Annotate your dataset¶
To experience the UI features firsthand, you can take a look at the Demo ↗.
Argilla UI offers many functions to help you manage your annotation workflow, aiming to provide the most flexible approach to fit the wide variety of use cases handled by the community.
Annotation interface overview¶
Flexible layout¶
The UI is responsive with two columns for larger devices and one column for smaller devices. This enables you to annotate data using your mobile phone for simple datasets (i.e., not very long text and 1-2 questions) or resize your screen to get a more compact UI.
At the right side of the navigation breadcrumb, you can customize the dataset settings and edit your profile.
This area displays the control panel on the top. The control panel is used for performing keyword-based search, applying filters, and sorting the results.
Below the control panel, the record card(s) are displayed one by one (Focus view) or in a vertical list (Bulk view).
This is where you annotate your dataset. Simply fill it out as a form, then choose to Submit, Save as Draft, or Discard.
This expandable area displays the annotation guidelines. The annotation guidelines can be edited by owner and admin roles in the dataset settings.
This expandable area displays your annotation progress.
Shortcuts¶
The Argilla UI includes a range of shortcuts. For the main actions (submit, discard, save as draft and selecting labels) the keys are showed in the corresponding button.
To learn how to move from one question to another or between records using the keyboard, take a look at the table below.
Shortcuts provide a smoother annotation experience, especially with datasets using a single question (Label, MultiLabel, Rating, or Ranking).
Available shortcuts
| Action | Keys |
|---|---|
| Activate form | ⇥ Tab |
| Move between questions | ↓ Down arrow or ↑ Up arrow |
| Select and unselect label | 1, 2, 3 |
| Move between labels or ranking options | ⇥ Tab or ⇧ Shift ⇥ Tab |
| Select rating and rank | 1, 2, 3 |
| Fit span to character selection | Hold ⇧ Shift |
| Activate text area | ⇧ Shift ↵ Enter |
| Exit text area | Esc |
| Discard | ⌫ Backspace |
| Save draft (Mac os) | ⌘ Cmd S |
| Save draft (Other) | Ctrl S |
| Submit | ↵ Enter |
| Move between pages | → Right arrow or ← Left arrow |
View by status¶
The view selector is set by default on Pending.
If you are starting an annotation effort, all the records are initially kept in the Pending view. Once you start annotating, the records will move to the other queues: Draft, Submitted, Discarded.
- Pending: The records without a response.
- Draft: The records with partial responses. They can be submitted or discarded later. You can’t move them back to the pending queue.
- Discarded: The records may or may not have responses. They can be edited but you can’t move them back to the pending queue.
- Submitted: The records have been fully annotated and have already been submitted.
Suggestions¶
If your dataset includes model predictions, you will see them represented by a sparkle icon ✨ in the label or value button. We call them “Suggestions” and they appear in the form as pre-filled responses. If confidence scores have been included by the dataset admin, they will be shown alongside with the label. Additionally, admins can choose to always show suggested labels at the beginning of the list. This can be configured from the dataset settings.
If you agree with the suggestions, you just need to click on the Submit button, and they will be considered as your response. If the suggestion is incorrect, you can modify it and submit your final response.
Focus view¶
This is the default view to annotate your dataset linearly, displaying one record after another.
Tip
You should use this view if you have a large number of required questions or need a strong focus on the record content to be labelled. This is also the recommended view for annotating a dataset sample to avoid potential biases introduced by using filters, search, sorting and bulk labelling.
Once you submit your first response, the next record will appear automatically. To see again your submitted response, just click on Prev.
Navigating through the records
To navigate through the records, you can use the Prev, shown as <, and Next, > buttons on top of the record card.
Each time the page is fully refreshed, the records with modified statuses (Pending to Discarded, Pending to Save as Draft, Pending to Submitted) are sent to the corresponding queue. The control panel displays the status selector, which is set to Pending by default.
Bulk view¶
The bulk view is designed to speed up the annotation and get a quick overview of the whole dataset.
The bulk view displays the records in a vertical list. Once this view is active, some functions from the control panel will activate to optimize the view. You can define the number of records to display by page between 10, 25, 50, 100 and whether records are shown with a fixed (Collapse records) or their natural height (Expand records).
Tip
You should use this to quickly explore a dataset. This view is also recommended if you have a good understanding of the domain and want to apply your knowledge based on things like similarity and keyword search, filters, and suggestion score thresholds. For a datasets with a large number of required questions or very long fields, the focus view would be more suitable.
With multiple questions, think about using the bulk view to annotate massively one question. Then, you can complete the annotation per record from the draft queue.
Note
Please note that suggestions are not shown in bulk view (except for Spans) and that you will need to save as a draft when you are not providing responses to all required questions.
Annotation progress¶
The global progress of the annotation task from all users is displayed in the dataset list. This is indicated in the Global progress column, which shows the number of records still to be annotated, along with a progress bar. The progress bar displays the percentage and number of records submitted, conflicting (i.e., those with both submitted and discarded responses), discarded and pending by hovering your mouse over it.
You can track your annotation progress in real time from the righ-bottom panel inside the dataset page. This means that, while you are annotating, the progress bar updates as you submit or discard a record. Expanding the panel, the distribution of Pending, Draft, Submitted and Discarded responses is displayed in a donut chart.
Use search, filters, and sort¶
The UI offers various features designed for data exploration and understanding. Combining these features with bulk labelling can save you and your team hours of time.
Tip
You should use this when you are familiar with your data and have large volumes to annotate based on verified beliefs and experience.
Search¶
From the control panel at the top of the left pane, you can search by keyword across the entire dataset. If you have more than one field in your records, you may specify if the search is to be performed “All” fields or on a specific one. Matched results are highlighted in color.
Order by record semantic similarity¶
You can retrieve records based on their similarity to another record if vectors have been added to the dataset.
To use the search by semantic similarity function, click on Find similar within the record you wish to use as a reference. If multiple vectors are available, select the desired vector. You can also choose whether to retrieve the most or least similar records.
The retrieved records are then ordered by similarity, with the similarity score displayed on each record card.
While the semantic search is active, you can update the selected vector or adjust the order of similarity, and specify the number of desired results.
To cancel the search, click on the cross icon next to the reference record.
Filter and sort by metadata, responses, and suggestions¶
Filter¶
If the dataset contains metadata, responses and suggestions, click on Filter in the control panel to display the available filters. You can select multiple filters and combine them.
Note
Record info including metadata is visible from the ellipsis menu in the record card.
From the Metadata dropdown, type and select the property. You can set a range for integer and float properties, and select specific values for term metadata.
Note
Note that if a metadata property was set to visible_for_annotators=False this metadata property will only appear in the metadata filter for users with the admin or owner role.
From the Responses dropdown, type and select the question. You can set a range for rating questions and select specific values for label, multi-label, and span questions.
Note
The text and ranking questions are not available for filtering.
From the Suggestions dropdown, filter the suggestions by Suggestion values, Score , or Agent.
Sort¶
You can sort your records according to one or several attributes.
The insertion time and last update are general to all records.
The suggestion scores, response, and suggestion values for rating questions and metadata properties are available only when they were provided.
Annotate in teams¶
Note
Argilla 2.1 will come with automatic task distribution, which will allow you to distribute the work across several users more efficiently.
Edit guidelines in the settings¶
As an owner or admin, you can edit the guidelines as much as you need from the icon settings on the header. Markdown format is enabled.
Tip
If you want further guidance on good practices for guidelines during the project development, check this blog post.