deberta-ai-detector-v3

DeBERTa-v3-Large AI vs Human Text Detector

This model is fine-tuned to classify whether a given piece of text was written by a human or generated by Artificial Intelligence (AI). It is based on the powerful microsoft/deberta-v3-large architecture, making it highly accurate in detecting synthetic text across various domains.

Model Details

Model Description

This model was trained to solve the growing need for academic integrity and content authenticity in the era of generative AI. By leveraging the advanced contextual understanding of DeBERTa-v3-large, the model distinguishes between human-written nuances and AI-generated patterns.

• Developed by: Md Rakib Ali & Tushar Imran
• Model type: Text Classification (Sequence Classification)
• Language(s) (NLP): English
• License: MIT
• Finetuned from model: microsoft/deberta-v3-large

Model Sources

• Base Model: microsoft/deberta-v3-large
• Dataset: acmc/multi_domain_ai_human_text

Uses

Direct Use

The model can be used directly for identifying AI-generated text. It is highly suitable for:

• Academic institutions (detecting AI-generated essays or assignments).
• Publishers and editors (verifying the authenticity of submissions).
• General users wanting to verify if a text is human-written or machine-generated.

Out-of-Scope Use

• The model should not be used as the sole evidence to penalize students or professionals without human review.
• It may not perform optimally on heavily edited AI text or languages other than English.

How to Get Started with the Model

Use the code below to get started with the model using the transformers library:

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

model_name = "YOUR_USERNAME/YOUR_MODEL_NAME"  # Replace with your actual repo name
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSequenceClassification.from_pretrained(model_name)

text = "In the era of artificial intelligence, take-home writing assignments have become difficult to police."
inputs = tokenizer(text, return_tensors="pt", truncation=True, max_length=256)

with torch.no_grad():
    logits = model(**inputs).logits

probs = torch.nn.functional.softmax(logits, dim=-1)
predicted_class = torch.argmax(probs).item()

# Assuming 0: Human, 1: AI (Adjust based on your label mapping)
labels = ["Human", "AI"]
print(f"Prediction: {labels[predicted_class]}")
print(f"Probabilities: {probs.tolist()}")

Training Details

Training Data

The model was fine-tuned using the acmc/multi_domain_ai_human_text dataset. This dataset contains a diverse mix of human-written and AI-generated texts from various domains, ensuring the model generalizes well across different writing styles.

Training Results (Summary)

The model achieved highly robust results across 3 epochs, showing strong generalization capabilities with an evaluation accuracy of 95.56%.

---

library_name: transformers tags:

• text-classification
• ai-detection
• deberta-v3
• pytorch datasets:
• acmc/multi_domain_ai_human_text

---

Technical Specifications

Model Architecture and Objective

• Architecture: DeBERTa-v3-Large
• Objective: Binary sequence classification (Human vs. AI).

Compute Infrastructure

The model was fine-tuned using a high-performance cloud GPU environment.

Hardware

• Cloud Provider: RunPod
• GPU: 1x NVIDIA A40
• VRAM: 48 GB GDDR6
• System RAM: 64 GB

Software

• Frameworks: PyTorch, Transformers (Hugging Face)
• Environment: CUDA-enabled environment

Bias, Risks, and Limitations

• False Positives: Like all AI detectors, this model may occasionally flag human-written text as AI.
• Recommendations: Always use the model's prediction as a signal rather than an absolute verdict. Human judgment should always be part of the final decision-making process.