refinery repository Python 3.9 pypi 1.4.0

This is the official Python SDK for refinery, the open-source data-centric IDE for NLP.

Table of Contents

If you like what we're working on, please leave a ⭐!


You can set up this SDK either via running $ pip install refinery-python-sdk, or by cloning this repository and running $ pip install -r requirements.txt.


Creating a Client object

Once you installed the package, you can create a Client object from any Python terminal as follows:

from refinery import Client

user_name = "your-username" # this is the email you log in with
password = "your-password"
project_id = "your-project-id" # can be found in the URL of the web application

client = Client(user_name, password, project_id)
# if you run the application locally, please use the following instead:
# client = Client(user_name, password, project_id, uri="http://localhost:4455")

The project_id can be found in your browser, e.g. if you run the app on your localhost: http://localhost:4455/app/projects/{project_id}/overview

Alternatively, you can provide a secrets.json file in your directory where you want to run the SDK, looking as follows:

  "user_name": "your-username",
  "password": "your-password",
  "project_id": "your-project-id"

Again, if you run on your localhost, you should also provide "uri": "http://localhost:4455". Afterwards, you can access the client like this:

client = Client.from_secrets_file("secrets.json")

With the Client, you easily integrate your data into any kind of system; may it be a custom implementation, an AutoML system or a plain data analytics framework 🚀

Fetching labeled data

Now, you can easily fetch the data from your project:

df = client.get_record_export(tokenize=False)
# if you set tokenize=True (default), the project-specific
# spaCy tokenizer will process your textual data

Alternatively, you can also just run rsdk pull in your CLI given that you have provided the secrets.json file in the same directory.

The df contains both your originally uploaded data (e.g. headline and running_id if you uploaded records like {"headline": "some text", "running_id": 1234}), and a triplet for each labeling task you create. This triplet consists of the manual labels, the weakly supervised labels, and their confidence. For extraction tasks, this data is on token-level.

An example export file looks like this:

    "running_id": "0",
    "Headline": "T. Rowe Price (TROW) Dips More Than Broader Markets",
    "Date": "Jun-30-22 06:00PM\u00a0\u00a0",
    "Headline__Sentiment Label__MANUAL": null,
    "Headline__Sentiment Label__WEAK_SUPERVISION": "Negative",
    "Headline__Sentiment Label__WEAK_SUPERVISION__confidence": "0.6220"

In this example, there is no manual label, but a weakly supervised label "Negative" has been set with 62.2% confidence.

Fetching lookup lists

In your project, you can create lookup lists to implement distant supervision heuristics. To fetch your lookup list(s), you can either get all or fetch one by its list id.

list_id = "your-list-id"
lookup_list = client.get_lookup_list(list_id)

The list id can be found in your browser URL when you're on the details page of a lookup list, e.g. when you run on localhost: http://localhost:4455/app/projects/{project_id}/knowledge-base/{list_id}.

Alternatively, you can pull all lookup lists:

lookup_lists = client.get_lookup_lists()

Upload files

You can import files directly from your machine to your application:

file_path = "my/file/path/data.json"
upload_was_successful = client.post_file_import(file_path)

We use Pandas to process the data you upload, so you can also provide import_file_options for the file type you use. Currently, you need to provide them as a \n-separated string (e.g. "quoting=1\nsep=';'"). We'll adapt this in the future to work with dictionaries instead.

Alternatively, you can rsdk push <path-to-your-file> via CLI, given that you have provided the secrets.json file in the same directory.

Make sure that you've selected the correct project beforehand, and fit the data schema of existing records in your project!


Sklearn Adapter

You can use refinery to directly pull data into a format you can apply for building sklearn models. This can look as follows:

from refinery.adapter.sklearn import build_classification_dataset
from sklearn.tree import DecisionTreeClassifier

data = build_classification_dataset(client, "headline", "__clickbait", "distilbert-base-uncased")

clf = DecisionTreeClassifier()["train"]["inputs"], data["train"]["labels"])

pred_test = clf.predict(data["test"]["inputs"])
accuracy = (pred_test == data["test"]["labels"]).mean()

By the way, we can highly recommend to combine this with Truss for easy model serving!

