Llama-3-Base-8B-SFT-CPO: benchmark results
Princeton NLP's 8B Llama 3 base checkpoint tuned with supervised fine-tuning and Contrastive Preference Optimization, released alongside their SimPO preference-optimization research. Provider: Princeton NLP. Released 2024-07-06. Access: Open.
Aggregate Rating 1370 ± 20, rank #1337 of 1661 rated models, from 21 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| Open Korean LLM - KorNAT-Harmless | 67.12 | Normalized accuracy (%) | 88.1 |
| Open Korean LLM - KorNAT-Helpful | 50.83 | Normalized accuracy (%) | 87.3 |
| Open Korean LLM - Ko-Winogrande | 61.4 | Accuracy (%) | 61.7 |
| Open Portuguese LLM - ENEM | 65.43 | Accuracy (%) | 61.7 |
| Open Portuguese LLM - OAB Exams | 45.6 | Accuracy (%) | 61.4 |
| Open Korean LLM - Ko-GSM8K (flexible extract) | 26.76 | Exact match, flexible extract (%) | 60.7 |
| Open Portuguese LLM - ASSIN2 RTE | 90.37 | Macro F1 (%) | 56.8 |
| Open Korean LLM Leaderboard | 36.2 | Average Score (%) | 56.3 |
| Open Ko-LLM Leaderboard | 36.2 | Average (%) | 56.2 |
| Open Portuguese LLM - BLUEX | 52.99 | Accuracy (%) | 54.5 |
| Open Korean LLM - Ko-GPQA-Diamond | 22.73 | Normalized accuracy (%) | 40.9 |
| Open Korean LLM - Ko-IFEval | 26.54 | Strict accuracy, prompt/instruction mean (%) | 39.2 |
Interactive version: theaggregate.ai/model?slug=llama-3-base-8b-sft-cpo · How It Works · Data refreshed daily, snapshot 2026-10-23.