File size: 7,875 Bytes
3482735 f4465a2 3482735 f4465a2 3482735 f4465a2 62b8194 f4465a2 62b8194 f4465a2 3482735 f4465a2 3482735 f4465a2 3482735 f4465a2 3482735 f4465a2 3482735 f4465a2 3482735 f4465a2 3482735 f4465a2 3482735 f4465a2 3482735 f4465a2 3482735 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 | ---
license: cc-by-sa-4.0
language:
- en
tags:
- knowledge-graph
- rdf
- wikidata
- preprocessed
- text-corpus
- world-model
size_categories:
- 1M<n<10M
task_categories:
- text-generation
- feature-extraction
pretty_name: Normalized Wikidata — clean text-form triples for world-model training
---
# Normalized Wikidata
A preprocessed text-form view of Wikidata, optimised for training language
models or knowledge-graph world models. The goal is a corpus where the
*semantic content* of Wikidata triples comes through cleanly, with the
catalog-and-identifier clutter that dominates raw Wikidata by volume stripped
out.
License inherits from Wikidata: **CC-BY-SA 4.0**.
This dataset is the input to a corresponding series of Loka world-model
checkpoints at [`EmmaLeonhart/loka`](https://huggingface.co/datasets/EmmaLeonhart/loka).
Each snapshot here is named to match the Loka model trained on it — e.g.
the `v11-50k` snapshot is the corpus the `v11` Loka model was trained on,
`v12-100k` corresponds to `v12`, and so on.
## Snapshots
| Tag | Entity rows | Output triples | File size | Trained Loka model |
|---|---|---|---|---|
| `v11-50k` (alias `v0.1-50k`) | 50,000 | **350,428** | 14.7 MB | [`EmmaLeonhart/loka@v11`](https://huggingface.co/datasets/EmmaLeonhart/loka/tree/v11) |
| `v12-100k` | 100,000 | **671,817** | 28.4 MB | [`EmmaLeonhart/loka@v12`](https://huggingface.co/datasets/EmmaLeonhart/loka/tree/v12) |
| `v13-500k` | 500,000 | **2,511,771** | 109 MB | (training in progress 2026-05-14) |
| `v14-1M` | 1,000,000 | **4,021,409** | 176 MB | (training queued behind v13) |
All four corpus tiers are shipped as of 2026-05-14. The latest pushed tag is
`v14-1M`. The total file-size sum across all four tiers is ~330 MB; pulling
just the largest gives you the deepest training signal.
**Pulling a specific snapshot:**
```python
from huggingface_hub import hf_hub_download
path = hf_hub_download(
repo_id="EmmaLeonhart/normalized-wikidata",
repo_type="dataset",
filename="triples_normalized.txt",
revision="v11-50k", # or v12-100k, v13-500k, ...
)
```
Each snapshot is **strictly larger than the previous** — same first-N rows from
the same upstream stream, just with N raised. The SQLite label cache at
`wikidata_labels.sqlite` also grows monotonically across snapshots (~7,300
curated property labels preloaded, plus all entity labels seen in the slice).
## What it is
One triple per line, tab-separated:
```
subject\tpredicate\tobject
```
All three positions are **English labels** — QIDs and PIDs are resolved to
their `rdfs:label@en`. Entity labels come from the entity's own row in the
source dump; **property labels come from a curated cache** of 7,312 manually-
resolved Wikidata properties, never from corpus `rdfs:label` rows on
properties (those are corrupted by an upstream RDF-star executor bug — see
"Known issues with raw Wikidata" below).
## What was stripped
Predicates whose Wikidata `datatype` falls into one of these classes are
**dropped entirely** — they teach the model catalog formats rather than world
knowledge, and v6 of the Loka world model demonstrated they leak format
shapes onto unrelated predicates:
- `external-id` (~10,206 properties) — Freebase ID, ISNI, GND, LCCN, Dewey, etc.
- `url` (~120 properties) — links to external sites
- `commonsMedia` (~91) — Wikimedia Commons filenames
- `math` (~36) — LaTeX formulae
- `wikibase-sense` / `-lexeme` / `-form` / `-entity-schema` (~47) — lexeme machinery
- `globe-coordinate` (~10) — `Point(lat lon)` strings
- `geo-shape` / `musical-notation` / `tabular-data` (~15) — rare, non-transferable
Predicates **kept**: `wikibase-item`, `wikibase-property`, `string`, `quantity`,
`time`, `monolingualtext`.
