app.py

import torch
import json
import os
from transformers import T5Tokenizer, T5ForConditionalGeneration, Trainer, TrainingArguments, DataCollatorForSeq2Seq
from datasets import Dataset

# ✅ 1. Ielādē treniņa datus
with open("train.json", "r", encoding="utf-8") as f:
    train_data = json.load(f)

# ✅ 2. Pārveido datus Hugging Face formātā
dataset = Dataset.from_list([{"input_text": d["question"], "target_text": d["answer"]} for d in train_data])

# ✅ 3. Izvēlies modeli (piemēram, `mT5-small` vai `facebook/opt-1.3b`)
model_name = "google/mt5-small"
device = "cuda" if torch.cuda.is_available() else "cpu"

tokenizer = T5Tokenizer.from_pretrained(model_name)
model = T5ForConditionalGeneration.from_pretrained(model_name).to(device)

# ✅ 4. Sagatavo datus treniņam
def preprocess_data(examples):
    model_inputs = tokenizer(examples["input_text"], max_length=128, truncation=True)
    labels = tokenizer(examples["target_text"], max_length=128, truncation=True)
    model_inputs["labels"] = labels["input_ids"]
    return model_inputs

dataset = dataset.map(preprocess_data, batched=True)

# ✅ 5. Definē trenēšanas parametrus
training_args = TrainingArguments(
    output_dir="./trained_model",
    evaluation_strategy="epoch",
    save_strategy="epoch",
    learning_rate=5e-5,
    per_device_train_batch_size=4,
    per_device_eval_batch_size=4,
    num_train_epochs=3,  # Trenē 3 epohas
    weight_decay=0.01,
    push_to_hub=True,  # Augšupielādē Hugging Face Hub
    logging_dir="./logs"
)

data_collator = DataCollatorForSeq2Seq(tokenizer, model=model)

# ✅ 6. Izveido `Trainer` klasi
trainer = Trainer(
    model=model,
    args=training_args,
    train_dataset=dataset,
    tokenizer=tokenizer,
    data_collator=data_collator
)

# ✅ 7. Sāc trenēšanu
trainer.train()

# ✅ 8. Saglabā trenēto modeli
model.save_pretrained("./trained_model")
tokenizer.save_pretrained("./trained_model")

# ✅ 9. Augšupielādē Hugging Face Hub (ja nepieciešams)
trainer.push_to_hub()