README.md

---
library_name: tf-keras
language:
- ro
base_model:
- readerbench/RoBERT-base
---

## Model description

BERT-based model for classifying fake news written in Romanian.

## Intended uses & limitations

It predicts one of six types of fake news (in order: "fabricated", "fictional", "plausible", "propaganda", "real", "satire").

It also predicts if the article talks about health or politics.

## How to use the model

Load the model with:

```python
from huggingface_hub import from_pretrained_keras

model = from_pretrained_keras("pandrei7/fakenews-mtl")
```

Use this tokenizer: `readerbench/RoBERT-base`.

The input length should be 512. You can tokenize the input like this:

```python
tokenizer(
    your_text,
    padding="max_length",
    truncation=True,
    max_length=512,
    return_tensors="tf",
)
```

## Training data

The model was trained and evaluated on the [fakerom](https://www.tagtog.com/fakerom/fakerom/) dataset.

## Evaluation results

The accuracy of predicting fake news was roughly 75%.

## Reference

[Romanian Fake News Identification using Language Models](https://grants.ulbsibiu.ro/fakerom/wp-content/uploads/8_Preda-et-al.pdf)

```bibtex
@inproceedings{inproceedings,
  author = {Preda, Andrei and Ruseti, Stefan and Terian, Simina-Maria and Dascalu, Mihai},
  year   = {2022},
  month  = {01},
  pages  = {73-79},
  title  = {Romanian Fake News Identification using Language Models},
  doi    = {10.37789/rochi.2022.1.1.13}
}
```