What’s Prodigy?

Prodigy is a modern annotation tool for creating training and evaluation data for machine learning models. You can also use Prodigy to help you inspect and clean your data, do error analysis and develop rule-based systems to use in combination with your statistical models.

The Python library includes a range of pre-built workflows and command-line commands for various tasks, and well-documented components for implementing your own workflow scripts. Your scripts can specify how the data is loaded and saved, change which questions are asked in the annotation interface, and can even define custom HTML and JavaScript to change the behavior of the front-end. The web application is optimized for fast, intuitive and efficient annotation.

Prodigy’s mission is to help you do more of all those manual or semi-automatic processes that we all know we don’t do enough of. To most data scientists, the advice to spend more time looking at your data is sort of like the advice to floss or get more sleep: it’s easy to recognize that it’s sound advice, but not always easy to put into practice. Prodigy helps by giving you a practical, flexible tool that fits easily into your workflow. With concrete steps to follow instead of a vague goal, annotation and data inspection will change from something you should do, to something you will do.

What Prodigy isn’t

1. Prodigy isn’t “software as a service” – it’s a tool that you can download, install and run yourself. This has several benefits: you can interact with it programmatically, import and extend it, and you don’t need to upload anything into the cloud. Prodigy never “phones home” or connects to our servers, and no data ever leaves your machine.

2. Prodigy isn’t free or open-source, although you do get some parts of the source code, and it does have great docs and an awesome community. Purchasing Prodigy does support open-source software though: we use revenue from Prodigy for a number of open-source projects, including our popular NLP library spaCy.

3. Prodigy doesn’t provide a lot of annotation-adjacent functionality, such as user accounts or project management. We designed Prodigy to be the best tool possible for small teams working on rapidly evolving projects. We’re also working on a second product, Prodigy Teams, which will help organizations keep all their annotation work in one place, and make it easier to collaborate on larger-scale projects.

First steps with Prodigy

Prodigy can be installed into your Python environment like any other library. The main way to interact with it is via the command line using the prodigy command, followed by the name of a recipe you want to run and optional settings. For example, to make sure everything it set up correctly and to view details about your installation, you can run the stats recipe:

Example

prodigy
stats
============================== Prodigy Stats ==============================

Version            1.11.0
Database Name      SQLite
Database Id        sqlite
Total Datasets     0
Total Sessions     0

python -m prodigy stats

Recipe scripts are what power the different annotation workflows. Prodigy ships with a variety of built-in recipes for different workflows and tasks. For example, the ner.manual recipe lets you annotate data for named entity recognition (NER) by manually highlighting the phrases and concepts. Clicking on a recipe name on this page takes you to the more detailed documentation.

1. Start the annotation server and label some examples

Let’s try it out and use ner.manual to highlight some entities in text. You can use your own data in one of the supported formats or download our example dataset of news headlines about Silicon Valley from the New York times.

Download news_headlines.jsonl

You can now run the following command in your terminal and hit enter. The command tells Prodigy to do run the ner.manual recipe with the following settings that it needs: the name of a dataset to save the annotations to (just make something up), the spaCy pipeline or language to tokenize the text (blank:en for a blank English model, more on that later), the path to the text you want to label and a comma-separated list of label options (PERSON, ORG, PRODUCT and LOCATION). By the way, you can also click the copy-paste button to copy the command to your clipboard and hover over the components for more info!

Recipe command

prodigy
ner.manual
ner_news_headlines
blank:en
./news_headlines.jsonl
--label PERSON,ORG,PRODUCT,LOCATION
✨ Starting the web server at http://localhost:8080 ...
Open the app in your browser and start annotating!
prodigy ner.manual ner_news_headlines blank:en ./news_headlines.jsonl --label PERSON,ORG,PRODUCT,LOCATION

When you open your browser at http://localhost:8080, you should see the Prodigy UI and an annotation question that looks something like this. You can now select a label at the top and click and drag to highlight a span of text – for example, you can highlight “Uber” as an ORG and “Silicon Valley” as a LOCATION.

Because the text is pre-tokenized and Prodigy knows where words start and end, you don’t have to select the exact letters – you just need to roughly hit the words and token span will be highlighted. For single words like “Uber”, you can also double-click on the word. To submit an annotation, click the green accept button or press the a key on your keyboard. If you don’t know the answer, you can click ignore or press space. Once you hit accept, reject or ignore, the example will move into your outbox. You can go back through the annotations in your outbox by clicking the undo arrow, or pressing backspace. As you annotate, batches of examples will be cleared from your outbox and sent back to the Prodigy server to be saved in the database. You can also hit the save button in the top-left corner or press command+s (on Mac) or ctrl+s to clear your outbox and manually save.

2. Look at the annotated data

If it’s your first time running Prodigy, you probably want to check out the data, just to see what you’ve created. The db-out command takes the name of a dataset and lets you export its contents to a file. Datasets are how annotations you collect are stored in Prodigy’s database. You typically want to use a new dataset for each new experiment.

