Contamination MapText / LLM / Dataset
MMLU-Pro
12,032 test questions in 14 disciplines with (typically) ten answer options, plus a 70-question validation split used for few-shot prompts. 6,810 questions come from the original MMLU (after removing trivial and noisy ones) and 5,222 are newly added from STEM websites, TheoremQA and SciBench.
- Also known as
- MMLU-Pro, MMLU Pro
- Domain
- knowledge
- License
- MIT (TIGER-Lab/MMLU-Pro HF card)
- Homepage
- github.com
Related datasets
-
derived from MMLU (Measuring Massive Multitask Language Understanding)
6,810 of 12,032 questions are filtered MMLU questions with distractors added (card table); the rest come from other sources.
Evidence
“**Original MMLU Questions:** Part of the dataset comes from the original MMLU dataset. We remove the trivial and ambiguous questions.”
huggingface.co · dataset card · 2. Dataset Summary · retrieved 2026-10-15 “Additionally, MMLU-Pro eliminates the trivial and noisy questions in MMLU.”
arxiv.org · paper · Abstract · retrieved 2026-10-15
Exposure by model
| Benchmark | Model | Level | Reason |
|---|---|---|---|
| Recent LLM benchmarks | LLM360 Amber | Unknown | training data not disclosed in enough detail |
| Recent LLM benchmarks | BLOOM | Unknown | training data not disclosed in enough detail |
| Recent LLM benchmarks | Claude 3.5 Sonnet | Unknown | training data not disclosed in enough detail |
| Recent LLM benchmarks | Claude 3.7 Sonnet | Unknown | training data not disclosed in enough detail |
| Recent LLM benchmarks | Claude 3 (Opus, Sonnet, Haiku) | Unknown | training data not disclosed in enough detail |
| Recent LLM benchmarks | Claude Opus 4 and Claude Sonnet 4 | Unknown | training data not disclosed in enough detail |
| Recent LLM benchmarks | Claude Fable 5 / Claude Mythos 5 | Unknown | training data not disclosed in enough detail |
| Recent LLM benchmarks | Claude Mythos Preview | Unknown | training data not disclosed in enough detail |
| Recent LLM benchmarks | Claude Opus 4.5 | Unknown | training data not disclosed in enough detail |
| Recent LLM benchmarks | Claude Opus 5 | Unknown | training data not disclosed in enough detail |
| Recent LLM benchmarks | Claude Sonnet 5 | Unknown | training data not disclosed in enough detail |
| Recent LLM benchmarks | DCLM-7B-IT (DCLM-IT-7B) | Test data seen | a contamination analysis found test examples in the pre-training data (16.20% of (downsampled) MMLU-Pro test questions are 'dirty' in DCLM-baseline.) [third-party study] (via DCLM-Baseline 1.0) |
| Recent LLM benchmarks | DCLM-7B (DCLM-Baseline-7B) | Test data seen | a contamination analysis found test examples in the pre-training data (16.20% of (downsampled) MMLU-Pro test questions are 'dirty' in DCLM-baseline.) [third-party study] (via DCLM-Baseline 1.0) |
| Recent LLM benchmarks | DeepSeek-R1-0528 | Unknown | training data not disclosed in enough detail |
| Recent LLM benchmarks | DeepSeek-R1-Distill | Unknown | training data not disclosed in enough detail |
| Recent LLM benchmarks | DeepSeek-R1 | Not seen | the authors state its test portion was excluded from training |
| Recent LLM benchmarks | DeepSeek-V3-0324 | Unknown | training data not disclosed in enough detail |
| Recent LLM benchmarks | DeepSeek-V3.1 | Unknown | training data not disclosed in enough detail |
| Recent LLM benchmarks | DeepSeek-V3.2-Exp | Unknown | training data not disclosed in enough detail |
| Recent LLM benchmarks | DeepSeek-V3.2 | Unknown | training data not disclosed in enough detail |
| Recent LLM benchmarks | DeepSeek-V3 | Unknown | training data not disclosed in enough detail |
| Recent LLM benchmarks | DeepSeek-V4.1-Flash | Unknown | training data not disclosed in enough detail |
| Recent LLM benchmarks | DeepSeek-V4 (Pro / Flash) | Unknown | training data not disclosed in enough detail |
| Recent LLM benchmarks | Gemini 1.5 (Pro, Flash) | Unknown | training data not disclosed in enough detail |
| Recent LLM benchmarks | Gemini 2.5 (Pro, Flash) | Unknown | training data not disclosed in enough detail |
| Recent LLM benchmarks | Gemini 3 Pro | Unknown | training data not disclosed in enough detail |
| Recent LLM benchmarks | Gemini 4 Argon | Unknown | training data not disclosed in enough detail |
| Recent LLM benchmarks | Gemma 2 | Unknown | training data not disclosed in enough detail |
| Recent LLM benchmarks | Gemma 3 | Unknown | training data not disclosed in enough detail |
