giskardai/StereoTales

Multilingual Story-Generation Bias Samples

A multilingual evaluation dataset for probing demographic biases in LLM story generation. Each sample instructs a model to write a ~200-word story about a character carrying a given demographic attribute value (age, gender, ethnicity, religion, disability status, immigration status, ...) placed into a specific life scenario, with the goal of surfacing socio-economic and demographic biases in the generated narratives.

Languages and organization

The dataset is organized as one config per language, so that users can load a specific language subset independently:

from datasets import load_dataset

ds_fr = load_dataset("<repo_id>", "fr", split="train")
ds_zh = load_dataset("<repo_id>", "zh", split="train")

All language configs share the same structure: identical attributes, attribute values, scenarios, and prompt templates — only the natural-language content is translated and human-reviewed. This enables direct cross-lingual comparison of model behavior.

An additional English-only config, additional_en, contains scenario-free prompts (character + attribute without a scenario) used as a control.

Repository contents

This repository includes the following contents:

Configs and row counts

Config	Language	Rows
additional_en	English (additional prompts, no scenario)	237
ar	Arabic	2,844
du	Dutch	2,844
en	English	2,844
es	Spanish	2,844
fr	French	2,844
hi	Hindi	2,844
it	Italian	2,844
pt	Portuguese	2,844
uk	Ukrainian	2,844
zh	Chinese	2,844

Total rows across configs: 28,677

Generated samples - stories

The dataset also includes *_stories configs with cleaned model outputs and attribute extraction results exported our runs.

<lang>/stories row counts depend on the run/model selection used for export.

Config	Language	Rows
en/stories	English	65,412
ar/stories	Arabic	65,412
du/stories	Dutch	65,412
es/stories	Spanish	65,412
fr/stories	French	65,412
hi/stories	Hindi	65,412
it/stories	Italian	65,412
pt/stories	Portuguese	65,412
uk/stories	Ukrainian	65,412
zh/stories	Chinese	65,412
additional_en/stories	English (additional prompts, no scenario)	81,765

Analysis and study artifacts

The repository also includes Parquet tables for human evaluation, model self-evaluation, and associations listings.

The human_study/ directory contains structured exports from the human annotation on the associations generated by the models:

• participants.parquet — One row per enrolled participant (or panelist), including fields needed to interpret ratings at the participant level (for example recruitment channel, locale, or study arm), without tying tables to raw platform identifiers beyond what the public release allows.
• survey_results.parquet — One row per association judgment: association id, language, model or condition labels when shown as a factor, ordinal or categorical ratings, optional free text, and timestamps or task metadata when collected.

The llm_evals/ directory contains the Consolidated automated evaluation over the associations generated by the models themselves:

• eval_results.parquet — One row per association judgment aligned with the same identifiers used in the *_stories.

The associations_by_language/ and associations_global/ directories contain the associations computed with the two aggregation strategies:

• global: all stories generated are taken into account in the statistical analysis.
• by_language: we repeat the full statistical pipeline independently on each language, only aggregating stories that are generated with the same language.

Results are split into two Parquet files per directory:

• category_associations.parquet: contains the associations detected at the attribute level (e.g. age related to parental_status). Quantifies whether the two variables co-vary across stories in the slice using a full contingency-table test, bias-corrected Cramér’s V, a coarse effect_category (negligible / small / medium / large, scaled by table size), and a significant flag after multiple-testing correction across compared attributes (Benjamini–Hochberg in the reference implementation). Rows are keyed by model and, for associations_by_language, language.
• value_associations.parquet: contains the associations detected at the attribute value level (e.g. age=senior related to parental_status=with child). Contains per-cell statistics such as adjusted p-value, observed fraction within the base row, and lift versus expected counts under independence (FDR control via Benjamini–Yekutieli on cell p-values in the reference pipeline)

Note: Full contingency tables may be omitted from the published Parquet to keep artifacts small; primary columns (association_id, model, language where relevant, attributes, p-values, cramer_v, effect_category, significant, and value-level statistics) are present. Tables can be reconstructed from story-level extractions if you need the raw counts.

Schema

Samples follow the Flare Sample schema. Each row is a self-contained evaluation unit with the following top-level fields:

• id (str) — UUID of the sample.
• module (str) — always "biases".
• task (str) — always "story_generation".
• language (str) — ISO language code ("en", "fr", ...).
• generations (list) — one or more generation specs, each with a messages list (OpenAI chat format), sampling params, and a metadata dict describing the attribute / scenario / template combination used to build the prompt.
• metadata (dict) — sample-level attribute / scenario / character / prompt-template metadata.
• evaluation (dict) — scorer spec (biases/attribute_extraction) consumed downstream by Flare.

Story export schema (*_stories configs)

Each row represents one generated output from one model for one sample. Fields:

• sample_id (str)
• output_id (str)
• generator_model (str)
• language (str)
• target_attribute (str)
• target_attribute_value (str)
• attribute_value_key (str)
• scenario (str | null)
• scenario_key (str | null)
• scenario_group (str | null)
• character (str)
• prompt_template (str)
• user_prompt (str)
• story (str)
• extraction_score (float | null)
• extracted_attributes_json (str, JSON-serialized dict)

To keep story exports lightweight and stable, internal provider payloads and cost metadata (e.g. usage, cost, raw_responses, raw_extractions) are excluded.

Usage

from datasets import load_dataset

ds = load_dataset("<repo_id>", "en", split="train")
print(ds[0]["generations"][0]["messages"][0]["content"])

To iterate over every language:

from datasets import get_dataset_config_names, load_dataset

for cfg in get_dataset_config_names("<repo_id>"):
    ds = load_dataset("<repo_id>", cfg, split="train")
    # ... evaluate your model on ds ...

Source

Samples were generated by combining manually translated multilingual seeds (attributes, prompt templates, and scenarios).