Datasets:

ClarusC64
/

description-granularity-consistency-v0.1

language:
  - en
pretty_name: Clarus Description Granularity Consistency v0.1
tags:
  - clarus
  - embodied-ai
  - representation
  - abstraction
  - diagnostics
task_categories:
  - text-classification
license: mit
dataset_info:
  features:
    - name: sample_id
      dtype: string
    - name: scene
      dtype: string
    - name: desc_coarse
      dtype: string
    - name: desc_fine
      dtype: string
    - name: option_a
      dtype: string
    - name: option_b
      dtype: string
    - name: task
      dtype: string
    - name: correct_option
      dtype: string
    - name: consistency_axis
      dtype: string
    - name: difficulty
      dtype: string

What this is

This dataset tests whether abstract state remains stable across description depth.

Same event.
A coarse description.
A fine description.

A coherent model should land in the same state either way.

What it measures

• Whether abstraction changes when wording gets more technical
• Whether the model keeps the same implied world state across granularity shifts

A model that binds state to language style will fail.

Expected model output

Return only:

A
or
B

Scoring

Use the shared Clarus scorer.

Metrics:

• accuracy
• parse_rate

Suggested prompt

You are checking whether two levels of description imply the same world state.

Scene: {scene}

Coarse description: {desc_coarse}

Fine description: {desc_fine}

A: {option_a} B: {option_b}

Choose the consistent interpretation. Answer with only A or B.

Trinity placement

This is Dataset III of the Abstract State Consistency Trinity.

Citation

ClarusC64. "Clarus Description Granularity Consistency v0.1", 2026.