WEBINAR Fine-Tuning Llama 3: Adapting LLMs for Specialized Domains 🚀

Custom script examples

The following examples work when custom scripts are enabled.

For details on implementing your own custom scripts, see Label Studio Interface (LSI) and Frontend API implementation details.

Plotly

Use Plotly to insert charts and graphs into your labeling interface. Charts are rendered in every annotation opened by a user.

Plotly should be loaded first from CDN: https://cdn.plot.ly/plotly-2.26.0.min.js. For security reasons, it’s better to use a hash for script integrity.

Screenshot of Plotly graph in Label Studio

Script

await LSI.import('https://cdn.plot.ly/plotly-2.26.0.min.js', 'sha384-xuh4dD2xC9BZ4qOrUrLt8psbgevXF2v+K+FrXxV4MlJHnWKgnaKoh74vd/6Ik8uF',);

let data = LSI.task.data;
if (window.Plotly && data) {
  Plotly.newPlot("plot", [data.plotly]);
}

Related LSI instance methods:

Labeling config

You need to add <View idAttr="plot"/> into your config to render the Plotly chart.

For example:

<View>
  <Text name="function" value="Is it increasing?" />
  <Choices name="slope" toName="function">
    <Choice value="Increasing" />
    <Choice value="Decreasing" />
    <Choice value="Non-monotonic" />
  </Choices>
  <View idAttr="plot"/>
</View>

Related tags:

Data

[
  {
    "plotly": {
      "x": [1, 2, 3, 4],
      "y": [10, 15, 13, 17],
      "type": "scatter"
    }
  },
  {
    "plotly": {
      "x": [1, 2, 3, 4],
      "y": [16, 5, 11, 9],
      "type": "scatter"
    }
  }
]

Custom validation

In this example, the script checks to ensure that the annotation does not include obscenity or disallowed words.

The following script displays a modal if a user tries to submit an annotation with the word “hate” added to any audio transcription.

Note that this is a “soft” block, meaning that the user can dismiss the modal and still proceed. For an example of a “hard” block, see Check that TextArea input is valid JSON below.

Screenshot of custom validation modal in Label Studio

Script

// Use Label Studio Interface to subscribe to events
// before Save Annotation is an event invoked before submitting and updating annotation
// returning "false" for this event prevents saving annotation

let dismissed = false;

LSI.on("beforeSaveAnnotation", (store, ann) => {
  // text in TextArea is always an array
  const obscene = ann.results.find(
  r => r.type === 'textarea' && r.value.text.some(t => t.includes('hate'))
  );
  if (!obscene || dismissed) return true;
  // select region to see textarea
  if (!obscene.area.classification) ann.selectArea(obscene.area);
  Htx.showModal("The word 'hate' is disallowed","error");
  dismissed = true;
  return false;
});

Related LSI instance methods:

Related frontend events:

Labeling config

<View>
  <Labels name="labels" toName="audio">
    <Label value="Speech" />
    <Label value="Noise" />
  </Labels>

  <Audio name="audio" value="$audio"/>

  <TextArea name="transcription" toName="audio"
    editable="true"
    perRegion="true"
    required="true"
  />
</View>

Related tags:

Data

[
  {
    "audio": "https://data.heartex.net/librispeech/dev-clean/3536/8226/3536-8226-0024.flac.wav"
  }
]

Bulk text labeling with regex

This script automatically applies the same label to all matching text spans. For example, if you apply the PER label to the text span Smith, this script will automatically find all instances of Smith in the text and apply the PER label to them.

Screenshot of bulk text labeling

Script

LSI.on('entityCreate', region => {
  if (window.BULK_REGIONS) return;
  window.BULK_REGIONS = true;

  const regionTextLength = region.text.length;
  const regex = new RegExp(region.text, "gi");
  const matches = Array.from(region.object._value.matchAll(regex));

  setTimeout(() => window.BULK_REGIONS = false, 1000);

  if (matches.length > 1) {
    const results = matches.reduce((acc, m) => {
      if (m.index !== region.startOffset) {
        acc.push({
          id: String(Htx.annotationStore.selected.results.length + acc.length + 1),
          from_name: region.labeling.from_name.name,
          to_name: region.object.name,
          type: "labels",
          value: {
            text: region.text,
            start: "/span[1]/text()[1]",
            startOffset: m.index,
            end: "/span[1]/text()[1]",
            endOffset: m.index + regionTextLength,
            labels: [...region.labeling.value.labels], 
          },
          origin: "manual",
        
        });
      }
      return acc;
    }, []);

    if (results.length > 0) {
      Htx.annotationStore.selected.deserializeResults(results);
      Htx.annotationStore.selected.updateObjects();
    }
  }
});

Related LSI instance methods:

Related frontend events:

Labeling config

<View>
  <Labels name="label" toName="text">
    <Label value="PER" background="red"/>
    <Label value="ORG" background="darkorange"/>
    <Label value="LOC" background="orange"/>
    <Label value="MISC" background="green"/>
  </Labels>

  <Text name="text" value="$text"/>
</View>

Related tags:

Bulk creation and deletion operations with keyboard shortcut

This script adds bulk operations for creating and deleting regions (annotations) based on the state of the Shift key:

  1. Shift Key Tracking
    • The script tracks the state of the Shift key using keydown and keyup event listeners. A boolean variable isShiftKeyPressed is set to true when the Shift key is pressed and false when it is released.
  2. Bulk Deletion of Regions
    • When a region (annotation) is deleted and the Shift key is pressed, the script identifies all regions with the same text and label as the deleted region.
    • It then deletes all these matching regions to facilitate bulk deletion.
  3. Bulk Creation of Regions
    • When a region is created and the Shift key is pressed, the script searches for all occurrences of the created region’s text within the document.
    • It creates new regions for each occurrence of the text, ensuring that no duplicate regions are created (i.e., regions with overlapping start and end offsets are avoided).
    • The script also prevents tagging of single characters to avoid unnecessary annotations.
  4. Debouncing Bulk Operations
    • To prevent rapid consecutive bulk operations, the script uses a debouncing mechanism with a timeout of 1 second. This ensures that bulk operations are not triggered too frequently.

