sphita/anthropic-crawl

🏛️ Anthropic Complete Web Corpus & Research Archive

High-fidelity web crawl, untouched DOM HTML, full Markdown corpora, technical papers, safety frameworks, and complete media assets from Anthropic's public web estate.

Overview · Snapshots · Languages · Token Statistics · Schema · File Listing · Quick Start · Crawler Engine

---

📌 Overview & Executive Summary

The Anthropic Complete Web Corpus & Research Archive is a curated, high-fidelity archive capturing the entirety of Anthropic's public-facing digital footprint (www.anthropic.com and claude.com). Published by Sphita AI Research, this dataset is designed for frontier AI safety research, web-agent grounding, language model pre-training, and longitudinal study of frontier model development.

Captured via Sphita's high-concurrency Rust crawler engine (queen-crawler), this archive preserves both the raw Server-Side Rendered (SSR) HTML DOM trees and structured, clean Markdown texts alongside all associated media assets (research PDFs, system cards, product diagrams, logos, audio, and presentation videos).

Key Highlights

• Dual-Corpus Representation: Every crawled page includes both clean, structured Markdown and the original untouched HTML DOM.
• Exact Token Auditing: Fully profiled using OpenAI and Claude standard cl100k_base BPE tokenizer, measuring over 21.14 Million total tokens in the latest release.
• Frontier Safety & Model Documentation: Exhaustive coverage of Claude model releases (Claude Mythos 5/5.1, Claude Fable 5/5.1, Claude Sonnet 5.5, Claude Opus 4.x/5, Claude Haiku), Responsible Scaling Policies (ASL-3 & ASL-4), and frontier threat intelligence reports.
• Dense Media Archives: Bundled into standalone, SHA256-verified .tar.xz archives with exact URL-to-asset mapping preserved in assets_metadata.parquet.

---

📅 Multi-Snapshot Releases

The dataset is partitioned into timestamped snapshot directories reflecting regular crawl intervals. Users can query the latest state or pin reproducible training pipelines to a historical snapshot.

Snapshot Directory	Crawl Date	Status	Unique Pages	cl100k Tokens (MD / Total)	Media Assets	Parquet Corpus Size
2026-10-15/	2026-10-15	Latest (Active)	528	3.27M / 21.15M	4,518 files (16 archives, 1.72 GB)	15.89 MB (Zstandard-7)
2026-09-29/	2026-09-29	Historical	586	~3.30M / ~20.5M	4,000+ files (19 archives, 2.30 GB)	13.88 MB (Zstandard-7)

---

🌐 Language & Multilingual Distribution

Languages and writing scripts are identified dynamically across DOM HTML attributes and Unicode script distributions:

Language	ISO Code	Page Occurrences	Notes & Context
English	en	528 pages (100.0%)	Primary language of corpus, technical papers, and web pages
Russian	ru	3 pages (0.6%)	Threat intelligence research and multilingual crawler guidance
Japanese	ja	3 pages (0.6%)	International events (e.g. Tokyo Builder Summit) and localized documentation
Korean	ko	2 pages (0.4%)	Multilingual crawler opt-out and API guidance
Chinese	zh	1 page (0.2%)	Threat intelligence indicators and localized articles
Arabic	ar	1 page (0.2%)	Threat intelligence research and multilingual security indicators

---

📊 Comprehensive Corpus & Token Statistics

Token volume calculated with OpenAI / Claude cl100k_base BPE tokenizer against the latest snapshot (2026-10-15).

Metric	Markdown Text	Raw HTML DOM	Combined Payload
Total Tokens (cl100k_base)	3,268,588	17,876,539	21,145,127
Total Words	1,487,597	—	—
Total Raw Character Size	13,600,469 chars (12.97 MB)	144,200,300 bytes (137.52 MB)	150.49 MB
Average Tokens per Page	6,190.5 tokens	33,857.1 tokens	40,047.6 tokens
Median Tokens per Page	5,014 tokens	31,805 tokens	—
Parquet Size (Zstandard-7)	—	—	15.89 MB
Columnar Compression Ratio	—	—	9.47x reduction

---

🗂️ Category & Section Breakdown

The dataset covers all major domains of Anthropic's web presence, categorized by primary URL routes:

