Setup conda environment
conda create -n aradialect-mt python=3.10
Install dependencies
pip install -r requirements.txt
If you want to run the m2m100 model, you can adjust the m2m100_madar.yaml then run
python main.py --config-name m2m100_madar
If you want to run the nllb model, you can adjust the nllb_madar.yaml then run
python main.py --config-name nllb_madar
For AraT5, we adjusted the notebook from here. To train the model, run the following command:
python AraT5/run_trainier_seq2seq_huggingface.py \
--learning_rate 2e-5 \
--max_target_length 128 --max_source_length 128 \
--per_device_train_batch_size 32 --per_device_eval_batch_size 32 \
--model_name_or_path "UBC-NLP/AraT5v2-base-1024" \
--output_dir "AraT5/FT" --overwrite_output_dir \
--num_train_epochs 22 \
--train_file "dataset/madar_train.tsv" \
--validation_file "dataset/madar_dev.tsv" \
--task "translate_into_MSA" --text_column "source" --summary_column "target" \
--load_best_model_at_end --metric_for_best_model "eval_bleu" --greater_is_better True --evaluation_strategy steps --logging_strategy steps --predict_with_generate \
--do_train --do_eval > log3.log 2>&1
To run inference on dev set, run the following command:
python AraT5/inference.py
For Jasmine, please do check NADI-Finetuning-LLM.ipynb notebook.
To evaluate your prediction you can just run like this
cd dataset
python eval_score.py ./val/NADI2024_subtask3_dev2_GOLD.txt ./predictions/araT5_dev.txt ./val/UBC_subtask3_dev_2.txt
To download our model checkpoints and predictions, you can download them here.