PyTorch Adapter

If you want to build a PyTorch network, you can build the train_loader and test_loader as follows:

from refinery.adapter.torch import build_classification_dataset
train_loader, test_loader, encoder, index = build_classification_dataset(
    client, "headline", "__clickbait", "distilbert-base-uncased"

Hugging Face Adapter

Transformers are great, but often times, you want to finetune them for your downstream task. With refinery, you can do so easily by letting the SDK build the dataset for you that you can use as a plug-and-play base for your training:

from refinery.adapter import transformers
dataset, mapping = transformers.build_dataset(client, "headline", "__clickbait")

From here, you can follow the finetuning example provided in the official Hugging Face documentation. A next step could look as follows:

small_train_dataset = dataset["train"].shuffle(seed=42).select(range(1000))
small_eval_dataset = dataset["test"].shuffle(seed=42).select(range(1000))

from transformers import (
  AutoTokenizer, AutoModelForSequenceClassification, TrainingArguments, Trainer
import numpy as np
from datasets import load_metric

tokenizer = AutoTokenizer.from_pretrained("distilbert-base-uncased")

def tokenize_function(examples):
    return tokenizer(examples["headline"], padding="max_length", truncation=True)

tokenized_datasets =, batched=True)
model = AutoModelForSequenceClassification.from_pretrained("distilbert-base-uncased", num_labels=2)
training_args = TrainingArguments(output_dir="test_trainer")
metric = load_metric("accuracy")

def compute_metrics(eval_pred):
    logits, labels = eval_pred
    predictions = np.argmax(logits, axis=-1)
    return metric.compute(predictions=predictions, references=labels)

training_args = TrainingArguments(output_dir="test_trainer", evaluation_strategy="epoch")

small_train_dataset = tokenized_datasets["train"].shuffle(seed=42).select(range(1000))
small_eval_dataset = tokenized_datasets["test"].shuffle(seed=42).select(range(1000))

trainer = Trainer(



Rasa Adapter

refinery is perfect to be used for building chatbots with Rasa. We've built an adapter with which you can easily create the required Rasa training data directly from refinery.

To do so, do the following:

from refinery.adapter import rasa


This will create a .yml file looking as follows:

  - intent: check_balance
    examples: |
      - how much do I have on my savings account
      - how much money is in my checking account
      - What's the balance on my credit card account

If you want to provide a metadata-level label (such as sentiment), you can provide the optional argument metadata_label_task:

from refinery.adapter import rasa


This will create a file like this:

  - intent: check_balance
      sentiment: neutral
    examples: |
      - how much do I have on my savings account
      - how much money is in my checking account
      - What's the balance on my credit card account

And if you have entities in your texts which you'd like to recognize, simply add the tokenized_label_task argument:

from refinery.adapter import rasa


This will not only inject the label names on token-level, but also creates lookup lists for your chatbot:

  - intent: check_balance
      sentiment: neutral
    examples: |
      - how much do I have on my [savings](account) account
      - how much money is in my [checking](account) account
      - What's the balance on my [credit card account](account)
  - lookup: account
    examples: |
      - savings
      - checking
      - credit card account

Please make sure to also create the further necessary files (domain.yml, data/stories.yml and data/rules.yml) if you want to train your Rasa chatbot. For further reference, see their documentation.


If you want to feed your production model's predictions back into refinery, you can do so with any version greater than 1.2.1.

To do so, we have a generalistic interface and framework-specific classes.

Sklearn Callback

If you want to train a scikit-learn model an feed its outputs back into the refinery, you can do so easily as follows:

from sklearn.linear_model import LogisticRegression
clf = LogisticRegression() # we use this as an example, but you can use any model implementing predict_proba

from refinery.adapter.sklearn import build_classification_dataset
data = build_classification_dataset(client, "headline", "__clickbait", "distilbert-base-uncased")["train"]["inputs"], data["train"]["labels"])

from refinery.callbacks.sklearn import SklearnCallback
callback = SklearnCallback(

# executing this will call the refinery API with batches of size 32, so your data is pushed to the app["train"]["inputs"], data["train"]["index"])["test"]["inputs"], data["test"]["index"])

