Meta-Llama-3-8B-Instruct-abliterated-v3: benchmark results
failspy's abliterated Llama 3 8B Instruct v3, with refusal-direction weights orthogonalized out to reduce the model's tendency to refuse. Provider: failspy. Released 2024-06-01. Access: Open.
Aggregate Rating 1405 ± 15, rank #1208 of 1656 rated models, from 45 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| Open Korean LLM - Ko-GPQA-Diamond | 27.27 | Normalized accuracy (%) | 92.4 |
| Open Medical LLM - MMLU Medical Genetics | 84 | Accuracy (%) | 91.4 |
| Open Korean LLM - Ko-IFEval | 45.55 | Strict accuracy, prompt/instruction mean (%) | 89.8 |
| Open LLM Leaderboard - IFEval | 72.45 | Score | 86.3 |
| Open Medical LLM - MedQA (USMLE) | 61.04 | Accuracy (%) | 79.4 |
| Open Portuguese LLM - ENEM | 70.68 | Accuracy (%) | 77 |
| Open Portuguese LLM - OAB Exams | 50.21 | Accuracy (%) | 76.2 |
| Open Medical LLM - MMLU College Medicine | 65.9 | Accuracy (%) | 74.7 |
| Open Portuguese LLM - FaQuAD NLI | 75.53 | Macro F1 (%) | 74 |
| Open Portuguese LLM - BLUEX | 58 | Accuracy (%) | 71.7 |
| Open Medical LLM - MedMCQA | 56.63 | Accuracy (%) | 69.3 |
| Open Chinese LLM - GSM8K | 49.13 | Accuracy (%) | 64.6 |
Interactive version: theaggregate.ai/model?slug=meta-llama-3-8b-instruct-abliterated-v3 · How It Works · Data refreshed daily, snapshot 2026-10-20.