Llama-3-Instruct-8B-SPPO-Iter3: benchmark results
UCLA-AGI's Llama-3-Instruct-8B-SPPO-Iter3, the third iteration of Self-Play Preference Optimization applied to Llama 3 8B Instruct, trained on synthetic UltraFeedback-derived responses. Provider: UCLA. Released 2024-06-25. Access: Open.
Aggregate Rating 1414 ± 15, rank #1175 of 1656 rated models, from 30 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| Open Korean LLM - Ko-GSM8K (flexible extract) | 54.81 | Exact match, flexible extract (%) | 88.7 |
| Open Korean LLM - KorNAT-Helpful | 50.75 | Normalized accuracy (%) | 86.9 |
| Open Korean LLM - Ko-IFEval | 40 | Strict accuracy, prompt/instruction mean (%) | 81.7 |
| Open LLM Leaderboard - IFEval | 68.34 | Score | 81 |
| Open Portuguese LLM - OAB Exams | 50.02 | Accuracy (%) | 75.6 |
| Open Portuguese LLM - ENEM | 69.91 | Accuracy (%) | 74.4 |
| Open Portuguese LLM - BLUEX | 58 | Accuracy (%) | 71.7 |
| Open Chinese LLM - GSM8K | 52.08 | Accuracy (%) | 71.3 |
| Open Korean LLM - Ko-GPQA-Diamond | 24.75 | Normalized accuracy (%) | 69.5 |
| Open Korean LLM - KorNAT-Harmless | 64 | Normalized accuracy (%) | 69.3 |
| Open Chinese LLM - TruthfulQA MC | 56.29 | Accuracy (%) | 68.6 |
| Open Portuguese LLM - ASSIN2 RTE | 91.12 | Macro F1 (%) | 63.8 |
Interactive version: theaggregate.ai/model?slug=llama-3-instruct-8b-sppo-iter3 · How It Works · Data refreshed daily, snapshot 2026-10-20.