Section Route	Pages	Markdown Words	cl100k Tokens	% of Corpus	Subject Matter & Content Scope
/news	182	343,704	765,025	23.4%	Model launches, product updates, system cards, partnerships
/research	114	392,805	752,924	23.0%	Frontier AI alignment, mechanistic interpretability, safety
/legal	43	169,873	282,700	8.6%	Terms of service, privacy policies, Responsible Scaling Policy (RSP)
/events	34	64,760	245,026	7.5%	Industry webinars, AI safety symposiums, developer sessions
/resources	31	46,478	147,651	4.5%	Enterprise deployment guides, developer prompts, whitepapers
/engineering	26	82,148	154,714	4.7%	Cluster infrastructure, distributed systems, internal tooling
/customers	10	16,892	51,119	1.6%	Case studies, production customer stories
/careers	8	63,014	199,312	6.1%	Technical competencies, research roles, organization structure
/solutions	8	20,139	57,032	1.7%	Enterprise offerings, specialized industry workflows
/beneficial-deployments	5	15,600	35,440	1.1%	AI for science, public interest projects, non-profit initiatives

---

🏆 Landmark Publications & Longest Documents

The corpus preserves critical foundational papers and regulatory filings in unclipped form:

Document Title	Section	Words	cl100k Tokens	Canonical Target URL
Countering Misuse of AI (Threat Intelligence)	/threat-intelligence-report	38,206	59,484	https://www.anthropic.com/threat-intelligence-report-september-2026
Claude’s Constitution	/constitution	30,802	38,373	https://www.anthropic.com/constitution
Anthropic Transparency Hub	/transparency	23,732	38,363	https://www.anthropic.com/transparency
Frontier Model Transparency Report	/transparency/model-report	23,732	38,363	https://www.anthropic.com/transparency/model-report
Does Anthropic Crawl Data From The Web And How Can Site Owners Block The Crawler	/en/articles	4,665	36,052	https://support.claude.com/en/articles/8896518-does-anthropic-crawl-data-from-the-web-and-how-can-site-owners-block-the-crawler
How Can I Access The Claude API	/en/articles	4,180	34,151	https://support.claude.com/en/articles/8114521-how-can-i-access-the-claude-api
Research & Engineering Career Competencies	/careers/jobs	8,744	27,865	https://www.anthropic.com/careers/jobs

---

🤖 Frontier Models & Safety Index

Index of tracked model generations, safety standards, and operational frameworks:

Frontier Model / Concept	Mentions	Architectural Significance
Claude Opus (4.5 / 4.6 / 4.8 / 5)	1,494	Deep cognitive reasoning, complex math, multi-step planning
Claude Fable (5 & 5.1)	790	Specialized agentic software engineering and theorem proving
Claude Mythos (5 & 5.1)	679	High-assurance cybersecurity analysis and computational biology
Claude Haiku (4.5 & 5.5)	251	Ultra-low latency edge inferencing and compact processing
Claude Sonnet 5.5	213	Balanced frontier capabilities and coding throughput
ASL-3 & ASL-4 (RSP Levels)	147	AI Safety Level containment standards & autonomous threat mitigations
Enterprise Frontier Safeguards (EFS)	59	Customer tenant isolation and cryptographic verification

---

📦 Media Archives Breakdown

All embedded assets discovered across the crawl were downloaded, deduplicated, and packaged into compressed .tar.xz archives:

Media Category	File Count	Archive Files	Total Size
Visual Assets (Images / Badges / Diagrams)	2,998 files	12 .tar.xz archives	951.46 MB
Documents (PDF Papers / Technical Reports)	269 files	2 .tar.xz archives	676.88 MB
Video Assets (Demos / Keynotes)	58 files	1 .tar.xz archive	90.66 MB
Audio Assets (Audio Clips / Recordings)	1 file	1 .tar.xz archive	0.59 MB
Total Media Volume	4,518 assets	16 archives	1,719.5 MB

---

🗃️ Dataset Schema

The primary data files are stored in Apache Parquet format with Zstandard (level 7) compression.

1. data/pages.parquet (Main Text Corpus)

Column Name	Apache Arrow Type	Description
url	string	Canonical URL of the web page
title	string	Extracted HTML title header
markdown	string	High-fidelity extracted markdown body text
html	string	Complete raw HTML DOM delivered by Server-Side Rendering
word_count	int64	Word count of the markdown content
char_count	int64	Character count of the markdown content
links	list<string>	Complete array of outbound hyperlinks discovered on the page
assets	list<struct<url, asset_type>>	All embedded media tags (images, videos, PDFs) linked on this page

2. data/assets_metadata.parquet (Media Catalog)

Column Name	Apache Arrow Type	Description
source_url	string	Webpage URL where the asset was discovered
asset_url	string	Absolute URL of the asset
asset_type	string	Category: image, document, video, or audio
is_downloaded	bool	Whether the asset was successfully fetched
local_filename	string	Path inside the respective .tar.xz archive
file_size_bytes	int64	Size of the asset file in bytes
extension	string	File format (.pdf, .png, .webp, .svg, .mp4, etc.)

