activations: Tanh
batch_size: 4
class_identifier: unified_metric
dropout: 0.1
encoder_learning_rate: 5.0e-06
encoder_model: XLM-RoBERTa
final_activation: null
hidden_sizes:
- 3072
- 1024
input_segments:
- src
- mt
- ref
keep_embeddings_frozen: true
layer: mix
layerwise_decay: 0.95
learning_rate: 1.5e-05
load_weights_from_checkpoint: null
nr_frozen_epochs: 0.3
optimizer: AdamW
pool: cls
pretrained_model: xlm-roberta-large
train_data: data/1719-da.csv
validation_data:
- data/qad-ende-newstest2020.csv
- data/qad-enru-newstest2020.csv
- data/wmt-ende-newstest2021.csv
- data/wmt-zhen-newstest2021.csv
- data/wmt-enru-newstest2021.csv