FolkMotif-270: a parallel cross-cultural entity set for cultural-bias evaluation

27 Thompson Motif-Index roles × 10 cultural traditions = 270 citation-anchored cells.

Each cell names the canonical entity that fills one structural folk-narrative role in one tradition — the thunder-god's weapon, the smith of the gods, the goddess of love — together with its native-script form, attested spelling variants, and a citation anchor to a primary text or an academic monograph.

Why parallel design matters

Most cultural-bias benchmarks compare performance across topics that differ in more than culture: different genres, different entity frequencies, different question formats. Here every cell instantiates the same structural role, so the only variable that separates the ten columns is cultural provenance. A model asked to distinguish cultures — or to name the correct entity — cannot succeed by topic matching, genre cues, or question form; only the culture-entity association can carry it.

That property is what makes the set usable as a controlled substrate for mechanistic work: because the role is held fixed, a probe trained to read culture from hidden states is reading culture, and a generation scored against the gold entity is scored against an unambiguous target.

Dataset at a glance

Cells	270 (27 motifs × 10 cultures, complete grid, no duplicates)
Cultures	Greek, Roman, Norse, Finnish, Ukrainian, Indian, Egyptian, Chinese, Japanese, Mesopotamian
Language families	6 (Indo-European, Uralic, Afro-Asiatic, Sino-Tibetan, Japonic, Akkadian-Sumerian)
Native-script forms	264 / 270 recorded
Citation anchors	270 / 270 rows carry a source anchor to a primary text or a monograph
Structural absences	4 cells flagged, excluded from accuracy scoring
Alternative spellings	17 cells carry additional attested forms
License	CC-BY-4.0

Motif coverage by index category: cosmology and the gods (A, 100 cells), magic (D, 40), marvels and nature spirits (F, 40), the dead (E, 30), tabu (C, 20), ogres and demons (G, 20), love and fertility (T, 10), religion and sacred trees (V, 10).

Loading

from datasets import load_dataset

ds = load_dataset("Aragoner/folkmotif", split="train")

# the ten parallel fillers of one role
thunder = ds.filter(lambda r: r["motif_id"] == "A1141.2")
print({r["culture"]: r["name"] for r in thunder})
# {'Greek': 'Zeus', 'Roman': 'Jupiter', 'Norse': 'Thor', 'Finnish': 'Ukko',
#  'Ukrainian': 'Perun', 'Indian': 'Vajra', 'Egyptian': 'Set',
#  'Chinese': 'Lei Gong (Duke of Thunder)', 'Japanese': 'Takemikazuchi',
#  'Mesopotamian': 'Adad'}

# exclude the cells that have no parallel in their tradition
# (structural_absence is loaded as a boolean)
scoreable = ds.filter(lambda r: not r["structural_absence"])   # 266 cells

Fields

field	description
motif_id	Thompson Motif-Index entry, e.g. A1141.2
motif_group	index category the motif belongs to
motif_description	English gloss of the structural role
culture	one of the ten traditions
name	canonical proper name, romanised
native	native-script form where one is recorded (— otherwise)
source	citation anchor: primary-text pointer (Kalevala runo, Pyramid Texts spell, Hypatian Codex annal, …) or monograph with page range
structural_absence	boolean; true for the four cells with no parallel filler
aliases	pipe-separated attested spellings and contested scholarly alternatives, canonical name first

Example row:

{
  "motif_id": "A1141.2",
  "motif_group": "A_MYTHOLOGICAL",
  "motif_description": "Thunder-god's weapon (thunderbolt, hammer, axe)",
  "culture": "Finnish",
  "name": "Ukko",
  "native": "Ukko",
  "source": "Pentikäinen 1999:227-229",
  "structural_absence": false,
  "aliases": "Ukko"
}

Curation

Source data. Entities are drawn from primary attestations where one exists — the Poetic and Prose Eddas, the Kalevala, the Rig Veda and Purāṇas, the Pyramid and Coffin Texts, the Kojiki and Nihon Shoki, Enuma Elish and the Epic of Gilgamesh, the Primary Chronicle and Hypatian Codex, Hesiod, Homer, Ovid, the Shan Hai Jing — and otherwise from a curated academic monograph for the tradition. Every row records which.

