Bakalarska_praca/trainingscript.py

from transformers import T5ForConditionalGeneration, T5Tokenizer, Trainer, TrainingArguments
from datasets import load_dataset

model_name = "t5-base"
tokenizer = T5Tokenizer.from_pretrained(model_name)
model = T5ForConditionalGeneration.from_pretrained(model_name)

def preprocess_function(examples):
    before_list = []
    after_list = []
    for ex in examples["before after"]:
        if ex is not None:
            splits = ex.split(" before after ")
            if len(splits) == 2:
                before_list.append(splits[0])
                after_list.append(splits[1])
            else:
                before_list.append(ex)
                after_list.append('')
        else:
            before_list.append('')
            after_list.append('')

    model_inputs = tokenizer(before_list, padding="max_length", truncation=True)
    labels = tokenizer(after_list, padding="max_length", truncation=True)
    model_inputs["labels"] = labels["input_ids"]
    return model_inputs


dataset = load_dataset("csv", data_files={"train": "converted.csv"}, delimiter=" ", column_names=["before after"])

tokenized_datasets = dataset.map(preprocess_function, batched=True)

training_args = TrainingArguments(
    output_dir="./results1",
    evaluation_strategy="epoch",
    save_strategy="epoch",
    learning_rate=2e-5,
    per_device_train_batch_size=64,
    per_device_eval_batch_size=64,
    num_train_epochs=1,
    weight_decay=0.01,
)

trainer = Trainer(
    model=model,
    args=training_args,
    train_dataset=tokenized_datasets["train"],
    tokenizer=tokenizer,
)

trainer.train()
model.save_pretrained("T5Autocorrection")
tokenizer.save_pretrained("T5TokenizerAutocorrection")