Model Card for roberta-large-openai-detector-custom
This model detects AI-generated vs human-written text using a fine-tuned RoBERTa-Large architecture trained on modern LLM outputs.
Model Details
Model Description
This model is a binary text classifier trained to identify AI-generated content from models such as GPT-4, GPT-3.5, Claude, and LLaMA. It improves over legacy GPT-2 detectors by adapting to modern generative patterns.
- Developed by: Daksh Thakuria
- Model type: Transformer-based sequence classification (RoBERTa-Large)
- Language(s): English
- License: Apache 2.0
- Finetuned from model: Community RoBERTa GPT-2 Detector
Model Sources
- Repository: https://huggingface.co/silentone0725/roberta-large-openai-detector-custom
- Training Code: https://github.com/silentone12725/Ai-Gen-Text-Detect
Uses
Direct Use
Detecting AI-generated text in research, moderation, and academic integrity systems.
Downstream Use
Integration into content filtering pipelines, analytics tools, or research benchmarks.
Out-of-Scope Use
- Legal/forensic authorship claims
- Fully automated high-stakes decisions
- Guaranteed detection after heavy paraphrasing
Bias, Risks, and Limitations
- May misclassify creative or structured human writing
- Performance drops under heavy paraphrasing
- English-focused
- Surface-text detector (no watermarking)
Recommendations
Use as a decision-support tool, not a final authority.
How to Get Started
from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch
model_name = "silentone0725/roberta-large-openai-detector-custom"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSequenceClassification.from_pretrained(model_name)
text = "Sample text to evaluate"
inputs = tokenizer(text, return_tensors="pt", truncation=True, padding=True, max_length=512)
outputs = model(**inputs)
prediction = torch.argmax(outputs.logits, dim=1)
print("AI-generated" if prediction.item() == 1 else "Human-written")
Training Details
Training Data
Dataset: https://huggingface.co/datasets/silentone0725/ai-human-text-detection-v1
Contains human text + GPT-4, GPT-3.5, Claude, LLaMA outputs.
Training Procedure
Fine-tuned on Google Colab GPUs using PyTorch + HuggingFace Transformers.
Training Hyperparameters
- Learning rate: 2e-5
- Batch size: 8 (effective 16)
- Epochs: 6
- Mixed precision: FP16
- Weight decay: 0.2
- Dropout: 0.3
Evaluation
Metrics
| Metric | Score |
|---|---|
| Accuracy | 0.5904 |
| Precision | 0.5087 |
| Recall | 0.7524 |
| F1 Score | 0.6070 |
| AUC | 0.690 |
Environmental Impact
- Hardware Type: NVIDIA T4 / A100
- Cloud Provider: Google Colab
- Compute Region: Global (Colab infrastructure)
Technical Specifications
Architecture
RoBERTa-Large transformer with classification head.
Software
PyTorch, Transformers, scikit-learn.
Citation
APA: Thakuria, D. (2026). AI-Generated Text Detection via Fine-Tuned RoBERTa-Large.
Model Card Authors
Daksh Thakuria
Model Card Contact
Via Hugging Face profile.
- Downloads last month
- 90