metadata
dataset_info:
features:
- name: id
dtype: string
- name: theme
dtype: string
- name: level
dtype: string
- name: problem
dtype: string
- name: answer
dtype: string
- name: solution
dtype: string
splits:
- name: combined
num_bytes: 58963
num_examples: 175
- name: cpge
num_bytes: 35624
num_examples: 75
- name: bac
num_bytes: 23339
num_examples: 100
download_size: 75701
dataset_size: 117926
configs:
- config_name: default
data_files:
- split: combined
path: data/combined-*
- split: cpge
path: data/cpge-*
- split: bac
path: data/bac-*
Kholle Benchmark
Kholle is a french benchmark designed for small language models (SLMs) to evaluate their academic and school-level knowledge in science, ranging from high school to French preparatory classes (classes prépa).
The benchmark consists of short course-style questions and small exercises, similar to those that could appear in exams.
All metrics and extraction code for evaluation are available in the official Luth repo.
Dataset Construction
- Built manually from official French high school and preparatory class curricula.
- Most mathematics questions are sourced from Bibmath.
- Covers three main scientific domains:
- Mathematics: 100 questions
- Physics & Chemistry: 50 questions
- Biology (SVT): 25 questions
Statistics
| Domain | # Questions | Level split (HS / Prépa) |
|---|---|---|
| Mathematics | 100 | 50 / 50 |
| Physics & Chemistry | 50 | 25 / 25 |
| Biology (SVT) | 25 | 25 |
| Total | 175 | 100 / 75 |
Citation
@misc{scholar2025kurakurai,
title = {Kholle},
author = {Kurakura AI Team},
year = {2025},
howpublished = {\url{https://huggingface.co/kurakurai/kholle}},
note = {French benchmark designed for **small language models (SLMs)** to evaluate their academic and school-level knowledge in science.}
}