mini-language-model

This repository has a simple pipeline for a mini language model, as well as some basic notes about the fundamental concepts of how a language model works which are collected across various sources. The main file, train_main.py, outlines the following steps:

Overview of train_main.py

1. Text Preprocessing: Converting a sample text into tokenized inputs and targets using a vocabulary mapping dict.
2. Embedding Layer: Mapping token indices to embedding vectors
3. Positional Encoding: Adding learnable positional information to the embeddings
4. Self-Attention: Computing self-attention over the sequence
5. Transformer Block: Applying a transformer pipeline
6. MiniLM Model: Combining the above components into a final model

1. Argument Parser

Uses argparse to require two parameters:

• --embed_dim: The size of the embedding vectors.
• --hidden_dim: The hidden dimension size used in the transformer block.
• --lr: The learning rate for the model
• --epochs: Total epochs to train the transformer
• --dataset: To train on full dataset or a subset

2. Some Notes to look for:

• Self-Attention: Click here
• Transformer: Click here

3. How to train the model?

Change the parameters as per your convinience

python train_main.py --embed_dim 16 --hidden_dim 64 --lr 0.01 --epochs 100 --dataset subset

4. How to use the trained model?

python generate.py --prompt *ENTER THE PROMPT HERE*