Datasets:
The dataset viewer is not available for this split.
Error code: TooBigContentError
Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
locus-1 encyclopedic pool
The encyclopedic pool for locus-1 - one of seven cluster datasets that are mixed into the pretraining corpus. Every pool shares one schema, so a mixture is a query rather than a rebuild.
- 6,498,683 documents, 7,314,613,330 tokens under
allenai/dolma2-tokenizer@5292e5d6c0f40b67cc765fe41bec991cf4345b5c - Built from: HuggingFaceFW/finewiki (
en@8bd13e72e6a0) - Build settings digest:
cbd87eae472c6b30 - Fingerprint compatibility digest:
6ee2a61c2f91d6a1
Licence
Recorded from the sources: cc-by-sa-4.0 AND gfdl. The current pipeline records licences for audit and does not filter rows by them.
Configurations
documents is the default, training-facing configuration. fingerprints is compact
construction metadata retained for auditing and a future cross-pool deduplication pass.
Both configurations contain the same document IDs in the same publication.
documents
One row per document. metadata is a nested struct; labels holds heuristic outcomes
that are recorded, not applied, and its keys are unstable by design. upstream is
the source's own metadata kept verbatim as a JSON string.
| field | type | notes |
|---|---|---|
id |
string | <source-slug>:<the source's own stable key> |
text |
string | normalized plain text |
metadata.source |
string | which dataset this row's text won from |
metadata.cluster |
enum | one of the seven |
metadata.quality_bin |
int? | 1-10, rank-based deciles |
metadata.dup_count |
int | duplicate source records discarded; 0 means unique |
metadata.dup_sources |
list | sources those copies came from |
metadata.language |
string | ISO code, from the source or our classifier |
metadata.language_confidence |
float? | classifier score; null when disabled |
metadata.provenance |
enum | native / ocr / crawl / synthetic / distilled |
metadata.n_tokens |
int | under allenai/dolma2-tokenizer@5292e5d6c0f40b67cc765fe41bec991cf4345b5c |
fingerprints
| field | type | notes |
|---|---|---|
id |
string | joins one-to-one with documents.id |
source |
string | source dataset |
cluster |
enum | one of the seven pools |
text_hash |
string | SHA-256 of normalized text |
minhash_sig |
fixed-size list | 128-value signature for near-matches |
crawl_date |
string? | normalized source date used by winner policies |
text_len |
int | normalized text length in Unicode code points |
dup_count |
int | duplicates already discarded inside this pool |
dup_sources |
list | sources already folded into this survivor |
How it was built
Normalize (NFC, invisibles stripped) -> Cheap filters (empty, undecodable) -> Optional fastText language filter (per source: finewiki=off) -> Fingerprint (SHA-256 + MinHash) -> Dedup (exact; near-dedup: none) -> Label -> Count -> Publish.
Quality scorer: encyclopedic_structural_v2. Bins are rank-based deciles over this cluster, so bin 5 is the median document here - not a score on any absolute scale.
Decontamination so far is source-level only: no configs known to ship benchmark data were ingested. The n-gram overlap sweep needs the eval suite and has not run yet.
Frozen settings
MinHash: 5-word shingles, 128 permutations, seed 1, xxh32 hash. Everything but the comparison threshold is frozen from the first pool onward - signatures built with different settings cannot be compared at all. Token counts are under allenai/dolma2-tokenizer at immutable revision 5292e5d6c0f40b67cc765fe41bec991cf4345b5c, no special tokens.
Settings digest: cbd87eae472c6b30.
quality_bin cutoffs
| boundary | raw score |
|---|---|
| 1 -> 2 | 2.2347 |
| 2 -> 3 | 2.5281 |
| 3 -> 4 | 2.8183 |
| 4 -> 5 | 3.0988 |
| 5 -> 6 | 3.3688 |
| 6 -> 7 | 3.6458 |
| 7 -> 8 | 3.9316 |
| 8 -> 9 | 4.2418 |
| 9 -> 10 | 4.6335 |
- Downloads last month
- 55