Annotation process. Each cell was assigned by matching the Thompson role definition against the tradition's canon, then anchored to a citation and given a verification status. Where the modern literature has revised an older attribution, the modern reading is followed and the revision recorded: Ukrainian T110 is Mokoš rather than Lada, following Łuczyński (2020), who shows Lada to be a folk-song refrain word reinterpreted as a theonym in nineteenth-century scholarship; the Khors / Dažbog distinction (sun-disk versus sun-as-deity) is preserved as two separate motif rows following Gieysztor (2006).

Cells we would like a specialist to read. Every row here is citation-anchored, but seven are flagged in our working ground truth as benefiting from a domain reading rather than a literature match: Finnish A1010 (Vedenpaisumus), Finnish E400 (Kalma), Finnish F470 (Pellon-Pekko), Finnish C310 (Lemminkäinen), Finnish V1 (World Tree), Ukrainian D1080 (Mech-kladenets), and Greek E400 (Eidolon). A further seven have a solid citation but an arguable entity-to-role fit. The release keeps the source anchor so any of these can be checked against the cited text directly.

Considerations for use

Structural absences are data, not errors. Japanese A1010 (flood), Finnish A1010 (flood), Mesopotamian A1415 (theft of fire) and Egyptian F460 (forest spirit) have no canonical parallel in the tradition. They carry structural_absence = true and should be excluded from accuracy scoring rather than counted as model failures. Recording absence as a first-class entry, instead of forcing a filler, keeps the grid honest about where traditions genuinely differ.

Contested boundaries remain contested. Canonical assignment is genuinely disputed at the edges — Odin and Thor are both Norse sky figures; Perun is shared between East and West Slavic traditions; the Finnish T110 forest-spirit boundary (Rauni versus Mielikki) is unsettled in current Finno-Ugric scholarship. Where an alternative is attested it is listed in aliases, so exact-match evaluation does not penalise a defensible alternative choice. Use aliases rather than name alone when scoring generations.

Two Mesopotamian cells record a descriptive phrase rather than a proper name (A710 "Utu's daily journey", V1 "Huluppu tree") and have no native form recorded.

Coverage is deliberately bounded. Sub-Saharan African, Native American, Polynesian, Australian and Arctic indigenous traditions are absent because the per-row citation discipline could not be sustained at that scale in one cycle — not because those traditions are less suitable. Extending the grid is the obvious next step, and the schema is designed to take new columns. Cross-tradition borrowing is not tracked: a Norse entity continuing a Proto-Indo-European figure shared with Vedic is recorded on the Norse side only.

No personal or sensitive data. All entities are mythological figures, objects, and places from published scholarly sources.

Files

file	contents
folkmotif_270.csv	the release table, 270 rows
folkmotif_270.jsonl	the same rows, one JSON object per line
build_release.py	derives both from the verified ground truth

Citation

If you use this dataset, please cite the accompanying paper:

@misc{chelombitko2026folkmotif,
  title  = {Cultural Awareness is Represented but Not Decoded:
            Tracing Mythological Knowledge across 18 Open-Source LLMs},
  author = {Chelombitko, Iaroslav and Chelombitko, Ekaterina and
            H\"am\"al\"ainen, Mika},
  year   = {2026},
  note   = {Preprint; dataset: \url{https://huggingface.co/datasets/Aragoner/folkmotif}}
}

and acknowledge the index the grid is built on: Stith Thompson, Motif-Index of Folk-Literature, Indiana University Press, 1955–1958.

Contributing

Corrections from specialists in any of the ten traditions are welcome, particularly on the seven NEEDS_EXPERT cells and the contested boundaries above. Open an issue or a pull request in the code repository with the proposed entity, the citation that supports it, and the reasoning.