---
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](https://github.com/roneysco/Fake.br-Corpus)** and **[FakeTrue.Br](https://github.com/jpchav98/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

```python
{
    "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).

```python
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](https://www.apache.org/licenses/LICENSE-2.0)

## Citation

Coming soon.