@mihagazvoda @abhi1thakur @rushworth_a I think he's more talking about labeling a dataset for a proper use-case. The drawdata tool is more meant for educational purposes.
That said, I have done a lot of work on a suite of tools that totally help me label in my day-to-day work.
A thread!
That said, I have done a lot of work on a suite of tools that totally help me label in my day-to-day work.
A thread!
@mihagazvoda @abhi1thakur @rushworth_a For starters, if you'd like to label ... possibly the lowest barrier to entry method is to check out pigeon for classification tasks. I didn't make the tool, but it's a nice jupyter widget that's quick to get started with.
calmcode.io/pigeon/introdu…
calmcode.io/pigeon/introdu…
@mihagazvoda @abhi1thakur @rushworth_a But you typically need more. For example; are you 100% sure that your labels are correct? Checkout labelerrors.com, bad labels are totally a common thing.
To aid here, I've made a small library with many tricks to find bad examples.
github.com/koaning/doubtl…
To aid here, I've made a small library with many tricks to find bad examples.
github.com/koaning/doubtl…
@mihagazvoda @abhi1thakur @rushworth_a Other times you may have a huge unlabelled dataset. Maybe there's a few examples in that huge dataset that you'd to get more samples of. In that case ... simsity might help!
Tutorial here:
koaning.github.io/simsity/quicks…
Github here:
github.com/koaning/simsity
Tutorial here:
koaning.github.io/simsity/quicks…
Github here:
github.com/koaning/simsity
@mihagazvoda @abhi1thakur @rushworth_a Finally, there are some tricks you could do with human-learn that can help as well. Part of labeling is a UI problem at times.
github.com/koaning/human-…
github.com/koaning/human-…
@mihagazvoda @abhi1thakur @rushworth_a In particular, you could try to do some "bulk labelling".
It's not a perfect trick (at all), but it may help you figure out clusters relatively quickly. It's explained in more detail here:
It's not a perfect trick (at all), but it may help you figure out clusters relatively quickly. It's explained in more detail here:
@mihagazvoda @abhi1thakur @rushworth_a In general, if folks have never labelled themselves before, please do so! I have so many anekdotes on how it saved the day.
In particular, I might recommend everyone to give this gem a glance:
explosion.ai/blog/supervise…
In particular, I might recommend everyone to give this gem a glance:
explosion.ai/blog/supervise…
@mihagazvoda @abhi1thakur @rushworth_a With so many bad labels in benchmarks, it's getting too easy to be optimal on paper but broken in reality.
• • •
Missing some Tweet in this thread? You can try to
force a refresh
