---
license: other
base_model: google/gemma-3-270m-it
tags:
- mlx
- finance
- gemma
- instruction-tuning
datasets:
- Josephgflowers/Finance-Instruct-500k
---

# Gemma-3-270M - Fine-tuned for Financial Instructions

This is a fine-tuned version of Google's `gemma-3-270m-it` model, adapted for financial instruction-following tasks.

## Model Description

This model was fine-tuned using the Apple MLX framework. The goal was to specialize the base model for financial reporting summary and decision-making assistance. It was trained on the `Josephgflowers/Finance-Instruct-500k` dataset.

## Intended Use

This model is intended for tasks related to the financial domain, such as:
*   Answering questions about financial concepts.
*   Summarizing financial reports.
*   Following instructions based on financial data.

## How to Use

You can use this model with the `transformers` library just like any other standard Hugging Face model.

```python
from transformers import AutoTokenizer, AutoModelForCausalLM

model_name = "tlgoa/tmr-ai-nano"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name)

# Note: Gemma 3 uses a specific chat template.
# For single-turn inference, you can format it like this:
prompt = "What is the difference between revenue and profit?"
formatted_prompt = f"### User:\n{prompt}\n\n### Assistant:"

inputs = tokenizer(formatted_prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=200)

response = tokenizer.decode(outputs[0], skip_special_tokens=True)
# Clean up the response to only show the assistant's part
assistant_response = response.split("### Assistant:")[1].strip()

print(assistant_response)
```

## Training Procedure

### Dataset
The model was fine-tuned on the `Josephgflowers/Finance-Instruct-500k` dataset. The data was preprocessed to fit the following format:
```
### User:
{user_prompt}

### Assistant:
{assistant_response}
```

### Fine-tuning
The model was fine-tuned directly (full parameter tuning) using an Adam optimizer. Due to challenges with LoRA implementation in the available MLX version, a full fine-tuning approach was chosen. The fine-tuned weights were originally saved in MLX's `.npz` format and subsequently converted back to Hugging Face `safetensors` format for distribution.

## Licenses

- **Base Model:** This model is based on Google's Gemma-3-270M, which is subject to the [Gemma Terms of Use](https://ai.google.dev/gemma/terms).
- **Dataset:** The training data from `Josephgflowers/Finance-Instruct-500k` is available under the Apache 2.0 License.