---
title: Finance-Llama-8B App
emoji: 📈
colorFrom: purple
colorTo: blue
sdk: docker
app_port: 7860
pinned: false
---

# Finance-Llama-8B Gradio App

Two-tab financial agent UI powered by `tarun7r/Finance-Llama-8B`.

- Price Prediction: pulls OHLCV candles from Finnhub and asks the model for a short-term forecast with brief rationale.
- Equity Research Report: fetches Alpha Vantage fundamentals (via RapidAPI) and generates a concise equity research note.

## Model

- Hugging Face model: `tarun7r/Finance-Llama-8B`
  - Fine-tuned `unsloth/Meta-Llama-3.1-8B` on `Josephgflowers/Finance-Instruct-500k`
- Loaded with `bitsandbytes` 4-bit quantization for efficient GPU memory usage.

## Tech Stack

- Python 3.12
- Gradio UI
- Transformers + Accelerate + BitsAndBytes (GPU)
- Data: Finnhub (market data), Alpha Vantage via RapidAPI (fundamentals)

## Prerequisites

- NVIDIA GPU with recent CUDA drivers (recommended)
- Python 3.12
- API keys:
  - `FINNHUB_API_KEY` from `https://finnhub.io`
  - `RAPIDAPI_KEY` with access to Alpha Vantage host `alpha-vantage.p.rapidapi.com`

## Setup

Install dependencies:

```
pip install -r requirements.txt
```

Set environment variables (PowerShell example):

```
$env:FINNHUB_API_KEY = "YOUR_FINNHUB_KEY"
$env:RAPIDAPI_KEY = "YOUR_RAPIDAPI_KEY"
```

Optional server overrides:

```
$env:GRADIO_SERVER_NAME = "0.0.0.0"
$env:GRADIO_SERVER_PORT = "7860"
```

## Run Locally

```
python app.py
```

Open `http://localhost:7860`.

## Hugging Face Spaces (Docker)

This folder includes a `Dockerfile` suitable for Spaces (Docker) runtime. Ensure Secrets are set in the Space:

- `FINNHUB_API_KEY`
- `RAPIDAPI_KEY`

The app will start on port 7860 by default.

## How It Works

- Price Prediction tab
  - Fetch last N candles from Finnhub.
  - Build a compact JSON context and prompt the LLM to provide a short-term view and bullet-point drivers.

- Equity Research Report tab
  - Call Alpha Vantage `OVERVIEW` via RapidAPI host `alpha-vantage.p.rapidapi.com`.
  - Prompt the LLM to synthesize a research note: business summary, performance, profitability, leverage, valuation, risks, and a Buy/Hold/Sell view.

## Troubleshooting

- CUDA/torch version mismatch: use compatible PyTorch wheels for your environment or the provided Dockerfile on Spaces.
- `bitsandbytes` load errors: ensure GPU is available; for CPUs, remove 4-bit loading and use standard FP16/FP32 (will be slower, higher memory).
- API errors or empty responses: verify `FINNHUB_API_KEY` and `RAPIDAPI_KEY`, and check provider rate limits.

## Notes

- Respect provider rate limits and terms of use.
- Outputs are model-generated and for research/education purposes only, not financial advice.