let's add a Plot.ruleY to what the humans have on average (info available in "entities--question" table)

a label mark with model names would be helpful

i added tickFormat: configToName in fx

when you have a situation like tickFormat: (d) => configToName(d), you don't have to spell it out fully like this in JS (and Python too, i hear) you can pass functions as regular variables, so instead of passing an anonymous function ()=>{}, inside of which you are calling configToName(), you can directly pass configToName and it will get called in place of your ()=>{}, hence tickFormat: configToName is sufficient this is googlable as "functions as first class citizens" functions can also get returned from functions, it's turtles all the way down. understanding how you can pass functions around and where and when they get called with which available context gives you a good foothold in JS and by extension web development

i have made the bottom chart show all models when nothing is selected above (by adding a condition in filter encoding !comparison.length || ) look at R1, both orange and blue! how do you interpret this?

basically it depends on how correct a model answers prompt variations labeled as worke or based compared to how it answers on average? regardless of the question nature, right? so it's for all questions

chatgpt appears in both sides, i wonder what makes 4o based, some model configs more than other, is there info about this in the dataset?

we have now added columns propmpt_family and prompt_specific for your convenience