hanseungwook/nemotron-specialized-100b-packed

Nemotron Specialized 100B — Nanochat packed

Training-ready, pretokenized shards derived from nvidia/Nemotron-Pretraining-Specialized-v1 at revision 9ed3718b5f2ae29074c5e34e64115432b7c4320f.

The data was selected, packed, and deterministically shuffled for Nanochat language-model pretraining. It is the exact corpus used by the matched AR, MTP-4, and RSM experiments in hanseungwook/rsm-llm.

Contents

Split	Segments	Sequences	Training tokens
train_50b	train_segment_000	24,254,720	49,673,666,560
train_100b	train_segment_000 + train_segment_001	48,509,440	99,347,333,120
validation	six source-specific shards	12,288	25,165,824

Each training row contains 2,049 little-endian uint16 token IDs: a 2,048-token context plus the final shifted target. Training shards are already globally shuffled. manifest.json records every shard path, size, SHA-256 digest, source mixture, preprocessing seed, tokenizer revision, and named split. The small .source.bin companions hold one source ID per packed row.

The included tokenizer/ directory is the pinned artifact from karpathy/nanochat-d32 at revision 016dba034c9c0ca9033ad1bc721bceff54680600.

Download

The complete repository is about 199 GB. snapshot_download is resumable.

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="hanseungwook/nemotron-specialized-100b-packed",
    repo_type="dataset",
    local_dir="/path/to/nemotron-specialized-100b-packed",
)

To download only the 50B training split, keep segment 0 and the supporting metadata:

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="hanseungwook/nemotron-specialized-100b-packed",
    repo_type="dataset",
    local_dir="/path/to/nemotron-specialized-100b-packed",
    allow_patterns=[
        "manifest.json",
        "provenance/*",
        "tokenizer/*",
        "train_segment_000/*",
        "validation/*/*",
        "verification/*",
    ],
)

Train with RSM-LLM

From a checkout of hanseungwook/rsm-llm, set DATA_DIR to the downloaded snapshot and launch a training variant. For example, this runs the 100B RSM configuration:

export DATA_DIR=/path/to/nemotron-specialized-100b-packed
export NANOCHAT_BASE_DIR=/path/to/nanochat-runtime

DATASET_MANIFEST="$DATA_DIR/manifest.json" \
TOKENIZER_DIR="$DATA_DIR/tokenizer" \
DATASET_SPLIT=train_100b \
bash runs/nemotron_mtp_baselines.sh rsm

The launcher also accepts ar, mtp, rsm-gamma, and rsm-gamma-film. Use DATA_CACHE_DIR to copy active shards to node-local storage during training.

Integrity and provenance

The completed local verification pass checked all 194 manifest-referenced binary files (198,890,550,784 bytes). The verification record is included at verification/complete.json; full construction metadata is under provenance/. Consumers can validate the portable manifest with nanochat.packed_data.load_manifest and use the shard digests in manifest.json for a full byte-level audit.

License and attribution

This is a format conversion and deterministic mixture of the NVIDIA dataset; it does not replace or narrow any upstream terms. Most source samples are CC BY 4.0. The upstream card states that Wiki-Rewrite is CC BY-SA 4.0 and Scientific-Coding is GFDL, and it identifies generator-model licenses that may apply to downstream models. The repository is therefore tagged license: other.

Review the nvidia/Nemotron-Pretraining-Specialized-v1 dataset card before redistributing this data or a trained model, and retain NVIDIA's attribution. If you use the data in research, cite the NVIDIA Nemotron 3 Nano technical report listed on the upstream card.