license: cc-by-sa-3.0
language:
- ar
- ru
- fr
- es
- zh
- en
size_categories:
- 1M<n<10M
dataset_info:
- config_name: default
features:
- name: id
dtype: string
- name: url
dtype: string
- name: title
dtype: string
- name: text
dtype: string
splits:
- name: pretrain
num_bytes: 3895403456
num_examples: 1200000
download_size: 2293060098
dataset_size: 3895403456
- config_name: instruct_tuning
features:
- name: instruct
dtype: string
- name: input
dtype: string
- name: output
dtype: string
splits:
- name: train
num_bytes: 11251187014
num_examples: 6720312
- name: valid
num_bytes: 590616379
num_examples: 353701
download_size: 6802172962
dataset_size: 11841803393
configs:
- config_name: default
data_files:
- split: pretrain
path: data/pretrain-*
- config_name: instruct_tuning
data_files:
- split: train
path: instruct_tuning/train-*
- split: valid
path: instruct_tuning/valid-*
pretty_name: sparrow
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
- 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")