Hugging Face

Supported labeling types: Span labeling, row labeling.

Datasaur integrates directly with Hugging Face, providing access to their 10k+ models.

After choosing Hugging Face as the option, you can go to Hugging Face website and choose the available models. You can also use your own private models hosted on Hugging Face.

Span labeling

For span labeling, enter the model name or provide an endpoint URL if using a self-hosted model. When using your own endpoint, a model name or API token is not required. You can also set a confidence score to adjust the prediction threshold.

Row labeling

In row labeling, select a column as the input in the Target text field and choose the target question as the output. To get started, enter either the model name or the Dedicated Inference Endpoint URL, along with your API token.

When choosing a model for label prediction, use a text classification model. The model should return a list of dictionaries (or objects) where each object contains all predictions (positive, negative, neutral), like this:

[
[ { label: "positive", score: 0.8 }, { label: "neutral", score: 0.15 }, { label: "negative", score: 0.05 } ],
[ { label: "negative", score: 0.6 }, { label: "neutral", score: 0.3 }, { label: "positive", score: 0.1 } ]
]

or just a single list/array that contains objects of single prediction (the highest score), like this:

[
{ label: "positive", score: 0.8 },
{ label: "negative", score: 0.6 }
]

There's an option for Faster prediction speed, which improves performance by processing entire rows at once. Note that this action cannot be undone. You can also adjust the Confidence score to set the prediction threshold.

Clicking Predict labels will automatically apply labels to the document based on the loaded model.