Contains

A ~$2 fine-tuned open model (Qwen3.5-4B + Tinker LoRA) plus a rules layer that finds the allergens hidden inside compound ingredients and prints a US-style "Contains:" line for the 9 US major allergens. Runs offline on the baker's laptop, so the recipe never leaves it.

Write-up: Contains: I built an allergen checker for home bakers that says "I'm not sure" instead of guessing (DEV).

Draft, not certification. Output is a draft a human must verify with their suppliers. Scope is the 9 US major allergens (FALCPA + FASTER Act). Not EU rules, not cross-contact, not nutrition, not legal certification.

Milestone M0 (setup + go/no-go smoke tests). See PLAN.md.

make setup      # uv venv + deps (Windows: ./make.ps1 setup)
make test       # unit tests
make smoke      # M0 go/no-go smoke tests (needs TINKER_API_KEY + Ollama)

Copy .env.example to .env and fill TINKER_API_KEY. The same key is used for LoRA training, LoRA sampling, and the gpt-oss-120b big-model baseline — there is no second inference account.

uv run uvicorn contains.app:app --host 127.0.0.1 --port 8000     # or: make app / ./make.ps1 app

Config via .env / environment: SAMPLER=tinker_lora|tinker_base|ollama|rules_only (swap the model layer with no code change), LORA_SAMPLER_PATH, TINKER_API_KEY. If the model layer is unreachable the app answers rules-only and shows a banner.

Reading a label photo runs Gemma 3 4B in Ollama on your own machine; the hosted demo has no Ollama (and not the RAM for it), so there the page says photo reading is local-only and you paste the text instead. To use photos: install Ollama, ollama pull gemma3:4b, run the app locally as above. Expect ~10-60 s per photo on CPU.

The hosted demo runs on Render's free tier. A free instance spins down after about 15 minutes of inactivity, and the first request after that can take 50 s or more (Render's documented behaviour) while the container restarts and the Tinker client warms up.

Measured on the live site (warm instance, 2026-10-18):

No cold-start figure is reported. I tried to measure one after 17+ minutes idle, but the instance had not been spun down (the in-memory counters were unchanged, so the process never restarted), and I chose not to trigger a redeploy just to time it. Treat the 50 s+ figure as Render's, not ours.

• remote backend: Tinker sampling, Qwen3.5-4B + our LoRA (hosted demo on Render free tier)
• local backend: llama.cpp, Qwen3.5-4B Q4 GGUF + LoRA (fully offline)
• baselines: keyword rules; Qwen3.5-4B zero-shot; Gemma 3 4B zero-shot (Ollama); gpt-oss-120b (Tinker sampling)
• photo → text: Gemma 3 4B vision via Ollama reads supplier label photos (local)

• tinker-cookbook — supervised training loop and renderers. Apache-2.0.
• llama.cpp — local GGUF inference and LoRA conversion. MIT.
• Open Food Facts — training/eval data (openfoodfacts/product-database). Open Database License (ODbL).
• Gemma (Google) — label-photo transcription and a zero-shot baseline, via Ollama, under the Gemma Terms of Use.
• Qwen3.5-4B (Alibaba) and gpt-oss-120b (OpenAI) — base / baseline models, served through Tinker.

Commits may continue after the hackathon deadline; see the DEV post for what was done by the deadline.