Migrate your legacy datasets to Argilla V2¶

This guide will help you migrate task specific datasets to Argilla V2. These do not include the FeedbackDataset which is just an interim naming convention for the latest extensible dataset. Task specific datasets are datasets that are used for a specific task, such as text classification, token classification, etc. If you would like to learn about the backstory of SDK this migration, please refer to the SDK migration blog post.

Note

Legacy Datasets include: DatasetForTextClassification, DatasetForTokenClassification, and DatasetForText2Text.

FeedbackDataset's do not need to be migrated as they are already in the Argilla V2 format.

To follow this guide, you will need to have the following prerequisites:

An argilla 1.* server instance running with legacy datasets.
An argilla >=1.29 server instance running. If you don't have one, you can create one by following the Argilla installation guide.
The argilla sdk package installed in your environment.

If your current legacy datasets are on a server with Argilla release after 1.29, you could chose to recreate your legacy datasets as new datasets on the same server. You could then upgrade the server to Argilla 2.0 and carry on working their. Your legacy datasets will not be visible on the new server, but they will remain in storage layers if you need to access them.

Steps¶

The guide will take you through three steps:

Retrieve the legacy dataset from the Argilla V1 server using the new argilla package.
Define the new dataset in the Argilla V2 format.
Upload the dataset records to the new Argilla V2 dataset format and attributes.

Step 1: Retrieve the legacy dataset¶

Connect to the Argilla V1 server via the new argilla package. The new sdk contains a v1 module that allows you to connect to the Argilla V1 server:

import argilla.v1 as rg_v1

# Initialize the API with an Argilla server less than 2.0
api_url = "<your-url>"
api_key = "<your-api-key>"
rg_v1.init(api_url, api_key)

Next, load the dataset settings and records from the Argilla V1 server:

dataset_name = "news-programmatic-labeling"
workspace = "demo"

settings_v1 = rg_v1.load_dataset_settings(dataset_name, workspace)
records_v1 = rg_v1.load(dataset_name, workspace)
hf_dataset = records_v1.to_datasets()

Your legacy dataset is now loaded into the hf_dataset object.

Step 2: Define the new dataset¶

Define the new dataset in the Argilla V2 format. The new dataset format is defined in the argilla package. You can create a new dataset with the Settings and Dataset classes:

First, instantiate the Argilla class to connect to the Argilla V2 server:

import argilla as rg

client = rg.Argilla()

Next, define the new dataset settings:

settings = rg.Settings(
    fields=[
        rg.TextField(name="text"), # (1)
    ],
    questions=[
        rg.LabelQuestion(name="label", labels=settings_v1.label_schema), # (2)
    ],
    metadata=[
        rg.TermsMetadataProperty(name="split"), # (3)
    ],
    vectors=[
        rg.VectorField(name='mini-lm-sentence-transformers', dimensions=384), # (4)
    ],
)

The default name for text classification is text, but we should provide all names included in record.inputs.
The basis question for text classification is a LabelQuestion for single-label or MultiLabelQuestion for multi-label classification.
Here, we need to provide all relevant metadata fields.
The vectors fields available in the dataset.

Finally, create the new dataset on the Argilla V2 server:

dataset = rg.Dataset(name=dataset_name, settings=settings)
dataset.create()

Note

If a dataset with the same name already exists, the create method will raise an exception. You can check if the dataset exists and delete it before creating a new one.

dataset = client.datasets(name=dataset_name)

if dataset.exists():
    dataset.delete()

Step 3: Upload the dataset records¶

To upload the records to the new server, we will need to convert the records from the Argilla V1 format to the Argilla V2 format. The new argilla sdk package uses a generic Record class, but legacy datasets have specific record classes. We will need to convert the records to the generic Record class.

Here are a set of example functions to convert the records for single-label and multi-label classification. You can modify these functions to suit your dataset.

For single-label classificationFor multi-label classificationFor token classificationFor Text generation

def map_to_record_for_single_label(data: dict, users_by_name: dict, current_user: rg.User) -> rg.Record:
    """ This function maps a text classification record dictionary to the new Argilla record."""
    suggestions = []
    responses = []

    if prediction := data.get("prediction"):
        label, score = prediction[0].values()
        agent = data["prediction_agent"]
        suggestions.append(rg.Suggestion(question_name="label", value=label, score=score, agent=agent))

    if annotation := data.get("annotation"):
        user_id = users_by_name.get(data["annotation_agent"], current_user).id
        responses.append(rg.Response(question_name="label", value=annotation, user_id=user_id))

