yuanchenyang/mathoptdataset-sdp

MathOptDataset SDP

Semidefinite optimization instances in MathOptFormat 1.9. Each instance is a model file (<name>.mof.json, or <name>.mof.json.zst when compressed with zstd) paired with a YAML sidecar (<name>.yaml) holding its metadata. Pairs live in hash-range directories (h00-1f/ through he0-ff/) and have sdp/<name> identifiers. metadata/instances.parquet indexes every pair.

Browse provenance, cone structure, scale, and solver results for every instance in the interactive overview. The solver results compare recorded solver runs by outcome, accuracy against the documented references, speed, and disagreements between solvers.

Using the dataset

The default dataset contains metadata only. model_path and sidecar_path are relative repository paths, not automatically downloaded attachments.

from datasets import load_dataset
from huggingface_hub import HfApi, hf_hub_download

repo = "yuanchenyang/mathoptdataset-sdp"
revision = HfApi().dataset_info(repo).sha
rows = load_dataset(repo, split="instances", revision=revision)
selected = rows.filter(lambda row: "control" in row["tags"])
for row in selected:
    path = hf_hub_download(repo, row["model_path"],
                           repo_type="dataset", revision=revision)

Downloads are cached. .zst models require zstandard decompression before parsing as JSON. model_sha256 hashes the stored bytes, including compression; model_size_bytes likewise measures the stored file. Pin metadata and model downloads to the same commit. To obtain original metadata, download sidecar_path in the same way.

Julia

Read one instance's metadata, then load and solve its model with JuMP and the open-source Clarabel solver (] add JuMP Clarabel CodecZstd YAML):

using Downloads, YAML, CodecZstd, JuMP, Clarabel

base = "https://huggingface.co/datasets/yuanchenyang/mathoptdataset-sdp/resolve/main"
name = "hc0-df/ai_hilbert_i3322_sos_bound"
getbytes(path) = take!(Downloads.download("$base/$path", IOBuffer()))

meta = YAML.load(String(getbytes("$name.yaml")))
println(meta["title"], " (", meta["provenance"]["license"], ")")
println("recorded: ", meta["solution"]["status"], " ", meta["solution"]["objective"])

source = MOI.FileFormats.Model(format = MOI.FileFormats.FORMAT_MOF)
read!(IOBuffer(transcode(ZstdDecompressor, getbytes("$name.mof.json.zst"))), source)
model = Model(Clarabel.Optimizer)
MOI.copy_to(model, source)
set_silent(model)
optimize!(model)
println(termination_status(model), " ", objective_value(model))

Uncompressed .mof.json models can be read with read_from_file(path) directly. Recorded solution_ results, where present, come from MOSEK; other solvers may need tolerances adjusted on harder instances.

Metadata

YAML sidecars are authoritative; metadata/instances.parquet is generated. Identity, descriptions, and tags are preserved. Provenance fields become provenance_kind, source_id, citation, source_url, and license. Expected outcomes use expected_ columns; solver observations use solution_ columns. Missing values are null. Dates remain source strings. The single instances split implies no training/evaluation partition.

Provenance and licensing

The card's license: other reflects that licenses differ by instance; consult each row's license, citation, and source URL. This collection has no blanket MIT license. It includes entries with unknown, noncommercial, copyleft, or otherwise restrictive terms. Inclusion here does not grant permission to redistribute an upstream instance; check each instance's terms before reuse.