Model Card: Fine-Tuned DistilBERT for Offensive/Hate Speech Detection

Model Description

The Fine-Tuned DistilBERT is a variant of the BERT transformer model, distilled for efficient performance while maintaining high accuracy. It has been adapted and fine-tuned for the specific task of offensive/hate speech detection in text data.

The model, named "distilbert-base-uncased," is pre-trained on a substantial amount of text data, which allows it to capture semantic nuances and contextual information present in natural language text. It has been fine-tuned with meticulous attention to hyperparameter settings, including batch size and learning rate, to ensure optimal model performance for the offensive/hate speech detection task.

During the fine-tuning process, a batch size of 16 for efficient computation and learning was chosen. Additionally, a learning rate (2e-5) was selected to strike a balance between rapid convergence and steady optimization, ensuring the model not only learns quickly but also steadily refines its capabilities throughout training.

This model has been trained on a proprietary dataset < 100k, specifically designed for offensive/hate speech detection. The dataset consists of text samples, each labeled as "non-offensive" or "offensive." The diversity within the dataset allowed the model to learn to identify offensive content accurately. THis dataset was combined from reseearch datasets on this topivc as well as news headliens. All duplicates were removed and meticulous effort was taken to ensure the dataset quality.

The goal of this meticulous training process is to equip the model with the ability to detect offensive and hate speech in text data effectively. The result is a model ready to contribute significantly to content moderation and safety, while maintaining high standards of accuracy and reliability.

Intended Uses & Limitations

Intended Uses

Offensive/Hate Speech Detection: The primary intended use of this model is to detect offensive or hate speech in text data. It is well-suited for filtering and identifying inappropriate content in various applications.
Of Special Note: The data suggests the word "like" is most often used as a comparative statement in the derogatory.
These have numerous instances within the "Offensive Speech Dataset". "You look like X" or "He smells like X" are quite common.
Also of note, the ABSENCE/LACK OF of punctuation lends itself heavily to the "Offensive" dataset.
Accordingly the model will identify these as well, based on their prominence in the training data.

How to Use

To use this model for offensive/hate speech detection, you can follow these steps:

from transformers import pipeline

classifier = pipeline("text-classification", model="Falconsai/offensive_speech_detection")
text = "Your text to classify here."
result = classifier(text)

Limitations

Specialized Task Fine-Tuning: While the model is adept at offensive/hate speech detection, its performance may vary when applied to other natural language processing tasks.
Users interested in employing this model for different tasks should explore fine-tuned versions available in the model hub for optimal results.

Training Data

The model's training data includes a proprietary dataset designed for offensive/hate speech detection. This dataset comprises a diverse collection of text samples, categorized into "non-offensive" and "offensive" classes. The training process aimed to equip the model with the ability to distinguish between offensive and non-offensive content effectively.

Training Stats

Evaluation Loss: 0.018403256312012672
Evaluation Accuracy: 0.9973234886940471
Evaluation Runtime: 85.0789
Evaluation Samples per Second: 127.352
Evaluation Steps per Second: 7.969

Note: Specific evaluation statistics should be provided based on the model's performance.

Responsible Usage

It is essential to use this model responsibly and ethically, adhering to content guidelines and applicable regulations when implementing it in real-world applications, particularly those involving potentially sensitive content.

References

Disclaimer: The model's performance may be influenced by the quality and representativeness of the data it was fine-tuned on. Users are encouraged to assess the model's suitability for their specific applications and datasets.