Hikaf143/Keural-MoE-14B-stage1-binary-Dataset

Keural Stage 1 — 69.5 Billion Tokens Binary Dataset

Pre-tokenized binary dataset for the Keural MoE Foundation Model. Ready to use directly for training — no tokenization step needed.

Dataset Overview

Property	Value
Total Tokens	69.5 billion (69,496,921,399)
Format	Binary (.bin) + Index (.idx) + Metadata (.meta)
Total Sequences	15,761,448
Sequence Length	4,096 tokens
Shards	158 shards
Archive Size	~242GB (binary_69B_tokens.tar)
Tokenizer	Keural SentencePiece Unigram, vocab=131,072
Last Updated	2026-04-03

Data Sources

Source	Language	Tokens (approx)
FineWeb	English	~20B
WanJuan Korean	Korean	~5B
Korean WebText	Korean	~4B
ArXiv	English Science	~4B
CC100 Korean	Korean	~3B
PubMed	English Medical	~3B
The Stack v1	Code	~8B
Wikipedia Korean	Korean	~1B
PG19 Literature	English	~1B
Other sources	Mixed	~20.5B

Archive Contents

The tar file contains a binary/ folder with:

• 158 .bin files: Pre-tokenized binary data (keural_000.bin to keural_157.bin)
• 158 .idx files: Index files for fast random access
• 158 .meta files: Metadata JSON for each shard
• build_stats.json: Complete build statistics

Binary Format Specification

File: keural_NNN.bin
─────────────────────────────────────────
HEADER (36 bytes):
  [0:8]   magic   = b"KEURAL\x00\x00"   (8 bytes)
  [8:12]  version = 1                    (uint32 LE)
  [12:20] num_seq                        (uint64 LE)
  [20:28] seq_len = 4096                 (uint64 LE)
  [28:36] padding = 0                    (uint64 LE)

BODY:
  num_seq × 4096 × 4 bytes (uint32 LE tokens)

File: keural_NNN.idx
─────────────────────────────────────────
  [0:4]  num_seq  (uint32)
  [4:8]  seq_len  (uint32)
  per sequence: 8-byte offset + 4-byte length

File: keural_NNN.meta (JSON)
─────────────────────────────────────────
  {"num_sequences": N, "seq_length": 4096, "source": "keural_NNN"}

How to Extract

# Download the tar file, then extract:
tar -xf binary_69B_tokens.tar

# This creates: binary/ directory with 158 shards

How to Use

import struct, mmap, torch

HEADER_FMT  = "<8sIQQQ"
HEADER_SIZE = struct.calcsize(HEADER_FMT)  # 36 bytes

with open("binary/keural_001.bin", "rb") as f:
    raw = f.read(HEADER_SIZE)
    magic, ver, num_seqs, seq_len, _ = struct.unpack(HEADER_FMT, raw)
    print(f"Sequences: {num_seqs}, Length: {seq_len}")

# Or use directly with training scripts:
# torchrun --nproc_per_node=2 train_keural_v2.py --data_dir ./binary

Build Statistics

{
  "documents_processed": 553,711,744,
  "tokens_processed": 69,496,921,399,
  "sequences_written": 15,761,448,
  "padding_added": 3,143,055,066,
  "shards_created": 158,
  "sequence_utilization": "95.13%"
}

Related Resources

• Model Training: github.com/mkd-hossain/Keural-Model-Training
• Tokenizer: huggingface.co/mkd-ai/keural-tokenizer
• Organization: huggingface.co/mkd-ai

Author

MKD CO., LTD.
Keural Foundation Model — Stage 1 pretraining dataset, 2026

License

Portfolio documentation: provenance, reproducibility, and limits

This repository hosts the MKD dataset originally released at mkd-ai/Keural-MoE-14B-stage1-binary-Dataset. Original authorship and license references are preserved above. The token counts, source composition, shard counts, and format description above are original-release claims; the archives have not been unpacked or independently audited for this portfolio.

Actual archive inventory

File	Bytes	Decimal GB
binary_69B_tokens.tar	258,425,139,200	258.4
binary_backup.tar	165,066,188,800	165.1

Use sufficient storage for both the selected archive and extracted contents. Archive names do not establish that both should be combined; inspect manifests and avoid double-counting a backup. Preserve an exact repository revision and SHA256 manifest. Server-side migration verified file sizes and LFS hashes; this verifies transfer integrity, not training quality, token counts, contents, or privacy properties.

License and data-governance boundaries

Metadata specifies license: other. The README links to a training-project license; that link alone does not establish permission for every upstream corpus. Review source-specific rights, redistribution terms, and intended use before reuse. This documentation does not change the original license or grant additional rights.

No real clinical-data de-identification audit, patient-level split, or clinical validation is claimed for these binary archives. PubMed appearing in the source list does not make the corpus an EMR dataset or validate medical generation.

For a small, fully inspectable safeguards example, use the separate synthetic EMR fixture dataset with its generator, patient-grouped partitions, tests, and deterministic report.