    vectors = (data.get("vectors") or {})
    return rg.Record(
        id=data["id"],
        fields=data["inputs"],
        # The inputs field should be a dictionary with the same keys as the `fields` in the settings
        metadata=data["metadata"],
        # The metadata field should be a dictionary with the same keys as the `metadata` in the settings
        vectors=[rg.Vector(name=name, values=value) for name, value in vectors.items()],
        suggestions=suggestions,
        responses=responses,
    )

def map_to_record_for_multi_label(data: dict, users_by_name: dict, current_user: rg.User) -> rg.Record:
    """ This function maps a text classification record dictionary to the new Argilla record."""
    suggestions = []
    responses = []

    if prediction := data.get("prediction"):
        labels, scores = zip(*[(pred["label"], pred["score"]) for pred in prediction])
        agent = data["prediction_agent"]
        suggestions.append(rg.Suggestion(question_name="labels", value=labels, score=scores, agent=agent))

    if annotation := data.get("annotation"):
        user_id = users_by_name.get(data["annotation_agent"], current_user).id
        responses.append(rg.Response(question_name="label", value=annotation, user_id=user_id))

    vectors = data.get("vectors") or {}
    return rg.Record(
        id=data["id"],
        fields=data["inputs"],
        # The inputs field should be a dictionary with the same keys as the `fields` in the settings
        metadata=data["metadata"],
        # The metadata field should be a dictionary with the same keys as the `metadata` in the settings
        vectors=[rg.Vector(name=name, values=value) for name, value in vectors.items()],
        suggestions=suggestions,
        responses=responses,
    )

def map_to_record_for_span(data: dict, users_by_name: dict, current_user: rg.User) -> rg.Record:
    """ This function maps a token classification record dictionary to the new Argilla record."""
    suggestions = []
    responses = []

    if prediction := data.get("prediction"):
        scores = [span["score"] for span in prediction]
        agent = data["prediction_agent"]
        suggestions.append(rg.Suggestion(question_name="spans", value=prediction, score=scores, agent=agent))

    if annotation := data.get("annotation"):
        user_id = users_by_name.get(data["annotation_agent"], current_user).id
        responses.append(rg.Response(question_name="spans", value=annotation, user_id=user_id))

    vectors = data.get("vectors") or {}
    return rg.Record(
        id=data["id"],
        fields={"text": data["text"]},
        # The inputs field should be a dictionary with the same keys as the `fields` in the settings
        metadata=data["metadata"],
        # The metadata field should be a dictionary with the same keys as the `metadata` in the settings
        vectors=[rg.Vector(name=name, values=value) for name, value in vectors.items()],
        # The vectors field should be a dictionary with the same keys as the `vectors` in the settings
        suggestions=suggestions,
        responses=responses,
    )

def map_to_record_for_text_generation(data: dict, users_by_name: dict, current_user: rg.User) -> rg.Record:
    """ This function maps a text2text record dictionary to the new Argilla record."""
    suggestions = []
    responses = []

    if prediction := data.get("prediction"):
        first = prediction[0]
        agent = data["prediction_agent"]
        suggestions.append(
            rg.Suggestion(question_name="text_generation", value=first["text"], score=first["score"], agent=agent)
        )

    if annotation := data.get("annotation"):
        # From data[annotation]
        user_id = users_by_name.get(data["annotation_agent"], current_user).id
        responses.append(rg.Response(question_name="text_generation", value=annotation, user_id=user_id))

    vectors = (data.get("vectors") or {})
    return rg.Record(
        id=data["id"],
        fields={"text": data["text"]},
        # The inputs field should be a dictionary with the same keys as the `fields` in the settings
        metadata=data["metadata"],
        # The metadata field should be a dictionary with the same keys as the `metadata` in the settings
        vectors=[rg.Vector(name=name, values=value) for name, value in vectors.items()],
        # The vectors field should be a dictionary with the same keys as the `vectors` in the settings
        suggestions=suggestions,
        responses=responses,
    )

The functions above depend on the users_by_name dictionary and the current_user object to assign responses to users, we need to load the existing users. You can retrieve the users from the Argilla V2 server and the current user as follows:

# For
users_by_name = {user.username: user for user in client.users}
current_user = client.me

Finally, upload the records to the new dataset using the log method and map functions.

records = []

for data in hf_records:
    records.append(map_to_record_for_single_label(data, users_by_name, current_user))

# Upload the records to the new dataset
dataset.records.log(records)

You have now successfully migrated your legacy dataset to Argilla V2. For more guides on how to use the Argilla SDK, please refer to the How to guides.