Nikhil Thorat issues

Results 53 issues of


Nikhil Thorat

Copy URL is broken when links end with a slash.

For example: https://lilacai-lilac.hf.space/datasets/#lilac/lmsys-chat-1m&viewPivot=true&pivot=%7B%22outerPath%22%3A%5B%22conversation__clusters%22%2C%22category_title%22%5D%2C%22innerPath%22%3A%5B%22conversation__clusters%22%2C%22cluster_title%22%5D%7D

Move the global project config into dataset-specific project configs.

Clustering input formats show up as <lambda> in the UI. We should show the human-readable string.

Now that we've propagated all the metadata for dataset-format specific inputs, we should show the string in the UI.

Clusters in the UI should show up next to the field it's enriching.

Currently they show up at the very end because of how we enrich the sibling field.

Add a back to clusters button from the group by mode.

Feature request: Copy link to selection

We have "Copy" and "Link to selection" in the right-click menu, let's add "Copy link to selection".

Single-item view with a single media field should be taller.

Currently one media field takes up half the page.. weird.

[UI] When searching a ShareGPT dataset, we should allow passing a filter of conversations.*.from= to the search.

It's currently weird that we search everything, when often users just want to search the human response. We can do this with dataset formats.

Environment flags that are boolean still are truthy when you set the string to "False"

From a user who set LILAC_AUTH_ENABLED to "False": ![image](https://github.com/lilacai/lilac/assets/1100749/cda3ede5-bc01-489b-8c61-08919bba3f3c)

Signals and embeddings are much slower

Text statistics on imdb takes 90+ seconds now, when it used to be much faster. gte-small used to be ~200ex/sec on my mac, now it's around 40.