Donoe/multilingual-nigerian-banking77

Multilingual Nigerian BANKING77

Overview

This dataset is a multilingual version of the BANKING77 intent-classification dataset containing English, Yoruba, Hausa, and Igbo versions of the same banking customer-service queries.

The dataset contains:

• 9,993 English samples
• 9,993 Yoruba samples
• 9,993 Hausa samples
• 9,993 Igbo samples

Total:

39,972 rows

There are 77 banking-related intent classes.

Dataset Structure

The dataset uses a long-format structure:

Column	Description
sample_id	Identifier for the original BANKING77 sample
text	Customer query
language	Language of the query
label	Numeric intent label
label_text	Human-readable intent name

Languages

The dataset contains:

• English
• Yoruba
• Hausa
• Igbo

Each language contains the same 9,993 sample IDs so that the multilingual versions remain aligned.

Intended Use

The dataset is intended for research and experimentation with multilingual NLP and banking intent classification.

Potential applications include:

• Multilingual intent classification
• Cross-lingual transfer learning
• Nigerian-language NLP
• XLM-R fine-tuning
• Multilingual banking conversational AI
• Evaluation of multilingual models

Translation

The Yoruba, Hausa, and Igbo versions were generated using NCAIR1/N-ATLaS.

The translations are machine-generated and should not be assumed to be perfectly accurate. Additional human evaluation and linguistic validation are recommended before production use.

Important Data Split Consideration

The dataset is intentionally provided without a final train/validation/test split.

When creating splits for machine learning, splitting should be performed by sample_id, rather than individual rows.

This prevents English, Yoruba, Hausa, and Igbo versions of the same original query from appearing in different splits and causing data leakage.

Source Dataset

The dataset is derived from BANKING77.

Original dataset:

BANKING77 by PolyAI.

The original BANKING77 dataset is distributed under the CC BY 4.0 license.

Users should review and comply with the original dataset's attribution requirements.

Translation Model

NCAIR1/N-ATLaS was used to generate the Nigerian-language translations.

Users should review the N-ATLaS license and attribution requirements before using these translations, particularly for commercial or large-scale deployment.

Limitations

The Nigerian-language translations are machine-generated. Translation quality may vary across intents, terminology, and language.

Human linguistic evaluation is recommended before deploying the dataset in safety-critical or customer-facing systems.

Citation

If you use this dataset, please acknowledge the original BANKING77 dataset and the N-ATLaS model used for translation.

mteb/banking77 and NCAIR1/N-ATLaS