Datasets:
dataset_info:
features:
- name: messages
list:
- name: content
dtype: string
- name: role
dtype: string
- name: images
list: string
- name: id
dtype: string
splits:
- name: NASPO_MMK12
num_examples: 2000
- name: hiyouga_geometry3k
num_examples: 601
- name: AI4Math_MathVerse
num_examples: 2180
- name: AI4Math_MathVista
num_examples: 1000
- name: We_Math
num_examples: 1740
- name: MMMU_MMMU_Pro
num_examples: 1730
- name: AI4Math_MathVerse_vision_dependent
num_examples: 1308
- name: lscpku_LogicVista
num_examples: 447
configs:
- config_name: default
data_files:
- split: NASPO_MMK12
path: data/NASPO_MMK12-*
- split: hiyouga_geometry3k
path: data/hiyouga_geometry3k-*
- split: AI4Math_MathVerse
path: data/AI4Math_MathVerse-*
- split: AI4Math_MathVista
path: data/AI4Math_MathVista-*
- split: We_Math
path: data/We_Math-*
- split: MMMU_MMMU_Pro
path: data/MMMU_MMMU_Pro-*
- split: AI4Math_MathVerse_vision_dependent
path: data/AI4Math_MathVerse_vision_dependent-*
- split: lscpku_LogicVista
path: data/lscpku_LogicVista-*
task_categories:
- image-text-to-text
language:
- en
library_name: datasets
tags:
- multimodal
- reasoning
- evaluation
NAS-PO evaluation benchmark
This repository contains the eight evaluation configurations reported by NAS-PO. They are prepared from MMK12, Geometry3K, MathVerse, MathVista, We-Math, MMMU-Pro, and LogicVista.
The release contains eight Hugging Face dataset splits under data/ and one
matching image ZIP for each split. The NAS-PO preprocessing command converts
the selected splits to ShareGPT JSON and extracts the archives into its local
data/images/ directory. Image references are portable paths relative to
that materialized data root.
Usage
from datasets import load_dataset
mmk12 = load_dataset("ANke121/NASPO_eval", split="NASPO_MMK12")
assert len(mmk12) == 2000
The complete inference and Accuracy@8 workflow is provided by NASPO-Eval
in the NAS-PO source repository.
Acknowledgements
The public PAPO release was used as a reference for organizing these files and the evaluation workflow. The individual datasets and benchmarks remain credited to their original creators and are subject to their respective licenses and terms.