README.md

metadata

language:
  - pt
license: apache-2.0
library_name: keras
pipeline_tag: text-classification
tags:
  - bilstm
  - portuguese
  - pt
  - fake-news
  - binary-classification
metrics:
  - accuracy
  - precision
  - recall
  - f1-score

BiLSTM for Fake News Detection (Portuguese)

Model Overview

This repository contains BiLSTM models for fake news detection in Portuguese. Models are trained and evaluated on corpora derived from Brazilian Portuguese datasets Fake.br and FakeTrue.Br, as well as a combined/aligned corpus.

• Architecture: Bidirectional LSTM (Keras)
• Task: Binary text classification (Fake vs. True)
• Language: Portuguese (pt)
• Framework: Keras / TensorFlow

---

Available Variants

• bilstm-combined Fine-tuned on the aligned combined corpus (data/corpus_* at project root).

• bilstm-fake-br Fine-tuned on Fake.br. The corresponding corpus is available in corpus/ (including preprocessed and size-normalized texts when applicable).

• bilstm-faketrue-br Fine-tuned on FakeTrue.Br. Includes aligned splits and the original CSV when available.

Each variant ships with:

• confusion_matrix.png
• final_classification_report.parquet
• final_predictions.parquet

---

Training Details

{
    "ngram_upper": 2,
    "units": 120,
    "dropout": 0.3374510345164157,
    "recurrent_dropout": 0.1588638491073387,
    "max_tokens": 96000,
    "embed_dim": 71,
    "embed_max_seq_len": 51,
    "learning_rate": 0.00011662663429277272,
    "batch_size": 16,
    "epochs": 8,
}

---

Evaluation Results

Evaluation metrics are stored in the repo as:

• confusion_matrix.png
• final_classification_report.parquet
• final_predictions.parquet

These files provide per-class performance and prediction logs for reproducibility.

---

Corpus

The corpora used for training and evaluation are provided in the corpus/ folder.

• Combined (root folder): corpus_train_df.parquet, corpus_test_df.parquet, corpus_df.parquet, corpus_alinhado_df.parquet.
• Fake.br: corpus_train_df.parquet, corpus_test_df.parquet, corpus_df.parquet, corpus_alinhado_df.parquet.
• FakeTrue.Br: corpus_train_df.parquet, corpus_test_df.parquet, corpus_df.parquet, corpus_alinhado_df.parquet and FakeTrueBr_corpus.csv.

---

How to Use

This model is a Keras model stored as final_bilstm_model.keras. Load it with TensorFlow and apply the same preprocessing (tokenization + padding).

import keras
import tensorflow as tf
from huggingface_hub import hf_hub_download

repo_id = "vzani/portuguese-fake-news-classifier-bilstm-combined"  # or fake-br / faketrue-br
filename = "final_bilstm_model.keras"

model_path = hf_hub_download(repo_id=repo_id, filename=filename)
model = keras.models.load_model(model_path)


def predict(text: str) -> tuple[bool, float]:
    input_data = tf.convert_to_tensor([[text]], dtype=tf.string)
    probs = model.predict(input_data)  # type: ignore
    prob = float(probs.flatten()[0])  # type: ignore
    pred = int(prob >= 0.5)
    return pred, prob  # type: ignore


if __name__ == "__main__":
    text = "BOMBA! A Dilma vai taxar ainda mais os pobres!"
    print(predict(text))

Expected output:

(False, 0.xxx)

Interpreting the results:

• LABEL_0: Fake news
• LABEL_1: True news

License

Apache 2.0

Citation

Coming soon.