Introducing TwelveGPT Scout
The last few months have been working on TwelveGPT: a large langusge model for football. And now we are ready with the first output from this work: the TwelveGPT Scout.
The starting point is producing succinct player reports, explaining what they are good (and less good at). Here is an example, from one of the queires we got when we opened the request-line on Twitter.
For those of us who are used to looking at radars (on the left) or distribution plots (on the right, which we use quite a bit at Twelve) the text should make sense: Andre is rated high in defence and involvement. What TwelveGPT adds is a simple, but engaging text, to explain what we are looking at.
It does more than just this. Our TwelveGPT Scout app allows the user to drill down and make detailed enquiries about how a metric like ‘Box Threat’ is constructed. This is essential if we want to trust the metrics we use to build up our understanding of a player.
Doing a live request-line allowed us to check how well TwelveGPT Scout did on its own: we didn’t have a chance to edit the texts before posting them. And in the case above it did get the player’s first name wrong: it is actually Tashreeq Matthews, Sirius player, and not Tommi. But other than that the bot did very well, accurately describing the skillsets of some very different players.
Check out the request answers to see for yourself.
A #TwelveGPTScout radar of two halves here! https://t.co/0YN3OkHh4U pic.twitter.com/8IOQIET91b
— Twelve football (@twelve_football) October 30, 2023
On Friday the 3rd of November at 7pm UK time/8pm in Europe we did a live stream demo on Youtube.
If you are working in football (for a club, national team or as a journalist) and would like to sign-up for a trial then click below.