In addition, object-level guards drop:
- URL-shaped values (`http://`, `https://`, `ftp://`, `irc://`, `mailto:`) that
slipped through with non-catalog predicates
- Long digit-only strings (8+ digits — GND/VIAF/ISNI shape) and DOIs
(`10.NNNN/...`) in the object position
- Rows where the subject *or* object is itself a property IRI
(`wdt:P\d+`) — these are RDF-star annotation rows surfacing in the wrong
slot, never legitimate
- System-reserved provenance triples (predicates under
`http://loka.dev/provenance/`)
## What was normalized
- **Time** values: `+YYYY-MM-DDTHH:MM:SSZ` → `YYYY-MM-DD` (or
`YYYY-MM-DDTHH:MM:SS` if time is non-zero). Leading `+` removed for CE
years; `-` preserved for BCE.
- **Quantity** values: leading `+` stripped from positive numbers
(`+1234` → `1234`).
- **Monolingualtext**: `@lang` tag stripped from the value. All languages kept;
the model sees `Tokyo` and `東京` as plain values, not as `Tokyo@en` and
`東京@ja`.
- **Datatype suffixes** on literals (`"value"^^<...>`): the suffix is parsed
off so it doesn't leak into training tokens. The datatype is consulted to
decide normalization rules and then dropped.
## Known issues with raw Wikidata that this corpus addresses
1. **Catalog / identifier explosion.** ~82 % of Wikidata's property types by
count are external identifiers, URLs, or other non-semantic catalog refs.
Training on them teaches the model catalog formats rather than world
knowledge. We strip them by datatype.
2. **Property `rdfs:label` corruption when materialised through some RDF-star
executors.** A `<<S P O>> rdfs:label "..."@en` annotation row, depending
on the executor, can surface as `wdt:Pnnn rdfs:label "object-value"@en`
— i.e. the property gets keyed against the inner triple's object value
instead of its real label. Entity labels are unaffected. We work around
this by sourcing property labels from a curated cache and never from
in-corpus `rdfs:label` rows on properties.
3. **Datatype suffix leakage.** `"2012-10-15T00:00:00Z"^^<...dateTime>` if
processed naively leaks tokens like `xmlschema`, `dateTime` etc. into the
training corpus. We strip these.
4. **Mixed-language values.** Wikidata's `monolingualtext` includes all
languages; we keep them but strip the `@lang` tag so values like `Tokyo`
and `東京` are plain strings.
## How it was built
The current preprocessor streams `philippesaade/wikidata` directly from
Hugging Face, with a SQLite label cache that persists across runs:
```bash
python tools/preprocess_from_hf.py \
--max-rows 100000 \ # entity-row count, sets the size tier
--label-db training/data/wikidata_labels.sqlite \
--output training/data/normalized/normalized_wikidata_v12_100k.txt
```
Two passes over the dataset:
- **Pass 1** scans every row to extract English `labels.en.value` into the
SQLite cache (constant memory regardless of corpus size).
- **Pass 2** streams again to emit the tab-separated text corpus, using the
cache for label lookups, applying the noise-datatype filter, normalising
time/quantity values, and dropping engine-bug-#2 RDF-star fallout at the
s/o level.
Source code: [`tools/preprocess_from_hf.py`](https://github.com/EmmaLeonhart/Loka/blob/main/tools/preprocess_from_hf.py),
[`tools/hf_push_normalized.py`](https://github.com/EmmaLeonhart/Loka/blob/main/tools/hf_push_normalized.py).
An earlier two-pass version that fetched from a Loka `.sdb` over SPARQL
(`tools/preprocess_streaming.py`) hit O(offset) cost at multi-million-triple
scale; the HF-direct version sidesteps that by streaming the upstream parquet.
## Provenance
See [`Loka` on GitHub](https://github.com/EmmaLeonhart/Loka) for the engine,
the preprocessor source, the trained model checkpoints, and the paper
describing the world-model training pipeline that motivated this corpus.
The Loka model series on Hugging Face:
[`EmmaLeonhart/loka`](https://huggingface.co/datasets/EmmaLeonhart/loka).
## Citation
Wikidata is the upstream source. Please cite Wikidata as well as this dataset
if you use the corpus.
|