metadata
dataset_info:
features:
- name: task_id
dtype: string
- name: prompt
dtype: string
- name: entry_point
dtype: string
- name: canonical_solution
dtype: string
- name: test
dtype: string
splits:
- name: train
num_bytes: 233557
num_examples: 82
download_size: 140824
dataset_size: 233557
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
This is the dataset for the paper https://arxiv.org/abs/2503.05860
HumanEvalNext
HumanEvalNext is an improved version of the HumanEval code generation benchmark. The improvements made are based on the framework EvalPro. This framework is aimed at enhancing benchmark quality through a rigorous improvement process along with peer reviews. An overview of the approach along with the specific applicability to HumanEval has been illustrated below.
Figure 1. The Process of improving HumanEval through the BenchFrame framework
The evaluation results of benchmark can be found here. A more detailed description of HumanEvalNext can be found here. We find the following results results when benchmarking 10 SOTA open-weight models.
Figure 2. Comparison of HumanEval with two enhanced versions. (HumanEvalNext and EvalPlus)