Release 0.3.1

Changes:

• New Demo Models without GIT-LFS dependency (to maximize compatability)
• Updated Correction Modal
• Updated UI State Logic
• Multi-Config Experiments (i.e. changing UI configs over training)
• Logging of meta events such as submit/reset for reaction time measurements

Release 0.3.0 - Text Feedback, Experiment Intro, Study URLs & more

Many updates & Code Cleanup:

This update adds text feedback as a new modality, brings a major re-design including an updated intro modal, color scheme, controls, etc. Saved setup configurations can now be accessed via custom URLs making deployment for custom study setups easy and scalable.

• Text Feedback
• Updated Design
• Setup Saving & Loading with custom URL for deployment
• Massively simplified data generation
• Prototype interface for keyboard shortcuts
• Code restructuring in frontend (reworked state management, and increased modaluarity)
• Updated demo models, compatible with gymnasium & newest StableBaselines3
• New & improved intro modal
• Updated Docs

Next steps:

• Finishing keyboard shortcuts
• Finishing input for multi-modal and text tasks/scenarios
• Reward model implementations

For questions and bug-reports, feel free to reach out to @ymetz

Pre-Release 0.3.0

First pre-release:

• Experiments to collect and log feedback (https://rlhfblender.readthedocs.io/en/latest/guide/quickstart.html)
• Registration of gym environments (https://rlhfblender.readthedocs.io/en/latest/guide/add_new_experiment.html)
• Fully functional user interface

To-Do's until full release:

• Reward Modeling components need testing and verificaiton
• User Tracking with Motomo
• Tutorial/Jupyter Notebook to showcase analyis of logged feedback