[davanstrien]

Adds run-on-jobs/, a static Space that lists uv-scripts recipes with copy-paste commands. It declares the models with working recipes in its README metadata, so the Hub lists it under "Spaces using this model" on those model pages.

What's here

• tools/build-catalog.py: generates index.html, recipes.json, AGENTS.md and README.md by parsing the scripts. It doesn't import or run them.
• tools/run-on-jobs-template.html: the page, including first-run examples for OCR, transcription, embeddings and fine-tuning.
• sync-run-on-jobs.yml: publishes the folder with the existing _sync-folder.yml.
• check-run-on-jobs.yml: fails a PR when the Space files are stale. It also runs weekly and checks that declared models still exist.

What gets declared
48 recipes are listed, and 26 models are declared. A recipe only counts towards models: if it has a header or its documented command passed a smoke Job. Training recipes, experimental OCR recipes and vllm/generate-responses are listed but don't count, because the default command for the last one needs --max-model-len.

Commands
Nothing is hard-coded. A recipe's command comes from its first documented example, keeping only mounts for header recipes. Otherwise it's built from the script's positional and required arguments, with mounts for path arguments. Every declared command passes --dry-run and resolves to a GPU flavor or its documented flavor.

Testing

• python3 tools/build-catalog.py --check --verify-models passes, and ruff is clean.
• Smoke Jobs with output checked: sam3/detect-objects, sam3/segment-objects and embeddings/generate-embeddings.
• classification/train-setfit: re-ran the first-run example on cpu-basic (0.77 accuracy vs 0.32 majority baseline, 100 test rows, one seed).
• gliner/extract-entities failed its smoke Job (the tokenizer needs tiktoken), so it's left out through BROKEN until it's fixed.
• The page was checked in a browser at desktop width: tabs, search, empty-search reset, recipes.json.

Not covered

• Three recipes have no command on the page yet: embeddings/launch-embedding-fleet, transcription/cohere-transcribe-vllm and transcription/moss-transcribe-diarize-server.
• Most header recipes rely on earlier sweeps, not a run from this PR.
• We can't count Jobs started from this page.

Rollout
uv-scripts/run-on-jobs exists as a private preview with these files. The sync keeps it private, so model pages only list it once it's made public.

🤖 Generated with Claude Code

https://claude.ai/code/session_016isSPpM5PRqNUzwsDntoG4