optional]

* Tokenized using `bert-base-uncased` tokenizer
* Truncation and padding to max\_length of 128 tokens

#### Training Hyperparameters

* **Training regime:** fp16 mixed precision
* **Epochs:** 3
* **Batch size:** 32
* **Optimizer:** AdamW with linear warmup
* **Learning rate:** 2e-5

#### Speeds, Sizes, Times \[optional]

* **Training time:** \~5 hours on Kaggle GPU
* **Model size:** \~420MB
* **Final Checkpoint:** `checkpoint-34371`

## Evaluation

### Testing Data, Factors & Metrics

#### Testing Data

* 10% hold-out split from the training set
* Similar distribution to training data

#### Factors

* Sentence structure
* Presence of explicit abusive terms
* Subtlety of intent

#### Metrics

* Accuracy, F1 score, Loss

### Results

* **Accuracy:** 95.6%
* **F1 Score:** 95.6%
* **Validation Loss:** 0.151

#### Summary

The model performs well for binary classification of bullying vs. non-bullying on general English text. Performance may degrade on ambiguous or culturally nuanced examples.

## Model Examination \[optional]

\[More Information Needed]

## Environmental Impact

Carbon emissions estimated via [ML CO2 calculator](https://mlco2.github.io/impact):

* **Hardware Type:** NVIDIA P100
* **Hours used:** \~5
* **Cloud Provider:** Kaggle
* **Compute Region:** North America
* **Carbon Emitted:** < 2 kg CO₂

## Technical Specifications \[optional]

### Model Architecture and Objective

* **Architecture:** BERT base uncased (12-layer, 768-hidden, 12-heads, 110M parameters)
* **Objective:** Binary sequence classification with cross-entropy loss

### Compute Infrastructure

#### Hardware

* Kaggle P100 GPU (free tier)

#### Software

* `transformers` 4.39.3
* `datasets` 2.19.1
* Python 3.11
* PyTorch 2.x

## Citation \[optional]

**BibTeX:**

```bibtex
@misc{bert-bullying-detector,
  title={BERT Bullying Detector},
  author={Davephoenix},
  year={2025},
  note={Fine-tuned BERT for binary text classification (bullying detection)},
  howpublished={\url{https://huggingface.co/Davephoenix/bert-bullying-detector}}
}
```

**APA:**

Davephoenix. (2025). *BERT Bullying Detector* \[Computer software]. Hugging Face. [https://huggingface.co/Davephoenix/bert-bullying-detector](https://huggingface.co/Davephoenix/bert-bullying-detector)

## Glossary \[optional]

* **BERT:** Bidirectional Encoder Representations from Transformers
* **FP16:** 16-bit floating point precision
* **F1 Score:** Harmonic mean of precision and recall

## More Information \[optional]

To request the training notebook or API wrapper, please contact the model author.

## Model Card Authors \[optional]

* Davephoenix

## Model Card Contact

* [https://huggingface.co/Davephoenix](https://huggingface.co/Davephoenix)

```

Let me know if you'd like this pushed directly to the Hub or edited from the UI.
```