On the one hand, this is impressive. TabPFN was already state of the art and is seriously shaking up Bayesian prediction for tabular data (which is almost everything).
On the other hand, perhaps it is just me, but I do not feel that this is an acceptable form of benchmark reporting in this domain. TabArena actually has multiple metrics, since ELO does not properly quantify the degree of improvement. The fact that these are not displayed here should give pause. Also the results section in the GitHub is a dumpster fire.
You don't need to. Just sample 1% of data, let the model do the feature engineering, ask model model to replicate everything in Pandas (or R, or whatever else), so you can run it for a full dataset.
On the other hand, perhaps it is just me, but I do not feel that this is an acceptable form of benchmark reporting in this domain. TabArena actually has multiple metrics, since ELO does not properly quantify the degree of improvement. The fact that these are not displayed here should give pause. Also the results section in the GitHub is a dumpster fire.
Results folder: Here's some undocumented parquet files
Definitely feels like they're hiding the ball lol.
If they had good benchmarks they'd talk about them.
Not comparing to tuned xgboost is also a warning sign.
https://news.ycombinator.com/item?id=48689744