Phi-3-medium-4k-instruct-abliterated-v3: benchmark results
failspy's abliterated version of Microsoft's Phi-3-medium-4k-instruct, with the refusal direction orthogonalized out of the weights. Provider: failspy. Released 2024-05-22. Access: Open.
Aggregate Rating 1375 ± 18, rank #1321 of 1661 rated models, from 27 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| Open LLM Leaderboard v1 - MMLU | 75.52 | Accuracy (%) (5-shot) | 96.9 |
| Open LLM Leaderboard v1 - GSM8K | 71.87 | Accuracy (%) (5-shot) | 96.8 |
| Open LLM Leaderboard - MuSR | 46.04 | Score | 87.9 |
| Open LLM Leaderboard - BBH | 63.05 | Score | 86.8 |
| Open LLM Leaderboard - MMLU-Pro | 44 | Score | 81.3 |
| Open Portuguese LLM - ASSIN2 RTE | 92.56 | Macro F1 (%) | 80.2 |
| Open Portuguese LLM - BLUEX | 59.25 | Accuracy (%) | 76.1 |
| Open LLM Leaderboard - IFEval | 63.19 | Score | 75.3 |
| Open Portuguese LLM - ENEM | 69.63 | Accuracy (%) | 73.5 |
| Open LLM Leaderboard - GPQA | 31.71 | Score | 73.2 |
| Open LLM Leaderboard v1 - HellaSwag | 85.07 | Normalized accuracy (%) (10-shot) | 71.5 |
| Open LLM Leaderboard v1 - ARC Challenge | 65.44 | Normalized accuracy (%) (25-shot) | 70 |
Interactive version: theaggregate.ai/model?slug=phi-3-medium-4k-instruct-abliterated-v3 · How It Works · Data refreshed daily, snapshot 2026-10-23.