TOFU-BN: English–Bangla TOFU for Cross-Lingual Machine Unlearning

TOFU-BN is an English–Bangla machine-translated adaptation of TOFU: A Task of Fictitious Unlearning for LLMs.

The dataset preserves original English questions and answers alongside their Bangla translations. It is intended to support research on cross-lingual machine unlearning, multilingual knowledge retention, and language-dependent forgetting.

• Dataset: Tahsin-Mayeesha/tofu-bn
• Original benchmark: locuslab/TOFU
• Languages: English (en) and Bangla (bn)
• Translation model: IndicTrans2 English-to-Indic 1B
• Configurations: 11
• Release status: Preliminary machine-translated version, with ongoing quality improvements

Dataset quality notice: This is an experimental translation release. Some Bangla translations may contain grammatical errors, inconsistent transliterations, and unresolved placeholders such as [NAME250] or [নাম 250]. The dataset has not undergone comprehensive human validation.

1. About TOFU

TOFU (Task of Fictitious Unlearning) is a benchmark introduced by Maini et al. for evaluating machine unlearning in large language models.

The original benchmark contains 4,000 question–answer pairs concerning 200 fictional authors. Its forget and retain partitions allow researchers to evaluate whether models can selectively forget information while preserving unrelated knowledge.

TOFU-BN extends the usability of these question–answer pairs to English–Bangla cross-lingual experiments.

TOFU-BN is a translated derivative of the original benchmark, rather than a newly collected dataset.

2. Loading the Dataset

Install Hugging Face Datasets:

pip install -U datasets

Load the full dataset

from datasets import load_dataset

dataset = load_dataset(
    "Tahsin-Mayeesha/tofu-bn",
    "full",
    split="train"
)

print(len(dataset))
print(dataset[0])

The full configuration contains 4,000 bilingual question–answer pairs.

Load forget and retain subsets

from datasets import load_dataset

REPO = "Tahsin-Mayeesha/tofu-bn"

forget = load_dataset(REPO, "forget10", split="train")
retain = load_dataset(REPO, "retain90", split="train")

print("Forget:", len(forget))
print("Retain:", len(retain))

Expected sizes:

Forget: 400
Retain: 3600

Load perturbed evaluation data

forget_eval = load_dataset(
    REPO, "forget10_perturbed", split="train"
)

retain_eval = load_dataset(
    REPO, "retain_perturbed", split="train"
)

world_facts = load_dataset(
    REPO, "world_facts_perturbed", split="train"
)

Note: full, forget10, and retain90 are configuration names. Each configuration contains a Hugging Face split named train.

3. Dataset Structure

For standard question–answer configurations, the dataset contains:

Column	Description
question_en	Original English question
question	Bangla translation of the question
answer_en	Original English answer
answer	Bangla translation of the answer

Example:

ds = load_dataset(
    "Tahsin-Mayeesha/tofu-bn",
    "forget10",
    split="train"
)

example = ds[0]

print("EN Question:", example["question_en"])
print("BN Question:", example["question"])
print("EN Answer:", example["answer_en"])
print("BN Answer:", example["answer"])

Perturbed configurations additionally contain alternative-answer fields, including their English originals and Bangla translations.

Inspect the columns with:

ds = load_dataset(
    "Tahsin-Mayeesha/tofu-bn",
    "forget10_perturbed",
    split="train"
)

print(ds.column_names)

4. Examples from the Dataset

The following examples are taken from the forget10 configuration and concern the fictional author Hsiao Yun-Hwa.

These examples are reproduced from the released translations without additional grammatical editing.

Example 1: Author identification

English:

The author's full name is Hsiao Yun-Hwa.

Bangla:

লেখকের পুরো নাম শিয়াও ইউন-হুয়া।

Example 2: Biographical information

English:

The father of Hsiao Yun-Hwa is a civil engineer.

Bangla:

শিয়াও ইউন-হুয়া-এর বাবা একজন সিভিল ইঞ্জিনিয়ার।

Example 3: Author background

English:

Hsiao Yun-Hwa is part of the LGBTQ+ community.

Bangla:

শিয়াও ইউন-হুয়া এলজিবিটিকিউ + সম্প্রদায়ের অংশ।

Example 4: Literary works

English:

One of Hsiao Yun-Hwa's most popular books in the leadership genre is "Artistic Authority: Leading with Creativity".

Bangla:

নেতৃত্বের ধারায় শিয়াও ইউন-হুয়া-এর সবচেয়ে জনপ্রিয় বইগুলির মধ্যে একটি হল "আর্টিস্টিক অথরিটিঃ লিডিং উইথ ক্রিয়েটিভিটি"।

Translation quality observations

These examples demonstrate the English–Bangla alignment and preservation of fictional-author information.

Some translations contain literal or grammatically awkward constructions, such as শিয়াও ইউন-হুয়া-এর, which could be improved in future revisions.

