Llama-3-8B-Instruct-abliterated-v2: benchmark results
QuixiAI's Llama-3-8B-Instruct-abliterated-v2, an orthogonalized Llama 3 8B Instruct with weights manipulated to inhibit refusal, based on the 'Refusal in LLMs is mediated by a single direction' method. Provider: failspy. Released 2024-06-01. Access: Open.
Aggregate Rating 1423 ± 14, rank #1136 of 1656 rated models, from 41 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| Open LLM Leaderboard v1 - GSM8K | 71.34 | Accuracy (%) (5-shot) | 95.7 |
| Open Korean LLM - Ko-IFEval | 49.82 | Strict accuracy, prompt/instruction mean (%) | 93.2 |
| Open LLM Leaderboard v1 - MMLU | 67.43 | Accuracy (%) (5-shot) | 89.3 |
| Open Korean LLM - Ko-EQ-Bench | 34.49 | EQ-Bench score | 84.3 |
| Open Medical LLM - MMLU College Biology | 79.17 | Accuracy (%) | 79.7 |
| Open Korean LLM Leaderboard | 40.8 | Average Score (%) | 78.9 |
| Open Ko-LLM Leaderboard | 40.8 | Average (%) | 78.7 |
| Open Medical LLM - MedQA (USMLE) | 60.96 | Accuracy (%) | 78.6 |
| Open Korean LLM - KorNAT-CKA | 26.81 | Normalized accuracy (%) | 66.8 |
| Open Medical LLM - MMLU Medical Genetics | 80 | Accuracy (%) | 66.4 |
| Open Medical LLM - MedMCQA | 56.32 | Accuracy (%) | 65.6 |
| Open Korean LLM - Ko-GSM8K (flexible extract) | 29.11 | Exact match, flexible extract (%) | 63.4 |
Interactive version: theaggregate.ai/model?slug=llama-3-8b-instruct-abliterated-v2 · How It Works · Data refreshed daily, snapshot 2026-10-20.