Synthetic Data

Synthetic examples and their intended contribution

Synthetic data is artificially constructed rather than collected from actual events. Here, a synthetic example means a record whose input, target, or both were constructed. A target is the answer or behavior supplied for learning. An observed request paired with a generated answer therefore contains synthetic supervision, even though its input is real. [1]

Training fits model behavior from examples; generalization concerns performance on new cases. These are distinct from generating a larger file. Training and inference and independent assessment explain the learning boundary. The useful question is which new behavior the additional examples should support. [2]

A constructed support-routing fixture provides a concrete task. Policy P1 sends explicit unauthorized-access reports to Security; otherwise, explicit duplicate-charge reports go to Billing. Other or unresolved requests go to Triage. Security takes precedence when both reports occur. The intended population contains English and Spanish requests. P1 defines this exercise, not an actual organization's policy. [3]

Fitness for purpose ties acceptance to the task and operating population. Missing cases may justify construction; missing labels may justify annotation. Missing facts require evidence. A fluent teacher answer can introduce details unsupported by the input, turning an information gap into incorrect supervision. [5]

Construction routes and sources of supervision

Augmentation modifies existing examples to add useful variation. Lowercasing a duplicate-charge request preserves Billing under P1. Adding an explicit unauthorized-access report changes its target to Security. The transformation contract depends on the task: preserving wording similarity is insufficient when the changed content crosses a decision boundary. [6]

Seeds are starting examples that guide subsequent construction. Self-Instruct illustrates a seed–generate–filter process for instructions and input–output records. Choosing a label before generating its input can improve label balance, but the requested label remains a generation condition. The completed input must still support it. [9]

Long records benefit from explicit intermediate structure. Fixed attributes can constrain an ordered event plan, which constrains individual documents and their preceding history. This coarse-to-fine construction localizes generation while preserving shared facts; it does not independently verify them. [10]

A teacher model supplies supervision to a student, the model learning from it. Distillation transfers selected teacher behavior or signals; supervision can be generated text or probability targets. The teacher need not generate the inputs. Teacher predictions express learned behavior, not independently established truth. [11]

Post-training adapts an already trained model using additional learning signals. Generated examples can supply those signals and support distillation. DeepSeek-R1's curated examples served both supervised fine-tuning and smaller-model distillation. Synthetic describes how examples originate; post-training and distillation describe how learning uses them. [12]

Task coverage and meaningful variation

Coverage means representation of task-relevant cases; diversity means variation along meaningful distinctions. Useful dimensions include intent, language, wording, missing information, ambiguity, and valid combinations. A constructed initial batch can contain many Billing paraphrases while missing every case that tests precedence. [13]

CheckList distinguishes changes that should preserve an output from changes that should alter it. That distinction separates robustness to wording from learning a new decision condition. Both may matter: paraphrases can test linguistic robustness, while adding a consequential fact tests policy reasoning. Filling either cell is a coverage achievement, not yet evidence of a training gain. [14]

Deliberately generating more rare cases can support diagnosis without reproducing their deployment frequency. Sampling and coverage explain why selected proportions do not automatically estimate population proportions. Report an enriched challenge set separately from a representative-use assessment; each answers a different question. [15][16]

Distribution shift changes input or outcome patterns between training and actual use. Clean generated requests can miss informal, incomplete, or combined requests. Distribution shift and prediction coverage explains why source-distribution performance need not transfer. Variation should follow relevant use conditions, not merely whatever the generator produces easily. [17][18]

Inherited defects and learned shortcuts

Generator bias is systematic distortion in produced cases or answers. Seeds, source material, instructions, and the model's learned behavior shape the result. Expanding a mistaken interpretation across many records can multiply its influence without adding independent support. Rephrasing a seed changes presentation while retaining much of its information. [19]

A shortcut is an incidental cue that predicts training labels without expressing the intended task. In the routing fixture, inserting team names only into their matching categories could teach name matching. Controlled image experiments demonstrate the broader mechanism: models exploited associations between generated-image origin and class, then failed when that association changed. [20]

Repetitive distractors create a similar risk. A vision-model developer reported generators repeatedly choosing dinosaurs and aliens for irrelevant questions. Those examples could teach rejection of particular topics instead of rejection of unsupported premises. Deliberate counterexamples must vary the incidental cue while preserving the intended distinction. [21]

Source grounding addresses unsupported additions only when the source supports the claim. LNQA constructs questions from facts extracted from image descriptions, yet an omitted object is not necessarily absent from the image. A shared source omission can survive both extraction and answer generation. [7]

