scibench / README.md
Mandy Wang
add data
d829b4e
|
raw
history blame
1.1 kB

SciBench

SciBench is a novel benchmark for college-level scientific problems consisting of 695 problems sourced from instructional textbooks. The benchmark is designed to evaluate the complex reasoning capabilities, strong domain knowledge, and advanced calculation skills of LLMs. Please refer to our paper for full description: SciBench: Evaluating College-Level Scientific Problem-Solving Abilities of Large Language Models

We developed an innovative evaluation protocol for a detailed analysis of reasoning abilities. This involves instructing LLMs to self-identify and categorize their errors within a predefined set of capabilities. This process offers a fine-grained understanding of where the models are falling short.

Data

Each file is list of dictionary and can be extracted using following scripts. Each file stands for one textbook, which is fully elaborated in the paper.

subject='atkins'
with open("./data/{}.json".format(subject), encoding='utf-8') as json_file:
    problems=json.load(json_file)

license: mit