Instruction-Following Evaluation
Measure whether the model obeys your stated constraints — format, length, vocabulary, style, content limits — separately from domain skill or prose quality, and build custom suites for the instructions your product actually uses.
The instruction-following bucket of
application evaluation criteria asks:
given your instructions, does the model comply — independent of whether
those instructions are good? A model can have domain skill (sentiment
analysis) and still fail the product contract by emitting HAPPY/ANGRY
instead of the required NEGATIVE/POSITIVE/NEUTRAL labels. Stronger
models generally follow instructions better, but that is not a substitute
for measuring the constraints you actually ship.
This axis is essential for structured output generation (JSON, regex-shaped fields) and for less obvious product rules (restrict vocabulary to what a child can read; ban phrases like "As a language model"). It is hard to cleanly separate from domain or generation capability: failure to write a constrained verse form may be missing domain knowledge or misreading the instruction. Evaluation is also confounded by instruction quality itself — a low score can mean a bad model or a bad prompt — so pair failures with explicit instruction design review before blaming the model.
Two public suites are useful templates for custom harness evals:
- Format-verifiable instructions (IFEval-style) — automatically checkable constraints: keyword inclusion/frequency, forbidden words, response language, paragraph/word/sentence counts, bullet/title/placeholder markers, JSON format, choose-from options. Score = fraction of instructions followed. Prefer these wherever a code-based grader can replace a judge.
- Broader constraint checklists (INFOBench-style) — content limits ("discuss only climate change"), linguistic guidelines ("Victorian English"), style rules ("respectful tone"). These need a yes/no criteria checklist per instruction; the model succeeds only if all criteria pass, with aggregate score = criteria met / criteria total. Model-based graders work here when calibrated; they can beat crowd annotators while still trailing experts.
Curate your own suite for the instructions your product actually emits — YAML schemas, banned phrases, length caps, persona rules — rather than relying only on public format benchmarks that may never exercise your failure modes.
Roleplay is a common instruction type (user-facing personas and roleplay-as-prompting). Automating it is hard; evaluate on two axes — style (does voice match the persona) and knowledge (does it stay inside what the persona would know, including "negative knowledge" checks so a game NPC cannot leak spoilers). Use similarity scores, model judges, or trained reward models as needed, but keep roleplay failures labeled as instruction-following rather than domain ignorance when the missing knowledge was deliberately out of scope.
Sources
- AI Engineering (Chip Huyen), ch. 4