README.md

metadata

title: TemperCheck
emoji: 😤
colorFrom: yellow
colorTo: red
sdk: gradio
sdk_version: 6.18.0
app_file: app.py
pinned: false
short_description: How short a temper does this profile look to have?
tags:
  - track:backyard
  - achievement:offgrid

😤 TemperCheck

Upload a social-media profile (or a screenshot of one) and TemperCheck gives a playful read on how short-tempered / difficult that person looks to deal with — a 0–100 score, a punchy verdict, and the "signals" it picked up on.

Built for the Build Small Hackathon using a small Gemma 4 E4B vision-language model (~8B params, well under the 32B limit).

⚠️ It's a party game. TemperCheck reads vibes in a picture for laughs. It is not a real personality test and makes no factual claim about any real person.

How it runs

On this Space it runs the transformers backend on ZeroGPU — the model is google/gemma-4-E4B-it, loaded on cuda at startup and run inside a @spaces.GPU function. (The Space needs an HF_TOKEN secret with access to the gated Gemma repo.) The backend is selected automatically from the SPACE_ID env, so there's nothing to configure.

Locally you can run the UI against a local Ollama instead (uv run app.py, opens http://localhost:7140), but note Gemma 4 vision is currently unreliable in Ollama, so the local path is for UI work — real verdicts come from the Space.

Configuration (env vars)

Var	Default	Purpose
TEMPER_BACKEND	transformers on a Space, else ollama	force a backend
TEMPER_HF_MODEL	google/gemma-4-E4B-it	transformers model id
TEMPER_OLLAMA_MODEL	huihui_ai/gemma-4-abliterated:e4b-q8_0	local Ollama model id
OLLAMA_HOST	http://127.0.0.1:11434	local Ollama server
TEMPER_PORT	7140	local Gradio port

Tests

uv run pytest

Project layout

• app.py — Gradio UI.
• tempercheck/inference.py — backend abstraction (Ollama ↔ transformers).
• tempercheck/prompt.py — system prompt + defensive JSON parsing.
• tests/test_parsing.py — output-parsing tests (no model needed).