| Recent LLM benchmarks | Gemma 3n | Unknown | training data not disclosed in enough detail |
| Recent LLM benchmarks | Gemma 4 | Unknown | training data not disclosed in enough detail |
| Recent LLM benchmarks | GPT-4 | a contamination analysis found test examples of the related dataset MMLU (Measuring Massive Multitask Language Understanding) (~0.6% of 1000 sampled MMLU test examples matched GPT-4 pre-training data (50-character substring match)) | |
| Recent LLM benchmarks | GPT-4o | Unknown | training data not disclosed in enough detail |
| Recent LLM benchmarks | GPT-5.2 | Unknown | training data not disclosed in enough detail |
| Recent LLM benchmarks | GPT-5 | Unknown | training data not disclosed in enough detail |
| Recent LLM benchmarks | GPT-6 Astra | Unknown | training data not disclosed in enough detail |
| Recent LLM benchmarks | gpt-oss | Unknown | training data not disclosed in enough detail |
| Recent LLM benchmarks | InternVL3.5 | Unknown | unknown for its base model Qwen3: training data not disclosed in enough detail |
| Recent LLM benchmarks | InternVL3 | Unknown | unknown for its base model Qwen2.5: training data not disclosed in enough detail |
| Recent LLM benchmarks | LLM360 K2-65B (K2 Diamond) | a contamination analysis found test examples of the related dataset MMLU (Measuring Massive Multitask Language Understanding) (MMLU test questions are also substantially 'dirty' in Pile-train: DCLM-baseline (27.70%) has only 1.1x more dirty entries than Pile-train. Exact Pile-train rate is in Table 2 (not reliably extracted from the PDF text).) [third-party study] (via The Pile) | |
| Recent LLM benchmarks | K2-Horizon | Unknown | training data not disclosed in enough detail |
| Recent LLM benchmarks | K2-V2 | Unknown | training data not disclosed in enough detail |
| Recent LLM benchmarks | Llama 2 | a contamination analysis found test examples of the related dataset MMLU (Measuring Massive Multitask Language Understanding) (MMLU-Humanities (L=50): 709 of 4,705 samples (15%) 'not clean' (>=20% token contamination) and 520 (11%) 'dirty' (>=80%); flagged as boosted by contamination. MMLU-Overall for 70B: 2,180 of 14,042 (15.5%) not clean, 1,536 (10.9%) dirty, with a small effect on the overall score (-0.9 clean-vs-mean delta).) | |
| Recent LLM benchmarks | Llama 3.1 | Test data seen | a contamination analysis found test examples in the pre-training data (8-gram overlap gave contamination scores too high to estimate a performance gain, even at higher thresholds; percentage not reported (dash in Table 15).) |
| Recent LLM benchmarks | Llama 3.2-Vision (11B, 90B) | Unknown | training data not disclosed in enough detail |
| Recent LLM benchmarks | Llama 3.2 (1B, 3B) | Unknown | training data not disclosed in enough detail |
| Recent LLM benchmarks | Llama 3.3 70B | Unknown | training data not disclosed in enough detail |
| Recent LLM benchmarks | Meta Llama 3 | Unknown | training data not disclosed in enough detail |
| Recent LLM benchmarks | Llama 4 (Scout, Maverick) | Unknown | training data not disclosed in enough detail |
| Recent LLM benchmarks | LLaVA-OneVision-1.5 | Unknown | unknown for its base model Qwen3: training data not disclosed in enough detail |
| Recent LLM benchmarks | LLaVA-OneVision | Unknown | unknown for its base model SigLIP: training data not disclosed in enough detail |
| Recent LLM benchmarks | Ministral 3 | Unknown | training data not disclosed in enough detail |
| Recent LLM benchmarks | Mistral 7B | Unknown | training data not disclosed in enough detail |
| Recent LLM benchmarks | Mistral Large 3 | Unknown | training data not disclosed in enough detail |
| Recent LLM benchmarks | Mistral Medium 3.5 | Unknown | training data not disclosed in enough detail |
| Recent LLM benchmarks | Mistral Small 4 | Unknown | training data not disclosed in enough detail |
| Recent LLM benchmarks | Mixtral 8x7B | Unknown | training data not disclosed in enough detail |
| Recent LLM benchmarks | Molmo (0924) | Unknown | unknown for its base model OpenAI CLIP: training data not disclosed in enough detail |
| Recent LLM benchmarks | Molmo2 | Unknown | unknown for its base model Qwen3: training data not disclosed in enough detail |
| Recent LLM benchmarks | OpenAI o1 | Unknown | training data not disclosed in enough detail |
| Recent LLM benchmarks | OpenAI o3 | Unknown | training data not disclosed in enough detail |
