suzume-llama-3-8B-multilingual-orpo-borda-half: benchmark results
Lightblue's Suzume ORPO (borda-half), an ORPO fine-tune of their multilingual Llama 3 8B Suzume model on the 50% most consistently ranked responses in the Mitsu preference dataset. Provider: Lightblue. Released 2024-04-25. Access: Open.
Aggregate Rating 1460 ± 15, rank #938 of 1656 rated models, from 30 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| Open Korean LLM - Ko-EQ-Bench | 44.55 | EQ-Bench score | 93.7 |
| Open Korean LLM Leaderboard | 44.8 | Average Score (%) | 91.1 |
| Open Ko-LLM Leaderboard | 44.8 | Average (%) | 90.9 |
| Open Korean LLM - KorNAT-CKA | 31.57 | Normalized accuracy (%) | 87.1 |
| Open Korean LLM - Ko-IFEval | 42.67 | Strict accuracy, prompt/instruction mean (%) | 86.7 |
| Open Korean LLM - KorNAT-Harmless | 65.91 | Normalized accuracy (%) | 83.6 |
| Open Korean LLM - Ko-GSM8K (flexible extract) | 46.4 | Exact match, flexible extract (%) | 80.8 |
| Open LLM Leaderboard - IFEval | 62.49 | Score | 74.5 |
| Open Portuguese LLM - ASSIN2 RTE | 92.03 | Macro F1 (%) | 73.4 |
| Open Chinese LLM - GSM8K | 49.05 | Accuracy (%) | 64.1 |
| Open Portuguese LLM - FaQuAD NLI | 69.49 | Macro F1 (%) | 59.4 |
| Ru Arena Hard | 57.13 | Win Rate (%) | 56.3 |
Interactive version: theaggregate.ai/model?slug=suzume-llama-3-8b-multilingual-orpo-borda-half · How It Works · Data refreshed daily, snapshot 2026-10-20.