VeritasOmega/veritas-claim-verification-dataset

---
language:
- en
license: apache-2.0
task_categories:
- text-classification
task_ids:
- fact-checking
pretty_name: VERITAS Claim Verification Dataset
size_categories:
- 100K<n<1M
---

# VERITAS Claim Verification Dataset

**775,294 labeled verification calls** (770,000 synthetic + 5,294 adversarial)
Multi-model consensus verdicts from the VERITAS Oracle across 8 domains.

## Paper
[VERITAS: Independence-Weighted Multi-Model Consensus for AI Output Verification](https://doi.org/10.5281/zenodo.19187611)

## Adversarial Robustness Findings (NEW)
5,294 adversarial variants generated via deterministic transforms:

| Transform | N | Accuracy |
|---|---|---|
| negation | 2,379 | **99.7%** |
| jurisdiction | 114 | 100.0% |
| quantifier | 337 | 98.8% |
| number_perturb | 85 | 98.8% |
| temporal | 2,379 | 79.8% |
| **Overall adversarial** | **5,294** | **90.7%** |
| Baseline | | 98.4% |
| Delta | | -7.7pp |

VERITAS is highly robust to negation attacks (99.7%) — a category where single-model verifiers typically drop 10-15pp.
Temporal hedging (79.8%) is the documented weakness: no time-indexed knowledge retrieval.

## Datasets
- eritas_oracle.synthetic_dataset — 770k synthetic claims
- eritas_oracle.synthetic_dataset_adversarial — 5,294 adversarial variants with transform labels

## Schema
claim_hash, claim_text, domain, ground_truth, source, confidence, consensus, flags, model_votes, trace_id, latency_ms, correct, ingested_at

## Citation
`
@misc{lopez2026veritas,
title={VERITAS: Independence-Weighted Multi-Model Consensus for AI Output Verification},
author={RJ Lopez},
year={2026},
doi={10.5281/zenodo.19187611}
}
`