README.md

---
configs:
- config_name: documents
  data_files:
  - split: train
    path: documents/train-*
- config_name: entities
  data_files:
  - split: train
    path: entities/train-*
- config_name: links
  data_files:
  - split: test
    path: links/test-*
- config_name: templates
  data_files:
  - split: train
    path: templates/train-*
- config_name: triples
  data_files:
  - split: train
    path: triples/train-*
dataset_info:
- config_name: documents
  features:
  - name: document_id
    dtype: string
  - name: entity_ids
    list: string
  - name: aliases
    list: string
  - name: index
    list:
      list: int64
  - name: entity_count
    dtype: int64
  splits:
  - name: train
    num_bytes: 20248546
    num_examples: 21845
  download_size: 18353071
  dataset_size: 20248546
- config_name: entities
  features:
  - name: qid
    dtype: string
  - name: label
    dtype: string
  - name: description
    dtype: string
  - name: alias
    list: string
  - name: wikidata_url
    dtype: string
  - name: wikipedia_url
    dtype: string
  - name: what_links_here
    list: string
  - name: properties
    list:
    - name: property_id
      dtype: string
    - name: value
      dtype: string
  - name: text
    dtype: string
  - name: abstract
    dtype: string
  - name: page_id
    dtype: int64
  - name: n_characters
    dtype: int64
  - name: origin
    list: string
  - name: n_mentions
    dtype: int64
  - name: pile_count
    dtype: int64
  - name: n_triples
    dtype: int64
  - name: n_tokens_pythia
    dtype: int64
  - name: n_tokens_llama
    dtype: int64
  splits:
  - name: train
    num_bytes: 200105053
    num_examples: 21970
  download_size: 137035367
  dataset_size: 200105053
- config_name: links
  features:
  - name: source
    dtype: string
  - name: document_id
    dtype: 'null'
  - name: entity_ids
    list: string
  - name: index
    list:
      list: int64
  - name: text
    dtype: string
  - name: wikipedia_ids
    list: int64
  - name: wikipedia_titles
    list: string
  splits:
  - name: test
    num_bytes: 5597904
    num_examples: 4765
  download_size: 3878170
  dataset_size: 5597904
- config_name: templates
  features:
  - name: relation_id
    dtype: string
  - name: relation_name
    dtype: string
  - name: templates
    list: string
  splits:
  - name: train
    num_bytes: 4766
    num_examples: 26
  download_size: 6424
  dataset_size: 4766
- config_name: triples
  features:
  - name: id
    dtype: int64
  - name: relation_id
    dtype: string
  - name: relation_name
    dtype: string
  - name: template
    dtype: string
  - name: template_id
    dtype: int64
  - name: subject
    dtype: string
  - name: subject_qid
    dtype: string
  - name: object
    dtype: string
  - name: object_qid
    dtype: string
  splits:
  - name: train
    num_bytes: 1309505
    num_examples: 9696
  download_size: 329450
  dataset_size: 1309505
---
# platovec — Wikipedia entities, documents, linearity triples & ZELDA links

A dataset for studying **entity representations in LLMs**: how
entities are encoded, how those representations form, and their causal effects.

Built from a seed entity list (the factual relations of Hernandez et al. 2024, by title, plus
ZELDA mention targets, by Wikipedia page id) by joining **`wikimedia/structured-wikipedia`**
(text + links) and **Wikidata** — labels / aliases / properties from either the Wikidata API or a
local `.json.bz2` dump.

## Configs

| config | granularity | description |
|---|---|---|
| `entities` | one row / entity | the entity, its text, abstract, aliases, properties, links |
| `documents` | one row / document | mention positions of entities within a document |
| `triples` | one row / linearity fact | (subject, object) + chosen template + QIDs + Wikidata property |
| `templates` | one row / relation | candidate prompt templates per relation (best-first) |
| `links` | one row / ZELDA document | entity-linking mentions for retrieval eval (split `test`) |

```python
from datasets import load_dataset
entities  = load_dataset("ykolo/entityconcepts", "entities",  split="train")
documents = load_dataset("ykolo/entityconcepts", "documents", split="train")
triples   = load_dataset("ykolo/entityconcepts", "triples",   split="train")
links     = load_dataset("ykolo/entityconcepts", "links",     split="test")
```

