v0.12.7 (beta)

Beta, everything from v0.12.6 plus:

Changed

• Retrained prediction model, with bread. Trained on 253 cooks (36 more shared cooks than before, including the first five bread bakes). Bread now has its own category, cut type and baked doneness in the model instead of being treated as a generic cook, so a bread estimate is fitted to how a loaf actually bakes. On the held-out bread bakes the average error fell from about 10 to 5 minutes; for other cooks the model was not trained on, accuracy is about the same as before (median about 10 minutes).
• The README's accuracy section is reworded: the replay figures are measured on cooks that are mostly in the training data, so they are a best case. It now also states the result on cooks the model had not seen.

Notes

• Five bread bakes (four brioche, one sourdough) is a small sample: yeasted, rye and with-inclusions loaves share one bread code until there is enough data to tell them apart. Please keep sharing your cooks.