test.py

from transformers import MT5ForConditionalGeneration, MT5Tokenizer
from datasets import load_dataset
import torch

# Путь к модели и данным
model_path = "./mt5-finetuned"
validation_file = "mt5_validation_data-1.jsonl"

# Загрузка модели и токенизатора
tokenizer = MT5Tokenizer.from_pretrained(model_path)
model = MT5ForConditionalGeneration.from_pretrained(model_path)
model.eval()

# Используем GPU если есть
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model = model.to(device)

# Загрузка валидационной выборки
dataset = load_dataset("json", data_files={"validation": validation_file})
val_data = dataset["validation"]

# Функция предсказания
def predict(text):
    inputs = tokenizer(text, return_tensors="pt", truncation=True, padding=True, max_length=256).to(device)
    outputs = model.generate(
        **inputs,
        max_length=64,
        num_beams=5,
        early_stopping=True
    )
    return tokenizer.decode(outputs[0], skip_special_tokens=True)

# Подсчёт точности
correct = 0
results = []

for idx, example in enumerate(val_data):
    text = example["text"]
    target = example["target"].strip()
    pred = predict(text).strip()

    results.append((text, pred, target))
    if pred == target:
        correct += 1

# Примеры
print("📋 Примеры предсказаний:\n")
for i, (text, pred, target) in enumerate(results[:80]): # кол-во примеров 
    print(f"#{i+1}")
    print(f"📝 Вход: {text}")
    print(f"✅ Target: {target}")
    print(f"🤖 Предсказание: {pred}")
    print("-" * 50)

# Accuracy
accuracy = correct / len(val_data)
print(f"\n✅ Accuracy: {accuracy:.4f} ({correct}/{len(val_data)})\n")