K-FinHallu - Hallucination Type Classification: leaderboard
Macro-F1 (0-100) at naming the type of a hallucinated response (contradictory, unverifiable, irrelevant, false acceptance or false refusal) in K-FinHallu (2026), a benchmark of multi-turn Korean financial RAG dialogues built from public financial documents, with injected hallucinations and refusal cases. Higher is better.
Metric: Macro-F1 (0-100) over five hallucination types (contradictory, unverifiable, irrelevance, false acceptance, false refusal) on the 404 hallucinated test responses, multi-turn Korean financial RAG dialogues, temperature 0 where applicable; higher is better. Source: arxiv.org. Saturation forecast: Estimated already saturated. 12 models tracked. Updated 7 Oct 2026.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 2.5 Pro | 72.5 | #276 |
| 2 | GPT-5 | 71.7 | #178 |
| 3 | Gemini 2.5 Flash | 66.9 | #463 |
| 4 | Qwen 3 32B | 49.2 | #972 |
| 5 | Qwen 3 14B | 44.2 | #1269 |
| 6 | Qwen 3 8B | 38.7 | #1873 |
| 7 | GPT-4o | 28.9 | #751 |
| 8 | Llama 3.1 70B | 13.3 | #1499 |
| 9 | Llama 3.3 70B | 10.3 | #1514 |
| 10 | Llama 3.1 8B | 8.7 | #5196 |
| 11 | EXAONE 4.0 32B | 3.9 | #1398 |
| 12 | kanana-2-30B-a3B-instruct | 2.2 |
No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.
Interactive version: theaggregate.ai/benchmark?slug=k-finhallu-hallucination-type-classification · How It Works · Data refreshed daily, snapshot 2026-11-21.