BanglaLM

Bigram Language Model

Overview

This repository contains a simple Bigram Language Model implemented in PyTorch. The model is trained to predict the next character in a sequence, given the current character. It's a character-level model and can be used for tasks like text generation.

Model Details

• Model Type: Character-level Language Model
• Architecture: Simple lookup table for character bigrams
• Training Data: [https://huggingface.co/datasets/csebuetnlp/xlsum/viewer/bengali]

Requirements

• Python 3.x
• PyTorch
• JSON (for loading the tokenizer)

Installation

First, clone this repository:

Loading the Model

To load the model, you need to initialize it with the vocabulary size and load the pre-trained weights:

import torch
from model import BigramLanguageModel 

vocab_size = 225  
model = BigramLanguageModel(vocab_size)

model.load_state_dict(torch.load('path_to_your_model.pth', map_location=torch.device('cpu')))
model.eval()

import json

with open('tokenizer_mappings.json', 'r', encoding='utf-8') as f:
    mappings = json.load(f)
    stoi = mappings['stoi']
    itos = mappings['itos']

# Example usage
encode = lambda s: [stoi[c] for c in s]
decode = lambda l: ''.join([itos[i] for i in l])


context = torch.tensor([encode("Your initial text")], dtype=torch.long)
generated_text_indices = model.generate(context, max_new_tokens=100)
print(decode(generated_text_indices[0].tolist()))