Groundtruth-Data/groundtruth-hallucination-bench-sample

Groundtruth Data Hallucination Benchmark Sample

This public teaser contains 180 representative, source-backed examples from Groundtruth Data products.

Groundtruth Data builds verified evaluation, remediation, and held-out validation datasets for AI models using authoritative source data. The commercial workflow is:

1. Find where a model fails.
2. Prove the failure with a larger verified evaluation.
3. Provide targeted remediation/training data.
4. Validate improvement on untouched held-out data.

What This Public Sample Demonstrates

The sample shows how Groundtruth Data records connect a model-visible question to a verified answer, a concise verification summary, and source provenance. It is intentionally small and broad rather than a full benchmark.

Domains Represented

• citation_openalex: 10 rows
• clinical_trials: 10 rows
• code_execution: 10 rows
• code_reality: 10 rows
• company_registry: 10 rows
• cross_source_factuality: 10 rows
• drug_indications: 10 rows
• fda_drug_safety: 10 rows
• finance_sec_edgar: 10 rows
• general_factuality: 10 rows
• geography: 10 rows
• government_contracts: 10 rows
• history: 10 rows
• language_literature: 10 rows
• legal: 10 rows
• math: 10 rows
• patent_ip: 10 rows
• science: 10 rows

Datasets Represented

• Citation / OpenAlex Metadata Grounding: 10 rows
• Clinical Trial Outcomes Verification: 10 rows
• Code Execution Verification: 10 rows
• Code Reality Verification: 10 rows
• Companies House Verification: 10 rows
• Cross-Source Factual Verification: 10 rows
• Drug Indication Grounding: 10 rows
• English Language Verification: 10 rows
• FDA Drug / Device Safety Verification: 10 rows
• Fact Verification: 10 rows
• Geography Facts Verification: 10 rows
• Government Contract Verification: 10 rows
• History Facts Verification: 10 rows
• Legal Verification: 10 rows
• Math Verification: 10 rows
• Patent & IP Verification: 10 rows
• SEC EDGAR Financial Verification: 10 rows
• Science Verification: 10 rows

Schema

Each line in train.jsonl is a JSON object:

{
  "id": "gtdata-source-slug-row-id",
  "dataset": "Human-readable dataset name",
  "domain": "domain_family",
  "question": "Model-visible question",
  "verified_answer": "Concise source-grounded answer",
  "verification_summary": "Short explanation of how the answer was verified",
  "source_urls": ["https://authoritative.source/..."],
  "reference_model": "actual model label when available",
  "model_response": "observed response excerpt when available",
  "verdict": "actual stored verdict when available",
  "grading_mode": "deterministic or source-aware grading mode",
  "difficulty": "easy|medium|hard|unspecified",
  "tags": ["domain", "capability"]
}

reference_model, model_response, and verdict are null when no actual observed model response is included for that row. Responses are not fabricated.

Verification Methodology

Rows are selected from existing verified Groundtruth Data product exports. Source families include official APIs, primary registries, direct code execution, mathematical computation, legal/source text, OpenAlex, SEC EDGAR, Companies House, FDA/openFDA, USASpending.gov, and other cited sources depending on the row.

The public schema renames internal fields for clarity:

• expected_answer becomes verified_answer.
• grading_context or internal explanations become verification_summary.

Original source datasets are not modified.

Proof, Remediation, and Held-Out Data

Groundtruth Data separates dataset roles:

• Diagnostic/evaluation data finds possible weaknesses.
• Proof evaluations test whether a weakness is systematic on fresh verified examples.
• Remediation/training data targets confirmed weaknesses.
• Held-out validation data remains untouched for post-training measurement.

This public sample is not the full proof, remediation, or held-out corpus. Full commercial packages are available separately.

Limitations

• This teaser is not sized to estimate model performance.
• Model results, where shown, are specific to the stored model label, prompt, grader, and dataset version.
• The sample does not claim universal model failure rates.
• The sample does not claim model improvement.
• Source and licensing terms vary by underlying provider; redistribution-sensitive source text and proprietary commercial-scale data are excluded.

Commercial Data

Commercial Groundtruth Data packages include larger verified proof evaluations, remediation/training datasets, held-out validation sets, and custom model failure studies. This public sample excludes full 10k remediation datasets, internal 25k assets, browser model-run artifacts, internal sales metadata, credentials, and private customer information.