Datasets:
File size: 3,098 Bytes
c635ede | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 | ---
license: apache-2.0
task_categories:
- text-generation
language:
- en
tags:
- web-crawl
- pretraining
- nlp
- text-corpus
pretty_name: Web Crawl 2026
size_categories:
- 10B<n<100B
---
# Web Crawl 2026
A large-scale web crawl dataset for language model pretraining, collected by the OpenTransformer project.
## Dataset Description
This dataset contains text extracted from web pages crawled directly from the internet using custom high-throughput crawlers. All data is freshly scraped.
### Data Format
Each record is a JSON line (gzipped) with fields:
- text: extracted text content (200-200,000 chars)
- url: source URL
- domain: source domain
- timestamp: crawl timestamp (ISO 8601)
- source: crawler identifier (crawl_rust_v1, crawl_go_v1, crawl_v5)
### Collection Methods
Three crawlers run in parallel on a Vast.ai GPU box (Titan Xp, $0.06/hr):
| Crawler | Language | Throughput | 1.2GB Chunk Time | Architecture |
|---------|----------|------------|-------------------|-------------|
| crawl_rust | Rust | 150-300 docs/s | 5-6 min | 500 async workers, tokio |
| crawl_go | Go | 11 docs/s | ~2 hrs | 150 goroutines |
| crawl_v5.py | Python | 0.8 docs/s | ~25 hrs | 20 async workers |
The Rust crawler is 27x faster than Go and 375x faster than Python.
### Rust Crawler Architecture
Source code: crawler/rust/
Key design decisions:
- 500 concurrent async workers via tokio + semaphore-based backpressure
- Background queue refiller: seed fetching runs in a separate task, never blocks crawling
- Pre-generated seed file: 593K URLs from Common Crawl index (12 crawl versions x 20 TLD patterns)
- Link discovery: extracts up to 50 links per crawled page, shuffled for domain diversity
- Content dedup: MD5 hash of first 500 chars, stored in DashMap (lock-free concurrent hashmap)
- Domain throttling: max 1000 pages per domain to ensure diversity
- Streaming gzip: writes compressed JSONL chunks (~1.2GB raw -> ~350MB compressed)
- Auto-upload: each completed chunk is uploaded to HuggingFace Hub
Seed sources:
1. Common Crawl URL index (CC-MAIN-2024-10 through CC-MAIN-2025-08)
2. Wikipedia random articles API (20K articles)
3. Sitemaps from 34 major sites (Reuters, BBC, Nature, StackOverflow, etc.)
4. Hacker News top/new/best stories
Performance on Titan Xp box ($0.06/hr):
- Phase 1: 562K seeds loaded in 28 seconds
- Phase 2: 150-300 docs/s sustained throughput
- ~1.2GB chunk every 5-6 minutes
- ~12-15 GB/hour of raw crawled text
- Cost: ~$0.004 per GB of crawled text
### Building and Running
Install Rust, then:
cd crawler/rust
cargo build --release
ulimit -n 65536
./target/release/crawl_rust > crawl.log 2>&1
### Quality Filtering
- HTML text extraction via scraper crate (article/main/body selectors)
- Minimum 200 chars, maximum 200K chars
- Content-type filtering (only text/html)
- URL filtering: blocks social media, login pages, media files, admin pages
- Deduplication via MD5 content hash
## Intended Use
Pretraining data for the AGILLM-3 language model (698M params, joint AR+SAT architecture).
## License
Apache 2.0
|