Skip to content

Potato on HuggingFace

Potato has a presence on the HuggingFace Hub: a small set of live demo Spaces you can try in the browser, a wider catalog of deployed-but-paused ones, sample annotation datasets, and guides for deploying your own. This page is the index.

The demo Spaces live under the Blablablab org and are grouped in a HuggingFace Collection. Each one is generated from a project in examples/ via the Spaces catalog tooling — adding a new one is a single manifest entry. Every example in the repo runs locally whether or not it has a Space.

Running a demo

HuggingFace caps a free CPU-basic organization at three Spaces running at once, and the 48-hour idle-sleep timer cannot be shortened on free hardware. A demo someone opens holds one of those three slots for two days after its last visit. Three demos are therefore kept live and the rest of the catalog is paused.

The three are marked · live below. They were picked from search traffic to potatoannotator.com:

Demo Why
video-annotation Video action and temporal segmentation is the most-searched topic on the site
agent-comparison LLM and agent evaluation is the second
image-bbox Image region annotation is the third

Search demand concentrates in video, agents and images, so a traffic-led pick of three leaves no text demo live. Span and classification annotation are in the catalog below but paused; run one of them locally to try it.

Every other demo is deployed but paused. A paused Space does not wake when you open it; only its owner can restart it. The rest of the catalog below is an index of what Potato supports, and those Space pages will not open.

Any of them runs locally from the example it was generated from. This needs no HuggingFace account and has no concurrency limit:

pip install potato-annotation
python potato/flask_server.py start examples/video/video-annotation/config.yaml -p 8000

"Duplicate this Space" no longer works on a free account. HuggingFace now requires PRO for personal accounts, or Team/Enterprise for organizations, to create or duplicate a Gradio or Docker Space. Potato demos are Docker Spaces. Running the example locally, as above, is the free path.

Demo catalog

Generated from deployment/huggingface-spaces/spaces_manifest.yaml. Entries marked · live are the ones kept reachable; the rest are deployed but paused. python deployment/huggingface-spaces/audit_spaces.py reports the current state.

Text classification (9)

  • sentiment-analysis — Single-choice radio labeling with sequential keybindings.
  • multi-label — Checkbox multi-select labeling for overlapping categories.
  • likert-scales — Likert-scale rating for agreement / intensity judgments.
  • slider-rating — Continuous slider scoring for fine-grained judgments.
  • best-worst-scaling — Best-worst scaling for robust relative ranking.
  • llm-preference — Pairwise preference judgments over LLM responses (RLHF-style).
  • pairwise-comparison — A/B comparison of two items side by side.
  • ranking — Drag-to-rank ordering of candidate items.
  • survey — Multi-question survey with mixed input types.

Span & structure (5)

  • ner-span — Highlight-and-label text spans for NER / extraction.
  • coreference — Link coreferent mentions across a document.
  • entity-linking — Link entity spans to a knowledge-base catalog.
  • dependency-tree — Annotate syntactic dependency arcs between tokens.
  • multi-span — Span annotation across multiple text fields.

Agent & GenAI evaluation (5)

  • agent-comparison · live — Side-by-side comparison of two agent trajectories.
  • anthropic-evaluation — Annotate Claude/Anthropic agent traces.
  • openai-evaluation — Annotate OpenAI agent / tool-use traces.
  • rag-evaluation — Rate retrieval relevance, faithfulness, citation accuracy.
  • web-agent-review — Review web-agent runs with screenshot + click/scroll overlays.

Multimodal (5)

  • image-bbox · live — Draw labeled bounding boxes on images.
  • audio-classification — Classify audio clips (emotion / event / speaker).
  • video-annotation · live — Video player with temporal segment labeling.
  • video-classification — Whole-clip video classification.
  • multimodal — Combine media + text fields in one task.

Advanced workflows (5)

  • codebook — Shared evolving codebook across annotators.
  • adjudication — Resolve multi-annotator disagreements.
  • quality-control — Attention checks and gold-standard items.
  • conditional-logic — Branching questions that adapt to prior answers.
  • mace — Annotator competence estimation with MACE.

Showcase (1)

  • all-annotation-types — Every schema type in a single task.

AI-assisted (5)

  • solo-mode — LLM auto-labels while you calibrate; human-in-the-loop.
  • judge-calibration — Calibrate an LLM judge against blind human labels.
  • ai-hints — LLM label suggestions inline in the annotation UI.
  • span-ai-keywords — LLM-suggested keyword spans to accept/reject.
  • llm-chat — Evaluate a live LLM chat conversation.

Domain layouts (3)

  • content-moderation — Custom moderation layout with policy categories.
  • medical-review — Domain layout for clinical text review.
  • dialogue-qa — Turn-level dialogue question-answering layout.

Deploying the catalog

See the Spaces deployment runbook:

python deployment/huggingface-spaces/build_space.py --list          # see all demos
python deployment/huggingface-spaces/build_space.py video-annotation
python deployment/huggingface-spaces/deploy_space.py video-annotation Blablablab