GlobalWheatSeedingsPrompt

Using vibe coding to develop head counting software

Codex setup

In chat:

follow this instructions in .codex/rules/setup.md to setup the development tools for this project

Download GWFSS dataset

from this link: https://huggingface.co/datasets/GlobalWheat/GWFSS_v1.0/tree/main

Download and unzip GWFSS_v1.0_labelled into data/GWFSS_v1.0_labelled. Keep the downloaded dataset outside git; data/ is ignored by this repo.

Expected folder structure:

data/
`-- GWFSS_v1.0_labelled/
    |-- images/
    |   |-- Arvalis/
    |   |-- CIMMYT/
    |   |-- ...
    |   |-- USASK/
    |   `-- UTokyo/
    `-- masks/
        |-- Arvalis/
        |-- CIMMYT/
        |-- ...
        |-- USASK/
        `-- UTokyo/

Images and masks are paired by source folder and file stem where the dataset provides annotations. The generated app should check this data path on startup and build an image/result index for fast browsing. Cache that index under a local cache path such as data/.cache/gwfss_index.json or data/.cache/gwfss_index.npy; do not commit cache files.

When the user selects Global Wheat data in the UI, the app should list indexed samples from the backend. The frontend should show source-folder tags for filtering, preview images from backend-provided URLs, 50 grid items per page, and pagination for additional samples. The selected sample should resolve the paired mask and metadata through the startup index.

Frontend stack

The generated demo frontend should use Svelte 5 with Vite and Tailwind CSS. Use TypeScript where practical. Do not use SvelteKit for this small local demo unless a later prompt explicitly changes the scope. Svelte components should use Svelte 5 runes such as $state, $derived, $effect, and $props, plus the current event syntax such as onclick.

Keep frontend package and build configuration at the project root, reusing package.json, pnpm-lock.yaml, and Vite/Tailwind/TypeScript config files there. Use app/frontend as the Vite source root rather than creating a nested frontend package under app/frontend.

Local model weights

The generated app must run local inference only. Backend scripts assume required model files have already been downloaded into a weights/ directory in the target app workspace. The scripts should validate these paths and skip model inference cleanly when files are missing; they should not download weights at runtime.

Do not commit weights/ to git.

Expected layout:

weights/
`-- sam3.1_multiplex.pt

SAM 3.1 through Ultralytics

SAM 3 is available through ultralytics>=8.3.237. Install or upgrade it in the generated app environment:

uv add ultralytics

SAM 3 weights are gated and are not automatically downloaded by Ultralytics. Request access on the SAM 3 Hugging Face model page linked from the Ultralytics SAM 3 installation docs. After approval (facebook/sam3.1 · Hugging Face), download sam3.1_multiplex.pt manually and place it at:

weights/sam3.1_multiplex.pt

For text or concept prompts, use Ultralytics SAM3SemanticPredictor, not the plain SAM("sam3.pt") visual-prompt wrapper. The plain wrapper is for points, boxes, or masks and segments a specific prompted object rather than all instances matching a concept.

If prediction raises TypeError: 'SimpleTokenizer' object is not callable, the Ultralytics docs recommend replacing the unrelated clip package with the Ultralytics CLIP package:

uv pip uninstall clip
uv pip install git+https://github.com/ultralytics/CLIP.git