🧠 AI Text Detector v1.0 (DeBERTa-v3-large)
🏷️ Model Details
| Field | Description |
|---|
| Model Name | ai-text-detector-v-n4.0 |
| Base Model | microsoft/deberta-v3-large |
| Task | Text Classification (Human-written vs AI-generated) |
| Language | English |
| Framework | PyTorch, Transformers |
| Trained by | Abhinav |
| Fine-tuned using | Hugging Face Trainer API with early stopping, mixed precision (fp16), and F1 optimization. |
📖 Model Description
This model fine-tunes DeBERTa-v3-large for detecting whether a given text is written by a Human or generated by an AI.
It was trained on a custom dataset containing 10,000+ samples of diverse text across multiple topics, labeled as:
0 → Human-written text1 → AI-generated text
The goal is to identify subtle linguistic differences and stylistic cues between natural human writing and machine-generated content.
⚙️ Training Configuration
| Parameter | Value |
|---|
| Epochs | 4 |
| Batch size | 8 |
| Learning Rate | 2e-5 |
| Max Sequence Length | 256 |
| Optimizer | AdamW |
| Scheduler | Linear decay |
| Weight Decay | 0.01 |
| Seed | 42 |
| Mixed Precision | ✅ Yes (fp16) |
| Gradient Accumulation Steps | 2 |
| Frameworks | PyTorch, Transformers, Datasets |
🧾 Dataset
| Field | Value |
|---|
| Source | Custom dataset (gpt_5_with_10k.csv) |
| Columns | text, label |
| Labels | 0 = Human, 1 = AI |
| Split | 90% Train / 10% Test |
| Cleaning | Removed special characters, normalized whitespace, and standardized punctuation. |
📊 Evaluation Results
| Metric | Score |
|---|
| Accuracy | ~0.97 |
| F1 Score | ~0.97 |
| Precision / Recall | Balanced |
Confusion Matrix Example:
[[4800 90] → True Human
[ 110 5000]] → True AI
🔍 Usage Example
from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch
model_id = "abhi099k/ai-text-detector-v-n4.0"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForSequenceClassification.from_pretrained(model_id)
text = "This article explores the evolution of large language models..."
inputs = tokenizer(text, return_tensors="pt", truncation=True, padding=True)
outputs = model(**inputs)
pred = torch.argmax(outputs.logits, dim=1).item()
print("🧑 Human" if pred == 0 else "🤖 AI")
📈 Intended Use
- Detect AI-generated content for moderation, academic integrity, or authenticity verification.
- Use as a foundation model for fine-tuning on domain-specific datasets (e.g., essays, reviews, research papers).
⚠️ Limitations
- May misclassify paraphrased AI text or human text with robotic phrasing.
- Primarily trained on English — not guaranteed for other languages.
- Should not be used for punitive or high-stakes decisions without human review.
🏆 Future Improvements
- Multi-language support (Hindi, Spanish, etc.)
- Add stylistic embeddings for cross-model generalization.
- Robustness testing against prompt-engineering and obfuscation.
🧩 Technical Summary
| Component | Library |
|---|
| Tokenization | AutoTokenizer |
| Model | AutoModelForSequenceClassification |
| Trainer | transformers.Trainer |
| Metrics | evaluate (accuracy, f1) |
| Visualization | matplotlib (confusion matrix) |
📬 Citation
If you use this model, please cite:
@model{abhinav_ai_text_detector_v1,
title = {AI Text Detector v1.0 – DeBERTa-v3-large Fine-tune},
author = {Abhinav},
year = {2025},
url = {https://huggingface.co/Abhinav/ai-text-detector-v-n4.0}
}