Data
Truncation
What is truncation?
Truncation is the silent cutting of an example that exceeds the configured sequence length. When it removes part of the target response, the model learns to produce answers that stop abruptly — a failure with no error message anywhere in the run.
It is worth measuring explicitly before training: the proportion of examples affected, and whether the cut falls in the prompt or the response. Truncating the start of a long prompt is often harmless; truncating the end of a target is not.
The fix is either a longer sequence length, at a memory cost, or shorter examples. Choosing requires knowing the distribution, which is why validation reports the token length percentiles rather than just the mean.
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.