IndoMultiDomain

IndoMultiDomain-Core V1 is a provenance-aware and harmonized multi-domain corpus of natural Indonesian text assembled from openly available research datasets.

The first frozen release contains 69,075 records, 1,418,001 words, 8 domains, 12 subdomains, 3 genres, and 9 source datasets.

Purpose

IndoMultiDomain is designed as shared infrastructure for Indonesian NLP research where domain, genre, and source provenance matter. The corpus is task-neutral: it does not introduce new semantic labels. Original labels from source datasets, when retained, are preserved only as provenance metadata.

Core principles

• natural, non-synthetic text;
• explicit source provenance;
• harmonized metadata;
• source-aware sampling;
• minimal and reversible cleaning;
• removal or redaction of direct identifiers detected during construction;
• exact duplicate control;
• deterministic train/validation/test splits;
• no new semantic annotation.

Corpus summary

Property	Core V1
Records	69,075
Words	1,418,001
Characters	9,718,645
Domains	8
Subdomains	12
Genres	3
Sources	9

The largest source contributes 14,999 records (21.71%). The largest domain is health_healthcare, with 20,008 records (28.97%).

Genres

Core V1 currently contains:

• user_review
• textbook_passage
• legal_norm

The corpus is intentionally heterogeneous. Future releases may add additional genres only when provenance and redistribution conditions are sufficiently clear.

Quality control

The final audit confirmed:

• 69,075 unique imd_id values;
• 69,075 unique exact text hashes;
• no empty texts;
• complete required schema;
• complete license and provenance fields;
• valid deterministic split values;
• 1 email address and 5 phone-number candidates redacted;
• no exact duplicates introduced by redaction.

A separate soft-normalization audit identified 294 possible near/trivial-duplicate groups. These were retained conservatively rather than automatically deleted because punctuation- or case-insensitive similarity does not necessarily imply semantic duplication.

Dataset schema

Each record contains:

imd_id, text, domain, subdomain, genre, source_id, source_repository, source_dataset, source_version, source_platform, original_id, language, year, original_task, original_label, original_label_type, license, provenance_note, char_count, word_count, quality_flag, duplicate_group, and split.

See DATASET_SCHEMA.md for field definitions.

Splits

Train, validation, and test splits are deterministic and derived from the final normalized text hash. Exact duplicates therefore cannot cross splits.

Privacy

Direct identifier columns were excluded during harmonization. A final privacy scan redacted detected email addresses and phone-number candidates. URLs are not automatically removed because they can constitute legitimate text content or provenance context.

Licensing

IndoMultiDomain is an aggregation of records derived from multiple source datasets. No single blanket license is asserted over all underlying texts. Users must comply with the original source licenses and attribution requirements listed in source_manifest_v1.csv and LICENSE_NOTES.md.

Intended uses

Suitable uses include:

• corpus characterization;
• domain and genre analysis;
• representation learning;
• continued pretraining;
• domain adaptation;
• cross-domain robustness studies;
• retrieval and embedding evaluation;
• Indonesian NLP benchmarking.

Limitations

IndoMultiDomain is not a population-representative sample of Indonesian language use. Source availability and redistribution permissions influence the domain and genre distribution. User reviews remain a large component of V1. Domain labels are assigned from source provenance and source entity, not from newly annotated semantic judgments.

Citation

A formal article citation will be added after publication of the IndoMultiDomain resource paper. Until then, cite the dataset release and the original source datasets listed in the manifest.

Version

Core V1 — frozen September 2026