FinFrontier Website

Static leaderboard + interactive HF Space demo for the FinFrontier benchmark. Total infra cost: $0 (see website_plan.md for paid upgrade tiers).

finfrontier-site/
├── index.html            # the whole site (leaderboard, frontier chart, heatmap, demo embed)
└── hf-space/             # Gradio ZeroGPU demo (deploy to Hugging Face Spaces)
    ├── app.py
    ├── requirements.txt
    └── README.md

Setup steps

1. Website (GitHub Pages) — 10 minutes

1. Create a repo (e.g. LabRAI/finfrontier-site), push this folder's contents to main.
2. Repo → Settings → Pages → Source: main / root → Save. Site goes live at https://labrai.github.io/finfrontier-site/.
3. In index.html, set the paper link (id="paper-link") and confirm the GitHub URL.

2. Demo (Hugging Face Spaces, ZeroGPU) — 30–45 minutes

1. On huggingface.co: accept the Meta Llama 3.1 license on the meta-llama/Llama-3.1-8B-Instruct model page (one time).
2. Upload each LoRA adapter from CARC to a private model repo (free; keeps your trained weights unexposed until you choose to release them):

   pip install -U huggingface_hub && huggingface-cli login
   huggingface-cli repo create finfrontier-fingpt --type model --private
   huggingface-cli upload LabRAI/finfrontier-fingpt ~/finfrontier_ckpts/fingpt .

   (repeat per method; adapters are ~55 MB each; flip to public post-acceptance)
3. Create a Space: New Space → SDK Gradio → Hardware ZeroGPU → upload hf-space/app.py, requirements.txt, README.md.
4. Space → Settings → Secrets → add HF_TOKEN. Use a fine-grained, read-only token (scope: "read access to contents of all repos you can access") from the account that accepted the Llama license. Secrets are server-side env vars: never visible to visitors and never committed to the repo. Do not hardcode tokens in app.py.
5. Edit ADAPTER_REPOS in app.py to your repo names. The Space boots even if some adapters are missing (those methods just don't appear).
6. Put the Space URL into index.html (id="demo-frame"), e.g. https://labrai-finfrontier-demo.hf.space.

3. Costs

$0 for all of the above. Optional: HF PRO ($9/mo) for larger ZeroGPU quota during the launch/promotion window; dedicated L4 endpoint (~$0.80/hr) only if the demo outgrows ZeroGPU. Cost table for Jackie is in website_plan.md.