metadata
license: cc-by-nc-4.0
language:
- pt
tags:
- ai-detection
- text-classification
- portuguese
- bert
- transformers
base_model: neuralmind/bert-base-portuguese-cased
pipeline_tag: text-classification
datasets:
- wiki40b-pt
- oscar-pt
- cc100-pt
- europarl-pt
- opus-books-pt
- Detecting-ai/ai_pt_corpus
model_type: bert
🇧🇷 pt-ai-detector
pt-ai-detector is a BERT-base model fine-tuned to decide whether a Portuguese sentence or paragraph was written by a human (label = 0) or generated by AI (label = 1).
| Metric | Value |
|---|---|
| Train data | 1 000 000 human + 1 000 000 AI |
| Balanced test set | 1 954 190 (½ human, ½ AI) |
| Accuracy | ≈ 99 % |
| F1 (macro) | ≈ 0.99 |
| Model size | 434 M parameters (≈ 430 MB) |
📖 Quick usage
Try it live at detecting-ai.com – our team at Detecting-ai built this model and demo so you can instantly test any Portuguese text online.
from transformers import pipeline
clf = pipeline(
"text-classification",
model="Detecting-ai/pt-ai-detector",
tokenizer="Detecting-ai/pt-ai-detector",
device=0 # set -1 for CPU
)
text = "A inteligência artificial está transformando a educação."
print(clf(text)) # → [{'label': 'AI', 'score': 0.987}]
| id | label |
|---|---|
| 0 | Human |
| 1 | AI |
🏋️♂️ Training details
- Base model:
neuralmind/bert-base-portuguese-cased - Epochs: 3 (fp16 on 1 × A100)
- Batch size: 32
- Optimizer/LR: AdamW 2 × 10⁻⁵
- Loss: Cross-entropy
Data sources
| Corpus | Lines used | Notes |
|---|---|---|
| Human corpus (wiki40b-pt, oscar-pt, cc100-pt, europarl-pt, opus-books-pt) | 1 M sampled | Diverse Portuguese web/news/books |
AI corpus (Detecting-ai/ai_pt_corpus) |
1 M | Generated with OpenAI models (various GPT-4 / GPT-3.5 variants); prompts cover essays, news, tweets, dialogs, paraphrases, T = 0.6–1.0 |
Datasets were balanced 1 : 1 and shuffled before training.
🚦 Intended use
Detect AI-generated Portuguese text in essays, articles, chats, support tickets, etc.
Limitations
- Not trained on code or non-Portuguese language.
- Accuracy may drop on texts < 10 tokens or heavily paraphrased AI.
- Commercial use is disallowed (CC-BY-NC-4.0).
⚠️ Future work
- Evaluate on adversarial paraphrases.
- Distill/quantize for edge deployment.
- Extend to multilingual detection.
📜 Citation
@misc{abdurazzoqov2025ptaidetector,
title = {pt-ai-detector: Detecting AI-generated Portuguese Text},
author = {Abdulla Abdurazzoqov},
year = {2025},
howpublished = {Hugging Face hub},
url = {https://huggingface.co/Detecting-ai/pt-ai-detector}
}
💬 License
Creative Commons CC-BY-NC-4.0 — free for research & personal use; commercial use requires written permission.