Example

prodigy
db-out
ner_news_headlines
>
./annotations.jsonl
prodigy db-out ner_news_headlines > ./annotations.jsonl

Here’s what the output could look like – each example now includes the "tokens" and has a "spans" property describing the entities you highlighted. Each span includes the "start" and "end" offset into the text and the "label" you selected.

annotations.jsonl (example){"text":"Uber\u2019s Lesson: Silicon Valley\u2019s Start-Up Machine Needs Fixing","meta":{"source":"The New York Times"},"_input_hash":1886699658,"_task_hash":-1952856502,"tokens":[{"text":"Uber","start":0,"end":4,"id":0},{"text":"\u2019s","start":4,"end":6,"id":1},{"text":"Lesson","start":7,"end":13,"id":2},{"text":":","start":13,"end":14,"id":3},{"text":"Silicon","start":15,"end":22,"id":4},{"text":"Valley","start":23,"end":29,"id":5},{"text":"\u2019s","start":29,"end":31,"id":6},{"text":"Start","start":32,"end":37,"id":7},{"text":"-","start":37,"end":38,"id":8},{"text":"Up","start":38,"end":40,"id":9},{"text":"Machine","start":41,"end":48,"id":10},{"text":"Needs","start":49,"end":54,"id":11},{"text":"Fixing","start":55,"end":61,"id":12}],"_session_id":null,"_view_id":"ner_manual","spans":[{"start":0,"end":4,"token_start":0,"token_end":0,"label":"ORG"},{"start":15,"end":29,"token_start":4,"token_end":5,"label":"LOCATION"}],"answer":"accept"}
{"text":"Pearl Automation, Founded by Apple Veterans, Shuts Down","meta":{"source":"The New York Times"},"_input_hash":1487477437,"_task_hash":-1298236362,"tokens":[{"text":"Pearl","start":0,"end":5,"id":0},{"text":"Automation","start":6,"end":16,"id":1},{"text":",","start":16,"end":17,"id":2},{"text":"Founded","start":18,"end":25,"id":3},{"text":"by","start":26,"end":28,"id":4},{"text":"Apple","start":29,"end":34,"id":5},{"text":"Veterans","start":35,"end":43,"id":6},{"text":",","start":43,"end":44,"id":7},{"text":"Shuts","start":45,"end":50,"id":8},{"text":"Down","start":51,"end":55,"id":9}],"_session_id":null,"_view_id":"ner_manual","spans":[{"start":0,"end":16,"token_start":0,"token_end":1,"label":"ORG"},{"start":29,"end":34,"token_start":5,"token_end":5,"label":"ORG"}],"answer":"accept"}
{"text":"How Silicon Valley Pushed Coding Into American Classrooms","meta":{"source":"The New York Times"},"_input_hash":1842734674,"_task_hash":636683182,"tokens":[{"text":"How","start":0,"end":3,"id":0},{"text":"Silicon","start":4,"end":11,"id":1},{"text":"Valley","start":12,"end":18,"id":2},{"text":"Pushed","start":19,"end":25,"id":3},{"text":"Coding","start":26,"end":32,"id":4},{"text":"Into","start":33,"end":37,"id":5},{"text":"American","start":38,"end":46,"id":6},{"text":"Classrooms","start":47,"end":57,"id":7}],"_session_id":null,"_view_id":"ner_manual","spans":[{"start":4,"end":18,"token_start":1,"token_end":2,"label":"LOCATION"}],"answer":"accept"}
{"text":"Women in Tech Speak Frankly on Culture of Harassment","meta":{"source":"The New York Times"},"_input_hash":-487516519,"_task_hash":62119900,"tokens":[{"text":"Women","start":0,"end":5,"id":0},{"text":"in","start":6,"end":8,"id":1},{"text":"Tech","start":9,"end":13,"id":2},{"text":"Speak","start":14,"end":19,"id":3},{"text":"Frankly","start":20,"end":27,"id":4},{"text":"on","start":28,"end":30,"id":5},{"text":"Culture","start":31,"end":38,"id":6},{"text":"of","start":39,"end":41,"id":7},{"text":"Harassment","start":42,"end":52,"id":8}],"_session_id":null,"_view_id":"ner_manual","answer":"accept"}

At any point, you can export your data and use it in a different process, e.g. to train a model outside of Prodigy. To train in Prodigy, all you need is the name of the dataset.

3. Keep annotating and run a training experiment

To keep annotating and label more data, you can re-run the ner.manual command and it’ll start again from where you left off. You might also want to create a new dataset and use your own data at this point. Once you’ve collected a few hundred examples, you can use the train command to run your first training experiment with spaCy. The arguments let you define the components to train from different, comma-separated datasets – in this case, --ner ner_news_headlines to train from the annotations collected in ner_news_headlines.

In this example, we’re training with 372 total annotations. If you don’t provide an evaluation dataset with the eval: prefix, Prodigy will hold back a percentage of examples for evaluation, so you can see how your model does on unseen data.

