đ answerdotai-ModernBERT-base-ai-detector
This model is a fine - tuned version of answerdotai/ModernBERT-base on the AI vs Human Text Classification dataset, DAIGT V2 Train Dataset. It can effectively distinguish between AI - generated and human - written texts, with a validation loss of 0.0036
on the evaluation set.
đ Quick Start
This model is a fine - tuned version of answerdotai/ModernBERT-base on the AI vs Human Text Classification dataset, DAIGT V2 Train Dataset.
It achieves the following results on the evaluation set:
⨠Features
đ Model Description
This model is based on ModernBERT-base, a lightweight and efficient BERT - based model. It has been fine - tuned for AI - generated vs Human - written text classification, allowing it to distinguish between texts written by AI models (ChatGPT, DeepSeek, Claude, etc.) and human authors.
đ¯ Intended Uses & Limitations
â
Intended Uses
- AI - generated content detection (e.g., ChatGPT, Claude, DeepSeek).
- Text classification for distinguishing human vs AI - generated content.
- Educational & Research applications for AI - content detection.
â ī¸ Limitations
- Not 100% accurate â Some AI texts may resemble human writing and vice versa.
- Limited to trained dataset scope â May struggle with out - of - domain text.
- Bias risks â If the dataset contains bias, the model may inherit it.
đĻ Installation
No specific installation steps are provided in the original document.
đģ Usage Examples
Basic Usage
from transformers import AutoModelForSequenceClassification, AutoTokenizer, pipeline
model_name = "answerdotai/ModernBERT-base-ai-detector"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSequenceClassification.from_pretrained(model_name)
classifier = pipeline("text-classification", model=model, tokenizer=tokenizer)
text = "This text was written by an AI model like ChatGPT."
result = classifier(text)
print(result)
đ Documentation
đ Training and Evaluation Data
- The model was fine - tuned on 35,894 training samples and 8,974 test samples.
- The dataset consists of AI - generated text samples (ChatGPT, Claude, DeepSeek, etc.) and human - written samples (Wikipedia, books, articles).
- Labels:
1
â AI - generated text
0
â Human - written text
âī¸ Training Procedure
Training Hyperparameters
The following hyperparameters were used during training:
Property |
Details |
Learning Rate |
2e-5 |
Train Batch Size |
16 |
Eval Batch Size |
16 |
Optimizer |
AdamW (β1 = 0.9, β2 = 0.999, Îĩ = 1e-08 ) |
LR Scheduler |
Linear |
Epochs |
3 |
Mixed Precision |
Native AMP (fp16) |
đ Training Results
Training Loss |
Epoch |
Step |
Validation Loss |
0.0505 |
0.22 |
500 |
0.0214 |
0.0114 |
0.44 |
1000 |
0.0110 |
0.0088 |
0.66 |
1500 |
0.0032 |
0.0 |
0.89 |
2000 |
0.0048 |
0.0068 |
1.11 |
2500 |
0.0035 |
0.0 |
1.33 |
3000 |
0.0040 |
0.0 |
1.55 |
3500 |
0.0097 |
0.0053 |
1.78 |
4000 |
0.0101 |
0.0 |
2.00 |
4500 |
0.0053 |
0.0 |
2.22 |
5000 |
0.0039 |
0.0017 |
2.45 |
5500 |
0.0046 |
0.0 |
2.67 |
6000 |
0.0043 |
0.0 |
2.89 |
6500 |
0.0036 |
đ Framework Versions
Property |
Details |
Transformers |
4.48.3 |
PyTorch |
2.5.1+cu124 |
Datasets |
3.3.2 |
Tokenizers |
0.21.0 |
đ License
The model is licensed under the apache - 2.0
license.