Mistral7B-PairRM-SPPO-Iter3: benchmark results
UCLA-AGI's Mistral7B-PairRM-SPPO-Iter3, the third iteration of Self-Play Preference Optimization applied to Mistral 7B Instruct v0.2, generating 5 responses per iteration on synthetic UltraFeedback-derived prompts. Provider: UCLA. Released 2024-05-04. Access: Open.
Aggregate Rating 1379 ± 17, rank #1302 of 1656 rated models, from 29 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| Open Korean LLM - KorNAT-Harmless | 67.91 | Normalized accuracy (%) | 90.9 |
| Open Chinese LLM - TruthfulQA MC | 60.84 | Accuracy (%) | 84.6 |
| Open Korean LLM - Ko-GPQA-Diamond | 25.76 | Normalized accuracy (%) | 81.3 |
| Open Korean LLM - KorNAT-Helpful | 49.5 | Normalized accuracy (%) | 79.9 |
| Open Korean LLM - Ko-EQ-Bench | 24.64 | EQ-Bench score | 70.4 |
| Open Korean LLM - Ko-IFEval | 31.79 | Strict accuracy, prompt/instruction mean (%) | 60.8 |
| Open Portuguese LLM - ASSIN2 RTE | 90.55 | Macro F1 (%) | 58.5 |
| Open Korean LLM - Ko-GSM8K (flexible extract) | 22.74 | Exact match, flexible extract (%) | 55.4 |
| Open Korean LLM Leaderboard | 35.52 | Average Score (%) | 52 |
| Open Ko-LLM Leaderboard | 35.52 | Average (%) | 51.9 |
| Open Portuguese LLM - FaQuAD NLI | 65.38 | Macro F1 (%) | 50 |
| Open Portuguese LLM - BLUEX | 51.04 | Accuracy (%) | 48.4 |
Interactive version: theaggregate.ai/model?slug=mistral7b-pairrm-sppo-iter3 · How It Works · Data refreshed daily, snapshot 2026-10-20.