Download annotated data

Example

prodigy
train
--ner ner_news_headlines
======================== Generating Prodigy config ========================
ℹ Auto-generating config with spaCy
✔ Generated training config
============================ Training pipeline ============================
Components: ner
Merging training and evaluation data for 1 components
  - [ner] Training: 298 | Evaluation: 74 (20% split)
Training: 298 | Evaluation: 74
Labels: ner (4)

ℹ Pipeline: ['tok2vec', 'ner']
ℹ Initial learn rate: 0.001
============================ Training pipeline ============================

E    #       LOSS TOK2VEC  LOSS NER  ENTS_F  ENTS_P  ENTS_R  SCORE
---  ------  ------------  --------  ------  ------  ------  ------
  0       0          0.00     55.22    0.00    0.00    0.00    0.00
  7     200        100.81   1831.94   70.91   81.25   62.90    0.71
 17     400         83.33    117.20   71.70   86.36   61.29    0.72
 28     600         13.83     17.33   75.93   89.13   66.13    0.76
 43     800          0.44      0.41   72.38   88.37   61.29    0.72
 60    1000          0.01      0.01   71.15   88.10   59.68    0.71
 81    1200         68.97     16.09   72.07   81.63   64.52    0.72
107    1400        138.88     19.77   75.47   90.91   64.52    0.75
139    1600        165.04     18.88   72.38   88.37   61.29    0.72
177    1800        135.35     15.34   75.73   95.12   62.90    0.76
223    2000        165.23     20.49   73.08   90.48   61.29    0.73
278    2200        297.13     26.26   71.03   84.44   61.29    0.71

If you want to export your model after training, you can specify an output directory:

Example

prodigy
train
./output_dir
--ner ner_news_headlines

Annotating data and training models is an iterative process. You often want to start over and try a different approach, try out new data or experiment with a different label scheme. Prodigy can make this process easier. For more details and strategies, including how to speed up annotation by preselecting entities and putting a model in the loop, check out the full documentation on named entity recognition with Prodigy.

Read the NER docs

4. Start a new project and assign categories to texts

Recipe command

prodigy
textcat.manual
news_topics
./news_headlines.jsonl
--label Technology,Politics,Economy,Entertainment
✨ Starting the web server at http://localhost:8080 ...
Open the app in your browser and start annotating!
prodigy textcat.manual news_topics ./news_headlines.jsonl --label Technology,Politics,Economy,Entertainment

When you open your browser at http://localhost:8080, you should see the Prodigy UI and an annotation question that looks something like this. You’re now able to select one or more options from the list of labels. Once you’ve made your selection, you can click accept or press a to submit the answer. Don’t forget to hit save when you’re done! Prodigy will also save periodically in the background.

In this example, the classes are not mutually exclusive, so one text can have multiple categories. If you want your labels to be mutually exclusive so you can only select one, you can set the --exclusive flag when you run textcat.manual. The answer will now be submitted automatically when you select an option, because there’s only one thing to select. If you make a mistake, you can always hit undo.

After annotation, you can run db-out to export your data (see the choice interface for details on the data format of multiple choice annotations), or use the train command to train a spaCy model. You can also train multiple components at the same time, using different datasets.

Example

prodigy
train
--ner ner_news_headlines
--textcat-multilabel news_topics

Read the text classification docs

Recipes and data flow

When you run Prodigy on the command line using the prodigy command, you execute a recipe script. Recipes are Python functions that define the annotation workflow – for example, the stream of examples to annotate, an optional callback to execute when new answers come in, how to load and preprocess the data, which annotation interface to use and where to save the annotations.

If a recipe defines these components and you run it, Prodigy will start the annotation server, including the REST API and web application. The web app will request batches of examples to annotate, and send back batches of answers in the background (or whenever you hit save). All answers are then saved to the database.

Recipes can also define optional callback functions that are executed whenever new answers are received by the server. This allows you to update a model in the loop. Streams of incoming examples are Python generators, and Prodigy will only ever ask for the next batch of examples. If you’re using a model to score the stream and decide what to send out for annotations, updates to the model weights will be reflected in future batches.

Here’s an example of a simple recipe function that uses the JSONL loader to load data from a file, renders it as plain text using the text interface, saves the annotations to a given dataset and executes a function update whenever new answers are received.

recipe.pyimport prodigy
from prodigy.components.stream import get_stream

@prodigy.recipe("custom-recipe")
def custom_recipe(dataset, source):
    stream = get_stream(source)

    def update(answers):
        print(f"Received {len(answers)} answers!")

    return {
        "dataset": dataset,
        "source": source,
        "update": update,
        "view_id": "text"
    }

Command-line usage

prodigy
custom-recipe
your_dataset
./data.jsonl
-F recipe.py

Read the custom recipes docs

Glossary

In the documentation, you’ll come across a variety of terms specific to collecting annotations with Prodigy. Here’s a quick overview of the most common ones, including a short description and links for more information.