Individual target and source sentences from the AI4Bharat Samanantar dataset. All 11 language sentences and their translations have been stacked and used for next character generation task.
Training Procedure
Trained on the next character generation task using cross-entropy loss.
Preprocessing [optional]
converted to raw UTF8 characters before training by using ByT5-large tokenizer
Training Hyperparameters
Training regime:
output_dir="mamba",
per_device_train_batch_size=1,
per_device_eval_batch_size=1,
num_train_epochs=4,
weight_decay=0.1,
lr_scheduler_type="cosine",
learning_rate=5e-4,
fp16=False,
Evaluation
A simple cross-entropy loss has been used to test the pipeline and working of the model.
