Sparrow

麻雀虽小 五脏俱全

Small as it is, the sparrow has all the vital organs

The pretraining dataset is collected from wikimedia/wikipedia

---

The sparrow project aims to help beginner to understand the base architecture of a large language model from scratch. Not only the model, but also the optimization methods that are widely use to shorten the training process.

• tokenizer from scratch & merge tokenizer
• model modules from scratch & train the stacked model
• supervised fine-tuning
• Reward Modelling

Data Preparation

1. For tokenizer and pretraining process, to simplify the data collection process, we use the data from wikimedia/wikipedia, ensuring that our training corpus is both rich in content and easily accessible. We use 10%-20% of the data with six official language of United Nation — Arabic, Chinese, English, French, Russian, and Spanish—providing a diverse and representative sample for training our tokenizer.

Tokenizer

A good tokenizer is vital as it is the first component that converts raw text into a structured format a model can understand. It determines the granularity of tokenization and ensures that important elements—such as special tokens marking the beginning and end of a sentence—are consistently incorporated, directly affecting the model's ability to learn and generate language accurately. In tokenizer/tokenizer.py, we provide a class SparrowTokenizer to help you understand the how a tokenizer been trained. This script demonstrates the complete pipeline—from preprocessing raw data and creating a training corpus, to training a BPE-based tokenizer with customized post-processing for adding special tokens, and finally, saving the vocabulary and configuration files. You can explore this workflow by running:

python tokenizer/tokenizer.py --args configs/tokenizers.yaml

Models Artitecture

Pretraining Set

from datasets import load_dataset
dataset = load_dataset("TerenceLau/sparrow", split="pretrain")