nirmalendu01/counterfactual_culture

Counterfactual Culture

Multilingual minimal-change counterfactual etiquette vignettes for five cultures, with conforming / violating pairs for factorization and representation studies.

Cultures	english (US norms), japan, china, india, russia
Languages	en, ja, zh, hi, ru (full cross: every culture × every language)
Samples	152,500 (76,250 pairs)
Seed samples	610 English seed vignettes (before variation expansion)
Norms	305 etiquette norms

Configs

Config	Rows	Description
samples (default)	152,500	L2 contexts × L1 lexical variations, all languages
seed_samples	610	English seed vignettes (1 pair per norm)
norms	305	Norm metadata / provenance

from datasets import load_dataset

ds = load_dataset("nirmalendu01/counterfactual_culture", "samples")
norms = load_dataset("nirmalendu01/counterfactual_culture", "norms")
seed = load_dataset("nirmalendu01/counterfactual_culture", "seed_samples")

How this was created

1. Norm collection

Etiquette norms were gathered for each culture from:

1. NormAd (akhilayerukola/NormAd) — Background column dos/don’ts (bullet lines) for united_states_of_america → english, plus japan, china, india, russia.
2. CultureBank (SALT-NLP/CultureBank) — American / Japanese / Chinese / Indian / Russian descriptors under Social Norms and Etiquette, filtered by agreement (≥ 0.8), embedding-deduplicated against that culture’s NormAd bullets, then GPT-curated for actionable interpersonal etiquette.

Each norm has source ∈ {normad, culturebank} in the norms config.

2. English seed vignettes

For each norm, gpt-4o-mini wrote a natural English scene with a minimal-change conforming / violating pair (light negation or small lexical flip). Country/nationality names are scrubbed from the surface text; culture is metadata only. Prompts end with an open stem (e.g. Socially, this is) for continuation-style eval.

3. Variations (Level 2 × Level 1)

Each seed norm was expanded with gpt-4o-mini into:

• Level 2 — 10 contextual instantiations of the same underlying norm
• Level 1 — 5 lexical paraphrases per context

Each cell is again a minimal conforming/violating pair → 50 pairs / 100 texts per norm.

4. Multilingual translation

English variations for every culture were translated with gpt-4o-mini via the OpenAI Batch API into Japanese (ja), Chinese (zh), Hindi (hi), and Russian (ru), preserving ids and pair alignment. This yields a full culture × language grid (not language-tied-only-to-its-culture).

Sample fields (samples)

Field	Meaning
culture	Norm culture slug
language	Surface language of text / prompt
role	conforming or violating
gold_label	acceptable / unacceptable
mate_id	Id of the opposite sample in the minimal pair
norm_id / pair_id	Join keys to norms
context_id / lexical_id	Variation indices (0–9 / 0–4)
context_note	Short situation label
text	Vignette body
prompt	text + completion stem
source	normad or culturebank
split	train / test

Credits / source datasets

Please cite / credit the upstream resources this corpus builds on:

• NormAd — Rao et al.; dataset akhilayerukola/NormAd
• CultureBank — Shi et al. (Findings of EMNLP 2024); dataset SALT-NLP/CultureBank

Synthetic vignettes and translations were generated with OpenAI gpt-4o-mini. This is a research pilot, not a validated cultural benchmark. Norms are simplified etiquette descriptors and may not represent all speakers of a culture.

Splits

Pair-level 80% / 10% / 10% train / val / test, stratified by culture. Conforming and violating mates always share a split. The same content pair uses the same split across all languages (parallel multilingual holdout).