Skip to content

AI uses for records

Fine-tuning dataset outcome coverage: are eval outcomes in training?

By SourceX Editorial · Updated

11 sections

Short answer

To audit fine-tuning dataset outcome coverage, list every outcome the evaluation scores, tag each training and evaluation example with its outcome, and count both splits per outcome within each failure cohort. Outcomes that are absent or thin in training are gaps. Fill them with real examples under a written sampling rule, then compare against a same-size baseline on untouched evaluation data and check other cohorts for regressions.

Evaluation split

Fine tuning dataset outcome

This page covers fine-tuning dataset outcome coverage: are eval outcomes in training.Illustrative

Key takeaways

  • A fine-tuning set built from successful cases often holds none of the escalations, refusals or rejections the evaluation scores.
  • Start from one named failure so the audit has an outcome to count and a cohort to slice.
  • Count after a grouped, outcome-stratified split, and never move evaluation examples into training.
  • Use four statuses, absent, thin, covered and overweighted, and a written sampling rule for each.
  • Prove the fix with a matched example and token budget, repeated runs and per-slice regression checks.
  • Some outcomes belong out of training on purpose: generalization tests, format errors and fast-changing facts.

How do you audit whether training examples include the outcomes a model will be evaluated on?

List every outcome your evaluation scores, tag each training and evaluation example with the outcome it ends in, and count both splits per outcome and per cohort. Any outcome the evaluation measures but training holds zero or a handful of times is a coverage gap. Fix gaps with an explicit sampling rule, then test the fix against a matched-budget baseline on evaluation examples nobody used for selection.

This sounds obvious and is routinely skipped. Teams assemble a fine-tuning set from whatever successful conversations, resolved tickets or approved documents they have, then evaluate on a test set that rightly includes escalations, refusals, rejected claims and partial refunds. The model is then judged on outcomes it was never shown, and the failure gets blamed on model size or learning rate.

The asset on this page is an outcome coverage and sampling table that lines up each evaluated outcome with its training count, evaluation count, coverage status and sampling rule. Around it sit the steps that keep the audit honest: a defined failure, leak-free counting, matched budgets and regression checks.

Start from one failure the model keeps making

Write down the specific failure before touching the data. A useful failure statement names the input cohort, the outcome expected and the outcome the model produces, such as: on billing disputes over a set age, the assistant offers a refund when policy requires escalation to a billing specialist.

A named failure turns a vague coverage question into a countable one. You now know which outcome label to search for in training (escalated to billing), which cohort to slice evaluation by (billing disputes over the age limit), and what counts as improvement. Without it, coverage audits drift into counting everything and fixing nothing.

Keep the failure list short. Two or three failure cohorts per audit cycle is enough to act on; a long list usually means the outcome definitions themselves need work.

What counts as an outcome in a fine-tuning example?

An outcome is whatever the evaluation compares against: the final answer, the action taken, or the state the case ended in. Define them from the evaluation rubric, not from the training data, because the point of the audit is to find outcomes the training data lacks.

At a glance

What counts as an outcome in a fine-tuning example

Training formats differ in where the outcome lives. In a prompt-completion dataset, as Hugging Face's TRL library documents it, the outcome sits inside the completion text and must be extracted or tagged. In an unpaired preference dataset, each completion also carries a label saying whether it is preferred, so negative outcomes are explicit. TRL's KTO trainer, which uses that label to mark completions desirable or undesirable, even provides separate weights to counter an imbalance between the two, a sign that the balance of outcomes is a known training concern.

  • Decision outcomes: the label, class, code or verdict, such as approve, deny, route to tier two.
  • Action outcomes: tool calls, escalations, refunds, handoffs to a person, or a deliberate no-action.
  • Terminal states: resolved, reopened, canceled, abandoned, recorded in the source system after the conversation ended.
  • Negative outcomes: refusals, requests for missing information, and answers the reviewer marked wrong.
  • Format outcomes: a required structure such as a JSON object or a specific number of fields, when the evaluation checks it.

The outcome coverage and sampling table

Fill one row per evaluated outcome within each failure cohort. The counts below are illustrative, for a customer support assistant; they show how the statuses and rules follow from the numbers, not real results.

Use four fixed statuses. Absent means zero training examples. Thin means below the floor you set for the cohort. Covered means at or above the floor. Overweighted means far above the outcome's share in the evaluation set and in production, which can push the model toward that outcome everywhere.

