Skip to content

Dataset

Lemutisme edited this page Nov 7, 2026 · 1 revision

Dataset

benchmarks/ holds the 505 instances behind the paper's results as one archive:

  • 265 from gams_world (GAMS World), and
  • 240 from minlplib_linear (MINLPLib).

Each instance is a PyTorch module (<name>_model.py, constants in <name>_model_state.pt) and a specification (<name>.vnnlib: variable bounds and the objective and constraint outputs).

File Content
gear-benchmarks.tar.xz the models: 1521 files, 987 MiB unpacked, 9.3 MB packed
MANIFEST.tsv family, path, bytes and sha256 of every unpacked file
instances.tsv family and name of the 505 instances
SHA256SUMS checksums of the archive, MANIFEST.tsv and instances.tsv

To verify and unpack, see the last command of Installation. Pass the benchmarks/ directory as --benchmark_root to nlp_example.py or --root to gear.cli (Quick Start).

The full description (file format, directory layout, conversion from the AMPL models in baselines/benchmark/, metadata, source collections and licenses, checksums) is in benchmarks/README.md. Sources: MINLPLib (CC BY 4.0) and GAMS World (MIT License).

baselines/benchmark/manifest.tsv lists, for each instance, the family, number of variables, number of constraints, number of discrete variables (0 for all 505), sense and the nonlinear functions that occur. 501 instances are minimizations and 4 are maximizations.

The same 505 instances were run by 12 AMPL solvers under a 180-second limit; see baselines/README.md.

Clone this wiki locally