MMLU-CF 5-shot Validation: leaderboard

MMLU-CF 5-shot Validation — Official MMLU-CF contamination-free benchmark; 5-shot validation cohort. Original MMLU column and test-minus-validation deltas are not independent capability panels. Native API/checkpoint labels retained. Metric: Accuracy (%). Source: https://github.com/microsoft/MMLU-CF

Metric: Accuracy (%). Source: github.com. Saturation forecast: Estimated already saturated. 44 models tracked. Updated 16 Nov 2026.

Top models

#ModelScoreOverall rank
1GPT-4o73.4#744
2Qwen 2.5 72B Instruct71.3#1095
3GPT-4 Turbo70.1#999
4Llama 3 70B Instruct68.8#1689
5Qwen 2.5 32B Instruct68.8#1232
6phi-4-14B68.5#1666
7Llama 3.1 70B Instruct68.1#1542
8Llama 3.3 70B Instruct67.8#1379
9Qwen 2.5 14B Instruct66.1#1750
10GPT-4o Mini65.1#1658
11Gemini 1.5 Flash64.9#1694
12Phi-3.5-MoE-instruct64.5#2199
13Qwen 2 72B Instruct64.3#1514
14Phi-3-medium-128k-instruct64.2#3321
15Gemma 2 27B63.5#2141

No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.

Interactive version: theaggregate.ai/benchmark?slug=mmlu-cf-5-shot-validation · How It Works · Data refreshed daily, snapshot 2026-11-16.