| Recent LLM benchmarks | OLMo (February 2024) | Unknown | training data not disclosed in enough detail |
| Recent LLM benchmarks | OLMo 2 (November 2024) | Test data seen | a contamination analysis found test examples in the pre-training data (16.20% of (downsampled) MMLU-Pro test questions are 'dirty' in DCLM-baseline.) [third-party study] (via DCLM-Baseline 1.0) |
| Recent LLM benchmarks | Olmo 3.1 | Unknown | training data not disclosed in enough detail |
| Recent LLM benchmarks | Olmo 3 | Unknown | training data not disclosed in enough detail |
| Recent LLM benchmarks | Olmo Hybrid 7B | Unknown | training data not disclosed in enough detail |
| Recent LLM benchmarks | Phi-3.5 (mini / MoE) | Unknown | training data not disclosed in enough detail |
| Recent LLM benchmarks | Phi-3 (mini / small / medium) | Unknown | training data not disclosed in enough detail |
| Recent LLM benchmarks | Phi-4-mini-flash-reasoning | Unknown | training data not disclosed in enough detail |
| Recent LLM benchmarks | Phi-4-mini-reasoning | Unknown | training data not disclosed in enough detail |
| Recent LLM benchmarks | Phi-4-mini | Unknown | training data not disclosed in enough detail |
| Recent LLM benchmarks | Phi-4-reasoning / Phi-4-reasoning-plus | Unknown | training data not disclosed in enough detail |
| Recent LLM benchmarks | Phi-4 | Not seen | the authors state its test portion was excluded from training |
| Recent LLM benchmarks | Phi-mini-MoE / Phi-tiny-MoE (SlimMoE) | Unknown | training data not disclosed in enough detail |
| Recent LLM benchmarks | Pixtral 12B | Unknown | training data not disclosed in enough detail |
| Recent LLM benchmarks | Pixtral Large | Unknown | training data not disclosed in enough detail |
| Recent LLM benchmarks | Pythia (deduplicated) | a contamination analysis found test examples of the related dataset MMLU (Measuring Massive Multitask Language Understanding) (MMLU test questions are also substantially 'dirty' in Pile-train: DCLM-baseline (27.70%) has only 1.1x more dirty entries than Pile-train. Exact Pile-train rate is in Table 2 (not reliably extracted from the PDF text).) [third-party study] (via The Pile) | |
| Recent LLM benchmarks | Pythia | a contamination analysis found test examples of the related dataset MMLU (Measuring Massive Multitask Language Understanding) (MMLU test questions are also substantially 'dirty' in Pile-train: DCLM-baseline (27.70%) has only 1.1x more dirty entries than Pile-train. Exact Pile-train rate is in Table 2 (not reliably extracted from the PDF text).) [third-party study] (via The Pile) | |
| Recent LLM benchmarks | Qwen2.5-VL | Unknown | unknown for its base model Qwen2.5: training data not disclosed in enough detail |
| Recent LLM benchmarks | Qwen2.5 | Unknown | training data not disclosed in enough detail |
| Recent LLM benchmarks | Qwen3-2507 (Instruct-2507 / Thinking-2507) | Unknown | training data not disclosed in enough detail |
| Recent LLM benchmarks | Qwen3-Next | Unknown | training data not disclosed in enough detail |
| Recent LLM benchmarks | Qwen3-VL | Unknown | unknown for its base model Qwen3: training data not disclosed in enough detail |
| Recent LLM benchmarks | Qwen3.5 | Unknown | training data not disclosed in enough detail |
| Recent LLM benchmarks | Qwen3.6 | Unknown | training data not disclosed in enough detail |
| Recent LLM benchmarks | Qwen3.8-Flash-Next | Unknown | training data not disclosed in enough detail |
| Recent LLM benchmarks | Qwen3.8 | Unknown | training data not disclosed in enough detail |
| Recent LLM benchmarks | Qwen3 | Unknown | training data not disclosed in enough detail |
| Recent LLM benchmarks | SmolLM2 | Test data seen | a contamination analysis found test examples in the pre-training data (16.20% of (downsampled) MMLU-Pro test questions are 'dirty' in DCLM-baseline.) [third-party study] (via DCLM-Baseline 1.0) |
| Recent LLM benchmarks | SmolLM3 | Unknown | training data not disclosed in enough detail |
Notes
Also incorporates questions from TheoremQA and SciBench (not registry datasets). Whether the MMLU-derived questions come from MMLU's test split only is not stated in the sources read.
Sources
- MMLU-Pro: A More Robust and Challenging Multi-Task Language Understanding Benchmark (arXiv v6) · paper · retrieved 2026-10-15
- TIGER-Lab/MMLU-Pro dataset card · dataset card · retrieved 2026-10-15
- TIGER-AI-Lab/MMLU-Pro README · readme · retrieved 2026-10-15