| # CashWise |
|
|
| ### Multilingual financial reasoning for South Africa |
|
|
| South Africans have increasing access to financial services, but access does not necessarily mean access to understandable financial guidance. The **2025 FinScope Consumer Survey**, based on 5,600 adults across all nine provinces, found that financial inclusion remains high while financial pressure is increasing, with more people relying on credit to cover everyday expenses and prioritising immediate needs over saving and investment. Meanwhile, the **World Bank's 2024 Global Findex** data shows that **8.82 million South African adults still lack a formal financial account**. |
|
|
| CashWise is a **multilingual financial Q&A dataset** designed to evaluate and fine-tune models that can reason about real financial decisions and explain them clearly across South African languages. |
|
|
| ## What it covers |
|
|
| **6 languages:** English, isiZulu, isiXhosa, Setswana, Xitsonga, and Afrikaans. |
|
|
| **Financial topics:** investing, savings, credit, debt, insurance, tax, banking, retirement, fraud, business finance, mortgages, and equity. |
|
|
| ## Dataset Structure |
|
|
| Each record contains: |
|
|
| | Field | Description | |
| |---|---| |
| | `question` | Financial question | |
| | `community_answer` | Informal/source response | |
| | `model_answer` | Adapted, plain-language financial response | |
|
|
|
|
| ## Evaluation |
|
|
| CashWise is evaluated across financial reasoning, safety, readability, and multilingual performance using established financial NLP evaluation frameworks, including: |
|
|
| - **FinQA** — financial numerical reasoning and question answering |
| - **FinanceBench** — open-book financial question answering |
| - **FinBen** — comprehensive financial LLM evaluation across 36 datasets and 24 financial tasks |
| - **Readability evaluation** — measuring clarity and accessibility of financial responses |
| - **Multilingual evaluation** — assessing semantic preservation and language fidelity |
|
|
| ### Readability Improvement |
|
|
| The adaptive responses show a measurable improvement in readability compared with the original community responses. |
|
|
| | Version | Score | Grade | |
| |---|---:|---| |
| | Original data responses | **5.0** | **D** | |
| | Adaptive responses (after adaption) | **7.3** | **B** | |
|
|
| This represents a **2.3-point improvement in readability**, reflecting the dataset's focus on converting informal or complex financial explanations into clearer, more accessible responses. |
|
|
| The evaluations measure more than financial knowledge alone: whether a model can **reason about financial problems, communicate the result clearly, and maintain appropriate safety and uncertainty across languages**. |
|
|
| ## Intended Use |
|
|
| CashWise can be used for: |
|
|
| - Financial LLM fine-tuning |
| - Financial QA evaluation |
| - Multilingual NLP research |
| - African-language model evaluation |
| - Financial accessibility research |
| - Financial AI safety evaluation |
|
|
| > **CashWise is a research dataset and should not be treated as a source of personalised financial, tax, or investment advice.** |
|
|
| ## Adaption Labs AutoScientist Challenge |
|
|
| CashWise was developed as part of the **Adaption Labs AutoScientist Challenge**, an initiative focused on using automated scientific workflows to explore and develop new datasets, models, and research directions. |
|
|
| ## Citation |
|
|
| ```bibtex |
| @dataset{cashwise2026, |
| title = {CashWise: A Multilingual South African Financial Guidance Dataset}, |
| author = {Nene, Ray}, |
| year = {2026}, |
| publisher = {Hugging Face}, |
| url = {https://huggingface.co/datasets/RayNene/CashWiseDataset} |
| } |