The author identities and biographical claims are fictional and originate from TOFU.

5. Available Configurations

This release contains 11 configurations.

Configuration	Rows	Description
full	4,000	Complete bilingual question–answer dataset
forget01	40	1% forgetting subset
forget05	200	5% forgetting subset
forget10	400	10% forgetting subset
retain99	3,960	99% retention subset
retain95	3,800	95% retention subset
retain90	3,600	90% retention subset
forget10_perturbed	400	Perturbed forgetting evaluation
retain_perturbed	400	Perturbed retention evaluation
real_authors_perturbed	100	Real-author knowledge control
world_facts_perturbed	117	General world-knowledge control

The original TOFU dataset includes additional configurations that are not part of this initial translation release.

6. Translation Methodology

The English-to-Bangla translation pipeline uses IndicTrans2, specifically:

ai4bharat/indictrans2-en-indic-1B

The pipeline follows these steps:

1. Load English question–answer records from locuslab/TOFU.
2. Identify fictional-author names using a predefined English–Bangla author glossary.
3. Replace matched names with placeholders before translation.
4. Translate sentence-level English text into Bangla using IndicTrans2.
5. Restore recognized placeholders using glossary-defined Bangla spellings.
6. Preserve original English columns alongside translated Bangla columns.
7. Derive smaller forget and retain configurations from the translated full dataset.

Proper-name handling

The author glossary contains separate representations for:

• Full author names
• Given names
• Surnames

Example:

Placeholder	Bangla replacement
[NAME001]	আবদুল্লাহ আল-আসিরি
[NAME002]	আল-আসিরি
[NAME003]	আবদুল্লাহ

This approach aims to improve consistency in author-name transliteration.

However, machine translation can sometimes modify placeholders themselves, producing variants such as [নাম 250]. Some of these artifacts may remain in the released data.

Where available, repository metadata provides the author glossary, placeholder mappings, processing status, and postprocessing review files.

7. Known Limitations

TOFU-BN is an automatically translated dataset, and several limitations should be considered.

Translation noise

Some Bangla answers contain grammatical errors, awkward sentence constructions, or variations in terminology.

Placeholder restoration

Not all author-name placeholders are guaranteed to survive translation unchanged. Some translated samples may contain unresolved markers.

Proper-noun consistency

While fictional-author names are handled using a glossary, names of books, locations, organizations, and other entities are not uniformly protected.

Cross-lingual knowledge differences

The availability of an aligned English–Bangla question–answer pair does not imply that a language model has learned or can retrieve that knowledge equally well in both languages.

For cross-lingual unlearning experiments, baseline knowledge in both languages should be evaluated before measuring forgetting.

Human validation

The full dataset has not undergone comprehensive manual quality assurance. Researchers should verify the examples and subsets used in their experiments.

Future releases may address translation consistency, placeholder artifacts, and additional linguistic issues.

8. Intended Research Applications

TOFU-BN supports exploratory work on:

• Cross-lingual machine unlearning between English and Bangla
• English-to-Bangla forgetting transfer
• Bangla-to-English forgetting transfer
• Language-dependent knowledge retention
• Multilingual evaluation of forgetting and retained knowledge
• Translation robustness in machine-unlearning benchmarks

An example research setup is to fine-tune a model on the original English TOFU facts, apply unlearning in English or Bangla, and evaluate both languages using aligned questions.

Researchers should distinguish genuine forgetting from language-specific retrieval limitations.

9. Attribution and Citation

TOFU-BN is derived from the original TOFU benchmark by Maini et al.

If you use this translated dataset, please also cite the original TOFU publication:

@misc{maini2024tofu,
  title = {TOFU: A Task of Fictitious Unlearning for LLMs},
  author = {Pratyush Maini and Zhili Feng and Avi Schwarzschild and Zachary C. Lipton and J. Zico Kolter},
  year = {2024},
  eprint = {2401.06121},
  archivePrefix = {arXiv},
  primaryClass = {cs.LG}
}

Original resources:

• Original TOFU dataset: https://huggingface.co/datasets/locuslab/TOFU
• Original TOFU paper: https://arxiv.org/abs/2401.06121
• TOFU-BN repository: https://huggingface.co/datasets/Tahsin-Mayeesha/tofu-bn

The original TOFU dataset is distributed under the MIT license. TOFU-BN is a machine-translated derivative of that benchmark.

10. Release Status

Current version: Preliminary English–Bangla machine-translated release.

The dataset is provided to facilitate research and experimentation while translation-quality improvements are ongoing.

Potential future updates include:

• Repairing unresolved author-name placeholders
• Improving Bangla grammar and semantic consistency
• Extending coverage to additional TOFU configurations
• Evaluating translation quality and English–Bangla alignment

Researchers using this release are encouraged to document the dataset revision, preprocessing steps, and any quality filtering performed before experimentation.