Ephraimmm/pidgin_bank_dataset

Nigerian Pidgin Bank Customer Support Dataset

Overview

This dataset contains 150,000 single-turn customer support conversations for a Nigerian retail bank, written primarily in Nigerian Pidgin English (pcm) with code-switched English. Each example simulates a customer inquiry about common banking issues — such as failed transfers, POS/ATM dispense errors, USSD (*737#) banking, and internet banking login/password resets — paired with an assistant response grounded in a bank help-centre style script. Personally identifiable information such as the bank's name has been redacted/blanked from the text (the file is named 150k_pidgin_pii_safe.jsonl).

Dataset Structure

• Total rows: 150,000 (verified by counting lines in the JSONL file)
• File: 150k_pidgin_pii_safe.jsonl
• Field: each JSON line has a single top-level key, messages, containing a list of exactly 3 chat-style turns:
  • role: "system" — a fixed instruction defining the assistant persona ("You are  Customer Support. Provide clear, safe, and accurate guidance..."). Only 1 unique system prompt appears across the whole dataset.
  • role: "user" — the customer's message, typically in Nigerian Pidgin (e.g. "POS debit fail but I was charge. Wetin I go do? Ref TRX5581265757"). There are 56,375 unique user messages (about 38% of rows), built from a set of query patterns parameterized with randomized transaction references (TRX##########), amounts (in ₦), and channel names (POS, ATM, USSD, Internet Banking).
  • role: "assistant" — the support response. Only 4 unique assistant response templates occur in the entire dataset, covering: (1) failed transfer/reversal guidance, (2) POS/ATM dispense-error timelines, (3) *737# USSD transfer codes, and (4) internet banking / app password reset steps. Some templates contain unfilled placeholders ({ref}, {amt}).

Every row has the identical structure [system, user, assistant] — there are no multi-turn conversations in this release.

Data Format

Single file in JSON Lines (.jsonl) format, one JSON object per line, in the OpenAI-style chat messages schema (role + content). File size is approximately 122 MB.

Intended Use

This dataset is intended for fine-tuning or evaluating conversational AI / customer-support chatbots that need to understand and respond appropriately to Nigerian Pidgin English banking queries — a low-resource language setting that is underrepresented in most NLP training data. It can support tasks such as intent classification, response generation, and instruction-tuning for Pidgin-aware assistants in the fintech domain.

Limitations

• Highly templated responses: with only 4 distinct assistant responses across 150,000 rows, the dataset offers very limited diversity on the answer side. Models trained on it as-is are likely to reproduce these 4 templates verbatim rather than generate genuinely varied, context-sensitive answers.
• Synthetic construction: the repetitive structure (fixed system prompt, parameterized user query templates, a handful of answer templates) strongly suggests the data was synthetically generated/templated rather than collected from real customer-agent transcripts. It should not be treated as naturalistic conversational data.
• Redacted bank identity: the bank's name and app name have been blanked out (leaving artifacts like double spaces and "the app"), which may affect fluency and downstream response quality if used verbatim.
• Some unfilled placeholders: a subset of assistant responses retain literal {ref} / {amt} placeholders instead of realized values.
• Single-turn only: there is no multi-turn dialogue or follow-up context in this release.
• No documented license or collection methodology was found in the repository at the time of this review.

Author

Curated by Ephraimmm