Model Card: SlovHC – Slovak Hate Speech Corrector

Model Overview

SlovHC is a fine-tuned model designed specifically for correcting hate speech in the Slovak language. Given the limited availability of robust language models for low-resource languages like Slovak, SlovHC aims to fill this gap by offering high-quality performance in identifying and masking offensive content.

Key Features

Tailored for the Slovak language
Focuses on masking hate speech while preserving sentence structure
Utilizes pre-trained Slovak BERT tokenizer for consistent tokenization

Example Outputs

Input: Ty si absolútny magor.
Output: Ty si absolútny *****.

Input: Priblblé električky stále meškajú.
Output: ********* električky stále meškajú.

Input: Opač jak ši sebe obľik tote nohavky, ši jak mantak.
Output: Opač jak ši sebe obľik tote nohavky, ši jak ******.

Tokenizer

We did not develop a new tokenizer for this model. Instead, we leveraged the high-quality tokenizer provided by gerulata/slovakbert, which aligns well with our model’s requirements.

How to Use

Here's a simple example demonstrating how to load and run inference with the model:

from transformers import RobertaTokenizer, AutoModelForSeq2SeqLM

# Load pretrained tokenizer and model
tokenizer = RobertaTokenizer.from_pretrained(
    'gerulata/slovakbert',
    weights_only=False,
    token="###YOUR_HF_TOKEN###"
)
model = AutoModelForSeq2SeqLM.from_pretrained(
    "timotejKralik/hate_speech_correction_slovak",
    weights_only=False,
    token="###YOUR_HF_TOKEN###"
)

# Input text containing potentially harmful language
input_text = "Opač jak ši sebe obľik tote nohavky, ši jak mantak."
print(f"Input: {input_text}")

# Tokenize input and generate output
inputs = tokenizer(input_text, return_tensors="pt")
outputs = model.generate(**inputs)
output_text = tokenizer.decode(outputs[0], skip_special_tokens=True)

print("Output:", output_text)