Marin 8B Instruct: benchmark results
Stanford CRFM's Marin 8B Instruct, an SFT-only fine-tune of the open Marin 8B Base on a mix of reasoning, math and instruction datasets including OpenThoughts, Tulu-3 and dolphin-r1. Provider: Stanford CRFM. Released 2025-05-14. Access: Open.
Aggregate Rating 1358 ± 48, rank #7507 of 10556 rated models, from 12 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| HELM Safety - Harm Bench | 75.06 | LM Evaluated Safety score | 53.5 |
| HELM Safety - Safety Scenarios | 87 | Mean score | 25.6 |
| HELM Safety - Anthropic Red Team | 97.02 | LM Evaluated Safety score | 23.3 |
| HELM Safety - Simple Safety Tests | 95.75 | LM Evaluated Safety score | 19.2 |
| SwitzerlandQA | 53.7 | Exact match (%) | 17.6 |
| HELM Safety - Xstest | 91.67 | LM Evaluated Safety score | 15.1 |
| HELM Safety - Bbq | 75.5 | BBQ accuracy | 14 |
| HELM Capabilities - Omni-MATH | 15.95 | Acc | 10.4 |
| HELM Capabilities - IFEval | 63.15 | IFEval Strict Acc | 4.5 |
| HELM Capabilities - MMLU-Pro | 18.8 | COT correct | 1.5 |
| HELM Capabilities - GPQA | 16.82 | COT correct | 0 |
| HELM Capabilities - WildBench | 47.71 | WB Score | 0 |
Interactive version: theaggregate.ai/model?slug=marin-8b-instruct · How It Works · Data refreshed daily, snapshot 2026-11-17.