README.md

metadata

annotations_creators:
  - no-annotation
language:
  - en
license: apache-2.0
task_categories:
  - text-generation
  - summarization
  - question-answering
  - feature-extraction
task_ids:
  - language-modeling
size_categories:
  - n<1K
tags:
  - openai
  - chatgpt
  - gpt-4
  - gpt-3
  - sora
  - dall-e
  - ai-safety
  - frontier-ai
  - cl100k-tokens
  - llm-pretraining
configs:
  - config_name: default
    data_files: data/train-*.parquet
  - config_name: full-corpus
    data_files: data/train-*.parquet

🌐 OpenAI Official News, Research & Landmark Publication Archive

High-fidelity web archive, structured clean texts, frontier AI safety whitepapers, architecture milestones (GPT-1 through GPT-4o, ChatGPT, Sora, DALL-E), and complete media assets from OpenAI's official news estate.

Overview · Corpus & Token Stats · Chunked Dataset Structure · Multimodal Assets · Section Distribution · Yearly Trends · Schema · Quick Start

---

📌 Overview & Executive Summary

The OpenAI Official News, Research & Landmark Publication Archive is a curated, high-fidelity archive capturing the entirety of OpenAI's official publication estate (openai.com/news/). Maintained and published by Sphita AI Research, this dataset is engineered for frontier LLM pretraining, domain-specific instruction tuning, alignment research, and technical retrieval-augmented generation (RAG).

Spanning 11+ years of transformative AI history (2015-12-11 to 2026-10-15), this archive documents OpenAI's evolution from an open non-profit research lab to the creator of GPT-1, GPT-2, GPT-3, GPT-4, GPT-4o, ChatGPT, DALL-E, Codex, Whisper, and Sora.

Key Highlights

• Zero Container Stutter: Custom parsing eliminates navigation noise, sidebar TOC anchors, and list container stutter.
• Exact Token Auditing: Profiled using OpenAI's standard cl100k_base BPE tokenizer, measuring 1,097,877 total tokens.
• Zero-Leak Sharded Parquet & Streaming JSONL: Distributed in standard Hugging Face Zstandard-compressed Parquet chunks (data/train-00000-of-0000N.parquet, 200 articles each) and streaming JSON Lines (data/openai_news.jsonl).
• Chunked Multimodal Media Archives (Max 500 MiB): Images and screenshots are split into size-bounded shards (media/images-0000X.tar.gz, media/screenshots-0000X.tar.gz) with immutable sealed historical shards for fast incremental updates.

---

📊 Comprehensive Corpus & Token Statistics

Token volume calculated using OpenAI cl100k_base BPE tokenizer against all 682 articles.

Metric	Clean Markdown Text	Raw Payload / Metadata	Combined Archive
Total Tokens (cl100k_base)	1,097,877	—	1,097,877
Total Words	838,614	—	838,614
Total Raw Character Size	5,476,619 chars (5.22 MB)	—	5.22 MB
Average Tokens per Article	1609.8 tokens	—	1609.8 tokens
Median Tokens per Article	1,267 tokens	—	1,267 tokens
Total Parquet Size (ZSTD-9)	—	—	3.13 MB (4 shards)
JSONL Streaming Size	—	—	11.81 MB
Columnar Compression Ratio	—	—	1.7x reduction

---

📦 Chunked Dataset Structure

Parquet tables are sharded into standard chunks (up to 200 rows per shard) for parallel streaming, instant loading, and fast network transfer:

Shard Path	Format	Rows	File Size	Description
data/train-00000-of-00004.parquet	Parquet (ZSTD-9)	≤ 200	1.09 MB	Canonical Hugging Face split shard
data/train-00001-of-00004.parquet	Parquet (ZSTD-9)	≤ 200	1.02 MB	Canonical Hugging Face split shard
data/train-00002-of-00004.parquet	Parquet (ZSTD-9)	≤ 200	0.75 MB	Canonical Hugging Face split shard
data/train-00003-of-00004.parquet	Parquet (ZSTD-9)	≤ 200	0.27 MB	Canonical Hugging Face split shard
data/openai_news.jsonl	JSON Lines (UTF-8)	682	11.81 MB	Complete streaming JSONL corpus

---

🖼 Multimodal Media & Visual Assets Archive

Images and full-page rendering screenshots are split into strictly bounded archives (≤ 500 MiB each). Completed historical shards are sealed and immutable, ensuring incremental updates only download new media:

Asset Archive	Description	Status	File Count	Archive Size	Hugging Face Repository Path
images-00000.tar.gz	Cover images & inline figures	🔒 Sealed	982 files	467.2 MB	media/images-00000.tar.gz
images-00001.tar.gz	Cover images & inline figures	🔒 Sealed	1,643 files	470.0 MB	media/images-00001.tar.gz
images-00002.tar.gz	Cover images & inline figures	🔒 Sealed	583 files	121.8 MB	media/images-00002.tar.gz
screenshots-00000.tar.gz	Full-page browser render screenshots	🔒 Sealed	194 files	96.4 MB	media/screenshots-00000.tar.gz
screenshots-00001.tar.gz	Full-page browser render screenshots	🔒 Sealed	200 files	80.2 MB	media/screenshots-00001.tar.gz
screenshots-00002.tar.gz	Full-page browser render screenshots	🔒 Sealed	200 files	90.0 MB	media/screenshots-00002.tar.gz
screenshots-00003.tar.gz	Full-page browser render screenshots	🔒 Sealed	88 files	33.0 MB	media/screenshots-00003.tar.gz

