Datasetpublic

Ephraimmm/translate_this

GTBank Customer Service Synthetic Dataset Overview This dataset contains 15,000 synthetic customer-service exchanges modeled on a Nigerian retail bank (GTBank), each pairing a customer user_query with an assistant_reply and an intent label (e.g. PIN change, account restriction, card activation, BVN issues). All content verified in this file is in English; no other languages (e.g. Nigerian Pidgin, Hausa, Igbo) were found in the data despite the repository name.… See the full description on the dataset page: https://huggingface.co/datasets/Ephraimmm/translate_this.

sourceHugging Faceunknownupdated 3mo agoView on Hugging Face

0likes17downloads

Dataset Card

GTBank Customer Service Synthetic Dataset

Overview

This dataset contains 15,000 synthetic customer-service exchanges modeled on a Nigerian retail bank (GTBank), each pairing a customer user_query with an assistant_reply and an intent label (e.g. PIN change, account restriction, card activation, BVN issues). All content verified in this file is in English; no other languages (e.g. Nigerian Pidgin, Hausa, Igbo) were found in the data despite the repository name.

Note on repository name: the repo is named translate_this, but on inspection the single data file it contains is not a translation corpus — there are no source/target language pairs. It is a monolingual (English) customer-support intent/response dataset. This card describes the data as it actually exists in the repository.

Dataset Structure

The dataset consists of a single JSON Lines file, gtbank_customer_service_synthetic_15k.jsonl, with 15,000 rows (verified by direct line count). Each row is a JSON object with the following fields:

Field	Type	Description
`id`	string	Unique record identifier (e.g. `gtb_cs_00001`). All 15,000 ids are unique.
`timestamp`	string	ISO-8601 datetime stamp on the record, ranging from `2022-01-01T01:53:00` to `2025-12-31T22:27:00`.
`year`	integer	Year extracted from the timestamp (2022–2025, roughly evenly distributed: ~3,660–3,810 rows per year).
`user_query`	string	A customer's banking support question (average length ~50 characters).
`assistant_reply`	string	The corresponding support response, sometimes containing markdown formatting (e.g. `bold`) (average length ~173 characters).
`intent`	string	One of 16 intent categories labeling the topic of the exchange.

The 16 observed intent categories are: account_opening, account_restriction, beneficiary_issues, bvn_issues, card_activation, card_block_unblock, chargeback_dispute, debit_alerts, failed_transfer_debit, nin_linking, password_reset, pin_change, statement_request, transfer_limits, usd_card_fx, ussd_help.

Language pair(s) observed: none — this is a monolingual English dataset, not a translation dataset.

Important duplication note: while there are 15,000 rows, there are only 1,816 unique `(user_query, assistant_reply)` pairs. Each underlying Q&A pair is repeated multiple times (roughly 8x on average) under different id/timestamp/year values. Users training on this data should be aware of this templated repetition rather than assuming 15,000 unique conversations.

Data Format

A single UTF-8 encoded JSON Lines (.jsonl) file, one JSON object per line, ~5.2 MB in size. No other files are present in the repository besides .gitattributes.

Intended Use

Given the verified content, this dataset is best suited for:

—Fine-tuning or evaluating conversational/customer-support response generation models for banking use cases.
—Intent classification training, using user_query (or user_query + assistant_reply) against the intent label.

It is not suitable as-is for machine translation model training, since no translation pairs exist in the data currently hosted in this repository.

Limitations

—Not a translation dataset: despite the repository name (translate_this), the hosted file contains no source/target language pairs. This appears to be a work-in-progress or mislabeled repository rather than a finished translation corpus.
—Synthetic data: both queries and replies appear to be synthetically generated (or templated) rather than sourced from real customer interactions, so it may not fully reflect real-world phrasing, noise, or edge cases.
—High duplication: only ~1,816 unique query/reply pairs underlie the 15,000 rows, so effective content diversity is much lower than the row count suggests.
—Single language: only English content was verified; no multilingual or code-switched (e.g. Pidgin) text was found.
—Unverified license and provenance: no license file or documentation accompanies the dataset, and GTBank does not appear to be an official data provider — treat content as illustrative/synthetic rather than authoritative banking guidance.

Author

Curated by Ephraimmm