Skip to content

Measuring Synthetic Eval Data Quality (SynAE)

Score synthetic tool-calling eval datasets on validity, fidelity, and diversity across four trace components before trusting them to gate deploys against production.

SynAE is a quality-measurement layer for synthetic eval datasets that test multi-turn tool-calling agents. It scores a synthetic set against a real production-trace reference along three pillars — validity, fidelity, and diversity — across four trace components. The result is a per-axis diagnostic, not a single pass or fail (arxiv.org/abs/2605.22564).

When the framework applies

Four preconditions must hold: a real production-trace reference exists, the agent is multi-turn and tool-calling, the synthetic set is large enough for distributional metrics (Vendi Score and embedding precision/recall are unstable on small samples), and you can absorb the judge cost. Validity scoring runs 450+ LLM-as-judge calls per dataset at F1 = 0.86 against humans (arxiv.org/abs/2605.22564).

If any fails, a golden query pair or incident-derived regression suite gives a stable signal more cheaply.

The three pillars

Three orthogonal axes apply to each of four trace components — task instructions and intermediate responses, tool calls, final outputs, downstream eval outcomes (arxiv.org/abs/2605.22564):

Pillar What it measures Sample metric
Validity Do synthetic tool calls and outputs successfully fulfill the given instructions? Validity Rate from LLM judge or rule checker
Fidelity How close is the synthetic distribution to the real one? Key Node Dependency, embedding precision/recall, downstream task-difficulty distance
Diversity How much of the real-data spread does the synthetic set cover? Vendi Score, Attribute Diversity

A set can score high on one axis and fail another. The decomposition exists because the authors found "no single metric is sufficient to fully characterize synthetic data quality" (arxiv.org/abs/2605.22564).

Why it works

Synthetic datasets diverge from production silently because the generator's prior is not the user's. Generators sample from model-induced distributions — templated prompts, in-context examples, fixed tool schemas — while production samples real intents, tool errors, and multi-step plans. A passing synthetic suite can still let a deploy regress where a golden query pair suite would not.

SynAE counters this by decomposition: scoring each trace component on each pillar attributes ranking distortion to a specific axis rather than a black-box verdict. The paper grounds this in controlled experiments where each generation failure mode moves a predictable pillar (arxiv.org/abs/2605.22564).

Diagnostic patterns

Five failure modes the authors injected, and the pillar signatures they produced (arxiv.org/abs/2605.22564):

Generation choice Validity Fidelity Diversity What to do
Blank Filling — mask tokens, resample Down Up Tighten the mask or constrain regeneration
Oversampling — duplicate frequent sequences Down Down Stratify against the real distribution
In-Context Generation with fixed examples Mixed Capped Rotate exemplars; watch fidelity per metric
Invalidation — traces that do not work Down Add validity filters before entry
Naive Relabeling — keyword substitution Down Up Use semantically grounded transformations

Read the pillar movements together — a single failing axis often points to one generation choice.

How to use it

  1. Pull a stable production-trace reference; incident-to-eval synthesis is one source.
  2. Run all three pillars on all four components — a single score hides cross-axis trade-offs.
  3. Match failure signatures to fixes using the table above.
  4. Recheck after each intervention; verify a fix does not move a non-target pillar (arxiv.org/abs/2605.22564).
  5. Do not collapse to a single score — the authors note averaging normalized metrics "is not necessarily the most meaningful way as they may scale differently."

When this backfires

Skip or replace the framework when:

  • No production reference. Fidelity and diversity reduce to noise, and a small golden query pair suite is cheaper.
  • Single-turn or non-tool-calling agents. The four-component decomposition collapses, and output grading is equally informative.
  • Frozen smoke-test usage. For a fixed gate on known issues, distributional fidelity is irrelevant, and a pass/fail suite costs a fraction of the judge calls.
  • Vendor platforms already bundle validity and fidelity scoring into the generator (Databricks, Tonic.ai), so a separate layer is redundant.
  • Uncalibrated LLM judge. Without judge calibration, the outputs mislead more than they help.
  • Interactive or multi-agent settings. The paper scopes SynAE to single-agent multi-turn tool-calling, and extensions are future work, not validated (arxiv.org/abs/2605.22564).

Key Takeaways

  • Synthetic eval data drifts silently from production — the same idealized-condition inflation benchmark contamination as eval risk describes; SynAE catches the drift before a deploy gate trusts the suite.
  • Score on three orthogonal pillars across four trace components — the matrix is the artefact, not an aggregated number.
  • Five known generation failure modes have predictable pillar signatures; read the signatures to pick the right fix.
  • Skip when there is no production reference, the agent is single-turn, the set is small, or the LLM judge is uncalibrated.
  • The validity check is itself bounded by LLM-judge F1 of 0.86 — treat the measurement layer as one more component that needs calibration.
Feedback