ADP — Model Adaptation Laboratory
An interactive, deterministic comparison of Full fine-tuning, LoRA and
QLoRA. Change the method, the data and the memory budget, and watch
parameter counts, estimated memory and synthetic training and
evaluation results change with them. Everything runs in your browser
as a teaching model.
Scope: no real model is trained or deployed.
Parameter counts are calculated, memory is estimated and learning
results are synthetic. Source verification does not verify simulator
numbers.
What it covers
-
Parameter counts and estimated training memory for Full FT, LoRA and
QLoRA.
-
Five authored experiments: domain QA, data quality, overfitting,
memory constraint and a rank explorer, plus a custom experiment.
-
Adaptation 101, the concepts behind the method choices, and the
assumptions behind every number.