This is a classification task for the Eduge dataset of 70 thousand Mongolian online news articles with labeled news categories. This project uses Fast.ai.

After training a language model using the Eduge dataset, a classifier accuracy of 93.5% is reached after 10 epochs. Training the classifier without the language model gives an accuracy of roughly 90.5% after 10 epochs.

• 01 - Mongolian Language Model: A language model is trained using the Eduge dataset. An accuracy of 33% is reached for the language model.
• 02 - Eduge Classifier: Using the vocabulary and encoder (i.e. language model) from notebook 01, a classifier is trained to predict news category. An accuracy of 93.5% is reached after 10 epochs.
• 03 - Eduge Classifier w_out Pretrained LM: The purpose of this notebook is to show the difference in accuracy achieved in a classifer model with a pre-trained language model vs without. Below we can see that after several epochs the accuracy flattens out around 90-91%.

Data is in the data folder. Google Drive link.

• news.csv: This is the Eduge news dataset as downloaded from the source. Used to train the language model and classifier.
• eduge_clean.csv: Not used in the notebooks. This is a cleaned verison of the Eduge dataset. A slightly better accuracy can be obtained with the cleaned data. For comparison reasons this dataset isn't used.

Outputs are in the models folder. Google Drive link.

• mn_eduge_lm.pth: The complete language model including encoder and decoder.
• mn_eduge_lm_encoder.pth: The saved encoder from the learner. This is used to prep our learner in notebook 02.
• mn_eduge_vocab.pkl: The vocabulary to correspdong with the language model. This is also needed in notebook 02 to create our dataloaders object.