peterkirby/mymirmdatasets

MUSE: MAC and Exorde authorship training data

Prepared by Peter Kirby for the Multilingual Universal Style Embeddings (MUSE) project. These are supplemental authorship training sources with account-grouped train/validation splits. Each row preserves one supplied source text after recorded outer trimming. No instruction prefixes, model outputs, or contrastive pairs have been inserted into the text.

Source	Train rows	Validation rows	Physical rows
MAC	5,971,263	663,473	6,634,736
Exorde	10,602,383	1,178,034	11,780,417
Total	16,573,646	1,841,507	18,415,153

Configurations and loading

• mac_authorship: MAC full v1, with 59 verified wiki language labels.
• exorde_authorship: the complete selected Exorde view.
• exorde_authorship__<language>__<platform>: 77 Exorde language/platform views across 41 languages. For example, exorde_authorship__en__x.com.

The aggregate Exorde view references the same shards as its individual configurations. Count physical rows once; do not concatenate the aggregate with its component views. The exact configurations, file inventories and split counts are in manifests/configurations.json.

from datasets import load_dataset

repo = "OWNER/REPO"  # Replace with this repository's ID.
mac = load_dataset(repo, "mac_authorship", split="train", streaming=True)
exorde = load_dataset(repo, "exorde_authorship", split="train", streaming=True)
english_x = load_dataset(repo, "exorde_authorship__en__x.com",
                         split="validation", streaming=True)
print(next(iter(english_x)))

Streaming avoids downloading the entire repository. A configuration's validation split is for development on held-out operational accounts from that source.

Shared training fields

Field	Meaning
text	One supplied MAC fragment or Exorde original text, with outer trimming only.
primary_id_int	Numeric account label within a source release.
secondary_category_int	MAC wiki/page category; Exorde uses -1 for unavailable shared-content groups.
source_record_id	Representative source occurrence identifier.
language	Verified wiki language for MAC; normalized upstream machine prediction for Exorde.
split_group_id	Canonical account grouping key used to assign splits.
text_length_chars	Unicode character count of the retained text.

Account labels are shared across all Exorde configurations and both splits. They are scoped to this release; use (source, primary_id_int) when combining MAC and Exorde. Exorde configuration mapping subsets can contain non-contiguous numeric IDs. Retain source URLs, attribution, hashes, file/row locators and auxiliary metadata. Feed text and the chosen instruction to the encoder; account IDs and other metadata support sampling and provenance.

Preparation and sampling

MAC uses wiki-scoped account identities, 2–100 distinct texts per author-language, a 500,000-text language cap, seed 42 and grouped train/validation assignment. Folder bn was verified as Bashkir Wikipedia and is labeled ba; the original folder label and verification evidence remain in the MAC package.

Exorde uses policy version 2: at least 50 characters after outer trimming, 2–100 distinct texts per account-language-platform, and a 500,000-row cap per combination. Every combination with at least 1,500 eligible accounts across the pinned month is included; there is no top-five restriction. The threshold applies before sampling and splitting. Platform/account identities are based on normalized URL namespaces and the publisher-supplied author hash. Seed-42 selection first admits two posts per account, then allocates remaining capacity round-robin. Accounts keep the same split across all their languages and configuration views. The source timestamps span 2024-11-14T00:00:00+00:00 through 2024-12-11T10:01:03+00:00.

Choose explicit source/language/platform and author sampling weights for training; row counts do not prescribe a training mixture. Keep each contrastive batch within one language. -1 is a missing content-group label. Exorde themes, sentiment, keywords and emotions remain auxiliary metadata rather than verified style labels.

Provenance, validation and attribution

This public copy preserves the complete validated source packages byte for byte. The root validation report verifies their checksums and local Hugging Face loading for all 79 configurations. The original source fidelity checks, complete occurrence accounting, mappings and preparation versions remain with each source. No source rows were reselected for publication.

• MAC card, source citation and attribution
• MAC language counts and exclusion accounting
• Exorde card and source information
• Original Exorde publisher card
• Exorde language/platform/split counts and length statistics
• Compact Exorde sampler index

MAC uses full v1 from Zenodo record 15538126. Exorde is pinned to d35b013f1db749076473a30e6de9ca5ca7f88bba. Keep each row's original attribution URL and the source package's notices when redistributing or deriving datasets.

Licensing and limitations

The root uses source-dependent licensing, with complete notices in LICENSE.md. MAC retains the user-specified CC BY-SA 4.0 basis for Wikipedia text; archive-specific licensing uncertainty is documented. Exorde's publisher declares MIT, and the source includes third-party platform content whose underlying rights were not independently verified. No blanket MIT or CC BY-SA license is asserted for both sources together.

Identities are operational accounts rather than verified people. Cross-wiki and cross-platform person linking is unavailable, so the same person can appear under different accounts. Exorde language predictions were not independently audited; MAC language labels describe wikis rather than language detection on every fragment. No complete quotation, repost, bot, or human-versus-AI writing labels are claimed. Cross-source overlap checks against MAC, STEB and teammate datasets were skipped at Peter's request. Within-source account separation does not establish cross-source separation or zero-shot performance. See the source validation reports for details.