Benchmark
Confabulation (Hallucination) for RAG (Lech Mazur)
Factuality benchmarks measure how often models get facts right. Scores are percentages. Lower is better.
- Models measured
- 54
- Scores recorded
- 60
- Sources
- 1
- IQ weight
- 2.46
Score distribution
- 18.8%
- Mean
- 17.1%
- Median
- 6.9%
- Std dev
- 12.3%
- 10th percentile
- 28.4%
- 90th percentile
- 10.3%
- Min
This benchmark has a weight of 2.46 for chat models, measured over 54 models. The weight is how much it counts toward llminfo IQ. We measure it from the data, not from the raw scores.
Leaderboard
| # | Model | Score |
|---|---|---|
| 51 | DeepSeek V3· Lech Mazur suite data | 26.2% |
| 52 | Gemini 2.0 Flash· Lech Mazur suite data | 26.9% |
| 53 | gemma2-27b· Lech Mazur suite data | 27.1% |
| 54 | GPT-4 Turbo· Lech Mazur suite data | 28.4% |
| 55 | Phi 4· Lech Mazur suite data | 29.4% |
| 56 | Nova Pro· Lech Mazur suite data | 30.1% |
| 57 | Claude 3 Haiku· Lech Mazur suite data | 34.2% |
| 58 | Claude Haiku 3.5· Lech Mazur suite data | 36.7% |
| 59 | GPT-4o mini· Lech Mazur suite data | 37.2% |
| 60 | Gemma 3 27B IT· Lech Mazur suite data | 40.3% |
Showing 51–60 of 60 results.