## `entities`

| column | type | description |
|---|---|---|
| `qid` | str | Wikidata QID (row key) |
| `label` | str | entity name (Wikipedia article title) |
| `description` | str | one-line description |
| `alias` | list[str] | alternative names (Wikidata "also known as", English) |
| `wikidata_url` | str | |
| `wikipedia_url` | str | |
| `what_links_here` | list[str] | QIDs of seed entities whose article links to this one |
| `properties` | list[{property_id, value}] | Wikidata claims (value = QID for items, rendered text otherwise) |
| `text` | str | flat article text (reconstructed from structured-wikipedia sections) |
| `abstract` | str \| null | article lead/abstract (from structured-wikipedia) |
| `page_id` | int \| null | English Wikipedia page id (structured-wikipedia `identifier`) |
| `n_characters` | int \| null | article length (structured-wikipedia `version.number_of_characters`) |
| `origin` | list[str] | seed origin(s): `linearity` and/or `zelda` |
| `n_mentions` | int | total mention occurrences across the `documents` table |
| `pile_count` | int \| null | frequency of the label in the Pile (if provided) |
| `n_triples` | int | number of `triples` where this entity is the subject |
| `n_tokens_pythia` | int \| null | tokens of `label` under the Pythia (GPT-NeoX) tokenizer |
| `n_tokens_llama` | int \| null | tokens of `label` under the Llama-3.x tokenizer |

## `documents`

One row = one document (an entity's Wikipedia article). Lists are aligned by position:
`entity_ids[i]` ↔ `aliases[i]` ↔ `index[i]`. The document text is `entities[document_id].text`
(not duplicated here). Use it to extract an entity's activations at its mention positions
across several documents (self-mention ⇔ `document_id` ∈ `entity_ids`).

| column | type | description |
|---|---|---|
| `document_id` | str | QID of the entity whose article is this document (→ `entities.qid`) |
| `entity_ids` | list[str] | QID of each mentioned entity (repeats for multiple mentions) |
| `aliases` | list[str] | surface form found in the text for each mention (hyperlink anchor for cross-page mentions; matched label/alias for self-mentions) |
| `index` | list[[int, int]] | character [start, end] offsets in the document's text |
| `entity_count` | int | number of **distinct** entities mentioned in the document |

## `triples`

Facts from the factual linearity relations (one row per (subject, object) sample).

| column | type | description |
|---|---|---|
| `id` | int | |
| `relation_id` | str \| null | mapped Wikidata property, e.g. `P36` (null if unmapped) |
| `relation_name` | str | human-readable name, e.g. "country capital city" |
| `template` | str | prompt template with `{}` as the subject placeholder |
| `template_id` | int \| null | index of the template used in `templates.templates[]` (0 = relation best; may differ per triple with per-triple selection) |
| `subject`| str | subject |
| `subject_qid` |  str\|null | qid of subject |
| `object`| str | object |
| `object_qid` |  str\|null | qid of object |

## `links`

Entity-linking documents from the **ZELDA** dataset, for retrieval evaluation. One row = one
document; each mention's ZELDA `wikipedia_id` is matched to the `page_id` column of `entities`
to recover its QID. Split `test` (the `source` column gives the ZELDA subset, e.g. `aida-b`;
`train` is available too if built). Lists aligned by position:
`entity_ids[i]` ↔ `index[i]` ↔ `wikipedia_ids[i]` ↔ `wikipedia_titles[i]`.

| column | type | description |
|---|---|---|
| `source` | str | split origin (`train`, or test set name e.g. `aida-b`) |
| `document_id` | str \| null | QID of the document's own Wikipedia page (train only; via `entities.page_id`) |
| `entity_ids` | list[str] | QID of each mention's gold entity |
| `index` | list[[int, int]] | character [start, end] offsets in `text` |
| `text` | str | document text |
| `wikipedia_ids` | list[int] | raw ZELDA Wikipedia page ids of the mentions (traceability) |
| `wikipedia_titles` | list[str] | ZELDA Wikipedia titles of the mentions (traceability) |

Mentions whose `wikipedia_id` is not present in `entities` (entity not scraped) are dropped (logged).

## Construction

- **seed entities**: subjects + objects of the factual relations (by title) + ZELDA mention
  targets (by `page_id`). `origin` records which source(s) each entity came from.
- **`entities` / `documents`**: text + link graph from hugging face `wikimedia/structured-wikipedia`
  (enwiki); aliases & properties from Wikidata (API or local dump); `abstract`/`text` from
  structured-wikipedia; disambiguation pages filtered out. Mentions: an entity in its **own** page
  is found by matching its label + aliases; mentions in **other** pages come from **hyperlink
  anchors** (exact surface form + offset) by default, which avoids false positives (e.g. matching
  "Paris" inside *Notre-Dame de Paris*).
- **`triples`**: linearity relations; `relation_id` is the Wikidata P-id from a static mapping.
  `template` is the best prompt — either per relation (max partial-match in generation) or, with
  per-triple selection, the candidate that maximises the gold object's teacher-forced log-prob for
  that specific (subject, object).
- **`links`**: ZELDA entity-linking documents; mention `wikipedia_id` matched to `entities.page_id` → QID.

Code: https://github.com/siemovit/platovec