Evaluate fine-tuned adapters and freeze a model selection
Compare Unsloth training budgets against the base model, inspect paired regressions, bind predictions to artifacts and keep test data outside selection.
Compare the base model and candidate adapters on validation data, choose one using a declared rule, then freeze its identity before evaluating a fresh held-out test set. Inspect baseline answers that regressed, even when total accuracy improves. This project checks prediction and artifact identity with CPU fixtures. Model training and actual prediction quality remain unmeasured; a previously inspected test set is development evidence.
A frozen validation choice, recorded prediction identities and a separate held-out test report.
Environment
CPU evaluation · existing Unsloth NVIDIA environment for model runs
Plan your session
2–3 hours plus two training and prediction runs. A planning estimate, not measured runtime.
Before you begin
Preserve the grouped prepared/ dataset and its manifest from the data chapter. Use the pinned Unsloth project scripts and environment; the evaluator tests themselves need only Python 3.12.
CPU selection checks executed · GPU run pending. Selection, regression rejection, prediction identity, group overlap and changed-artifact tests executed with synthetic fixtures. Prediction wrapper exercised with a CPU stand-in. Real CUDA training, adapter predictions and quality improvement remain unmeasured.
01Compare on validation
Score the base and declared candidates on the same examples.
↓
02Freeze the chosen artifact
Save the rule, dataset identity, prediction hashes and adapter bytes.
↓
03Report held-out results
Check the frozen choice on test data without selecting another candidate.
Follow the changing state through the project. Each boundary has a check in the walkthrough.
Continue with Premium
Premium includes this guide’s complete walkthrough and downloadable project files. The technical overview, prerequisites and verification notes above are public.