---
license: mit
task_categories:
- text-generation
language:
- en
tags:
- code
- medical
- biology
- chemistry
- finance
---


# Orion-Spark-2 Dataset

## Overview
The **Orion-Spark-2 Dataset** is a text corpus curated for training the Orion-Spark-2 transformer language model. It consists of a diverse collection of sentences extracted from multiple sources including Wikipedia articles, technology news sites, developer resources, and other open-access web pages. The dataset is designed to provide broad coverage of general knowledge, programming topics, artificial intelligence, space, popular culture, and current events.

## Structure
- **File:** `corpus.txt`
- **Format:** Plain text, one sentence per line.
- **Encoding:** UTF-8
- **Line Count:** Approximately 60,000+ lines
- **Checkpoint:** `corpus_checkpoint.txt` to track downloaded lines for resuming corpus collection.

## Sources
The dataset draws content from:
- Wikipedia pages (various topics including AI, programming languages, mathematics, astronomy, and historical events)
- News and tech sites (BBC Technology, TechCrunch)
- Open-source repositories (GitHub)
- Educational and community platforms (Fast.ai)
- Hugging Face datasets

## Processing
- Each line in the dataset is cleaned to remove excessive whitespace.
- Sentences shorter than 30 characters are discarded.
- HTML content is parsed using BeautifulSoup to extract text from paragraph and header tags (`<p>`, `<h1>`, `<h2>`, `<h3>`).
- Sentences are split on punctuation marks (`.`, `?`, `!`) to ensure individual sentence granularity.

## Usage
1. Load the dataset:
   ```python
   from torch.utils.data import DataLoader
   from dataset import TextDataset
   dataset = TextDataset(texts, tokenizer)
Use TextDataset for training or evaluation in PyTorch.

Pad sequences using the collate_batch function when forming batches for model training.

Notes
The dataset is intended for educational and research purposes.

It contains only publicly available information; no private or copyrighted content has been included beyond fair use.

Designed for training medium-sized language models (30M parameters) efficiently with maximum sequence length of 128 tokens.