PyTorch Callback

For PyTorch, the procedure is really similar. You can do as follows:

from refinery.adapter.torch import build_classification_dataset
train_loader, test_loader, encoder, index = build_classification_dataset(
    client, "headline", "__clickbait", "distilbert-base-uncased"

# build your custom model and train it here - example:
import torch.nn as nn
import numpy as np
import torch

# number of features (len of X cols)
input_dim = 768
# number of hidden layers
hidden_layers = 20
# number of classes (unique of y)
output_dim = 2
class Network(nn.Module):
    def __init__(self):
        super(Network, self).__init__()
        self.linear1 = nn.Linear(input_dim, output_dim)

    def forward(self, x):
        x = torch.sigmoid(self.linear1(x))
        return x

clf = Network()
criterion = nn.CrossEntropyLoss()
optimizer = torch.optim.SGD(clf.parameters(), lr=0.1)

epochs = 2
for epoch in range(epochs):
    running_loss = 0.0
    for i, data in enumerate(train_loader, 0):
        inputs, labels = data
        # set optimizer to zero grad to remove previous epoch gradients
        # forward propagation
        outputs = clf(inputs)
        loss = criterion(outputs, labels)
        # backward propagation
        # optimize
        running_loss += loss.item()
        # display statistics
        print(f'[{epoch + 1}, {i + 1:5d}] loss: {running_loss / 2000:.5f}')
        running_loss = 0.0

# with this model trained, you can use the callback
from refinery.callbacks.torch import TorchCallback
callback = TorchCallback(

# and just execute this, index["train"]), index["test"])

HuggingFace Callback

Collect the dataset and train your custom transformer model as follows:

from refinery.adapter import transformers
dataset, mapping, index = transformers.build_classification_dataset(client, "headline", "__clickbait")

# train a model here, we're simplifying this by just using an existing model w/o retraining
from transformers import pipeline
pipe = pipeline("text-classification", model="distilbert-base-uncased")

# if you're interested to see how a training looks like, look into the above HuggingFace adapter

# you can now apply the callback
from refinery.callbacks.transformers import TransformerCallback
callback = TransformerCallback(
)["train"]["headline"], index["train"])["test"]["headline"], index["test"])

Generic Callback

This one is your fallback if you have a very custom solution; other than that, we recommend you look into the framework-specific classes.

from refinery.callbacks.inference import ModelCallback
from refinery.adapter.sklearn import build_classification_dataset
from sklearn.linear_model import LogisticRegression

data = build_classification_dataset(client, "headline", "__clickbait", "distilbert-base-uncased"0)
clf = LogisticRegression()["train"]["inputs"], data["train"]["labels"])

# you can build initialization functions that set states of objects you use in the pipeline
def initialize_fn(inputs, labels, **kwargs):
    return {"clf": kwargs["clf"]}

# postprocessing shifts the model outputs into a format accepted by our API
def postprocessing_fn(outputs, **kwargs):
    named_outputs = []
    for prediction in outputs:
        pred_index = prediction.argmax()
        label = kwargs["clf"].classes_[pred_index]
        confidence = prediction[pred_index]
        named_outputs.append([label, confidence])
    return named_outputs

callback = ModelCallback(
    client: Client,

# executing this will call the refinery API with batches of size 32
callback.initialize_and_run(data["train"]["inputs"], data["train"]["index"])["test"]["inputs"], data["test"]["index"])


Contributions are what make the open source community such an amazing place to learn, inspire, and create. Any contributions you make are greatly appreciated.

If you have a suggestion that would make this better, please fork the repo and create a pull request. You can also simply open an issue with the tag "enhancement".

  1. Fork the Project
  2. Create your Feature Branch (git checkout -b feature/AmazingFeature)
  3. Commit your Changes (git commit -m 'Add some AmazingFeature')
  4. Push to the Branch (git push origin feature/AmazingFeature)
  5. Open a Pull Request

And please don't forget to leave a ⭐ if you like the work!


Distributed under the MIT License. See LICENSE.txt for more information.


This library is developed and maintained by Kern AI. If you want to provide us with feedback or have some questions, don't hesitate to contact us. We're super happy to help ✌️