Exeaon/anite-corpus-5t

ANITE 5-Trillion Token Preprocessed Foundation Corpus

Welcome to the ANITE 5-Trillion Token Preprocessed Corpus, curated and tokenized for training next-generation foundation models with sub-quadratic architectures.

Overview

• Tokens: ~5,000,000,000,000 (5 Trillion)
• Tokenizer: ANITE Production BPE (100,278 vocabulary tokens, padded to 100,352 for Tensor Core warp alignment)
• Format: Zero-copy memory-mapped uint32 token arrays (.npy and .bin)
• License: Apache 2.0

Curriculum & Provenance Sources

The corpus aggregates curated, decontaminated, and high-quality open pretraining sources:

1. Academic & Research: PeS2o (40M open-access STEM papers) and ArXiv STEM preprints.
2. Mathematics & Reasoning: OpenWebMath and Algebraic-Stack (ProofPile 2).
3. Software & Code: StarCoder v1 curated source code across 80+ programming languages.
4. General Knowledge: Wikipedia & Wikibooks (decontaminated).
5. Web & Discourse: DCLM (DataComp-LM) high-quality web shards and filtered Common Crawl.
6. Literature: Project Gutenberg and open books corpus.

Preprocessing Pipeline

Every shard has passed through multi-stage quality filtering:

• HTML artifact and script stripping
• Personally Identifiable Information (PII) redaction with specialized tags (|||PHONE_NUMBER|||, |||EMAIL_ADDRESS|||, |||IP_ADDRESS|||)
• Exact and MinHash Locality-Sensitive Hashing (LSH) near-deduplication
• 13-gram benchmark decontamination against downstream evaluation tasks

Usage with ANITE

from anite.data import MemMapDataset, DataCollator
from torch.utils.data import DataLoader
import numpy as np

# Stream directly with zero memory copy overhead
dataset = MemMapDataset("anite_train.npy", chunk_size=2048, dtype=np.uint32)
loader = DataLoader(dataset, batch_size=32, collate_fn=DataCollator(pad_token_id=100277))

---

Maintained by Exeaon / Zenux Plimver Technologies LTD.