  • Set the floor per cohort before counting, and record it with the table.
  • Record where each added example came from and its source period, so lineage survives the next refresh.
  • Keep a column for the evaluation share and the production share when you know it; overweighting is judged against both.
The outcome coverage and sampling table
Outcome in billing disputesTraining count, illustrativeEvaluation count, illustrativeStatusSampling rule
Refund issued within policy2,40060OverweightedCap at a recorded multiple of its evaluation share; drop near-duplicates first
Escalated to billing specialist045AbsentAdd from real escalated tickets to the cohort floor; never copy evaluation tickets
Asked for missing invoice number3830ThinRaise to the floor with real cases from other months
Refused out-of-policy refund1225ThinRaise to the floor; check reviewer agreement on these labels
Partial credit applied21020CoveredKeep as is
Ticket reopened after reply015AbsentNeeds terminal-state data from the help desk; flag if unavailable
Malformed structured reply010Absent by designKeep out of training; cover with format validation instead

Counting outcomes without leaking the evaluation set into training

Split first, then count, and never move evaluation examples into training to fill a gap. scikit-learn's guidance on leakage states the rule plainly: test data should never be used to make choices about the model, and data should be split before any preprocessing. Choosing which outcomes to oversample by looking at individual evaluation failures and then pulling near-identical cases into training is the fine-tuning version of that mistake.

Split by group, not by example, wherever one customer, account, thread or document family produces several examples. scikit-learn's GroupKFold keeps a group from appearing in both training and testing sets; the same idea applied to a fine-tuning split stops a model from memorizing one customer's dispute and getting credit for it on the evaluation side.

Stratify the split on outcome, especially for rare outcomes. The scikit-learn cross-validation guide notes that with rare classes, splitting can produce folds with no occurrence of a class at all, and that stratified splitters roughly preserve class frequencies. An evaluation set with zero escalations cannot tell you whether escalation coverage worked.

Matched budgets: proving the added examples did the work

Compare the rebalanced training set against a baseline of the same size and token budget. If the new run simply has more data, any gain might come from volume rather than from the outcomes you added.

At a glance

Matched budgets: proving the added examples did the work

Count examples and tokens per arm before training. OpenAI's cookbook notebook for chat fine-tuning data shows the kind of pre-flight checks worth running: format validation, counts of conversations missing system or user messages, and the distribution of messages and tokens per example. Run those per outcome, because added escalation examples are often longer than the refunds they displace, which quietly changes the budget.

  • Arm A, baseline: the original training set, downsampled or padded with randomly drawn in-distribution examples to the agreed budget.
  • Arm B, coverage fix: the same budget, with the sampling rules from the table applied.
  • Optional arm C, cap only: the overweighted outcome capped, nothing added, to see how much of the change comes from removing excess.
  • Same base model, hyperparameters, seed policy and number of runs in every arm; report the spread across runs, not a single score.

Did the failure cohort improve, and did anything regress?

Score every arm on the same untouched evaluation set, sliced by cohort and outcome, and read the regression slices before the headline. Slice-level reporting is standard practice in evaluation tooling; TensorFlow Model Analysis, for example, is built to compute metrics over different slices of data.

Three readings are common. The target cohort improves and nothing else moves: the coverage gap was the cause. The target improves and a neighboring outcome degrades, such as escalations rising on disputes that should have been refunded: the sampling rule overcorrected, so lower the floor or add contrasting examples. Nothing improves: the outcome was covered in name only, often because its labels are unreliable or the input never contained the signal needed to choose it.

Did the failure cohort improve, and did anything regress?
CheckSliceWhat to compareAction if it fails
Target improvementFailure cohort, target outcomeArm B against arm A on the same examplesInspect labels and inputs for the added examples
Neighbor regressionSame cohort, other outcomesPer-outcome accuracy or rubric scoreLower the floor or add contrasting cases
Capability regressionCohorts outside the failureOverall and per-cohort scoresReduce the share of added examples or mix in prior data
Format regressionAll examples with structured outputParse and schema pass ratesCheck whether added examples used a different format
Run-to-run spreadEvery slice aboveRange across repeated runsTreat differences inside the spread as no change

Illustrative case: an assistant that never escalated

Everything in this case is illustrative. A support team fine-tunes an assistant on 6,000 resolved tickets. Evaluation on 400 held-out tickets shows billing disputes failing badly: the model offers refunds where policy requires escalation. The audit finds that the training set holds no escalated billing disputes at all, because the export only included tickets closed by the first agent.

The team pulls escalated billing tickets from a different six-month period, splits by customer account so no account appears on both sides, and sets a floor for escalation within the billing cohort. To keep the budget fixed, it caps the refund outcome and drops near-duplicate refunds first. The rebalanced arm and a baseline of the same size each train three times.

