This model fine-tunes facebook/mbart-large-50 for automatic correction of Urdu text, addressing spelling mistakes, grammatical errors, and improving text fluency.
The Urdu Text Correction model is designed to automatically detect and correct errors in Urdu text, making it valuable for content editors, publishers, educational institutions, and applications requiring high-quality Urdu text.
Performance Metrics
The model achieves the following results on our evaluation set:
Metric
Score
BLEU
0.6996
METEOR
0.8296
WER (Word Error Rate)
0.1795
CER (Character Error Rate)
0.0761
ROUGE-1
0.2025
ROUGE-2
0.0699
ROUGE-L
0.2023
Exact Match
0.1096
Generation Length
28.4033
Loss
0.4305
Usage
bash
1from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
23# Load model and tokenizer4tokenizer = AutoTokenizer.from_pretrained("mahwizzzz/urdu_text_correction")5model = AutoModelForSeq2SeqLM.from_pretrained("mahwizzzz/urdu_text_correction")67# Example text with errors8incorrect_text ="یہہ ایک اچھی بات ہے"910# Tokenize and generate correction11inputs = tokenizer(incorrect_text, return_tensors="pt", max_length=128, truncation=True)12outputs = model.generate(**inputs, max_length=128)13corrected_text = tokenizer.decode(outputs[0], skip_special_tokens=True)1415print(f"Original: {incorrect_text}")16print(f"Corrected: {corrected_text}")
Model Description
This is an encoder-decoder model based on mBART-large-50, specifically fine-tuned on Urdu text correction pairs. The model learns to transform incorrect Urdu text into its corrected form, addressing issues such as:
Spelling mistakes
Grammar errors
Word order issues
Missing diacritics
Punctuation errors
Training Details
Data
The model was trained on a dataset of incorrect-correct Urdu text pairs.
Hyperparameters
The model was trained with the following hyperparameters:
Learning rate: 3e-05
Batch size: 128 (32 per device with gradient accumulation of 4)
Optimizer: AdamW with betas=(0.9, 0.999)
LR scheduler: Cosine with 500 warmup steps
Training epochs: 3
Mixed precision: Native AMP
Limitations
The model may not perform well on highly domain-specific text (technical, medical, etc.)
Very long texts may need to be split into smaller chunks due to the model's maximum sequence length
The model may sometimes over-correct dialectal variations or stylistic choices
Performance is dependent on the quality and diversity of the training data
Future Work
Expanding the training dataset with more diverse text sources
Domain adaptation for specific use cases (legal, medical, etc.)
Performance optimization for faster inference
Improved handling of complex grammatical structures
Citation
If you use this model in your research or project, please cite:
@model{urdu_text_correction,
title = {Urdu Text Correction Model}
author = {Mahwiz Khalil},
year = {2025},
publisher = {HuggingFace},
howpublished = {\url{https://huggingface.co/mahwizzzz/urdu_text_correction}}
}
