drozado/llms_epistemic_consistency

LLMs Epistemic Consistency Dataset

This dataset artifact contains the stimuli and prompt templates used for experiments on epistemic consistency and political-cue sensitivity in LLM evaluations.

Dataset URL: https://huggingface.co/datasets/drozado/llms_epistemic_consistency

Contents

• croissant.json: root-level copy of the completed Croissant metadata for NeurIPS 2026 Evaluations and Datasets submission.
• metadata/croissant.json: same Croissant metadata, kept with the generated metadata tables.
• metadata/prompts_catalog.csv: normalized prompt templates parsed from vpei/epistemic_consistency/prompts.py and vpei/epistemic_consistency/prompts_strict.py.
• metadata/data_inventory.csv: file-level inventory of included stimulus/data files.
• metadata/source_datasets.csv: provenance and license notes for synthetic and external-derived components.
• scripts/build_croissant_metadata.py: generator used to create the Croissant file and metadata tables.
• vpei/epistemic_consistency/**/data/**: experiment stimuli and processed source data.
• vpei/epistemic_consistency/prompts.py, prompts_strict.py, and active_prompts.py: prompt source files referenced by the metadata.

Model-output result CSVs are intentionally not part of the dataset record sets. They are hosted separately as a companion archive on Zenodo due to size:

Experimental Results Archive: https://doi.org/10.5281/zenodo.20033483

That archive contains the raw model-response CSVs from all experiment runs and is sufficient to reproduce the statistical analyses in the paper without re-querying any LLM APIs.

Intended Use

Use this artifact to reproduce or audit benchmark-style evaluations of whether LLMs apply consistent evidentiary standards when substantively similar evidence is paired with different political cues.

The artifact is intended for evaluation methodology research, benchmark auditing, and reproducibility. It is not intended for model training or operational decision-making.

Provenance

The artifact combines synthetic stimuli generated for this project with processed or sampled components derived from external datasets. Consult metadata/source_datasets.csv before redistributing or reusing any component, especially external-derived data with noncommercial or share-alike terms.

Responsible AI Metadata

The canonical Responsible AI metadata is in croissant.json and metadata/croissant.json, including:

• rai:dataLimitations
• rai:dataBiases
• rai:personalSensitiveInformation
• rai:dataUseCases
• rai:dataSocialImpact
• rai:hasSyntheticData
• prov:wasDerivedFrom
• prov:wasGeneratedBy

High-level summary:

• The dataset intentionally contains political labels, activist-group labels, politically valenced contexts, and other socially salient cues as experimental manipulations. These labels should not be interpreted as claims about real individuals.
• The dataset contains synthetic data. Synthetic CVs/job descriptions, generated posts, generated articles, generated art images, and generated policy or legal stimuli are benchmark stimuli, not real application records.
• Some external-derived components include real-world public claims, speaker names, locations, source URLs, and inherited annotations. Consult metadata/source_datasets.csv before reuse or redistribution.
• Recommended uses are LLM evaluation-methodology research, auditing relative consistency across controlled prompt variants, and reproducing the associated experiments.
• Not recommended uses include model training, hiring, education, legal, medical, financial, moderation, public-policy decisions, demographic inference, or factual claims about real groups, people, policies, or events.

Regenerating Metadata

Run:

python scripts/build_croissant_metadata.py --repo-root . --artifact-dir . --dataset-url https://huggingface.co/datasets/drozado/llms_epistemic_consistency

This writes the Hugging Face dataset card to README.md and the NeurIPS Croissant files to croissant.json and metadata/croissant.json. Run the command from the dataset repository root. In this source repository, that root is the dataset/ directory.

When uploading to Hugging Face, use the contents of dataset/ as the dataset repository root. Then validate croissant.json with the Croissant validator before NeurIPS submission.