CuratedCorpus-5B

A 5-Billion Token Sharded Pretraining Corpus Engineered for Compute-Optimal Linear, Hybrid & Transformer Language Models.

---

1. Executive Summary

CuratedCorpus-5B is an open-access, production-grade language modeling pretraining dataset comprising 5,000,000,000 tokens (4,882,800 packed sequences of length $L=1024$).

Designed by MiamnusAI, CuratedCorpus-5B addresses the fundamental bottleneck in modern small and medium language model pretraining: data quality density versus volume. While web-scraped dumps often suffer from high noise, duplicate fragments, and low signal-to-noise ratios, CuratedCorpus-5B applies rigorous heuristic filtering, MinHash LSH deduplication, and a balanced domain mixture optimized for reasoning, factual recall, and syntactic fluency.

The dataset is partitioned into 10 independent, pre-mixed, pre-shuffled Apache Parquet shards (500M tokens each) plus a dedicated validation set, enabling zero-overhead distributed streaming and seamless checkpoint resumption without data repetition.

---

2. Dataset Composition & Domain Mixture

CuratedCorpus-5B implements an empirically validated 4-way domain mixture formulated to maximize learning signal per compute flop:

Domain	Allocation	Token Budget	Primary Knowledge Sources & Objectives
Educational Web	60.0%	3.00 Billion	Filtered knowledge corpora, encyclopedic domains, and curated high-signal web text stripped of navigation and boilerplate.
High-Quality Textbooks	25.0%	1.25 Billion	University-grade and secondary education materials spanning computer science, physics, biology, history, and humanities.
STEM & Mathematical Reasoning	10.0%	500 Million	Mathematical problem formulations, algorithmic logic, formal step-by-step reasoning, and proof structures.
Synthetic Narratives	5.0%	250 Million	High-coherence narrative progressions with controlled vocabulary, multi-agent character interactions, and explicit moral/logical consistency.

CuratedCorpus-5B Domain Mixture:
┌─────────────────────────────────────────────────────────────┐
│ ████████████████████████ Educational Web (60%)              │
│ ██████████ Textbooks (25%)                                  │
│ ████ STEM / Math (10%)                                      │
│ ██ Synthetic Narratives (5%)                                │
└─────────────────────────────────────────────────────────────┘

---

3. Technical Specifications

• Total Tokens: 5,000,000,000 (pre-tokenized)
• Total Packed Sequences: 4,882,800 sequences ($1024$ tokens/sequence)
• Tokenizer: Custom Byte-Pair Encoding (BPE)
• Vocabulary Size: 4,096 tokens (tokenizer_4k.json included in repo)
  • Designed for high token compression while minimizing embedding table parameters, compute overhead, and memory footprint in edge and hybrid architectures.
• Storage Format: Apache Parquet (Snappy compressed, columnar input_ids)
• Sharding Schema:
  • data/train-00000-of-00010.parquet to data/train-00009-of-00010.parquet (488,280 sequences / 500M tokens each)
  • data/validation-00000-of-00001.parquet (25,000 sequences / 25.6M tokens)

---

4. Data Quality & Safety Pipeline

CuratedCorpus-5B was constructed through an automated end-to-end data curation pipeline adhering to enterprise responsible AI standards:

1. Deduplication: Document-level MinHash Locality-Sensitive Hashing (LSH) with 128 permutations across 5-gram shingles at a Jaccard threshold of 0.85 to eliminate repetitive web spans.
2. Quality & Perplexity Filtering: Documents filtered using statistical text heuristics (mean word length, symbol-to-alphabet ratio, line-level repetitive n-grams, and language identification confidence > 0.98).
3. Personally Identifiable Information (PII) Scrubbing: Automated regex and named-entity recognition filters scrubbed sensitive PII, including private email addresses, telephone numbers, IP addresses, and personal identification numbers.
4. Toxicity & Harm Mitigation: Strict elimination of explicit NSFW content, hate speech, malicious code fragments, and non-educational spam.

---

5. Quickstart & Usage

Method A: Hugging Face Datasets

from datasets import load_dataset

# Stream directly without downloading the entire 5.5GB
dataset = load_dataset("MiamnusAI/CuratedCorpus-5B", split="train", streaming=True)

for sample in dataset.take(5):
    print("Sequence length:", len(sample["input_ids"]))

Method B: PyArrow / Polars (High-Throughput Distributed Loading)

import pyarrow.parquet as pq

# Read a specific 500M token shard directly
table = pq.read_table("data/train-00000-of-00010.parquet", columns=["input_ids"])
input_ids = table["input_ids"].to_numpy(zero_copy_only=False)
print(f"Loaded {len(input_ids)} sequences of length {len(input_ids[0])}")

Method C: PyTorch CUDA Training Loop

import torch
import numpy as np
import pyarrow.parquet as pq

def get_batch(shard_path, batch_size=32, block_size=1024, device="cuda"):
    table = pq.read_table(shard_path, columns=["input_ids"])
    flat_tokens = np.stack(table["input_ids"].to_numpy()).ravel().astype(np.int64)
    max_idx = len(flat_tokens) - (block_size + 1)
    idxs = np.random.randint(0, max_idx, size=batch_size)
    x = torch.from_numpy(np.stack([flat_tokens[i : i + block_size] for i in idxs])).to(device)
    y = torch.from_numpy(np.stack([flat_tokens[i + 1 : i + 1 + block_size] for i in idxs])).to(device)
    return x, y

---

6. Licensing, Legal Governance & Fair Use

• Primary License: Apache License 2.0. You are free to use, modify, distribute, and build commercial or research applications upon this corpus.
• Attribution Notice: Users of this dataset must provide appropriate credit to MiamnusAI in technical reports, documentation, or published artifacts.
• Disclaimer: The dataset is provided "as is", without warranty of any kind. MiamnusAI disclaims liability for any downstream generations or usage of models trained on this corpus.
• Content Removal & DMCA Inquiries: To request content review or redaction, please contact: legal@miamnus.ai.

---

7. Citation

If you use CuratedCorpus-5B in your research, training runs, or benchmarking studies, please cite:

@misc{miamnusai2026curatedcorpus5b,
  title={CuratedCorpus-5B: A High-Throughput Sharded Pretraining Corpus for Linear, Hybrid, and Transformer Language Models},
  author={MiamnusAI Research Team},
  year={2026},
  publisher={Hugging Face},
  howpublished={\url{https://huggingface.co/datasets/MiamnusAI/CuratedCorpus-5B}}
}