LLaMA3-iterative-DPO-final: benchmark results
RLHFlow's LLaMA3-iterative-DPO-final, trained with an iterative online DPO-based RLHF recipe on open-source datasets only, from the 'RLHF Workflow' paper's reward-modeling-to-online-RLHF pipeline. Provider: RLHFlow. Released 2024-05-17. Access: Open.
Aggregate Rating 1455 ± 19, rank #970 of 1656 rated models, from 27 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| Open Korean LLM - Ko-GSM8K (flexible extract) | 62.77 | Exact match, flexible extract (%) | 95.6 |
| Open Ko-LLM Leaderboard | 45.23 | Average (%) | 91.4 |
| Open Korean LLM Leaderboard | 45.23 | Average Score (%) | 91.4 |
| Open Korean LLM - KorNAT-Harmless | 67.68 | Normalized accuracy (%) | 90.2 |
| Open Korean LLM - Ko-IFEval | 38.69 | Strict accuracy, prompt/instruction mean (%) | 80 |
| Open Korean LLM - Ko-EQ-Bench | 30.55 | EQ-Bench score | 79.4 |
| Open Portuguese LLM - ASSIN2 RTE | 92.28 | Macro F1 (%) | 77.3 |
| Open Korean LLM - Ko-GPQA-Diamond | 25.25 | Normalized accuracy (%) | 75.9 |
| Open Korean LLM - KorNAT-CKA | 27.98 | Normalized accuracy (%) | 73 |
| Open Portuguese LLM - BLUEX | 58 | Accuracy (%) | 71.7 |
| Open Portuguese LLM - ENEM | 67.39 | Accuracy (%) | 66.7 |
| Open Korean LLM - KorNAT-Helpful | 48.23 | Normalized accuracy (%) | 65.8 |
Interactive version: theaggregate.ai/model?slug=llama3-iterative-dpo-final · How It Works · Data refreshed daily, snapshot 2026-10-20.