---
pretty_name: "ChronoCoexist-WD: Multi-Version Temporal Knowledge Timelines"
license: cc0-1.0
language:
- en
task_categories:
- question-answering
tags:
- knowledge-editing
- temporal-knowledge
- temporal-question-answering
- wikidata
- model-editing
- continual-learning
configs:
- config_name: cases
  default: true
  data_files:
  - split: source
    path: data/cases/*.parquet
- config_name: timelines
  data_files:
  - split: source
    path: data/timelines/*.parquet
- config_name: versions
  data_files:
  - split: source
    path: data/versions/*.parquet
- config_name: prompts
  data_files:
  - split: source
    path: data/prompts/*.parquet
- config_name: provenance
  data_files:
  - split: source
    path: data/provenance/*.parquet
- config_name: relations
  data_files:
  - split: source
    path: data/relations/*.parquet
- config_name: forecast_scheduled
  data_files:
  - split: source
    path: data/forecast_scheduled/*.parquet
---

# ChronoCoexist-WD

## Multi-Version Temporal Knowledge Timelines

ChronoCoexist-WD (technical dataset identity: PTC-WD-2026) is an automatically
curated source pool of real temporal
Wikidata histories. Each timeline keeps one subject and one relation fixed
while the answer changes across dated, non-overlapping intervals. It supports
research on whether an edited language model can learn a newly arriving value
without losing earlier values that remain correct for historical dates.

**Important scope:** this release is automatic **silver** data. It is not
human-verified gold data and is not a frozen model-specific evaluation split.
Every row remains `human_approved=false`, `gold=false`,
`model_eligible=false`, and `experimental_eligible=false`.

## Snapshot and exact scale

The source is the complete Wikidata JSON dump frozen on **2026-07-20**.

| Quantity | Count |
|---|---:|
| Timelines | 8,468 |
| Dated versions | 25,946 |
| English prompts | 110,049 |
| Relations | 20 |
| Unique subject QIDs | 8,378 |
| Unique answer QIDs | 17,207 |
| Chain depth | 2–52 |
| Timelines with a repeated answer such as A→B→A | 977 |

The structural strata are:

- `PRIMARY_K_GE_3_INTERVAL_VERSIONS`: 3,322 timelines and 15,654 versions;
- `PAIRWISE_CONTROL_K_EQ_2_INTERVAL_VERSIONS`: 5,146 timelines and 10,292 versions.

These are structural labels, not train/test splits.

## Configurations

| Configuration | Unit | Rows | Intended use |
|---|---|---:|---|
| `cases` **(default)** | one readable dated fact | 25,946 | Start here: subject, relation, time and answer first |
| `timelines` | one subject–relation history | 8,468 | Main nested source pool |
| `versions` | one dated answer interval | 25,946 | Flat analysis and model screening |
| `prompts` | one safe, unique English query | 110,049 | Prompt generation and scoring preparation |
| `provenance` | one source statement | 25,946 | Raw Wikidata statement/qualifier/reference audit |
| `relations` | one Wikidata property | 20 | Relation counts and coverage |
| `forecast_scheduled` | one timeline | 5 | Starts after the snapshot; diagnostic only |

Use the technical `source` split in each configuration. It only exposes the
source rows through the Hugging Face API; it is not a training split. No
experimental train/validation/test split is claimed.

The default viewer is deliberately human-readable. Its first columns are
`case_id`, `subject_label`, `property_label`, `object_label`, `time_scope`,
dates, and the canonical question. Long SHA-256 identities are retained only
in later audit columns.

## Example

For subject `Q1799` (Wrocław) and relation `P6` (head of government), the
timeline contains:

| Valid interval | Answer |
|---|---|
| 2002-11-10 to 2018-11-19 | Rafał Dutkiewicz (`Q2127075`) |
| 2018-11-19 onward | Jacek Sutryk (`Q57691411`) |

One released prompt is:

> On 2010-11-14, the head of government of Wrocław was

with accepted answer `Rafał Dutkiewicz`. See `SAMPLE.json` for the complete
timeline record.