3. data/errored.parquet (Crawler Telemetry)

Column Name	Apache Arrow Type	Description
url	string	Target URL that encountered an exception
error	string	HTTP error code or parser failure detail
timestamp	int64	Epoch timestamp of failure
is_permanent	bool	Whether failure was 404/permanent or transient

---

📁 Complete Repository File Listing

The repository contains the following structured hierarchy across both releases:

sphita/anthropic-crawl/
├── .gitattributes
├── README.md                                          # Master Dataset Card & Documentation
│
├── 2026-10-15/                                        # [LATEST RELEASE]
│   ├── README.md                                      # Snapshot-specific Dataset Card (11.28 KB)
│   ├── SHA256SUMS.txt                                 # Cryptographic checksum manifest (2.46 KB)
│   ├── data/
│   │   ├── pages.parquet                              # 528 crawled pages (15.89 MB)
│   │   ├── assets_metadata.parquet                    # 4,518 asset catalog records (0.87 MB)
│   │   └── errored.parquet                            # 127 crawler dropout logs (4.0 KB)
│   ├── metadata/
│   │   ├── head.json                                  # Crawler execution telemetry (pages, bytes, timing)
│   │   ├── manifest.json                              # Stream manifest configuration
│   │   ├── config.json                                # Engine configuration (seeds, depth, rate limits)
│   │   └── discovered_subdomains.txt                  # Discovered endpoints (www.anthropic.com)
│   └── media_archives/
│       ├── audio/
│       │   └── Audio_loose.tar.xz                     # Audio assets (0.59 MB)
│       ├── documents/
│       │   ├── Document_1792067320.tar.xz             # Research papers & legal filings (211.84 MB)
│       │   └── Document_loose.tar.xz                  # General documents & whitepapers (465.04 MB)
│       ├── images/
│       │   ├── Image_1792067478.tar.xz                (40.16 MB)
│       │   ├── Image_1792067580.tar.xz                (42.88 MB)
│       │   ├── Image_1792067654.tar.xz                (46.80 MB)
│       │   ├── Image_1792067742.tar.xz                (49.65 MB)
│       │   ├── Image_1792067810.tar.xz                (55.06 MB)
│       │   ├── Image_1792067963.tar.xz                (82.80 MB)
│       │   ├── Image_1792068099.tar.xz                (91.24 MB)
│       │   ├── Image_1792068203.tar.xz                (99.98 MB)
│       │   ├── Image_1792068330.tar.xz                (104.58 MB)
│       │   ├── Image_1792068447.tar.xz                (109.31 MB)
│       │   ├── Image_1792068544.tar.xz                (111.36 MB)
│       │   └── Image_loose.tar.xz                     (117.54 MB)
│       └── videos/
│           └── Video_loose.tar.xz                     # Product demonstrations & keynotes (90.66 MB)
│
└── 2026-09-29/                                        # [HISTORICAL RELEASE]
    ├── README.md                                      # 2026-09-29 Dataset Card (3.19 KB)
    ├── SHA256SUMS.txt                                 # Cryptographic checksum manifest
    ├── data/
    │   ├── pages.parquet                              # 586 crawled pages (13.88 MB)
    │   ├── assets_metadata.parquet                    # Media catalog (0.77 MB)
    │   └── errored.parquet                            # Error logs (4.0 KB)
    ├── metadata/
    │   ├── head.json
    │   ├── manifest.json
    │   ├── config.json
    │   └── discovered_subdomains.txt
    └── media_archives/
        ├── audio/
        │   └── Audio_loose.tar.xz                     # Audio assets (0.59 MB)
        ├── documents/
        │   ├── Document_1790688970.tar.xz             (168.21 MB)
        │   └── Document_loose.tar.xz                  (537.82 MB)
        ├── images/                                    # 17 image archives (.tar.xz, ~1.48 GB total)
        └── videos/
            └── Video_loose.tar.xz                     (337.40 MB)

---

🚀 Quick Start Examples

1. Load via Hugging Face datasets (Python)

from datasets import load_dataset

# 1. Load the latest default snapshot (2026-10-15 pages corpus)
pages = load_dataset("sphita/anthropic-crawl")
print(f"Loaded {len(pages['train'])} pages.")
print("Sample Title:", pages['train'][0]['title'])
print("Sample URL:", pages['train'][0]['url'])

# 2. Load with streaming (zero RAM overhead)
streaming_pages = load_dataset("sphita/anthropic-crawl", split="train", streaming=True)
first_page = next(iter(streaming_pages))
print("First stream page length:", len(first_page['markdown']))

