Skip to content

Upgrade 2.0 - hotloss and hotopt added#1

Merged
sidphbot merged 7 commits intomasterfrom
upgrade_2.0
Mar 11, 2026
Merged

Upgrade 2.0 - hotloss and hotopt added#1
sidphbot merged 7 commits intomasterfrom
upgrade_2.0

Conversation

@sidphbot
Copy link
Copy Markdown
Owner

@sidphbot sidphbot commented Mar 1, 2026

No description provided.

sidphbot added 7 commits March 1, 2026 21:37
works mostly and amazing results so far

current issues:
View in recipe editor open the editor/recipe-tab but doesn't load up the recipe(saved now), also there should be an import option there in recipe tab.
autopilot seems to be on by default why is that? are we using any non-deterministic logic there, no api calls I hope? the mutation log should be sacred to the run and should report any un-authorised change promptly in it, ideally any cli hotcb command based mutations too
the manifold is nice to see, but why only two color gradient? and lets have some info, what manifold is it? training? features? outputs? how is it compressed/dimensionally-reduced?
also this training configuration creation can it be better on an assisted(formatting placeholders config instructions, callback/hotcallback files/recipes/) session like on the app a different page/tab.
the saved entries tab stays after even reset should fade away immediately
can we include some live metrics in training health with a lambda, optionally?
also when frozen once, there should be a lock and shouldnt be allowed to unfreeze, same happens when you start training with replay mode selected.
the panes themselves are still overflowing without sliders, they need to expand as per content uptill a limit and then start being scrollable
page refresh loses complete memory of panels, thats not good combined with lack of experiment tracking worsens it a bit

how to compare now? i know that is a whole different feature(probably with experiment hotcb logdir path hook).
next pending:
the mutations on graph can come a down and have linebreaks to avoid overlap.
the right pane cards are still squeezing, they need to expand to a point and then be scrollable
2 questions regarding the new config setup,
- the training function args, can we make steps/step_delay etc inferred from metric entry timestamps?
- what is the stop_event? what does the user need to import/implement to do that call?

the compare feature is great but does need some more love,
- the main graph pane should have a zoom(hide other panes with run info overlayed translucent - or something like the metrics graph pane, that is still properly visible, although both need some control and filtering options)
- there should be all metrics available for comparison(like in metrics graph pane again with toggles/pins) not just hardcoded train_loss
- the mutation events are not distinguished from run to run anywhere, log or curve
- goal is to compare two 'recipes' and their independent/cumulative trend benefits/losses on same training scope on a filterable metric space where user can zoom in on certain metrics and the mutation effect on the two runs from that perspective or in combination with some others
lets remove the replay mode and introduce a toggle in training pane asking if recipe present with config is to be loaded or not.
the projection system is great mostly but a lot of times just hooks on to some weird coordinate, related to last run i guess
the pinned tabs when zoomed sometimes stick to zoomed state across resets, although the pinned tabs do disappear correctly, so there is no way left to zoom out and see all metrics as usual
and when i start the demo, please don't start any run already
next issues:
the json string for mutations is gone should come back as tiny capsule/tags( I guess this might be my request to make those lists concise, I see the same effect on the dotted line tag as well - my suggestion, lets leave the curve dotted line tags concise as they are, only expand the mutation table and comparison descriptions
the controls and training window are still ot reflecting what is there for this run itsall set of default golden-demo/multitask and change only with changing the training config to 1 of those 3 demos. both still don't show the current run. related is recipe, should also show currently used mutations
mutations applied sh
the ai keeps sleeping exponentially longer and overestimating the step lengths

minor: make the hover tool tip thing a bit more reactive, they are points you need to search for. also the training health pane, just make it collapsible with details on card expand/un-collapse.
@sidphbot sidphbot merged commit d10cfd5 into master Mar 11, 2026
0 of 6 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant