Dataset and baselines for the Moral Uncertainty
benchmark. See the Moral Uncertainty website for details.
You can download the dataset files from the links in the table:
Split | Download | # examples |
---|---|---|
Train | train.csv | 13,910 |
Test | test.csv | 2,771 |
The code for training our baseline models is in moral-uncertainty/baselines
. Our specific baselines can be reproduced with:
# bert-base-uncased
python tune.py --ngpus 1 --model bert-base-uncased --learning_rate 3e-5 --batch_size 16 --nepochs 4 --gradient_acc_steps 1 --verbose
# bert-large-uncased
python tune.py --ngpus 1 --model bert-large-uncased --learning_rate 5e-6 --batch_size 16 --nepochs 4 --gradient_acc_steps 1 --verbose
# roberta-large
python tune.py --ngpus 1 --model roberta-large --learning_rate 1e-5 --batch_size 16 --nepochs 4 --gradient_acc_steps 1 --verbose
# albert-xxlarge-v2
python tune.py --ngpus 1 --model albert-xxlarge-v2 --learning_rate 3e-5 --batch_size 16 --nepochs 4 --gradient_acc_steps 1 --verbose
# microsoft/deberta-v3-large
python tune.py --ngpus 1 --model microsoft/deberta-v3-large --learning_rate 1e-5 --batch_size 16 --nepochs 4 --gradient_acc_steps 1 --verbose
# microsoft/deberta-v2-xxlarge
python tune.py --ngpus 1 --model microsoft/deberta-v2-xxlarge --learning_rate 1e-6 --batch_size 8 --nepochs 4 --gradient_acc_steps 1 --verbose
For the GPT-3 baseline, see moral-uncertainty/baselines/finetune_gpt3.ipynb
.