## Construction

1. Freeze the 20-relation universe before editor evaluation.
2. Parse the complete 2026-07-20 Wikidata dump.
3. Group statements by exact subject QID and property PID.
4. Preserve statement GUIDs, ranks, qualifiers, qualifier order, references,
   source indices, and raw statement payloads.
5. Normalize exact-day `P580` (start), `P582` (end), and `P585`
   (point-in-time) qualifiers while retaining their raw forms.
6. Reject duplicate intervals, strict overlaps, equal starts, multiple open
   intervals, and open intervals followed by a later version.
7. Retain shared boundaries, gaps, and recurrent answers as explicit
   diagnostics instead of silently collapsing them.
8. Select bounded per-relation samples using output-blind SHA-256 timeline
   identities.
9. Join frozen labels and aliases without changing membership.
10. Reject aliases shared by distinct answers in the same timeline,
    canonicalize answer whitespace, and block globally ambiguous subject
    labels.
11. Choose deterministic interior probe dates and render relation-aware
    prompts. Require prompt IDs and UTF-8 prompt bytes to be globally unique.

Other benchmarks informed relation selection and design; their rows were not
copied into this release.

## Prompt quality status

| Timeline status | Count | Meaning |
|---|---:|---|
| `AUTOMATIC_PROMPT_READY_FOR_NEXT_GATE` | 7,269 | Safe automatic prompt packet exists |
| `AUTOMATIC_PROMPT_BLOCKED_UPSTREAM_ALIAS` | 1,086 | Missing or ambiguous label/alias |
| `AUTOMATIC_PROMPT_BLOCKED_ALIAS_LEAKAGE` | 38 | A prompt contains an answer surface |
| `AUTOMATIC_PROMPT_EXCLUDED_TEMPORAL` | 75 | Excluded by temporal routing |

Only ready rows contribute to the `prompts` configuration. A prompt being
ready does not make the timeline model-eligible: tokenizer checks and a
pre-edit knowledge screen are model-specific and remain pending.

## How to create an evaluation subset

For each exact model name and revision:

1. verify that the unedited model retrieves the historical anchor;
2. verify that it does not already retrieve each arriving target;
3. inject later values in chronological order;
4. test every visible dated value after every injection;
5. keep all timelines for one subject in the same experimental split;
6. report free-generation complete-chain success, per-version accuracy,
   forgetting, paraphrase robustness, locality, and save/reload persistence.

Do not reuse one model's eligibility decisions for another model.

## Known limitations

- Wikidata records database state, not guaranteed truth. Facts may be stale,
  disputed, incomplete, or later corrected.
- All labels and prompts are English-only in this release.
- 10,644 of 25,946 versions have no embedded reference block, although the
  statement GUID, dump identity, and raw statement provenance are retained.
- Five timelines contain starts after 2026-07-20. They are isolated in the
  `forecast_scheduled` configuration and must not be treated as factual
  current-answer evaluation on the snapshot date.
- Temporal gaps are retained. Use the supplied interior `probe_date`, not a
  date in a gap or exactly at an uncertain boundary.
- Relation frequencies are long-tailed. `P355` has only three pairwise
  timelines and no admitted chain of depth three or greater.
- Automatic alias matching may miss valid surface forms. Manual review may
  improve coverage but would require a separately versioned gold release.
- No implicit-event-time questions or multi-hop questions are included.

## Personal and sensitive information

The data contains public Wikidata statements, including statements about
living people and relations such as spouse, employer, residence, political
party, and position held. Do not use it to infer private information, profile
people, or make consequential decisions. Verify sensitive claims against
authoritative and current sources.

## Licensing and provenance

Wikidata structured data is provided under
[CC0 1.0](https://creativecommons.org/publicdomain/zero/1.0/). See the
[Wikidata licensing page](https://www.wikidata.org/wiki/Wikidata:Licensing).
The derived data tables are released under CC0-1.0. `LICENSE` contains the
license text. File hashes and all admitted upstream identities are in
`checksums.sha256`, `release_manifest.json`, and `metadata/`.

## Reproducibility

The package was built with a fail-closed join and independently validated
alias/prompt gates. Parquet files were read back and checked for exact row
counts and schemas. The package includes:

- a checksum-bound release manifest;
- the frozen extraction, label, alias, temporal, and prompt manifests;
- the complete prompt-template registry;
- a data dictionary and one complete example.

## Citation

The paper citation is not frozen yet. Until then, cite this dataset version,
the 2026-07-20 Wikidata snapshot, and Wikidata itself. Please record the exact
Hugging Face repository revision used in experiments.