Biogod/mas_reasoning

MAS Reasoning Formal Study Archive

This private dataset repository preserves the complete formal research artifacts used by the associated multi-agent reasoning manuscript. It is a versioned data archive, not a repackaging of the original benchmark datasets.

Visibility requirement: keep this full archive private. It contains generated candidate answers and rationales, and some records are derived from benchmarks whose examples should not be redistributed publicly. A separately built, content-reduced release is required before public access is enabled.

Scope

The archive contains every file under the study's two formal evidence roots, except the operational files listed in EXCLUDED_FILES.csv:

• original generated sample banks;
• natural and composition-controlled candidate pools;
• selector rankings and replay records;
• model-task logs and parse audits;
• intermediate analysis artifacts;
• manuscript-facing summaries and source tables;
• derived figure data needed by the submission repository.

The archive does not include API credentials, process-ID files, macOS metadata, the original third-party benchmark distributions or unrelated pilot outputs. SOURCE_SYMLINKS.csv records links from the source archive; linked files are materialized in this release so that every uploaded payload is self-contained.

Layout

data/external/evidence/formal/   # manuscript-facing evidence and audits
data/external/results/formal/    # raw, intermediate and summary formal results
data/derived/                    # compact derived tables used by figures
DATA_MANIFEST.csv                # source, role, size and SHA-256 per payload
CHECKSUMS.sha256                 # digest for every released file
RELEASE_METADATA.json            # release scope and aggregate counts
SOURCE_SYMLINKS.csv              # source link reconstruction map
EXCLUDED_FILES.csv               # explicit exclusion ledger
THIRD_PARTY_NOTICES.md           # benchmark licences and restrictions

Models and settings

The manuscript-facing labels are:

• DeepSeek V4 Flash (medium-thinking);
• Mimo v2.5 (no-think);
• Mimo v2.5 (medium-thinking).

The generator and judge settings are encoded in file paths and table fields. Different generator-judge cells should not be treated as interchangeable.

Download a pinned snapshot

Use the revision recorded in data/dataset.lock.json in the code repository:

uvx --from huggingface_hub hf download Biogod/mas_reasoning \
  --repo-type dataset \
  --revision REVISION_FROM_DATASET_LOCK \
  --local-dir data/releases/REVISION_FROM_DATASET_LOCK

To use the downloaded data with the plotting and analysis code, expose its two data roots at data/external/ or pass the downloaded paths to the documented analysis commands. The code-repository data/README.md provides exact commands.

Integrity verification

From the downloaded snapshot:

cd data/releases/REVISION_FROM_DATASET_LOCK
sha256sum -c CHECKSUMS.sha256

On macOS, use shasum -a 256 -c CHECKSUMS.sha256.

The repository also provides a structural verifier:

cd code
uv sync
uv run python -m reasoning_meme.io.public_dataset verify \
  --root ../data/releases/REVISION_FROM_DATASET_LOCK

Data interpretation

• rank AUC is the probability that the judge places a correct candidate above an incorrect candidate within the evaluated pair set.
• p_gen is question-level correct-answer availability in the stated generation bank.
• Natural pools preserve the observed answer distribution; composition-controlled pools fix the correct/wrong count.
• Cost fields are paid-equivalent estimates reconstructed from token accounting and the stated price model, not provider invoices.
• Some analyses use different strict-common question intersections. Compare or subtract rows only when their cohort identifiers match.

Licence and reuse

The authors release their newly derived measurements and metadata for research audit. This archive does not relicense benchmark content or model-generated text. Access to the private archive does not grant permission to publish its contents. Consult THIRD_PARTY_NOTICES.md and the original benchmark terms before redistributing any file.