Using NeyabAI:

Direct Use:

import torch
from transformers import GPT2LMHeadModel, GPT2TokenizerFast

model = GPT2LMHeadModel.from_pretrained("XsoraS/NeyabAI")
tokenizer = GPT2TokenizerFast.from_pretrained("XsoraS/NeyabAI")

def generate_response(prompt):
    inputs = tokenizer(prompt, return_tensors='pt') # You can add .to(torch.device("cuda")) to use GPU acceleration.
    return tokenizer.decode(model.generate(inputs.input_ids, max_length=512, do_sample=True,top_p=0.8, temperature=0.7, num_return_sequences=1,attention_mask=None)[0],skip_special_tokens=True)

prompt = "Hello"
response = ' '.join(map(str, str(generate_response("### Human: "+prompt+" \n### AI:")).replace("</s>","").split()))
print(response)

Fine-Tuning:

This repository demonstrates how to fine-tune the NeyabAI(GPT-2) language model on a custom dataset using PyTorch and Hugging Face's Transformers library. The code provides an end-to-end example, from loading the dataset to training the model and evaluating its performance.

Requirements

• Python 3.6+
• PyTorch
• Transformers (Hugging Face)
• NumPy

You can install the required packages using pip:

pip install torch transformers numpy

Fine-Tuning Script

The following script outlines the steps for fine-tuning GPT-2 on a custom dataset:

from transformers import GPT2LMHeadModel, GPT2TokenizerFast, AdamW
import torch
from torch.utils.data import DataLoader, TensorDataset
import numpy as np

# Load pre-trained model and tokenizer
model_name = "XsoraS/NeyabAI"
model = GPT2LMHeadModel.from_pretrained(model_name)
tokenizer = GPT2TokenizerFast.from_pretrained(model_name)
tokenizer.pad_token = tokenizer.eos_token

# Example dataset
dataset = ["Your custom dataset goes here."]  # Replace with your actual dataset

# Tokenization function
def tokenize_function(examples):
    return tokenizer(examples, padding='max_length', truncation=True, max_length=512)

# Tokenize the dataset
tokenized_inputs = [tokenize_function(text) for text in dataset]
input_ids = [input['input_ids'] for input in tokenized_inputs]
attention_masks = [input['attention_mask'] for input in tokenized_inputs]

# Convert to torch tensors
input_ids = torch.tensor(input_ids)
attention_masks = torch.tensor(attention_masks)
labels = input_ids.clone()

# Create DataLoader
batch_size = 8
dataset = TensorDataset(input_ids, attention_masks, labels)
dataloader = DataLoader(dataset, batch_size=batch_size, shuffle=True)

# Configure device
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model.to(device)
model = model.half()

# Set up optimizer
optimizer = AdamW(model.parameters(), lr=3e-5)

# Define accuracy calculation
def calculate_accuracy(preds, labels):
    pred_flat = np.argmax(preds, axis=-1).flatten()
    labels_flat = labels.flatten()
    return np.sum(pred_flat == labels_flat) / len(labels_flat)

# Training loop (simplified)
for epoch in range(3):  # Adjust the number of epochs as needed
    for batch in dataloader:
        batch = tuple(t.to(device) for t in batch)
        input_ids, attention_masks, labels = batch

        outputs = model(input_ids, attention_mask=attention_masks, labels=labels)
        loss = outputs.loss
        logits = outputs.logits

        loss.backward()
        optimizer.step()
        optimizer.zero_grad()

        preds = logits.detach().cpu().numpy()
        label_ids = labels.to('cpu').numpy()
        acc = calculate_accuracy(preds, label_ids)

        print(f"Loss: {loss.item()}, Accuracy: {acc}")

print("Training complete!")

Notes

• Dataset: Replace the dataset variable with your actual dataset.
• Max Length: Adjust the max_length parameter in the tokenize_function as needed based on the length of your input texts.
• Batch Size and Learning Rate: You may need to tune the batch_size and learning rate (lr) according to your dataset and hardware capabilities.
• Epochs: Adjust the number of epochs based on your convergence criteria.

Acknowledgments

• This project uses the Transformers library by Hugging Face.
• Inspired by various fine-tuning examples and tutorials from the Hugging Face community.