Blenderbot successfully trained but still gives wrong answer

Models
1 
from transformers import AutoModelForSeq2SeqLM, AutoTokenizer
import torch

# Load the model and tokenizer
model_name = "facebook/blenderbot-3B"

# Define the training data
training_data = [
    {"input": "Who is Raj Sanpui?", "output": "Raj Sanpui works at ABC"},
    {"input": "What does Raj Sanpui do?", "output": "Raj Sanpui works at ABC"},
    {"input": "Where does Raj Sanpui work?", "output": "Raj Sanpui works at ABC"},
]

class RajSanpuiDataset(torch.utils.data.Dataset):
    def __init__(self, training_data1, tokenizer1):
        self.training_data = training_data1
        self.tokenizer = tokenizer1

    def __getitem__(self, idx):
        input_text = self.training_data[idx]["input"]
        output_text = self.training_data[idx]["output"]

        input_ids1 = self.tokenizer.encode(input_text, return_tensors="pt")
        decoder_input_ids1 = self.tokenizer.encode(output_text, return_tensors="pt")

        return {"input_ids": input_ids1, "decoder_input_ids": decoder_input_ids1}

    def __len__(self):
        return len(self.training_data)

if __name__ == "__main__":
    tokenizer = AutoTokenizer.from_pretrained(model_name)
    model = AutoModelForSeq2SeqLM.from_pretrained(model_name)

    dataset = RajSanpuiDataset(training_data, tokenizer)
    batch_size = 1
    data_loader = torch.utils.data.DataLoader(dataset, batch_size=batch_size)

    device = "cpu"

    # Define the optimizer (Doesnt work with lr=1e-2 as well)
    optimizer = torch.optim.Adam(model.parameters(), lr=1e-5) 

    print("Starting training")

    # Train the model
    for epoch in range(1):
        print(epoch)
        model.train()
        total_loss = 0
        for batch in data_loader:
            input_ids = batch["input_ids"].to(device)
            decoder_input_ids = batch["decoder_input_ids"].to(device)

            # Zero the gradients
            optimizer.zero_grad()

            # Forward pass
            outputs = model(input_ids=input_ids, labels=decoder_input_ids)
            loss = outputs.loss

            # Backward pass
            loss.backward()

            # Update the model parameters
            optimizer.step()

            # Accumulate the loss
            total_loss += loss.item()

            print(f"Epoch {epoch+1}, Loss: {total_loss / len(data_loader)}")

    # Use the trained model to generate an answer
    input_ids = tokenizer.encode("Who is Raj Sanpui?", return_tensors="pt").to(device)
    outputs = model.generate(input_ids=input_ids)
    answer = tokenizer.decode(outputs[0], skip_special_tokens=True)
    print(answer)

Gives this wrong answer:

C:\Users\Raj\PycharmProjects\PythonProject\.venv\Scripts\python.exe C:\Users\Raj\PycharmProjects\PythonProject\Train-chatbot3.py 
Starting training
0
Epoch 1, Loss: 3.5551576614379883

Epoch 1, Loss: 6.134128411610921

Epoch 1, Loss: 7.840777079264323
 raj sanpui is an indian singer and songwriter.

Process finished with exit code 0
1 Like