microsoft/benchpress-score-matrix

BenchPress Score Matrix

This dataset contains the public model-by-benchmark score matrix used by BenchPress. The release includes the lossless audited JSON, benchmark cost evidence, flat model and benchmark metadata, one row per observed score, and the paper-canonical dense subset used in the BenchPress experiments.

The source repository is microsoft/benchpress.

Canonical artifacts

data/llm_benchmark_data.json is the authoritative rich score-matrix artifact. It preserves nested alternative candidates, audit provenance, source URLs, and normalized benchmark cost fields that cannot be represented losslessly in CSV.

data/benchmark_cost_evidence.json is the authoritative raw public cost-evidence artifact. metadata.json records the SHA-256 digest and byte size of both canonical JSON files. The CSV and Parquet files are deterministic flat exports from the score-matrix JSON.

Files

File	Contents
data/llm_benchmark_data.json	Lossless audited score matrix: models, benchmarks, scores, candidates, and audit provenance.
data/benchmark_cost_evidence.json	Raw public token, dollar, and run-budget evidence used by benchmark cost metadata.
data/scores_all.csv / .parquet	Flat numeric score rows in the audit pool.
data/scores_paper.csv / .parquet	Long-form rows for the paper-canonical matrix.
data/models.csv / .parquet	Model metadata and canonical evaluation settings.
data/benchmarks.csv / .parquet	Benchmark metadata and canonical benchmark settings.
data/score_matrix_paper_wide.csv	Wide model x benchmark matrix for the paper-canonical subset.
data/README.md, data/SCHEMA.md	Dataset conventions and the canonical JSON schema.
data/LICENSE-CDLA-2.0.md	Dataset license text.
metadata.json	Export counts, matrix construction metadata, file inventory, and canonical JSON hashes.

Quick start

from datasets import load_dataset

scores = load_dataset("microsoft/benchpress-score-matrix", "scores_paper")["train"].to_pandas()
models = load_dataset("microsoft/benchpress-score-matrix", "models")["train"].to_pandas()
benchmarks = load_dataset("microsoft/benchpress-score-matrix", "benchmarks")["train"].to_pandas()

For the lossless audit artifact:

import json
from urllib.request import urlopen

url = (
    "https://huggingface.co/datasets/microsoft/benchpress-score-matrix/"
    "resolve/main/data/llm_benchmark_data.json"
)
with urlopen(url) as response:
    matrix = json.load(response)

Schema

The flat score tables include:

• model_id, benchmark_id, score
• reference_url, source_type, audit_status, matches_canonical
• reported_setting_json, notes

The lossless JSON additionally preserves candidates, audit-rule identifiers, audit notes, timestamps, and benchmark-level cost evidence.

models and benchmarks include an in_paper_matrix flag that identifies rows retained by the paper-canonical threshold filter.

Matrix construction

The paper-canonical matrix applies the BenchPress construction pipeline: audit-status filtering, canonical representative selection, and the iterative threshold filter. Current export counts:

• audit pool: 283 models, 712 benchmarks, 8713 score rows
• paper matrix: 129 models x 253 benchmarks, 4905 observed cells (15.0% fill)

License

The dataset files are released under CDLA-Permissive-2.0. The source repository's MIT license applies to code and documentation, not to this dataset license grant.

Caveats

Scores come from heterogeneous public sources: model cards, official blogs, technical reports, benchmark leaderboards, and third-party aggregators. Each score retains source and audit metadata so downstream users can choose their own filtering policy.