On the untouched evaluation tickets, escalation accuracy in the billing cohort rises in every run of the coverage arm while scores outside billing stay inside the run-to-run spread. One neighbor slice regresses: out-of-policy refund refusals start being escalated instead. The team adds real refusal cases to the floor in the next cycle rather than declaring success.

When an outcome should stay out of training on purpose

Not every absent outcome is a gap to fill. Some evaluations deliberately test outcomes the model has never seen, such as a new product line or a policy introduced after the training period, to measure generalization. Filling those cells would turn a generalization test into a memorization test; mark them absent by design and report them separately.

Other outcomes are better handled outside fine-tuning. A malformed structured reply is an outcome to prevent with validation or constrained decoding, not one to teach. Outcomes driven by facts that change weekly, such as current prices or stock levels, belong in retrieval or tools rather than in model weights.

And some outcome labels are not trustworthy enough to sample on. If a help desk sets resolved automatically when a macro is sent, a resolved count says nothing about whether the customer's problem was solved. Check how each outcome field is written in the source system before you trust its counts.

Signals that a coverage fix will not solve the failure

Stop rebalancing and rethink the approach when any of these appear. Real examples of the outcome are so rare that the floor can only be met by duplicating them. The input never contains the information needed to choose the outcome, so no amount of examples can teach it. Reviewers disagree on the outcome label for a large share of cases. Or the matched-budget comparison shows no change across repeated runs.

Bring in a post-training researcher when sampling rules start interacting, for example when raising one floor forces a cap elsewhere that hurts another cohort. Bring in the evaluation lead when the rubric's outcome definitions turn out to be ambiguous; fixing the definition usually does more than any sampling rule. And bring in the business owner of the process when a policy change explains the failure, since the data then needs a period boundary rather than more examples.

Requesting outcome-labeled records for a coverage gap

A gap in an outcome cell is often a gap in what was exported, not in what the business recorded. When you request data to fill one, specify the outcome codes, the terminal-state fields, the period and the cohort floors from your table, and ask for the field definitions so you can judge label reliability.

When outcome-labeled records are not already in hand, a request to SourceX starts a search among U.S. companies rather than a pull from inventory; attach the outcome codes and cohort floors from your table next to volume, format, timing and licensing needs. Escalation and refusal histories then arrive only after the company holding them has signed off on what is in scope.

Frequently asked questions

What is outcome coverage in a fine-tuning dataset?

It is the extent to which training examples include each outcome the evaluation scores, such as escalations, refusals or specific decision codes, within each input cohort. A dataset can be large and still have zero examples of an outcome the model is tested on.

How many examples of each outcome do I need?

There is no universal number. Set a floor per cohort before counting, test it with a matched-budget comparison, and raise or lower it based on whether the target cohort improves without regressions elsewhere.

Can I copy failing evaluation examples into training?

No. That leaks the evaluation set and inflates the score you use to judge the fix. Pull new real examples of the same outcome from other periods, accounts or sources, and split by group so related examples stay on one side.

What should the outcome coverage and sampling table include?

One row per evaluated outcome within each failure cohort, with training count, evaluation count, a status of absent, thin, covered or overweighted, and the sampling rule applied. Record the floor, the sources of added examples and their periods with it.

Is it a problem if one outcome dominates the training data?

It can be. An outcome far above its evaluation and production share can push the model toward it across cohorts. Cap it, removing near-duplicates first, and check neighbor slices for regressions after retraining.

Sources

  • scikit-learn's cross-validation guide warns that rare classes can leave folds with no occurrence of a class, recommends stratified splitters that roughly preserve class frequencies, and describes GroupKFold, which keeps a group out of both train and test. Source
  • scikit-learn defines data leakage as using information unavailable at prediction time, says test data should never be used to make choices about the model, and advises splitting before preprocessing. Source
  • Hugging Face TRL documents prompt-completion datasets and unpaired preference datasets, where each completion carries a label saying whether it is preferred. Source
  • TRL's KTO trainer expects a boolean label marking each completion desirable or undesirable and offers separate weights to counter an imbalance between the two. Source
  • TensorFlow Model Analysis computes evaluation metrics over different slices of data. Source
  • OpenAI's cookbook notebook for chat fine-tuning data checks format errors, counts examples missing system or user messages, and reports message and token counts per example. Source

Tell us what data you need

Share scope, volume, format, timing and licensing needs. Datasets are sourced on request.

See if your company qualifies