Screenshot of bulk actions with keyboard shortcut

Script

// Track the state of the shift key
let isShiftKeyPressed = false;

window.addEventListener('keydown', (e) => {
  if (e.key === 'Shift') {
    isShiftKeyPressed = true;
  }
});

window.addEventListener('keyup', (e) => {
  if (e.key === 'Shift') {
    isShiftKeyPressed = false;
  }
});


LSI.on('entityDelete', region => {
  if (!isShiftKeyPressed) return; // Only proceed if the shift key is pressed

  if (window.BULK_REGIONS) return;
  window.BULK_REGIONS = true;
  setTimeout(() => window.BULK_REGIONS = false, 1000);

  const existingEntities = Htx.annotationStore.selected.regions;
  const regionsToDelete = existingEntities.filter(entity => {
    const deletedText = region.text.toLowerCase().replace("\\\\n", " ")
    const otherText = entity.text.toLowerCase().replace("\\\\n", " ")
    console.log(deletedText)
    console.log(otherText)
    return deletedText === otherText && region.labels[0] === entity.labels[0]
  });

  regionsToDelete.forEach(r => {
    Htx.annotationStore.selected.deleteRegion(r);
  });

  Htx.annotationStore.selected.updateObjects();
});


LSI.on('entityCreate', region => {
  if (!isShiftKeyPressed) return; 

  if (window.BULK_REGIONS) return;
  window.BULK_REGIONS = true;
  setTimeout(() => window.BULK_REGIONS = false, 1000);

  const existingEntities = Htx.annotationStore.selected.regions;

  setTimeout(() => {
    // Prevent tagging a single character
    if (region.text.length < 2) return;
    regexp = new RegExp(region.text.replace("\\\\n", "\\\\s+").replace(" ", "\\\\s+"), "gi")
    const matches = Array.from(region.object._value.matchAll(regexp));
    matches.forEach(m => {
      if (m.index === region.startOffset) return;

      const startOffset = m.index;
      const endOffset = m.index + region.text.length;

      // Check for existing entities with overlapping start and end offset
      let isDuplicate = false;
      for (const entity of existingEntities) {
        if (startOffset <= entity.globalOffsets.end && entity.globalOffsets.start <= endOffset) {
          isDuplicate = true;
          break;
        }
      }

      if (!isDuplicate) {
        Htx.annotationStore.selected.createResult({
            text: region.text,
            start: "/span[1]/text()[1]",
            startOffset: startOffset,
            end: "/span[1]/text()[1]",
            endOffset: endOffset
          }, {
            labels: [...region.labeling.value.labels]
          },
          region.labeling.from_name,
          region.object
        );
      }
    });

    Htx.annotationStore.selected.updateObjects();
  }, 100);
});

Related LSI instance methods:

Related frontend APIs:

Related frontend events:

Labeling config

<View>
  <Header>
    Labels
  </Header>
 <View style="padding: 0em 1em; background: #f1f1f1; margin-right: 1em; border-radius: 3px">
  <View style="position: sticky; top: 0; height: 50px; overflow: auto;">
    <Labels name="label" toName="text">
      <Label value="type_1" background="#3a1381"/>
      <Label value="type_2" background="#FFA39E"/>
      <Label value="type_3" background="#46ae19"/>
      <Label value="type_4" background="#8ab1c1"/>
     </Labels>
        </View>
  </View>
  <Header>
    Document
  </Header>
  <View style="height: 600px; overflow: auto;">
    <Text name="text" value="$text"/>
  </View>
</View>

Related tags:

Check that TextArea input is valid JSON

This script parses the contexts of a TextArea field to check for valid JSON. If the JSON is invalid, it shows an error and prevents the annotation from being saved.

This is an example of a “hard” block, meaning that the user must resolve the issue before they can proceed. For an example of a “soft” block, see Custom validation above.

Screenshot of JSON error message

Script

 LSI.on("beforeSaveAnnotation", (store, annotation) => {
  const textAreaResult = annotation.results.find(r => r.type === 'textarea' && r.from_name.name === 'answer');
  if (textAreaResult) {
    try {
      JSON.parse(textAreaResult.value.text[0]);
    } catch (e) {
      Htx.showModal("Invalid JSON format. Please correct the JSON and try again.", "error");
      return false;
    }
  }
  return true;
});

Related LSI instance methods:

Related frontend events:

Labeling config

<View>
  <View>
    <Filter toName="label_rectangles" minlength="0" name="filter"/>
    <RectangleLabels name="label_rectangles" toName="image" canRotate="false" smart="true">
      <Label value="table" background="Blue"/>
      <Label value="cell" background="Red"/>
      <Label value="column" background="Green"/>
      <Label value="row" background="Purple"/>
    </RectangleLabels>
  </View>
  <View>
    <Image name="image" value="$image" />
  </View>
  <View style=".htx-text { white-space: pre-wrap; }">
    <TextArea name="answer" toName="image"
              editable="true"
              perRegion="true"
              required="false"
              maxSubmissions="1"
              rows="10"
              placeholder="Parsed Row JSON"
              displayMode="tag"/>
  </View>
</View>

Related tags:

In this article

  1. Plotly
  2. Custom validation
  3. Bulk text labeling with regex
  4. Bulk creation and deletion operations with keyboard shortcut
  5. Check that TextArea input is valid JSON
Unlock more with Enterprise Get a Demo