My summary of lessons learned applying data science.
Pinned Tweet
Follow
Click to Follow fishnets88
Vincent D. Warmerdam
@fishnets88
Prefer common sense over hype. Employed , builds calmcode.io, dearme.email, blog koaning.io. he/him
@koaning@fosstodon.org
Vincent D. Warmerdam’s Tweets
Nice tutorial on the coref in by the awesome ! Happy to hear you've found it easier to use and more accurate than other systems ☺️! Best of luck with your project finding out about women who worked at the Smithsonian!
5
10
This might deserve retweeting.
Quote Tweet
Replying to @fishnets88 and @TonyPlaysGuitar
But it also seems that running it multiple times gives different output too. Entropy plays a role.
1
One trick that sometimes seems to work is to add examples to the prompt of behavior. This seems to work "ok" if you have a set of examples that OpenAI got wrong before.
But this moves the problem into a new one; which examples are good/bad for the prompt?
1
5
Show this thread
Words that seem really minor to me can totally tick the output the other way.
I might be comparing it too much to a bag of words model, which isn't fair. But I really would like to understand/debug my prompt better here.
2
3
Show this thread
I'm exploring prompt engineering a bit in OpenAI, and it turns out to be tricky!
The difference between these two input examples is "television ad" vs "television appearance".
Just to ping folks; got the same experience? I'm trying to understand this better.
7
2
16
Show this thread
NEW RELEASE: Introducing spaCy v3.5!
Since v3.4, we've added:
⌨️ Three new CLI commands
💥 Fuzzy matching
🚀 Improvements to our entity linking functionality
💫 A range of language updates and bug fixes!
spacy.io/usage/v3-5
read image description
ALT
3
29
116
TIL that there was a pandemic on World of Warcraft in 2005.
It was called the "corrupted blood incident" and it was caused by a disease that spread from a raid boss, to players, to their pets, to capital city NPCs and from there to low level players.
koaning.io/til/social-dis
1
18
Announcing: Orc of the Office.
A totally non-official set of MTG cards. Made for laughs, the tech enterprise, and for the horde!
koaning.io/posts/orcs-of-
1
9
37
Work in progress.
4
3
39
Note that folks are also more than welcome to ping , she's collaborating with me on this one.
1
2
Show this thread
But maybe that A/B compared dataset can be super useful for something else! Varification!
Still a work in progress, but I'm pretty interested in how it'll turn out.
If folks did something similar, please ping!
1
1
Show this thread
One idea is to train two models. A classifier and a sorter.
But technically, if I have a classifier that can predict a score between 0 and 1 then I can already use that to sort ....
1
1
Show this thread
But I also made a much simpler one that just lets me deem an example as "would click".
This feels a lot simpler to reason about, but I'm wondering how I might be able to combine annotations from both.
1
Show this thread
Working on a system that can generate my own front page of the internet.
So I'm toying around with annotation interfaces. The first one allows me to compare two options.
2
3
13
Show this thread
Zugzug.
1
7
Whenever people tell me that "AI is Magic", my mind immediately creates another one of these.
3
2
23
She's amazing and she knows it.
1
21
It is finally here, the v0.13 release of BERTopic! 🎉
Explore multi-topic assignments, supervised topic modeling, outlier reduction, light- and heavyweight options, and much more in this bigger-than-expected release!
Changelog:
maartengr.github.io/BERTopic/chang
An overview thread👇🧵
5
54
276
Show this thread
Quote Tweet
Announcing finetuners for embetter!
There are scikit-learn components that can update the embedded space to make bulk labelling easier!
I wrote a full tutorial here: koaning.github.io/embetter/finet
Show this thread
10
As always, I try to stick to the scikit-learn API. So you can just go and pipeline away!
1
4
Show this thread
Here's a before and after on the PCA space of a sentence-transformer. The color is the class in the test set.
After finetuning, the space is much more "polarised" but also makes it easier to select candidates of interest.
1
3
Show this thread
The trick is to re-use a neural network but as a sklearn transformer. We use the `y`-label, but only to use the gradient signal to update the representation.
1
1
5
Show this thread
Announcing finetuners for embetter!
There are scikit-learn components that can update the embedded space to make bulk labelling easier!
I wrote a full tutorial here: koaning.github.io/embetter/finet
2
23
78
Show this thread
is there an app to turn blogposts into audio/podcast segments?
i recall folks talking about it a while ago and id really like to add more blogs into my morning wall with the stroller.
3
2
10
This is a nice example! Thanks for sharing.
Quote Tweet
@fishnets88 @jeremyjordan this was just posted on Medium (human-learn and a link to jeremy's normconf video :) towardsdatascience.com/human-learn-ru
4
The talks are live!
Mine is about "groupby statements that save the day".
Turns out, a group-by can be really help prevent a bad ML situation.
2
42
The past year has been filled with new videos! has released a ton of content around and prodi.gy, including two "shorts" series 🎉
has also made his debut, releasing videos on the spaCy components spancat and coref!
1
2
20
Show this thread
It's true. The box of stickers is a standard in my backpack these days.
Quote Tweet
@fishnets88 attended a few conferences talking about techniques and tools for data and machine learning. He's always handing out stickers!
@DamianRomero_CL and Magda also discussed data annotation at several events, and @victorialslocum presented on spancat at PyData Global!
Show this thread
2
It could be a whole series!
1
3
15
Show this thread
Should I start a collectible card series?
8
20
171
Show this thread
One of the cooler parts of the paper is rarely the appendix, but in this case it’s different …
Quote Tweet
Here's the link again for the paper: arxiv.org/abs/2212.09255
The hyperparameters are in the appendix as part of the config file, allowing easy reproducibility. We will also release a spaCy project for this very soon! (4/4)
Show this thread
1
6
😅
Quote Tweet
@fishnets88’s recent @normconf talk came to mind
“you can’t get certificate in common sense and critical reasoning” twitter.com/StephaneMaarek…
🆕 Typeset plugin ✏️
Insiders 4.27.0 ships a new built-in plugin, which allows to preserve formatting in parts of the site that MkDocs normally doesn't 😏 This has been requested soooo many times and finally, it's possible!
Documentation:
squidfunk.github.io/mkdocs-materia
2
3
15
Show this thread
It’s true. When you only have one hand at your disposal, this feature is a life saver!
Quote Tweet
Did you know that you can use Prodigy from your mobile phone?
In our latest Prodigy Short, @fishnets88 (who has a newborn) shows how to set up Prodigy so that you can use swipe gestures and need only one hand to annotate 
Watch it here: youtu.be/-Bx-3DaE64A
Watch it here: youtu.be/-Bx-3DaE64A
read image description
ALT
2
So instead of starting big ... maybe it's better to start small. That way, you're more nible and you can iterate when you spot an issue.
Iteration really is a beautiful thing.
For details, check the blogpost!
koaning.io/posts/go-emoti
3
Show this thread
That means that you shouldn't be surprised if these annotators end up not agreeing with each other. Especially when you're giving them a subjective task like emotion detection.
Imagine training an algorithm on a dataset full of disagreement.
read image description
ALT
1
3
Show this thread