Atlas

Benchmarks

← All benchmarks

Global PIQA - Non-Parallel (Log-Likelihood Accuracy)

General QA · 2025-10-28

Global PIQA's non-parallel split contains locally authored physical-commonsense questions with two answer options. This definition compares answer-option log likelihoods without requiring the model to generate a choice token.

Top models (higher is better)

ModelScore
Gemma 3 12B PT64.2
Llama 3.1 70B64.0
Gemma 2 27B63.4
Gemma 2 9B60.5
Qwen2.5 72B60.2
Qwen2.5 32B57.7
Llama 3.1 8B56.9
Mixtral 8x7B56.2
Falcon 40B53.6
Mistral 7B v0.1 Base53.3
Falcon 7B49.8
Gemma 3 270M49.8
Loading Atlas data…