Unpacking Media Assets

# 1. Download and unpack all chunked images (preserves images/<slug>/... layout)
huggingface-cli download sphita/openai-news --include "media/images-*.tar.gz" --local-dir ./
for f in media/images-*.tar.gz; do tar -xzf "$f"; done

# 2. Download and unpack screenshots (preserves screenshots/<slug>.png layout)
huggingface-cli download sphita/openai-news --include "media/screenshots-*.tar.gz" --local-dir ./
for f in media/screenshots-*.tar.gz; do tar -xzf "$f"; done

---

🗂️ Section & Topic Distribution

Section Name	Article Count	Clean Words	cl100k Tokens	% of Corpus	Subject Matter & Scope
Company	230	199,502	256,810	33.7%	Corporate governance, philanthropic partnerships, educational initiatives, infrastructure milestones
Product	181	241,205	330,417	26.5%	Public foundation model releases, ChatGPT capabilities, API platform tooling, and enterprise features
Safety	147	178,404	226,742	21.6%	Frontier AI safety evaluations, red teaming, preparedness framework, watermarking, and cyber defense
Research	43	80,913	105,166	6.3%	Fundamental deep learning theory, RL algorithms, reasoning, alignment, and model interpretability
Security	34	45,779	57,672	5.0%	Threat intelligence, counter-espionage investigations, disruption of malicious influence campaigns
Engineering	20	62,246	83,108	2.9%	Large-scale infrastructure engineering, distributed supercomputing clusters, inference optimization
Global Affairs	16	16,411	20,404	2.3%	International policy, multilateral governance, educational equity, economic impact studies
AI Adoption	9	9,364	11,848	1.3%	Real-world institutional deployments across scientific labs, enterprise systems, and classrooms
new platform	1	2,951	3,507	0.1%	Frontier technical documentation and releases
this blog	1	1,839	2,203	0.1%	Frontier technical documentation and releases

---

📅 Historical Yearly Distribution

Year	Article Count	Total Words	cl100k Tokens	Major Historic Milestones
2026	396	566,702	742,544	Rosalind Academic Plan, Sora-2 System Cards, AI-enabled False Front Disruptions, Teens Learning & Planning
2025	85	91,505	120,816	Sora 2 General Release, GPT-4.5 System Card, GPT-4.1 Architecture, Disney Collaboration, Sycophancy Studies
2024	180	166,685	216,929	GPT-4o Omnimodal Model Launch, Sora World Simulator Debut, Advanced Voice Mode, Strawberry/o1 Reasoning
2023	13	8,624	10,976	GPT-4 Landmark Launch, DALL-E 3 ChatGPT Integration, Custom GPTs, Superalignment Mission
2020	8	5,098	6,612	GPT-3 (175B Parameters) Milestone, Scaling Laws for Neural Language Models, OpenAI API Commercial Launch

---

📐 Dataset Schema

Column	Type	Description
url	string	Canonical permanent URL to the official OpenAI publication
slug	string	Normalized URL slug identifier (e.g. introducing-the-rosalind-academic-plan)
title	string	Publication headline
description	string	SEO summary / meta description
authors	list<string>	Documented authors or corporate research collective
date_published	string	ISO 8601 publication timestamp
date_modified	string	ISO 8601 modification timestamp (if available)
categories	list<string>	Editorial categories (Research, Product, Safety, etc.)
tags	list<string>	Associated topic and year tags
word_count	int64	Word count of the cleaned body text
character_count	int64	Total character count of cleaned body text
cl100k_tokens	int64	Exact OpenAI cl100k_base BPE token count
body_text	string	Normalized plain text content
body_markdown	string	Structured markdown with headings, code blocks, lists
featured_image	string	Remote CDN URL for high-resolution cover image
featured_image_path	string	Local relative filesystem path to downloaded cover image (images/<slug>/featured.*)
inline_images	list<string>	Remote CDN URLs for inline architectural illustrations
inline_image_paths	list<string>	Local relative filesystem paths to downloaded inline images (images/<slug>/inline_*.*)
screenshot_path	string	Local relative filesystem path to full-page render screenshot (screenshots/<slug>.png)

---

🚀 Quick Start Examples

Load via Hugging Face datasets

from datasets import load_dataset

# Automatically loads and concatenates all train-*.parquet shards
dataset = load_dataset("{repo_id}", split="train")

print(f"Total articles loaded: {len(dataset)}")
sample = dataset[0]
print(f"Latest: {sample['title']} ({sample['date_published']})")
print(f"Tokens: {sample['cl100k_tokens']:,} cl100k")

Streaming Iteration

dataset = load_dataset("{repo_id}", split="train", streaming=True)
for article in dataset:
    print(f"Title: {article['title']}")
    print(f"Hero Image: {article['featured_image_path']}")
    print(f"Screenshot: {article['screenshot_path']}")
    break

---

⚖️ License & Attribution

Corpus texts and assets are property of OpenAI. This dataset compilation and pipeline is archived under Apache 2.0 by Sphita AI Research.