liveBook · Manning

Evaluation dataset

As explained in LLM Customization and Fine-Tuning

A subset of examples reserved for evaluating a model rather than updating its parameters.

An evaluation dataset is a set of examples reserved to measure model behavior rather than update model parameters. Without a representative held-out set, a promising development score may reflect fit to familiar examples instead of performance on production inputs. Establish a baseline before training, then compare adaptation techniques on the same evaluation cases. section 1.7 section 4.4

A golden test set is a curated evaluation set for a release or model gate. Its inputs should reflect the production workload, and its reference answers should be fully human-annotated. Freeze and version both inputs and answers; keep them out of training, prompt design, and synthetic-data generation. Synthetic evaluation examples can measure imitation of a teacher’s style rather than generalization to real inputs. section 3.7.1 section 3.4.1 section 3.6.6

Build a golden set around the decisions it must support:

  • Define three to five user-facing tasks and give each a separate slice, so an overall score does not hide task-specific failures. section 3.7.2
  • Collect realistic cases from appropriately redacted and consented production logs, past tickets, or domain experts. Include common cases and consequential edge cases; aim for 10–40 examples per task slice. section 3.7.2
  • Have a domain expert supply or verify each reference answer and record its rationale. Save the set as version-controlled JSONL; document changes and record the version used in every evaluation run. section 3.7.2
  • Keep a separate red-team set for adversarial requests, refusals, and escalation cases, so safety and representative quality remain distinct measurements. section 3.7.3

Size the set for the decision, not an arbitrary split. Twenty examples can expose a broken prompt but cannot reliably rank close variants. Fifty can reveal large gaps; a few hundred are needed to report differences of a few points. For a fine-tuning go/no-go decision, a target of 200–500 examples is more appropriate. When a larger set is unaffordable, compare variants on the same items and count their disagreements. section 4.4.2 section 1.7

Repeatedly selecting prompts or models based on the same evaluation cases can overfit those cases. Use fresh unseen examples for an additional check, and reserve a separate test set from training and model selection when a final comparison is needed. section 5.10 section 6.1.1

Clinical assistant splits

A clinical assistant used a 520-example validation set, stratified by category, and a separate 200-example test set drawn from a different source-data time period. The later-period test set guarded against temporal leakage. section 6.12.1

Comparing distilled models

On the same 50-example held-out split, a base model, teacher, and student scored 0.153, 0.168, and 0.165 Token-F1 respectively. The student retained about 98% of the teacher’s score, but Token-F1 measures word overlap, so it was read alongside behavioral and safety checks. Category-level results were especially noisy with only a few examples per category. section 7.2.2

How LLM Customization and Fine-Tuning covers it

How deep it goes
Details sourcing, annotation, versioning, leakage controls, sample-size tradeoffs, and separate safety evaluation.
Examples
Uses clinical validation and temporal test splits, plus a shared distillation test comparing base, teacher, and student models.
What you can do after
Practice building and versioning a golden set, sizing it for a decision, and comparing models without confusing lexical scores with behavior.
How it connects
Connects held-out evaluation with prompt engineering, supervised fine-tuning, synthetic data, distillation, and red teaming.

Also explained in

Compare how 41 books explain evaluation dataset →

Related concepts

This explainer was generated by Manning's knowledge graph from the text of LLM Customization and Fine-Tuning. The section links point to the parts of the book it draws on.