activations: Tanh
batch_size: 2
class_identifier: speech_regression_metric
dropout: 0.1
encoder_learning_rate: 1.0e-06
encoder_model: XLM-RoBERTa
encoder_model_audio: sonar_speech_encoder_eng
final_activation: null
fuse_emb_strategy: avg
hidden_sizes:
- 2048
- 1024
input_modality: text
keep_embeddings_frozen: true
layer: mix
layer_norm: false
layer_transformation: sparsemax
layerwise_decay: 0.95
learning_rate: 1.5e-05
load_pretrained_weights: true
local_files_only: false
loss: mse
nr_frozen_epochs: 0.3
optimizer: AdamW
pool: avg
pretrained_model: xlm-roberta-base
train_data:
- maikezu/iwslt2026-metrics-shared-train-dev
validation_data:
- maikezu/iwslt2026-metrics-shared-train-dev
warmup_steps: 0