ericrcwu/week1-general-20b-dolma2-v1

Week-One General 20B Dolma2

This is a deterministic, pretokenized 20-billion-token baseline corpus for controlled language-model architecture and training experiments. It contains nested 100M, 1B, 5B, and 20B views; each larger view is an exact ordered extension of the previous view. It also includes a dataset-only 370m-1.25xc view: the first 9,281,564,672 packed tokens of the verified 20B order, sized for the 1.25xC target of the OLMo-ladder 370M parameter count.

The repository contains token IDs, manifests, validation data, and audits. It does not contain expanded source text. The files named .npy are headerless little-endian uint32 memmaps, matching the ai2-olmo 0.6 raw memmap loader contract; they are not NumPy files with .npy headers.

Views

View	Exact aligned tokens	Sequences of 4,096
100M	100,003,840	24,415
1B	1,000,001,536	244,141
5B	5,000,003,584	1,220,704
370M 1.25xC	9,281,564,672	2,266,007
20B	20,000,002,048	4,882,813

Use views/<view>/manifest.json as the authoritative ordered file list. Do not sort object paths yourself. The 1.25xC view name is 370m-1.25xc. commits/<view>.json is the publication-ready marker for the 20B and 1.25xC views, and audits/publication-sha256.json records the byte size and SHA-256 digest of every published artifact.

The 370m-1.25xc manifest reuses the same immutable packed/objects/ paths as the 20B manifest. Only one content-addressed boundary object was added because the target ends inside a 20B object. The approximately 37.1 GB prefix is not duplicated in Hub storage.

Selecting and downloading a view

Set view to either "20b" or "370m-1.25xc", download its manifest, and then download exactly the object paths listed by that manifest. This avoids fetching unrelated packed objects.

import json
from pathlib import Path

from huggingface_hub import hf_hub_download

repo_id = "ericrcwu/week1-general-20b-dolma2-v1"
view = "370m-1.25xc"  # or "20b"
local_dir = Path("week1-corpus")

manifest_file = hf_hub_download(
    repo_id=repo_id,
    repo_type="dataset",
    filename=f"views/{view}/manifest.json",
    local_dir=local_dir,
)
manifest = json.loads(Path(manifest_file).read_text())
for item in manifest["objects"]:
    hf_hub_download(
        repo_id=repo_id,
        repo_type="dataset",
        filename=item["key"],
        local_dir=local_dir,
    )

Mixture

Category	Source	Tokens in 20B view	Pinned revision
DCLM web	allenai/dolmino-mix-1124:dclm	10B	a319f19eef1e257417b11ea8c30da266ae175557
FineWeb-Edu	HuggingFaceFW/fineweb_edu_100BT-shuffled	1B	be6b2a50d3a9c60d330c45384e80c7863cd3a25d
Academic/STEM	allenai/dolmino-mix-1124:pes2o	2B	a319f19eef1e257417b11ea8c30da266ae175557
Code	common-pile/stackv2_edu_filtered	2B	c354dbe88469a1153e97c6a63ac50591849654de
Math	allenai/dolmino-mix-1124:math	2B	a319f19eef1e257417b11ea8c30da266ae175557
Wikipedia	allenai/dolmino-mix-1124:wiki	1B	a319f19eef1e257417b11ea8c30da266ae175557
StackExchange	allenai/dolmino-mix-1124:stackexchange	1B	a319f19eef1e257417b11ea8c30da266ae175557
FLAN	allenai/dolmino-mix-1124:flan	1B	a319f19eef1e257417b11ea8c30da266ae175557

Exact per-view category counts and complete source-shard provenance are in each view manifest.

Tokenization and construction

• Tokenizer: allenai/dolma2-tokenizer at 5292e5d6c0f40b67cc765fe41bec991cf4345b5c
• Vocabulary size: 100,278; embedding size: 100,352
• EOS: 100,257; PAD: 100,277
• EOS is appended between documents.
• Documents are packed contiguously into 4,096-token sequences without padding.
• Storage is headerless little-endian uint32 in immutable objects of about 512 MiB, except at view boundaries.
• Deterministic seed: 6,198 for inventory, sampling, packing, and view order.
• Exact normalized-text deduplication is applied across selected sources.
• A deterministic 0.1% per-category validation sample is excluded before training selection.
• Exact-match decontamination covers the pinned OLMo-ladder evaluation bundle plus GSM8K test. Expensive semantic or near-duplicate filtering was not applied.

Minimal loading example:

import json
from pathlib import Path

import numpy as np

root = Path(".")
view = "370m-1.25xc"  # or "20b"
manifest = json.loads((root / f"views/{view}/manifest.json").read_text())
for item in manifest["objects"]:
    tokens = np.memmap(root / item["key"], mode="r", dtype="<u4")
    sequences = tokens.reshape(-1, manifest["sequence_length"])
    # Feed `sequences` to the training loader in manifest order.

Licensing and attribution

This is a mixed-source derivative database, so the repository is marked license: other; no single license replaces the upstream terms. Users must review and comply with every applicable upstream license and source-site term.

• The DOLMino Mix card labels the aggregate as ODC-By and reports component licenses including CC-BY-4.0 for DCLM, ODC-By for FLAN/peS2o/Wikipedia and several math components, CC-BY-SA 2.5/3.0/4.0 for StackExchange, Apache-2.0, MIT, and CC-BY-SA-4.0 for other math components.
• FineWeb-Edu is published as ODC-By.
• Stack V2 Edu is filtered to repositories whose detected licenses are on the Blue Oak Council certified list. Its upstream examples contain per-document license metadata, but this packed token stream does not retain document boundaries or per-document license fields. Upstream warns that license detection and metadata may be inaccurate.

The pinned source repository IDs and revisions above provide source-level attribution. This card is not legal advice, and the absence of raw text does not waive upstream obligations.

Intended use and limitations

The corpus is intended as a stable baseline for controlled pretraining A/B tests, especially OLMo-ladder-style experiments. It is not a neutral sample of the web: half of the mixture is DCLM, and the remaining categories and quotas were chosen deliberately. Source errors, personal information, harmful text, benchmark leakage not covered by exact decontamination, and license-metadata errors may remain. FLAN also makes this a partially instruction-like mixture, not a pure from-scratch web corpus.

The validation set is useful for within-project comparisons, not as an independent benchmark. Do not treat category labels as document-level ground truth beyond the source-level routing used to build the mixture.

Reproducibility

The manifests record source revisions, inventory, tokenizer fingerprint, ordered objects, SHA-256 values, category counts, validation exclusion, decontamination bundle, seed, and parent-view identity. OLMo-ladder was pinned at 67a3f440f787d020da35e3ca8eeae475fae754f5 for compatibility checks.