K-FinHallu - Four-Class Detection: leaderboard

Accuracy (%) at labelling each dialogue response as faithful, hallucinated, a false refusal or a true refusal given the retrieved passages 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: Overall accuracy (%), micro-averaged over 808 test responses, at labelling each response faithful, hallucinated, false refusal or true refusal given the retrieved passages, 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

#ModelScoreOverall rank
1GPT-585.5#178
2Gemini 2.5 Pro85.3#276
3Gemini 2.5 Flash84.5#463
4Qwen 3 32B74.3#972
5Qwen 3 14B74.1#1269
6Llama 3.1 70B69.3#1499
7GPT-4o69.2#751
8Qwen 3 8B68.8#1873
9Llama 3.3 70B68.7#1514
10EXAONE 4.0 32B60.1#1398
11kanana-2-30B-a3B-instruct51.5
12Llama 3.1 8B34.2#5196

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-four-class-detection · How It Works · Data refreshed daily, snapshot 2026-11-21.