Secured / SETUP.md
gowtham0992's picture
Promote MiniCPM5 1B LoRA v4
928990e
|
Raw
History Blame Contribute Delete
2.25 kB

A newer version of the Gradio SDK is available: 6.19.0

Upgrade

Setup

Local Project

Project path:

cd jawbreaker

Current local repo:

git status
git log --oneline -1

GitHub Repo

The OpenAI Codex Track expects a public GitHub repo linked from the Space README, with Codex-attributed commits.

Because GitHub CLI is not installed locally, create the repo in the GitHub web UI:

  • Repo name: jawbreaker
  • Visibility: public
  • Do not initialize with README, license, or gitignore

Then connect this local repo:

git remote add origin git@github.com:YOUR_GITHUB_USERNAME/jawbreaker.git
git push -u origin main

If SSH is not configured, use the HTTPS remote GitHub shows after repo creation:

git remote add origin https://github.com/YOUR_GITHUB_USERNAME/jawbreaker.git
git push -u origin main

Hugging Face Space

The Space can stay private during development, but must be public before the June 15, 2026 deadline.

Log in locally:

hf auth login
hf auth whoami

Create the private Space under the hackathon org:

hf repos create build-small-hackathon/jawbreaker --type space --space-sdk gradio --private

Add the Space as a second git remote:

git remote add space https://huggingface.co/spaces/build-small-hackathon/jawbreaker
git push space main

ZeroGPU Mode

For responsive model inference without paid hardware:

  • set Space hardware to ZeroGPU
  • set Space variable JAWBREAKER_BACKEND=zerogpu
  • keep JAWBREAKER_TRANSFORMERS_MODEL_ID=openbmb/MiniCPM5-1B for the OpenBMB/Tiny Titan-targeted build
  • keep JAWBREAKER_ADAPTER_ID=build-small-hackathon/jawbreaker-minicpm5-1b-lora-v4; the app also defaults to this adapter when the variable is omitted
  • keep JAWBREAKER_TRUST_REMOTE_CODE=true; MiniCPM requires custom model code through Transformers
  • keep JAWBREAKER_ATTENTION_IMPLEMENTATION=eager; this remains the conservative MiniCPM Transformers path
  • fallback model if latency is unacceptable: Qwen/Qwen3-0.6B

CPU fallback remains available with JAWBREAKER_BACKEND=llama-cpp.

Before final submission:

  • make the Space public
  • add demo video link to README.md
  • add social post link to README.md
  • add public GitHub repo link to README.md
  • document bonus badges and model choice