---
title: DeepSeek-R1-Distill-Qwen-14B
emoji: 🤖
colorFrom: indigo
colorTo: purple
sdk: gradio
sdk_version: 5.47.0
#python_version: 3.12
app_file: app.py
pinned: false
---

# Finance Space

Gradio app with two tabs:
- Price Prediction
- Equity Research Report

Model: `tarun7r/Finance-Llama-8B` (fine-tuned Llama-3.1-8B for finance). Runs on GPU when available.

Data sources:
- Finnhub (candles, profile, news) — set `FINNHUB_API_KEY`
- Alpha Vantage via RapidAPI — set `RAPIDAPI_KEY`, host `alpha-vantage.p.rapidapi.com`

## Setup

```bash
pip install -r requirements.txt
# set env vars
$env:FINNHUB_API_KEY="..."      # PowerShell
$env:RAPIDAPI_KEY="..."
python app.py
```

In Hugging Face Spaces (Python 3.12 GPU), just add the secrets and run `app.py`.

## Notes
- If Finnhub key is absent, the app falls back to Alpha Vantage daily adjusted via RapidAPI.
- Prompts are in `prompts.py`. Adjust temperature/top_p in `app.py` as needed.

## ZeroGPU (no local PyTorch) setup
- Default path uses Hugging Face Inference API to run `tarun7r/Finance-Llama-8B` remotely.
- Set secrets on your Space:
  - `HF_TOKEN` (recommended; required for higher rate limits)
  - `FINNHUB_API_KEY`
  - `RAPIDAPI_KEY`
- Environment toggle:
  - `USE_HF_API=1` (default) → remote inference (works on ZeroGPU)
  - `USE_HF_API=0` → try local inference (requires GPU/torch)
- If you later switch to GPU runtime, uncomment `torch==2.3.1` in `requirements.txt` and set `USE_HF_API=0` if you prefer local inference.