train.yaml

seed: 42
epochs: 3
batch_size: 2
num_workers: 0
device: cpu
lr: 1e-4
clip_grad: 1.0
scheduler: none

# Tokenizer path
tokenizer_path: "tokenizer.json"

# Data paths
train_csv: "dataset/train.csv"
val_csv: "dataset/val.csv"

# Image settings
image_height: 64
image_width: 512
max_seq_length: 150

# Saving
output_dir: "checkpoints/"
save_every: 1