aieng-lab/genter-ajibawa-name-filled

GENTER Ajibawa Name-Filled

This dataset expands aieng-lab/genter-ajibawa by inserting concrete names into each template. It provides nested Hugging Face configs with 1, 2, 5, or 10 names per gender and template. For every template sentence, names are sampled independently from NAMEXACT (matching split; frequency-weighted), using K female and K male names in config nK. It is intended for experiments that need concrete text rather than [NAME]/[MASK] placeholders while still preserving a link back to the template dataset. Together with the gradiend Python package (pip install gradiend), this dataset can be used for GRADIEND training without manually filling names first.

Usage

pip install datasets

from datasets import load_dataset

REPO = "aieng-lab/genter-ajibawa-name-filled"

# Config names `nK`: for each template, K female and K male names are sampled
# from NAMEXACT (frequency-weighted; split-matched). Examples:
#   n1  — 1 name per gender per template (smallest; default Hub config)
#   n2  — 2 names per gender per template
#   n5  — 5 names per gender per template
#   n10 — 10 names per gender per template
# Available configs on this repo: `n1`, `n2`, `n5`, `n10`.
# Configs are nested subsets: every row in n1 also appears in n2, n2 in n5, etc.
#
# The second argument to load_dataset is the *config name*, not a split.
# Splits are always train / validation / test.

# All splits for the default config (n1):
ds = load_dataset(REPO)                 # n1 is default
ds = load_dataset(REPO, "n1")           # explicit; equivalent to above

print(ds)                               # DatasetDict(train, validation, test)
train = ds["train"]
validation = ds["validation"]
test = ds["test"]

# Single split:
train = load_dataset(REPO, "n1", split="train")
validation = load_dataset(REPO, "n1", split="validation")

# Larger config (more name-filled variants per template):
ds_n5 = load_dataset(REPO, "n5")

# pandas (all splits concatenated):
import pandas as pd

frames = [split_ds.to_pandas().assign(split=name) for name, split_ds in ds.items()]
df = pd.concat(frames, ignore_index=True)

GRADIEND training

pip install gradiend datasets

from gradiend import TextPredictionTrainer, TrainingArguments

trainer = TextPredictionTrainer(
    model="bert-base-cased",
    hf_dataset="aieng-lab/genter-ajibawa-name-filled",
    hf_subset="n1",
    target_classes=["M", "F"],
    masked_col="masked",
    label_col="pronoun",
    label_class_col="label",
    args=TrainingArguments(max_steps=100, train_batch_size=32),
)
trainer.train()

Dataset Details

Dataset Description

Each row is one name-filled version of a GENTER Ajibawa template. The template_id column references the corresponding id in aieng-lab/genter-ajibawa. For each template, config nK contains K female-name and K male-name variants.

Dataset Sources

• Template Data: aieng-lab/genter-ajibawa
• Name Data: aieng-lab/namexact
• Repository: https://github.com/aieng-lab/gradiend-bias
• Paper: https://arxiv.org/abs/2502.01406
• Source Licenses: Apache-2.0 for Ajibawa-derived template text, CC-BY-4.0 for NAMEXACT names

Dataset Structure

• id: stable name-filled example id
• template_id: id of the corresponding row in aieng-lab/genter-ajibawa
• text: name-filled text with a concrete name and matching pronoun
• masked: name-filled cloze text with [MASK] at the pronoun position(s); use as GRADIEND masked_col
• label: gender label of the inserted name (F or M)
• name: inserted name
• pronoun: pronoun token for the row (she for F, he for M); factual target at each [MASK] site
• original_gender: whether the inserted gender matches the original template gender
• original_label: original template gender label
• original_name: original name found during template extraction
• original_pronoun: original pronoun found during template extraction
• pronoun_count: number of pronoun placeholders in the template

Examples

id	template_id	masked	pronoun	label	name	original_gender
train-002011-F-00	train-002011	Meanwhile, Elena didn't receive the offer as [MASK] fell short in certain crucial areas needed for the role.	she	F	Elena	True
train-008901-F-00	train-008901	Rachel sighed. That was the part [MASK] dreaded the most.	she	F	Rachel	True
train-025205-M-00	train-025205	But here comes the twist. Just when things were going great, Larry revealed that due to work commitments, [MASK] might have to move away.	he	M	Larry	True
train-000027-F-00	train-000027	Devastated, Dianne feared that [MASK] would go back to struggling academically again.	she	F	Dianne	True
train-012141-M-00	train-012141	Shaun was deeply moved by their words. [MASK] realized that the Highland Games were not merely a showcase of strength and skill but also a celebration of heritage and identity.	he	M	Shaun	False

Config And Split Sizes

config	split	examples
n1	train	73860
n1	validation	2110
n1	test	8440
n2	train	147720
n2	validation	4220
n2	test	16880
n5	train	369300
n5	validation	10550
n5	test	42200
n10	train	738600
n10	validation	21100
n10	test	84400

Source Notes

Templates come from aieng-lab/genter-ajibawa, which is based on ajibawa-2023/General-Stories-Collection. Names come from aieng-lab/namexact; the train/validation/test split of NAMEXACT is respected when filling templates.

Bias, Risks, and Limitations

This dataset multiplies each template with a per-template sample of names (frequency-weighted from NAMEXACT) and is therefore denser in common names than a uniform name draw. Use original_gender to separate examples that preserve the original template gender association from counterfactual name/pronoun instantiations.

License

This derived dataset is released under CC-BY-4.0, matching the existing GENTER/NAMEXACT dataset family. The Ajibawa-derived template text comes from an Apache-2.0 source dataset, and inserted names come from aieng-lab/namexact, which is declared as CC-BY-4.0.

Citation

@inproceedings{drechsel2026gradiend,
  title     = {{GRADIEND}: Feature Learning within Neural Networks Exemplified through Biases},
  author    = {Drechsel, Jonathan and Herbold, Steffen},
  booktitle = {Proceedings of the International Conference on Learning Representations},
  year      = {2026},
  url       = {https://arxiv.org/abs/2502.01406}
}

Dataset Card Authors

jdrechsel