MILP-Evolve is a large-scale dataset of Mixed Integer Linear Programming (MILP) problem classes and instances generated by an evolutionary framework based on large language models (LLMs), released by Microsoft Research in 2025. The related research paper is "Towards Foundation Models for Mixed Integer Linear Programming", aimed at promoting research on MILP foundation models that can generalize across problem classes.

The dataset contains diverse MILP class code representations generated by the evolutionary framework (divided into Tab 1 and Tab 2 corresponding to the paper's tables), as well as up to 1,000 MILP instances in MPS format generated under each class using different seeds. The data supports various learning tasks such as integrality gap prediction, learning to branch, and alignment of MILP instances with natural language descriptions, providing rich data resources for model training and evaluation in related fields.

Dataset Composition

• milp_code: MILP class codes, where tab 1 and tab 2 under evolve/seed correspond to Table 1 and Table 2 in the paper, respectively.

• instances: MILP instance data in MPS format, with up to 1,000 instances per class. Due to the large data volume, some instances are provided in compressed form as tab1_compressed and tab2_compressed, and need to be decompressed before use.

• gap_data_example.tar.gz: Example data for the Integrality Gap prediction task.

• language_data_example.tar.gz: Example data for the language-MILP contrastive learning task.

• Learning to Branch: Due to the large data volume, the dataset is not directly provided and needs to be generated from the GitHub source code based on Ecole.

Citation

@article{li2024towards,
  author    = {Li, Sirui and Kulkarni, Janardhan and Wu, Cathy and Menache, Ishai and Li, Beibin},
  title     = {Towards Foundation Models for Mixed Integer Linear Programming},
  booktitle = {The Thirteenth International Conference on Learning Representations},
  year      = {2025}
}