train-cl-multilingual.sh

#!/bin/bash

python3 run_cl.py \
    --output_dir output/mwp-cl-multilingual \
    --bert_pretrain_path pretrained_models/bert-base-multilingual-uncased \
    --data_dir data \
    --train_file Math_23K-MathQA_mbert_token_train.json \
    --dev_file_1 Math_23K_mbert_token_val.json \
    --test_file_1 Math_23K_mbert_token_test.json \
    --dev_file_2 MathQA_mbert_token_val.json \
    --test_file_2 MathQA_mbert_token_test.json \
    --contra_pair pairs/Math_23K-MathQA-sample.json \
    --learning_rate 5e-5 \
    --n_epochs 20 \
    --n_val 1 \
    --hidden_size 768 \
    --beam_size 3 \
    --contra_loss_func margin \
    --contra_common_tree_pair \
    --contra_loss_margin 0.2 \
    --neg_sample 1 \
    --neg_sample_from_pair_file \
    --neg_no_expr_loss \
    --alpha 5 \
    --alpha_warmup \
    --warmup_begin 1000 \
    --warmup_end 6000 \
    --dropout 0.5 \
    --seed 17