Training
Epoch
What is epoch?
An epoch is one complete pass through the training dataset. Fine-tuning typically uses between one and three; more than that on a small dataset usually produces memorisation rather than learning.
The instinct to train for many epochs comes from training models from scratch, where dozens are normal. Fine-tuning starts from a model that already works, so the amount of change required is small and the number of passes needed is correspondingly small.
If more capacity to learn is genuinely needed, more distinct data is almost always a better answer than more passes over the same data. The second pass teaches the model the examples; the fifth teaches it the specific wording.
Related terms
Start with the free tier
A magic link creates your account, your tenant and your first API key. No card until you ask for compute.