---
base_model: togethercomputer/Meta-Llama-3.3-70B-Instruct-Reference
library_name: peft
---
# CashWise

### A Multilingual Financial Reasoning Model for South Africa 🇿🇦

**CashWise** is a finance-specialised language model fine-tuned from **Meta Llama 3.3 70B Instruct** using the [RayNene/CashWise](https://huggingface.co/datasets/RayNene/CashWise) dataset.

The model targets a practical gap in financial AI: models may possess broad financial knowledge but struggle to turn that knowledge into **clear, safe, context-aware guidance for users outside English-first financial environments**.

CashWise combines large-model reasoning with multilingual financial adaptation across **English, isiZulu, isiXhosa, Setswana, Xitsonga, and Afrikaans**, covering investing, savings, credit, debt, insurance, tax, banking, retirement, fraud, mortgages, business finance, and equity.

## Model

- **Base model:** Meta Llama 3.3 70B Instruct
- **Fine-tuning dataset:** [RayNene/CashWise](https://huggingface.co/datasets/RayNene/CashWise)
- **Domain:** Financial reasoning and financial Q&A
- **Languages:** 6 South African languages
- **Focus:** Financial reasoning, safety, multilingual communication, and readability

## Evaluation Protocols

CashWise was evaluated using a held-out evaluation set designed to measure whether domain adaptation improved financial performance without relying solely on training-set examples.

Evaluation was conducted across four dimensions:

### 1. Financial Reasoning

The model was tested on unseen financial questions covering:

- Investing
- Savings
- Credit and debt
- Insurance
- Tax
- Banking
- Retirement
- Business finance
- Equity and compensation

Responses were compared against expected answers using financial QA and reasoning evaluation criteria.

### 2. Financial Benchmark Evaluation

The evaluation protocol was aligned with established financial NLP benchmarks, including:

- **FinQA** — numerical and logical reasoning over financial information
- **FinanceBench** — open-book financial question answering
- **FinBen** — broader financial evaluation covering reasoning, QA, risk, and decision-making

The same evaluation methodology was applied to the base model and the fine-tuned CashWise model to measure the effect of domain adaptation.

### 3. Multilingual Evaluation

The model was evaluated across the six languages represented in the training data to assess:

- Semantic preservation
- Financial terminology
- Response relevance
- Natural language quality
- Cross-lingual consistency

Particular attention was given to whether financial concepts remained intact rather than being translated literally.

### 4. Readability & Safety

Generated responses were additionally evaluated for:

- Clarity
- Unnecessary financial jargon
- Appropriate uncertainty
- Risk disclosure
- Avoidance of guaranteed financial outcomes
- Distinction between general financial information and personalised advice

## Results

Fine-tuning produced a substantial improvement over the underlying Llama 3.3 70B Instruct model.

| Metric | Base Llama 3.3 70B | CashWise | Change |
|---|---:|---:|---:|
| Financial evaluation win rate | **<15%** | **87%** | **+72pp** |
| Financial reasoning | Baseline | Improved | ↑ |
| Financial QA | Baseline | Improved | ↑ |
| Multilingual financial response | Baseline | Improved | ↑ |
| Readability | Baseline | Improved | ↑ |

### Financial Win Rate

The headline result is the increase in financial evaluation win rate from **under 15% for the base model to 87% after CashWise fine-tuning**.

This represents an improvement of **more than 5× the original win rate** and demonstrates that targeted financial training can substantially change the behaviour of a large general-purpose model on specialised financial tasks.

### Readability

The underlying CashWise dataset also demonstrated a measurable improvement in response readability:

| Response type | Score | Grade |
|---|---:|---|
| Original / community responses | **5.0** | D |
| Adaptive responses | **7.3** | B |

The **2.3-point improvement** reflects the dataset's emphasis on converting financial explanations into clearer, more accessible responses.

## Key Finding

The results suggest that **domain-specific data can substantially improve the financial capabilities of a large general-purpose model without requiring training a foundation model from scratch**.

CashWise therefore provides both:

1. A specialised dataset for multilingual financial reasoning.
2. A 70B-parameter model demonstrating the effect of targeted domain adaptation.

> **From <15% to 87% financial evaluation win rate.**

## Intended Use

CashWise can be used for:

- Financial QA and reasoning
- Multilingual financial assistants
- Financial NLP research
- African-language model evaluation
- Domain-specific LLM fine-tuning
- Research into financial accessibility and AI safety

## Limitations

CashWise is a research model and **is not a substitute for a licensed financial, tax, or investment professional**.

Financial regulations, tax rules, products, and individual circumstances change. Model outputs should therefore be independently verified before being used for consequential financial decisions.

## Dataset

The model was trained using:

**[RayNene/CashWise](https://huggingface.co/datasets/RayNene/CashWise)**

CashWise is a multilingual financial Q&A dataset developed as part of the **Adaption Labs AutoScientist Challenge**.

## Acknowledgements

Developed as part of the **Adaption Labs AutoScientist Challenge**, exploring how targeted datasets and domain adaptation can improve frontier language models on specialised tasks.

## Citation

```bibtex
@misc{cashwise2026,
  title  = {CashWise: A Multilingual Financial Reasoning Model for South Africa},
  author = {Nene, Ray},
  year   = {2026},
  url    = {https://huggingface.co/RayNene/CashWise}
}