IbukunGracey/Hausa_English_MT

# A Transformer Based Translation Model for Hausa Language using Pretrained Models This repository contains a machine translation model that translates english to hausa languages. This was trained using pre-trained models MBART and M2M100 - "facebook/m2m100_418M" like architectures from Masakhane git repo and fine-tuned on MAFAND_MT datasets. ### Model description mt-ha-en is a machine translation model from English language to Hausa language based on a fine-tuned MBART-base model. It establishes a strong baseline for automatically translating texts from English to Hausa. ### Training data This model was fine-tuned on MAFAND_MT en_hau datasets with * Train data - 5865 * Test data- 1500 * Validation data - 1300 ### Training procedure This model was trained on Colab - NVIDIA V100 GPU **Eval results on Test set (BLEU score)** Fine-tuning the model achieves 12.16 BLEU on the test set and 12.18 BLEU on the validation set with a training runtime of 37 mins and prediction runtime of approx 10mins ### Limitations and bias This model is limited by its training dataset. This may not generalize well for all use cases in different domains. Also the Bleu score can be improved on, if some of the hyper-parameters are tuned. We used 3 epoch and 2 beams due to the computational resources available at the time. As such, more work can be done to improve the model. ### Benefits/Recommendations Using pre-trained model is highly efficient compared to training the model from scratch, as the later can be more time consuming and costly with respect to time and computational resorces. ### Citation ``` @inproceedings{adelani-etal-2022-thousand, title = "A Few Thousand Translations Go a Long Way! Leveraging Pre-trained Models for {A}frican News Translation", author = "Adelani, David and Alabi, Jesujoba and Fan, Angela and Kreutzer, Julia and Shen, Xiaoyu and Reid, Machel and Ruiter, Dana and Klakow, Dietrich and Nabende, Peter and Chang, Ernie and Gwadabe, Tajuddeen and …