epic: Improved Hub Experience for Models

[dan-menlo]

Motivation

To address user feedback and enhance the Hub experience by improving model recommendations, refining filtering options, and providing clearer model version visibility.

Specs

Major user stories:

1. Recommended Models by Hardware:

   • As a user, I want to receive model recommendations tailored to my hardware specifications for optimal performance.

2. Filtering Models by Attributes:

   • As a user, I need to filter models based on tags, supported tools, maximum model size, parameters, and computer RAM compatibility.

3. Filtering Downloaded Models:

   • As a user, I want the ability to filter out models that I have already downloaded for more streamlined browsing.

4. Hardware-Based Recommended Starter List:

   • As a user, I expect a recommended starter list of models aligned with my hardware capabilities to kickstart my exploration.

5. Efficient Model Search:

   • As a user, I should be able to search for models by their name, ID, author, or URL, ensuring ease of discovery.

6. Model Grouping with Version Information:

   • As a user, I want models and their different versions displayed within the same group, simplifying model version comparisons and selections.

Design

https://www.figma.com/file/ytn1nRZ17FUmJHTlhmZB9f/Jan-App?type=design&node-id=1699-105656&mode=design&t=78l4n8TLR3sJeIAn-4

Tasklist

Design: @imtuyethan

• [x] #1008
• [x] #1205

Development:

• [ ] @urmauur FE #1386
• [ ] @louis-jan @namchuai BE #1387
• [ ] @hahuyhoang411 Metadata #1388

[hahuyhoang411]

Got the point. I think we should remove the "recommended" tag for >33B models.

[hahuyhoang411]

@imtuyethan This issue address for the category on the Hub too.

Related task: https://github.com/janhq/jan/issues/839

[hahuyhoang411]

Realated task: Featured model https://github.com/janhq/jan/issues/860

[dan-menlo]

I also think we should have nicer cover images for Models, e.g.

Mistral: Paris, "Malayalam MN"

OpenHermes Neuralchat 7b "Major Mono Display"

[hahuyhoang411]

Related: https://github.com/janhq/jan/issues/897

[hantran-co]

HUB REVAMP

Figma link: https://www.figma.com/file/ytn1nRZ17FUmJHTlhmZB9f/Jan-App?type=design&node-id=479-21691&mode=design&t=uV0D1oQLPoF6kzy2-4

Changes that address all the feedbacks & comments:

• 2 features model to get started
• List all models with sticky navs, and allow users to sort by "All models, Recommended, Downloaded" as well as quick selection of model type and parameters on the left side panel
• Changes in model card based on the current data we have right now & shortcomings:

Preview:

[dan-menlo]

Feedback from user: https://www.reddit.com/r/LocalLLaMA/comments/18pyul4/comment/ketw7vp/?utm_source=share&utm_medium=web2x&context=3

[hantran-co]

Archive Daniel's original comments - @imtuyethan to work on the latest specs:

Kanban

https://github.com/orgs/janhq/projects/5/views/7?filterQuery=milestone%3A%22Improved+Hub+Experience+for+Models%22+

Problem

• Hub is currently a poorly-ordered list of models that user has to go through
• We should have a better way of structuring it, such that it feels more "appstore"-like, and less initimidating to a new user

Ideas

• Hub has "Best-of" lists, e.g. LLM Leaderboard
• Sorting
• Starter Models ("get started"), e.g. NeuralChat 7b models that are a fast download
• Recommender Algorithm?
• Content should be centered in screen (vs. stretched too wide)
• Animations when scrolling/loading

As an aside, I find that it's difficult to download the larger models due to Huggingface server timing out. This should be a consideration in recommending models. It's better to get user to download a fast, quick model and get them using it quickly.

Tasklist

• https://github.com/janhq/jan/issues/848
• https://github.com/janhq/jan/issues/835
• #871
• #777
• #786
• #936
• #1008
• #1205
• #1008
• #1384
• [ ] Recommended Models should pull from remote, auto-update when Jan is launched
• @tikikun's comment: https://github.com/janhq/jan/pull/829#issuecomment-1838035448

[freelerobot]

@imtuyethan gentle reminder to close Github issues because they remain open even if you move them to done