Measure base vs prompt-only vs adapter quality

A plausible result is not yet evidence. Evaluate with format adherence and tone scores improve vs base without factual holdout regression. The test fixture should contain an easy positive case, an easy negative or baseline case, and the boundary case most likely to flip. Separate assertions about software contracts from claims about model quality.

Compare base, prompt-only, and adapter on format adherence and a factual holdout that should remain RAG-backed. Ship only if format lifts and facts do not regress beyond tolerance.

The artifact's user-facing goal is specific: stabilize response format and tone with LoRA while leaving facts to retrieval. Its accepted input is curated instruction examples split by intent into train, validation, and held-out test. Those statements are intentionally narrower than “build an AI system.” Narrow scope lets us inspect every input and expected result, and it prevents a toy result from being presented as a production claim. System shape for this chapter: a curated dataset is split by intent and source into train, validation, and untouched test sets; a pinned instruct model receives a low-rank adapter; evaluation compares base, prompt-only, and adapted outputs; serving loads the adapter as a versioned artifact beside retrieval. Keep model calls behind adapters, keep authorization and validation in deterministic code, and carry stable IDs and versions through every response. That separation lets you decide whether a bad result came from input handling, retrieval, inference, validation, or deployment. This page's job is the evaluation step: a plausible result is not yet evidence. Setup baseline for the chapter (run once per machine, not secrets in git):

python -m venv .venv && source .venv/bin/activate
# Training stack varies by GPU; pin versions in requirements.
pip install pyyaml
mkdir -p data/train data/val data/test adapters

If hardware or a hosted provider differs, preserve the interface and expected behavior. Do not present provider syntax as universal—when a vendor adapter is unavoidable, keep it behind a thin boundary and test with a fake first. The deliverable is not “it ran once”; it is a reproducible artifact another developer can inspect, including expected output and one deliberate failure related to adapter memorizes shifting facts or regresses general quality. Operationally, write down the owner of this stage, the command you ran, the observed output, and the next page's dependency on that output. If you cannot point to a file, fixture, metric, or config key, the stage is not done. Prefer small, reviewable increments: one contract, one path, one metric, one failure, one gate. When tradeoffs appear—latency versus quality, hit rate versus false hits, local privacy versus cloud quality—record both numbers instead of moving the threshold until the report looks green. The chapter ships only when evidence for format adherence and tone scores improve vs base without factual holdout regression and a rehearsed recovery path exist beside versioned LoRA adapter loaded at serve time with ADAPTER_ID rollback.

Save this as lesson.py and run python3 lesson.py. Prefer the standard library or the pinned packages from the setup block so the example stays reproducible.

scores={"base_format":0.62,"prompt_format":0.78,"adapter_format":0.91,"fact_holdout_base":0.88,"fact_holdout_adapter":0.87}
print(scores)

Expected output: scores showing format lift. Exact floating-point formatting may vary slightly, but the asserted behavior must not. Read the output as evidence about this stage, not merely proof that the interpreter started.

Print per-case expected versus observed values. Aggregates that look green can still hide a severe slice failure. At the evaluation stage, save the smallest failing fixture beside the expected result. Change one cause at a time and rerun the exact command printed above; that makes the repair reviewable and keeps this chapter's progressive artifact reproducible.

Require the stated thresholds for format adherence and tone scores improve vs base without factual holdout regression. Do not delete hard cases to green the report. For this evaluation page, preserve the fixture and result as evidence for the next page. Label observations separately from conclusions: a passing assertion establishes the behavior it names, while broader usefulness requires the chapter's full evaluation set and stated operating limits. Primary metrics for the chapter remain format adherence and tone scores improve vs base without factual holdout regression.