CDR-Dataset v1

Dataset details

Five domains, fixed entity features, and context-dependent reference labels.

from datasets import load_dataset

train = load_dataset("leo-bjpark/CDR-Dataset", "v1", split="train")
test = load_dataset("leo-bjpark/CDR-Dataset", "v1", split="test")
revision_1 = load_dataset("leo-bjpark/CDR-Dataset", "v1", split="revision_1")
revision_2 = load_dataset("leo-bjpark/CDR-Dataset", "v1", split="revision_2")

The public export contains source identities and raw attributes. Reproduce the original experiments with the preserved anonymous student exports. Public target labels are strings; numeric feature judgments retain their original 0 / 0.5 / 1 encoding.

Dataset structure

DomainsQueries / domainContexts / domainEntities / domainFeatures / entityTotal samples
51510072,500

Feature counts refer to the seven experimental judgments. Public records also preserve raw profile attributes. The homepage examples show four selected contexts; v1 contains five per domain.

Queries, contexts, and features

Splits

HF splitEntities / domainSamples / domainTotal samples
train402001,000
test20100500
revision_120100500
revision_220100500

Seed 2027; instance-level splitting keeps all five contexts together. Revision splits add entities, not contexts. No validation samples are included. The split does not guarantee unseen species or feature vectors.

Data Generation Statement

v1 combines curated source identities, constructed entity states, fixed feature and context definitions, and model-generated reference labels. The following describes the procedure recorded in the preserved code and run manifests.

  1. Domains and Queries

    Five domains were specified in the dataset design, each with one query: Birds (CanFly), Legal Cases (IsGuilty), Humans (CanRun), Mammals (CanSurvive), and Plants (CanSurvive). Mammal and plant survival use separate query IDs. Queries have explicit target definitions, including the supplied hypothetical offence for Legal Cases.

  2. Entities

    Each domain contains 100 entity profiles. Birds and Plants combine six source species per domain with constructed individual states. Mammals use 100 distinct species from the curated species catalog, with stipulated developmental stage, health, and energy reserves. Human and legal-case profiles are fictional. Source identities and attributes are distinguished from constructed states in the input records.

    Entity generation uses seed 2027. Original student prompts use anonymous IDs that remain consistent across contexts; the public export also preserves source identities and raw attributes. All five contexts for an entity remain in one split.

  3. Feature Categories and Labeling

    Seven feature predicates per domain were defined in the implementation, with explicit descriptions and value mappings. They describe properties such as wing condition, training, species adaptations, plant condition, or elements of the hypothetical offence. The same feature values are reused across an entity’s contexts.

    Human and legal-case feature values are projected from the stipulated profiles by code. Bird, mammal, and plant feature judgments in v1 are estimated by GPT-5.4 under the fixed feature definitions. Values are 0 (false), 0.5 (ambiguous), or 1 (true). Missing observations, intermediate states, and induced ambiguity are recorded separately. An ambiguous feature judgment is not a probability or a third target label.

  4. Context Categories and Labeling

    Five context categories per domain were specified in the implementation. They vary flight distance and plumage moisture, running distance and pace, mammal exposure conditions, plant watering and temperature, or the situation-specific legal rule. Each entity is paired with every context in its domain, yielding 100 × 5 samples per query.

    Context IDs identify these fixed conditions; they are not separately predicted labels. Full descriptions state the task, exposure duration, available resources, and any applicable exception. The legal contexts additionally supply explicit rules. Contexts change the conditions under which the query is judged while retaining the entity’s stipulated feature values.

  5. Query Labeling

    GPT-5.4 assigns a binary reference to each entity–context pair using the target definition, original identity, attributes and their provenance, stipulated state, full context, and applicable rule. The teacher receives the full profile; it does not infer its target label solely from the public feature vector. For legal cases, it is instructed to use only the supplied hypothetical rule and facts.

    The response schema permits only {"label": 0} or {"label": 1}, without explanations. The public export maps these to "false" and "true". All 2,500 pairs are resolved in the imported snapshot, with no excluded entities or recorded human reviews. These are teacher-generated reference judgments, not independently verified factual or normative ground truth.