LOTD - Lord Of The Datasets

LOTD (Lord Of The Datasets) is an efficient and flexible NLP dataset preprocessing library designed to simplify working with large-scale text datasets. Its goal is to make dataset preparation, filtering, tokenization, and batching seamless for research.


Motivation

Preparing NLP datasets for model training can be tedious and error-prone. Tasks such as:

  • Tokenizing text or chat dialogs
  • Applying templates for instruction tuning
  • Filtering sequences by length
  • Padding and batching for PyTorch

often require custom scripts that are hard to maintain and reuse. LOTD aims to solve this by providing reusable building blocks and ready-to-use utilities, so developers and researchers can focus on modeling instead of boilerplate preprocessing.


Features Overview

LOTD provides:

  • Flexible tokenizers for text and chat datasets, including support for system, user, and assistant roles.
  • Collators and filters that handle padding, batching, and sequence length constraints efficiently.
  • Dataset utilities for splitting, caching, and creating PyTorch DataLoaders from HuggingFace datasets.
  • Prebuilt loaders for popular datasets such as Alpaca, ready for instruction fine-tuning.

LOTD is designed to reduce boilerplate, improve reproducibility, and make dataset preprocessing more intuitive, so you can spend more time experimenting with models and less time writing tedious preprocessing code.