Demo: Fine-tuning and post train chat eval POC

[nerdalert]

• Demonstrates an end to end knowledge submission, generate, train and a post-training side-by-side model chat comparison, enabling the user to validate their knowledge submission is included in the newly trained checkpoint.
• All aspects of the pipeline are done via the UI. No command-line operations are required at any stage.
• For this to function for the frontend to make REST API calls to Instructlab via an api-server that frontends ilab until REST endpoints are provided via InstructLab natively. The code is here https://github.com/nerdalert/ilab-api-server
• The demo was run on a 24GB GPU leveraging the simple pipeline. Will get an example accelerated/full pipeline demo with some hardware soon (which also produces a better tuned model).
• Integrates the new Patternfly Chatbot component. Note: the node that ilab is running inference needs to have ports open for connectivity since as I haven't gotten the chat streaming working with Next.js app router server side rendering with the chatbot component yet.
• Training and generation for the demo took around ~30-45m or so.
• All functionality is decoupled from the system via REST making it serviceable out of the gate and enabling the UI functionality.
• The knowledge submission was just a random new wiki that docling converted for the submission. It could be any topic of knowledge with accompanying documentation in the submission.

e2e-fine-tune-demo.mp4

[Misjohns]

@nerdalert Here are some initial comments from UX. Happy to create a new UX issue and create more mockups if you need additional design direction. JLMK

Knowledge wizard

• Update to the new InstructLab masthead and use figma design tokens for colors, etc.
• We should update the left navigation once we have the final direction and microcopy
• Update the wizard steps based on new MVP knowledge designs.
• Use steps from MVP knowledge designs to reduce the # of steps and group related fields
• No need to display the download/view function until the Review step. Will add a sample in MVP knowledge designs
• Assuming Auto-fill button is just a placeholder for the test NOT part of the final implementation
• We do not display 2 levels of buttons. The Next button swaps to Submit on the last wizard step.

Model chat evaluation

• Use the design from Red Hat Composer AI Studio demo
• UX will provide a design that shows the pattern for InstructLab
• UX can design a feedback mechanism (ie. Thumbs up/down) for each response.
• Does the user needs to 2 separate chats for the models? I thought the point was to compare responses.

Fine-tuning jobs (Empty state & Toolbar)

• This is a Getting started empty state and should only display when the Fine Tuning Job features haven’t been used yet.
• We should show a No Results empty state in cases where the system doesn’t find any data to show, such as when a user’s search criteria doesn’t yield any results.
• Let’s include a toolbar that allows the user to filter by status and job type, providing a count.

Fine-tuning jobs (Cards)

• Place labels above data instead of beside
• Using horizontal cards takes up a lot of space. Recommend removing the card frame and use a simple line separating the jobs or use the PF expandable table component
• This status takes up a lot of space. Recommend showing status inline

[Misjohns]

@williamcaban FYI

[Misjohns]

@andybraren @beaumorley Pls feel free to add anything you think I missed.