Skip to content

Latest commit

 

History

History
61 lines (46 loc) · 1.39 KB

README.md

File metadata and controls

61 lines (46 loc) · 1.39 KB

Team arthur-schopenhauer for ValueEval'24

Link to paper: https://ceur-ws.org/Vol-3740/paper-339.pdf

Local execution

Training the models

Example:

python train.py \
       --num-epochs 10 \
       --random-seed 66 \
       --loss-fn-type non-weighted \
       --target-column hv_label \
       --model-langs non-english \
       --ds-dir /path/to/dataset \
       --ckpt-dir ./checkpoints

Note: All the training checkponts will be saved in ckpt-dir. Un-needed checkpoints have to be manually deleted.

Reproducing the submission file

# takes around 15 minutes with V100S
python predict.py \
       --sentences-file /path/to/test/sentences.tsv \
       --output-file ./run.tsv \
       --models-dir /path/to/finetuned_models

Dockerization

Training

# build
docker build -f Dockerfile_train -t valueeval24-arthur-schopenhauer-train-ensemble:1.0.0 .

# run
docker run --rm \
  -v "$PWD/valueeval24:/dataset" -v "$PWD/models:/models" \
  valueeval24-arthur-schopenhauer-train-ensemble:1.0.0

Prediction

# build
docker build -f Dockerfile_predict -t valueeval24-arthur-schopenhauer-ensemble:1.0.0 .

# run
docker run --rm \
  -v "$PWD/valueeval24/test:/dataset" -v "$PWD/output:/output" \
  valueeval24-arthur-schopenhauer-ensemble:1.0.0

# view results
cat output/run.tsv