This is a patch release with some helpful features and bug fixes including:
- training with chat templates: automatically apply chat templates when training on datasets of conversations
- VLM training: use combined image and text datasets on LoRA finetunes
- A number of UI improvements to Training, Generate and Evaluation modals
- fix for docker image so llama.cpp plugin works properly
- plus many more stability and UI fixes
Detailed Log
- Make export page look exactly like evals and generate by @deep1401 in #649
- Fix Close button on Workflow Runs Output Modal by @aahaanmaini in #636
- Fixed Workflow runs Training Jobs Streaming by @aahaanmaini in #634
- Add chat column into config by @mina-parham in #643
- Fix ModelProviderWidget to have an option to also specify loader plugin by @deep1401 in #652
- Fix editing task names in Train and Evaluate by @deep1401 in #647
- Include evaluation steps in GEval Widget by @deep1401 in #658
- Add Progress Bar to Workflow Runs by @aahaanmaini in #656
- Make Workflow JSON Editable in UI by @aahaanmaini in #666
- Use the job streaming route for train outputs too by @deep1401 in #668
- Fix recipe install logic to improve performance for plugins and dataset install by @deep1401 in #665
- Fix selection of tasks with duplicate names and set default mode as TRAIN by @deep1401 in #663
- Fix system message reset by @mina-parham in #638
- Fix blank system message when starting a model by @deep1401 in #669
- Add/activate-recipes-for-everyone by @aliasaria in #601
- Remove Training Recipes from the App by @aliasaria in #660
- Fix Workflow JSON in dark mode by @aahaanmaini in #671
- Workflows can only add tasks within experiment by @aahaanmaini in #648
- Show Apply Chat Template toggle only for specific plugin types by @mina-parham in #657
- Fix Editing, Saving on TRAIN, EVAL and GENERATE by @deep1401 in #674
Full Changelog: v0.20.1...v0.20.2