Starling-LM-7B-alpha: benchmark results
UC Berkeley's Starling-LM-7B-alpha, an RLAIF-trained fine-tune of Openchat 3.5 (Mistral 7B-based), aligned via APA with the Starling-RM-7B-alpha reward model on the GPT-4-labeled Nectar dataset. Provider: UC Berkeley. Released 2023-11-25. Access: Open.
Aggregate Rating 1411 ± 12, rank #1186 of 1656 rated models, from 89 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| Open Korean LLM - Ko-IFEval | 42.94 | Strict accuracy, prompt/instruction mean (%) | 87.2 |
| Open Portuguese LLM - FaQuAD NLI | 78.42 | Macro F1 (%) | 84 |
| Open Portuguese LLM - ASSIN2 RTE | 92.73 | Macro F1 (%) | 82.8 |
| Open Korean LLM Leaderboard | 41.03 | Average Score (%) | 80.2 |
| Open Ko-LLM Leaderboard | 41.03 | Average (%) | 80.1 |
| Open Korean LLM - Ko-EQ-Bench | 29.99 | EQ-Bench score | 78.4 |
| Open PL LLM - PSC (multiple choice, 5-shot) | 83.31 | Binary F1 (%) | 76 |
| Open PL LLM - PPC (multiple choice, 5-shot) | 76.7 | Accuracy (%) | 73.1 |
| Open Korean LLM - Ko-GSM8K (flexible extract) | 37.68 | Exact match, flexible extract (%) | 71.9 |
| Open PL LLM - DYK (multiple choice, 5-shot) | 60.69 | Binary F1 (%) | 69.8 |
| Open PL LLM - Multiple Choice | 51.16 | Average Multiple-Choice Score (%) | 69.2 |
| Open PL LLM - PolQA Reranking (multiple choice, 5-shot) | 75.73 | Accuracy (%) | 68.9 |
Interactive version: theaggregate.ai/model?slug=starling-lm-7b-alpha · How It Works · Data refreshed daily, snapshot 2026-10-20.