50 billion tokens, pretokenized with cl32k, split by source family into training, validation and test. This is a Megatron indexed binary dataset, not a raw-text dataset for datasets.load_dataset(). The .bin files are uncompressed and ready to load after downloading. The viewer is disabled because the Parquet files contain provenance, not training examples.

Split	Tokens	Documents	Token files
Train	49,749,849,599	61,304,861	32
Validation	200,120,258	239,403	3
Test	50,030,143	62,746	3
Validation-small	5,001,214	Included in validation	3

The fixed small validation set overlaps full validation intentionally. The union of train, validation and test contains exactly 50,000,000,000 tokens and 61,607,010 documents. Counts include one EOS per document. Small deviations from the 200M/50M heldout targets preserve whole families.

Source	Unique tokens	Share
UltraData-Code-L2 Python	29,992,673,704	59.985%
UltraData-Code-L3 Python	4,603,363,531	9.207%
FineWeb-Edu English	15,403,962,765	30.808%

Each split approximates this mix. Python-source tokens include comments and English task descriptions, especially in L3; these percentages describe dataset sources, not token-level language classification.

Download and load

Download the complete package with the Hugging Face CLI, then verify it on a shared training filesystem:

uvx --from huggingface_hub hf download YOUR_DATASET_ID --repo-type dataset --local-dir dataset
cd dataset
sha256sum -c SHA256SUMS

The package is approximately 102.5GB including provenance. Token binaries occupy 100.01GB including validation-small; indexes and IDs add about 1.73GB. No decompression or conversion is required. Allow additional space for training index caches and checkpoints.

With a compatible NeMo AutoModel checkout installed:

from pathlib import Path
from transformers import PreTrainedTokenizerFast
from nemo_automodel.components.datasets.llm.megatron.indexed_dataset import IndexedDataset

root = Path("dataset").resolve()
dataset = IndexedDataset(str(root / "train/train-00000-of-00032"))
tokenizer = PreTrainedTokenizerFast.from_pretrained(root / "tokenizer")
print(tokenizer.decode(dataset[0].tolist()))

Every prefix has a .bin, .idx and .ids.npy. The index stores sequence/document boundaries; IDs are aligned int64 provenance identifiers. Token data are little-endian uint16. Training context length is chosen by the loader; documents are packed into fixed-length samples at runtime.

train/manifest.json is the input to AutoModel's RankShardedMegatronPretrainingConfig. Set the recipe's dataset path to this manifest and tokenizer path to tokenizer/. All paths inside manifests are relative. Assign shard i to data-parallel rank i % world_size; supported world sizes are 1, 2, 4, 8, 16 and 32. Eight ranks own four shards each. Use equal sample budgets and avoid applying a second cross-rank sampler after exclusive shard partitioning. Stage the files before launching training; the loader does not automatically download them.

Evaluation prefixes are {validation,test,validation-small}/{l2,l3,english}. Use MegatronPretrainingConfig with explicit split distributions, for example paths={"validation": [str(root / "validation-small/l2")]}, splits_to_build="validation", and trainer_limit_val_batches=1.0 for one epoch. Do not apply another percentage split. Evaluate each source separately and aggregate summed negative log-likelihood over scored tokens. Use validation-small routinely, full validation at milestones and test for final comparisons.

Tokenizer and preparation

cl32k uses a cl100-style pretokenization pattern and byte-level BPE: exactly 32768 vocabulary entries, including 64 reserved IDs. EOS is 0. Reserved/control IDs do not appear in document bodies. The tokenizer is unchanged from the source 50B package; its exact files and SHA-256 are included.

Preparation screens legacy heldouts and benchmark snippets, removes global whitespace-normalized text/solution duplicates, and applies lexical near deduplication: five-token shingles, 64 MinHash values, 16 bands of four, verified Jaccard similarity at least 0.8. Candidate retrieval is approximate. Original whitespace is preserved for tokenization. L3 uses its content field.

All retained Python is included. English was sampled without replacement using proportional quotas across 56 source shards (selection seed 20260914); 56 boundary documents were truncated with EOS preserved to reach exactly 50B tokens. This re-split introduces no further truncation, repetition or retokenization.

Source-family split and limitations

Split seed: 20260916. L2 groups are repositories; English groups are URL hosts; L3 groups are normalized raw-source digests. Groups are connected using eligible exact text/solution matches, L3-to-L2 source links, and historical verified near-duplicate components, including copies removed during deduplication. Whole connected families are assigned to test then validation with deterministic random priority and greedy per-source token quotas. Training documents are then balanced across 32 shards by source and document length and shuffled.

One connected family contains 13.446B tokens and remains entirely in training. The heldouts therefore measure generalization to smaller unseen source families rather than uniform document sampling. No known family, exact text/solution group or historical near-duplicate component crosses train/validation/test. This does not guarantee absence of every similar passage or semantic overlap. Benchmark screening is lexical, not a guarantee of uncontaminated benchmark evaluation. The full filtering/removal policy has not received a separate independent review.

Models trained on the earlier unsplit 50B corpus may have seen the new heldouts. Use fresh scratch runs for clean comparisons. Vocabulary learning was not restricted to these new training families, so this is a neural-model-training holdout, not a tokenizer-training holdout.

provenance/document-splits.parquet records original row, UID, source code, family, split code and length. Source codes: English=0, L2=1, L3=2; split codes: train=0, validation=1, test=2. source-families.parquet maps original groups to connected families. The split plan and manifests record exact counts and checksums.

Sources and terms

UltraData-Code, pinned revision 85182d8: all 119 L2 Python and 147 L3 Python shards.
FineWeb-Edu, pinned revision 87f0914: 56 shards from sample/100BT.

This package does not assign a new blanket license to the source material. Consult the pinned upstream dataset cards and underlying source terms.

Verification

All 41 token binaries match their original SHA-256; indexes, IDs and tokenizer are also verified. The complete original UID union and all token bytes were preserved, with index-length, EOS and token-range checks. The same token/index bytes passed an eight-rank CPU loader test at 4K/8K, with exclusive shard ownership and exact dataloader resume. verification.json records test coverage. SHA256SUMS verifies all distributed files.

Downloads last month: 310

Total file size:

103 GB