ArmoRM-Llama3-8B-v0.1: benchmark results
RLHFlow's ArmoRM-Llama3-8B-v0.1, an Absolute-Rating Multi-Objective reward model built on Llama 3 8B with Mixture-of-Experts aggregation of reward objectives, fine-tuned from FsfairX-LLaMA3-RM-v0.1. Provider: RLHFlow. Released 2024-05-23. Access: Open.
Aggregate Rating 1345 ± 16, rank #1394 of 1656 rated models, from 56 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| RubricBench - Chat | 50.8 | Preference accuracy (%) on the chat pairs on RubricBench's 1 | 100 |
| RewardBench Reasoning | 97.35 | Accuracy (%) | 93.5 |
| RewardBench Prior Sets (0.5 weight) | 74.29 | Score (%) | 92.8 |
| Open Korean LLM - Ko-GPQA-Diamond | 27.27 | Normalized accuracy (%) | 92.4 |
| RubricBench - Instruction Following | 44.4 | Preference accuracy (%) on the instruction-following pairs o | 88.9 |
| RewardBench Safety | 90.54 | Accuracy (%) | 83.5 |
| RewardBench Chat | 96.93 | Accuracy (%) | 83 |
| RewardBench | 88.6 | Score (%) | 79 |
| RewardBench Chat Hard | 76.75 | Accuracy (%) | 75.6 |
| RewardBench 2 Precise IF | 41.88 | Score (%) | 72.8 |
| RewardBench 2 Math | 66.12 | Score (%) | 67.5 |
| RewardBench 2 | 66.46 | Score (%) | 67.3 |
Interactive version: theaggregate.ai/model?slug=armorm-llama3-8b-v0-1 · How It Works · Data refreshed daily, snapshot 2026-10-20.