Releases: label-sleuth/label-sleuth
Releases · label-sleuth/label-sleuth
v0.9.2
v0.9.1
Features
- Support training/inference with Apple M-series chips
- Improve transformers inference time by sorting examples
- Check for model-language compatibility
- Add a multilingual XLM-R transformer model that supports 100 languages
- Display weak labels for the current category
Bug Fixes
- Fix issues related to iteration flow following #373
- Fix scrolling issues In Safari
- Improve backdrop logic to prevent it from blinking
- Prevent the right panel from jumping back to the first page
Development
- Continue migration of the frontend to Typescript
v0.9.0
Features
- Extended language support: enable support for languages that rely on FastText representations; add support for Hebrew.
- Model support: add Naive Bayes model policies; facilitate adding new transformer-based classifiers.
- Performance improvements for the label import functionality.
- Collecting the set of examples for training now happens in a background thread.
- It is now possible to define active learning policies, i.e. a predefined policy for dynamically determining which active learning strategy should be used.
Development
- Initiated migration of the frontend to Typescript (#205)
v0.8.6
What's Changed
- Sidebar elements are no longer scrolled into view when clicked.
- The system version is displayed on the bottom-left corner of the menu page.
Bug fixes
- Dashes are replaced in imported document ids.
- Invalid regex in search no longer makes the UI to crash.
v0.8.5
v0.8.4
Features
- Add ability to override the configuration file using a command line argument with the relevant parameter name (e.g., changing language using
--language ITALIANO
) - Add Romanian support
Bug fixes
- Return only standard labels in the document view (without weak labels)
- Determine label counts according to the label types used in
training_set_selector
v0.8.3
v0.8.2
Features
- Support import of weakly labeled data using
Weak
inlabel_type
column - Add training set selectors which use weakly labeled data
UI changes
- Avoid scrolling into an element when labeling it
- Improve notification of internal server errors (500 errors)
Bug fixes
- Fix get model predictions endpoint
- Fix element ids not properly converted into document ids when sending to the frontend
v0.8.1
UI changes
- Unfocus text input when search is done.
- Add icons for accessing Github, Slack and Webpage.
- Decrease version font size.
- Expose shortcuts information so users are aware of them.
Performance improvements
- Positive predictions are loaded in batches instead of the whole corpus's positive predictions at once.
- Dataset is preloaded when the user enters a workspace.
Bug fixes
- Category name's whitespaces are replaced with underscore when importing labels into the workspace.
- Labeling a main element scrolls to top of the doc.
- Unexpected behavior due to shared state across tabs.
v0.8.0
- Improved performance in backend and UI
- Display the system version in the UI
- Added keyboard shortcuts for labeling using the arrow keys
- Update requirements.txt to support newer Mac hardware
- Fixed a major bug in BERT model which caused all the predictions to be positive
- Various bug fixes