Skip to content

Benchmark

Estonian Native Benchmark open-ended human preference win rate

Multilingual benchmarks compare model performance across languages. Higher scores are better.

Models and systems measured
12
Scores recorded
12
Sources
1
IQ weight
No weight published

Score distribution

0.528
Mean
0.545
Median
0.161
Std dev
0.327
10th percentile
0.746
90th percentile
0.211
Min
0.2110.754
The band covers the middle 80% of scores. The line marks the median.

Results

The source marks these results as descriptive, not capability scores, so we list them by name instead of ranking them.

Model or systemScore
Claude 3.7 Sonnetinstruction-tuned modelstable: 10, evaluation: pairwise-open-ended· Estonian Native Benchmark (arXiv:2510.21193v2) data0.754
Deepseek Chat V3instruction-tuned modelstable: 10, evaluation: pairwise-open-ended· Estonian Native Benchmark (arXiv:2510.21193v2) data0.746
Gemini 2.0 Flash 001instruction-tuned modelstable: 10, evaluation: pairwise-open-ended· Estonian Native Benchmark (arXiv:2510.21193v2) data0.699
Gemma 2 27B Itinstruction-tuned modelstable: 10, evaluation: pairwise-open-ended· Estonian Native Benchmark (arXiv:2510.21193v2) data0.642
Gemma 2 9B Itinstruction-tuned modelstable: 10, evaluation: pairwise-open-ended· Estonian Native Benchmark (arXiv:2510.21193v2) data0.211
GPT-4oinstruction-tuned modelstable: 10, evaluation: pairwise-open-ended· Estonian Native Benchmark (arXiv:2510.21193v2) data0.545
GPT-4o Miniinstruction-tuned modelstable: 10, evaluation: pairwise-open-ended· Estonian Native Benchmark (arXiv:2510.21193v2) data0.533
Llama 3.1 405B Instructinstruction-tuned modelstable: 10, evaluation: pairwise-open-ended· Estonian Native Benchmark (arXiv:2510.21193v2) data0.5
Llama 3.1 70B Instructinstruction-tuned modelstable: 10, evaluation: pairwise-open-ended· Estonian Native Benchmark (arXiv:2510.21193v2) data0.552
Llama 3.1 8B Instructinstruction-tuned modelstable: 10, evaluation: pairwise-open-ended· Estonian Native Benchmark (arXiv:2510.21193v2) data0.389
Mistral Nemo Instructinstruction-tuned modelstable: 10, evaluation: pairwise-open-ended· Estonian Native Benchmark (arXiv:2510.21193v2) data0.327
Qwen2 72B Instructinstruction-tuned modelstable: 10, evaluation: pairwise-open-ended· Estonian Native Benchmark (arXiv:2510.21193v2) data0.442

Data credits: Lillepalu and Alumäe, Estonian Native Large Language Model Benchmark, arXiv:2510.21193v2. Cited numeric facts only. arXiv non-exclusive distribution is not a results reuse grant. Emitted and committed payloads exclude tasks, examples and generations. Human preference models were available in March 2025; judge evaluation was in March 2025.