Recipes

Annotate coreference, i.e. links of ambiguous mentions like "her" or "the woman" back to an antecedent providing more context about the entity in question.

Annotate syntactic dependencies between words, starting from a trained spaCy model's predictions.

Accept or reject a label on a stream of generated examples.

Annotate named entities (NER) like people, organizations, places or products, manually or with suggestions from a spaCy model or match patterns.

Annotate part-of-speech (POS) tags like verbs or nouns, starting from a trained spaCy model's tagger.

Annotate relations between entities, spans or tokens. Also supports joint span and relation annotation.

Correct a spaCy model's sentence boundaries to create gold data for sentence segmentation.

Annotate overlapping and nested text spans (spancat), manually or with suggestions from a spaCy model or match patterns.

Bootstrap a terminology list (gazetteer) from spaCy word vectors. Terminology lists can be converted into match patterns to pre-select entity spans during annotation.

Assign categories to whole texts or sentences, e.g. topic, sentiment or intent. Single- or multi-label, optionally with suggestions from a spaCy model.

Browse an input asset or dataset and accept or reject examples, e.g. to filter data or check annotations before training.

Task with tunable delays and errors for debugging.

Review annotations from multiple annotators, resolve conflicts and disagreements, and save one final gold-standard annotation per example.

Annotate bounding boxes and segments on PDF pages rendered as images, e.g. tables, figures or signatures.

Run Tesseract OCR on the bounding boxes annotated with PDF Image Annotation and correct the extracted text.

Annotate text spans, e.g. named entities, in PDFs pre-processed into a Layout asset by the Extract PDF layout action.

Draw bounding boxes, polygons or freehand segments on images (object detection, segmentation), or classify images into categories.

Label regions in audio files (e.g. speakers or sound events), classify recordings into categories, or transcribe them.

Label regions in video files, classify videos into categories, or transcribe them.

Publish a trained spaCy pipeline or fine-tuned Hugging Face model to a model repo, private by default.

Fine-tune a Hugging Face transformers model for NER or text classification on your annotations (CPU). The result can run as an annotation agent or be pushed to the Hub.

Fine-tune a Hugging Face transformers model for NER or text classification on your annotations on a GPU worker, for models too large to train on CPU.

Train a spaCy pipeline on your annotations: named entities, spans, text classification, tagging, parsing, sentence segmentation or coreference.

Train a spaCy pipeline using spacy-curated-transformers for GPU-accelerated training and inference, loading BERT, RoBERTa, CamemBERT or XLM-RoBERTa weights from a Hugging Face checkpoint.

Run inference with a trained spacy-curated-transformers model over an input dataset or asset, writing predictions to an output dataset.

Copy or merge annotated data into a new dataset.

Download and install one or more spaCy models to shared storage so they can be loaded with spacy.load().

Download a dataset split from the Hugging Face Hub and save it as an Input asset, ready to annotate.

Publish an annotated dataset to a Hugging Face dataset repo, private by default.

Download a model checkpoint from the Hugging Face Hub into the cluster's shared cache, so agents and training start without downloading it again.

Get text classification predictions from an OpenAI model, set up in a spacy-llm config, and save them to a dataset.

Convert a dataset's examples from before structured storage to the structured format. The originals are backed up as a JSONL asset on your cluster first, then removed from the dataset.

Generate diverse synthetic examples from a plain-English task description using Google Gemini Flash, write them as a JSONL file and register the result as an Input asset on the cluster.

Label texts with TypeSafe's Jev and keep the answers it's confident about. Exclude the output dataset from a text classification task to leave the rest to people.

Render a published PDF asset's pages into images with pypdfium2, making it ready to annotate. Run via ellf publish data --kind pdf --render, or directly against an already-published asset.

Extract text and layout from a PDF asset with spacy-layout and Docling, and save as a Layout asset for fast annotation with PDF Span Annotation.

Write a small text file to the given path and register it as an Input asset, to try out actions and assets on a cluster.

Print 'hello world'.

Annotation agent powered by Google Gemini. Annotates named entities (NER), spans, part-of-speech tags and text categories, and accepts or rejects binary suggestions.

Annotation agent powered by an LLM on Hugging Face Inference Providers, e.g. Llama. Annotates named entities (NER), spans, part-of-speech tags and text categories, and accepts or rejects binary suggestions.

Annotation agent running a fine-tuned Hugging Face transformers model on your cluster. Annotates named entities (NER) and spans with a token-classification model, or text categories with a text-classification model.

Accept every task it is given.

Deterministic annotation agent for tests and development. Annotates named entities and spans in manual NER and span categorization tasks, using spaCy match patterns.

Live Streamlit dashboard for an annotation dataset: label distribution, conflicting annotations, a browsable table of every annotation, and health checks. Opens in the browser for logged-in project members.

Run editable marimo notebooks on your cluster, with access to the Prodigy database and authentication handled by Ellf. Opens in the browser for logged-in project members.

spaCy NER pattern generation as an MCP server. Uses Google Gemini to propose and iteratively refine spaCy matcher patterns given example text and a description of the entity types.

Serve a 'hello world' HTTP endpoint.