🪐 spaCy Project: Name Entity Recognition for text-to-image prompting

Project to train a evaluate a name entity recognition model to analyzing text-to-image prompts. The entities comprise 17 categories 7 main categories and 11 subcategories, extracted from a topic analysis made with BERTopic. The topic analysis can be explored the following visualization.

Specifier taxonomy
├── medium/
│   ├── photography
│   ├── painting
│   ├── rendering
│   └── illustration
├── influence/
│   ├── artist
│   ├── genre
│   ├── artwork
│   └── repository
├── light
├── color
├── composition
├── detail
└── context/
    ├── era
    ├── weather
    └── emotion

Prompt data are from the diffusionDB database and were annotated by hand using Prodigy.

📋 project.yml

The project.yml defines the data assets required by the project, as well as the available commands and workflows. For details, see the spaCy projects documentation.

⏯ Commands

The following commands are defined by the project. They can be executed using spacy project run [name]. Commands are only re-run if their inputs have changed.

Command	Description
install	Install dependencies, log in to Hugging Face and download a model
preprocess	Convert the data to spaCy's binary format
pretrain	Pretrain the embedding on all prompts
train	Train a named entity recognition model
evaluate	Evaluate the model and export metrics
package	Package the trained model so it can be installed
push_to_hub	Push the model to the Hub

⏭ Workflows

The following workflows are defined by the project. They can be executed using spacy project run [name] and will run the specified commands in order. Commands are only re-run if their inputs have changed.

Workflow	Steps
all	preprocess → pretrain → train → evaluate → package → push_to_hub

🗂 Assets

The following assets are defined by the project. They can be fetched by running spacy project assets in the project directory.

File	Source	Description
assets/diffusiondb_raw_prompts.jsonl	URL	JSONL-formatted of all diffusionDB prompts
assets/ner_prompting.jsonl	URL	JSONL-formatted development data exported from Prodigy, annotated with 17 entities