# 3. Load historical snapshot (2026-09-29)
hist_pages = load_dataset("sphita/anthropic-crawl", "pages-2026-09-29")
print(f"Historical pages: {len(hist_pages['train'])}")

# 4. Filter research papers
research = pages['train'].filter(lambda x: "/research/" in x['url'])
print(f"Total Research Publications: {len(research)}")

2. Instant SQL over Parquet with DuckDB

You can run direct SQL analytics against remote or local Parquet files without ingesting into a database:

-- Top 5 longest research papers by markdown word count
SELECT 
    title, 
    word_count, 
    url 
FROM '2026-10-15/data/pages.parquet' 
WHERE url LIKE '%/research/%' 
ORDER BY word_count DESC 
LIMIT 5;

-- Count total assets categorized by extension
SELECT 
    extension, 
    count(*) AS total_files, 
    sum(file_size_bytes) / (1024 * 1024) AS total_mb
FROM '2026-10-15/data/assets_metadata.parquet'
GROUP BY extension
ORDER BY total_files DESC;

3. Load with Pandas & PyArrow (Zero-Bloat Column Projection)

import pyarrow.parquet as pq

# Load only necessary metadata columns into memory
table = pq.read_table(
    "2026-10-15/data/pages.parquet", 
    columns=["url", "title", "word_count", "char_count"]
)
df = table.to_pandas()
print(df.sort_values(by="word_count", ascending=False).head(10))

4. Download via Hugging Face CLI (hf)

# Download only the latest 2026-10-15 parquet data files
~/.venv/bin/hf download sphita/anthropic-crawl \
  --repo-type dataset \
  --include "2026-10-15/data/*" \
  --local-dir ./anthropic-crawl-data

# Download all research PDFs and documents
~/.venv/bin/hf download sphita/anthropic-crawl \
  --repo-type dataset \
  --include "2026-10-15/media_archives/documents/*" \
  --local-dir ./anthropic-docs

5. Extracting Media Archives

# Extract all downloaded research PDFs
mkdir -p ./unpacked_documents
tar -xf 2026-10-15/media_archives/documents/Document_loose.tar.xz -C ./unpacked_documents/

# Extract audio archive
mkdir -p ./unpacked_audio
tar -xf 2026-10-15/media_archives/audio/Audio_loose.tar.xz -C ./unpacked_audio/

# Extract visual assets
mkdir -p ./unpacked_images
tar -xf 2026-10-15/media_archives/images/Image_loose.tar.xz -C ./unpacked_images/

---

⚙️ Crawler Engine & Pipeline Architecture

flowchart TD
    A["Target Domains<br/>(anthropic.com, claude.com)"] --> B["queen-crawler Engine<br/>(Rust Concurrency Node)"]
    B --> C["Politeness & Proxy Fleet<br/>(TLS Impersonation & 1-3s delays)"]
    C --> D["HTML DOM Capture & Asset Ingestion"]
    D --> E["Deduplication & Markdown Extraction"]
    E --> F["Parquet Conversion<br/>(Zstandard Level 7)"]
    D --> G["Media Packaging<br/>(.tar.xz Archives)"]
    F --> H["Hugging Face Hub<br/>(sphita/anthropic-crawl)"]
    G --> H

• Crawler Engine: queen-crawler — custom multi-threaded Rust crawler node with async runtime.
• Network & Security: Dynamic rotating proxy pool, Cloudflare TLS fingerprint preservation, and domain-scoped polite delays (1.0s – 3.0s).
• DOM Transformation: High-precision HTML5 parser with automated layout normalization, extracting clean Markdown while preserving original SSR HTML payloads.
• Storage Compression: Columnar Apache Parquet with Zstandard level 7 compression for high read throughput and 9.47x compression factor.

---

🔒 Cryptographic Verification

All files published in each release are cataloged with SHA-256 hashes to guarantee data integrity against transport corruption or tampering.

To verify a downloaded snapshot:

cd 2026-10-15
sha256sum -c SHA256SUMS.txt

---

📜 License

This dataset is published under the GNU General Public License v3.0 (GPL-3.0).
The scraped HTML and Markdown contents remain subject to Anthropic's public terms and copyright where applicable.

---

🌐 Sphita Research Collective & Leadership

Sphita is an open-first AI collective dedicated to engineering datasets, advancing agentic reasoning, and providing high-fidelity infrastructure for the open-source community.

• Lead & Creator: Adhyansh Verma (GitHub: @AVadhyanshverma)
• Organization Hugging Face: huggingface.co/sphita
• Organization GitHub: github.com/sphita
• Discord Community: Join the Sphita Server
• Profile & Hub Space: sphita/README