DoRA (Defense Documents) - Hallucination (Oracle Evidence): leaderboard
DoRA (2026) builds a retrieval-augmented question answering benchmark from defense-related documents, with test questions in five intent styles and auditable evidence. RAGEval keypoint hallucination rate in percent with the gold evidence as context, judged by an LLM; lower is better.
Metric: RAGEval keypoint hallucination rate (%) of grounded answers to the 1,259-question DoRA defense-document test split with the gold evidence passages as context (oracle retrieval): per-question share of the reference answer's keypoints the answer addresses incorrectly, judged by GPT-4o-mini and averaged over questions; lower is better. Source: arxiv.org. Saturation forecast: Estimated already saturated. 8 models tracked. Updated 7 Oct 2026.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-3.5 Turbo | 2.69 | #3005 |
| 2 | GPT-4o Mini | 3.4 | #1667 |
| 3 | Qwen 2.5 7B Instruct | 3.53 | #2723 |
| 4 | Ministral-8B-Instruct-2410 | 3.93 | #4555 |
| 5 | Claude 3.5 Haiku | 3.93 | #1580 |
| 6 | Llama 3 8B Instruct | 4.2 | #4852 |
| 7 | GPT-4o | 4.23 | #751 |
| 8 | Llama 3.1 8B Instruct | 4.67 | #4600 |
No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.
Interactive version: theaggregate.ai/benchmark?slug=dora-defense-documents-hallucination-oracle-evidence · How It Works · Data refreshed daily, snapshot 2026-11-21.