sphita/nvidia-blogs

🚀 NVIDIA Official Blogs Tech Dataset

A high-performance, structured, and cleaned multimodal-ready dataset comprising all 2,269 articles published on the official NVIDIA Blog from 2014-11-14 through 2026-10-17.

This dataset is distributed in ZSTD-compressed columnar Parquet chunks (train-*.parquet) and streaming JSONL, offering rapid ingestion, low-memory footprint, and native compatibility with Hugging Face datasets, PyTorch, LangChain, and LlamaIndex.

---

📑 Key Dataset Facts

• Total Articles: 2,269
• Total Word Count: 1,841,409 words
• Average Article Length: 811.6 words
• Archive Time Span: 2014-11-14 to 2026-10-17 (17 years of NVIDIA tech history)
• Sharding: 5 Parquet shards (train-00000-of-00005.parquet to train-00004-of-00005.parquet)
• Multimodal Assets: 97.4% posts include featured images; 56.3% include inline body diagrams; 40.3% link embedded videos.

---

⚡ Quickstart & Tutorials

1. Standard Ingestion with Hugging Face datasets

from datasets import load_dataset

# Load full train split across all shards
dataset = load_dataset("AdhyanshVerma/nvidia-blogs", split="train")

print(f"Loaded {len(dataset)} articles.")
sample = dataset[0]
print(f"Title: {sample['title']}")
print(f"Author: {sample['author_name']} | Published: {sample['date']}")
print(f"Categories: {sample['categories']}")
print(f"Text Snippet:\n{sample['clean_text'][:300]}...")

2. Zero-RAM Streaming Mode (Ideal for Constrained Hardware / Colab)

from datasets import load_dataset

# Stream records on-demand without downloading the entire dataset
stream_ds = load_dataset("AdhyanshVerma/nvidia-blogs", split="train", streaming=True)

for post in stream_ds.take(5):
    print(f"• [{post['id']}] {post['title']} ({post['word_count']} words)")

3. Category Filtering (e.g. AI & Deep Learning)

ai_posts = dataset.filter(lambda x: "AI" in x["categories"] or "Deep Learning" in x["categories"])
print(f"Found {len(ai_posts)} AI-specific articles.")

4. Vector Search & RAG Pipeline Setup (LangChain Example)

from langchain.text_splitter import RecursiveCharacterTextSplitter
from langchain_community.vectorstores import FAISS
from langchain_community.embeddings import HuggingFaceEmbeddings

# Chunk text for retrieval-augmented generation (RAG)
splitter = RecursiveCharacterTextSplitter(chunk_size=500, chunk_overlap=50)
documents = []

for post in dataset:
    chunks = splitter.create_documents(
        texts=[post["clean_text"]],
        metadatas=[{"title": post["title"], "url": post["link"], "date": post["date"]}]
    )
    documents.extend(chunks)

print(f"Generated {len(documents)} RAG chunks.")

---

🗂️ Dataset Schema

Each row contains 28 structured fields:

Field	Type	Description
id	int64	Unique WordPress Post ID
post_type	string	WordPress post type (post)
status	string	Publication status (publish)
title	string	Unescaped human-readable article title
slug	string	URL slug
date	string	Initial publication date (ISO 8601)
date_gmt	string	Initial publication date in GMT
modified_date	string	Last modified / revised timestamp (ISO 8601)
modified_gmt	string	Last modified timestamp in GMT
link	string	Canonical URL to the original post
author_id	int64	Unique author ID
author_name	string	Unescaped author / contributor name
author_url	string	Link to author archive or profile
categories	list<string>	Taxonomy category tags
tags	list<string>	Editorial topic tags
word_count	int32	Word count of the cleaned article body
character_count	int32	Character count of the cleaned article body
seo_description	string	Yoast SEO meta description
seo_title	string	Yoast SEO page title
featured_image_url	string	Direct URL to high-resolution featured image
featured_image_local_path	string	Deterministic local path for downloaded image
image_alt_text	string	Alt text for the featured hero image
inline_images_count	int32	Total count of inline body images
inline_images_urls	list<string>	Direct URLs of all inline images in article
inline_videos_urls	list<string>	URLs of embedded YouTube / Vimeo videos
excerpt	string	Clean plain text summary / lead excerpt
clean_text	string	Cleaned plain text body with normalized whitespace
raw_html	string	Original HTML content (preserved for markup fidelity)

---

📊 Corpus Distributions & Analytics

Top Article Categories

Rank	Category Name	Article Count
1	AI	548
2	Deep Learning	449
3	AI Infrastructure	435
4	Pro Graphics	350
5	Gaming	328
6	Corporate	248
7	Robotics	229
8	Software	171

Top Authors & Contributors

Rank	Author Name	Article Count
1	GeForce NOW Community	270
2	Brian Caulfield	145
3	Gerardo Delgado	126
4	Angie Lee	80
5	Isha Salian	75
6	Rick Merritt	71

---

🔄 Post Revision Tracking & Incremental Updates

An important empirical finding from analyzing the NVIDIA blog archive: 98.5% of posts are updated after initial publication.

NVIDIA frequently amends articles to include updated benchmark scores, new driver releases, OEM laptop availability, partner links, and editorial errata.

This pipeline tracks both initial creation timestamps (date) and modification timestamps (modified_date). When running incremental syncs (python cli.py --task update), the pipeline polls orderby=modified&order=desc to catch revisions to historic posts without redownloading static content.

---

⚖️ Licensing & Attribution

• Dataset curation & pipeline code: Apache-2.0
• Content Attribution: All article text and images are copyright © NVIDIA Corporation. This dataset is provided for academic research, language modeling evaluation, and non-commercial educational use in accordance with Fair Use guidelines.