NaijaMultilingualBank

NaijaMultilingualBank is a multilingual banking intent-classification model fine-tuned from FacebookAI/xlm-roberta-base for banking queries in English, Yoruba, Hausa, and Igbo.

It classifies queries into 77 banking intents and is intended to serve as the intent-detection component of a broader multilingual Nigerian banking assistant.

Model Details

• Base model: FacebookAI/xlm-roberta-base
• Languages: English, Yoruba, Hausa, Igbo
• Number of intents: 77
• Maximum sequence length: 128 tokens
• Training epochs: 3
• Learning rate: 2e-5
• Training batch size: 8
• Gradient accumulation: 2
• Random seed: 42

Dataset

The model was trained on a multilingual Nigerian banking dataset derived from BANKING77.

• Total samples: 39,972
• Unique source queries: 9,993
• Languages: 4
• Intents: 77

Dataset repository: Donoe/multilingual-nigerian-banking77

Final Test-Set Evaluation

The final model was evaluated on 4,000 held-out test samples.

Metric	Score
Accuracy	65.87%
Macro Precision	66.60%
Macro Recall	65.49%
Macro F1	64.51%
Weighted Precision	67.17%
Weighted Recall	65.87%
Weighted F1	65.28%

Macro F1 is emphasized because this is a 77-class classification problem and macro averaging gives each intent equal weight.

Additional aggregate evaluation artifacts are provided in evaluation/.

Intended Use

This model is intended for:

• multilingual banking intent detection;
• NLP research and experimentation;
• intent routing in conversational systems;
• downstream multilingual banking assistants.

It can serve as the intent-classification layer of NaijaBankBot.

Limitations

This model is a research/portfolio baseline and should not be treated as a fully production-ready banking decision system.

The test results show useful multilingual classification capability, but performance is not perfect across all 77 intents. Similar banking intents can be confused, so downstream systems should use appropriate confidence handling, testing, monitoring, and business validation.

The model performs intent classification. It does not authorize transactions, access bank accounts, or independently execute banking operations.

Evaluation Methodology

A deterministic split using random seed 42 was used:

• Training: 31,976 samples
• Validation: 3,996 samples
• Test: 4,000 samples

The saved final model was reloaded and verified before evaluation.

Attribution

This work builds on the BANKING77 task and the FacebookAI/xlm-roberta-base architecture. Users should review and comply with the applicable licenses and attribution requirements for the underlying model and dataset before redistribution or commercial use.

Project

The broader project is being developed as NaijaBankBot, a multilingual Nigerian banking assistant integrating intent classification, multilingual translation, and an LLM response layer.

Release prepared on 2026-09-10.