Repeated agreement is also weaker than independent verification. A model judge can reproduce an error suggested by a candidate answer; different model names alone establish no independence. Checks grounded in separately established references address this specific failure more directly than additional votes. [22]

Teacher selection should follow student outcomes. OpenThoughts found that QwQ-32B supervision outperformed DeepSeek-R1 supervision in tested settings despite the opposite ordering of standalone teacher performance. A stronger answerer is not necessarily a better source of learnable targets. [23]

Verification and justified acceptance

A test oracle is a procedure for deciding correctness; a rubric states assessment criteria. Generating an input does not supply its oracle. Oracles, rubrics and human judgments explains how specifications, execution evidence, and expert judgments support different checks. [24]

One generated record contains the request “I was charged twice, and someone accessed my account without permission,” with target Billing. A schema, the contract for permitted fields and types, can accept this record. P1 still supports Security. Structure and policy meaning require separate checks. [25]

Execution checks establish tested behavior. Incomplete tests can miss defects, and a test encoding the generator's mistaken policy interpretation can faithfully confirm the same mistake. A stale policy similarly checks the wrong contract. More executions do not repair an incorrect expectation. [27]

An LLM-as-a-judge applies assessment criteria using a language model. Compare its decisions with domain-expert judgments before relying on its filtering, as discussed in automated judge reliability. A request correctly assigned Triage because information is missing differs from unresolved annotation disagreement. The latter needs review or exclusion from accepted supervision. [29]

Self-Instruct's one-annotator audit of 200 sampled instructions, each with one instance, judged 92% of instructions valid but only 54% of complete records valid. Those scoped findings distinguish a sensible task description from acceptable supervision. [9]

Selection effects and training exposure

Rejection sampling, as used in these pipelines, generates candidates and retains those meeting criteria. DeepSeek-R1 combined correctness judgments with additional presentation filters. Such rules define the retained distribution; they are recipe choices, not universal definitions of useful data. [12]

Unequal retention can hide behind an acceptable overall score. The figure holds candidate policy-branch proportions equal across languages, then exposes different verifier decisions. These counts identify composition change. They cannot establish false rejection without independently reviewing rejected examples. [30]

Inspect both sides of the filter. C4's blocklist audit found benign material among rejected documents and disproportionate removal in model-assigned dialect categories. In a synthetic pipeline, analogous audits can reveal valuable cases discarded by a quality proxy. Generating more candidates under an unchanged exclusion rule can preserve the same omission. [30]

Stricter filtering need not improve learning. In OpenThoughts' math experiments, random selection outperformed tested answer-quality filters. Its unfiltered arm retained twice as many examples and was not compute-controlled. Selection quality and added training exposure therefore require separate interpretation. [23]

Retained membership also differs from training exposure: sampling can repeatedly show a learner some records and rarely show others. Exact duplicates, differently worded seed descendants, and genuinely distinct cases require different handling. Duplicates and meaningful repetition provides the underlying identity distinctions; inspect topic coverage after filtering rather than treating deduplication as sufficient. [31]

Generation provenance and affected descendants

Provenance records origins and production activities. Collection, identity and permitted use introduces this foundation. For generated examples, shared source, seed, and verifier dependencies identify families needing reassessment even when their wording differs. [32]

Corrections create new record versions and membership records. Preserve the identity of historical snapshots instead of silently changing what an earlier evaluation used. [32]

Recorded settings support traceability, not guaranteed regeneration from a changing service. DataDreamer's authors explicitly acknowledge remaining reproducibility limits with closed models. [33]

Generation does not automatically remove privacy risk. Experiments on synthetic tabular releases found that some source individuals remained vulnerable to inference. Remaining identification risk addresses this boundary. Permitted use likewise depends on actual source and generator restrictions; rights and decision authority explains the separate governance decision. [34][33]

Protected assessment and indirect contamination

Contamination occurs when protected assessment information influences generation, selection, training, or development choices contrary to the assessment's independence assumptions. A benchmark case used as a generation demonstration can influence training through its descendants. A separate test filename does not establish a separate information source. [2]

Establish independent data boundaries before expanding seeds. Keep related descendants together where the claim requires unseen families. Paraphrased benchmark cases have evaded tested overlap detectors while inflating subsequent benchmark scores. Exact and near-match checks are partial defenses; different wording does not restore independence, and a teacher's previous exposure may remain unknown. [35][2]

Final results can leak through decisions without supplying any copied text. Repeatedly changing prompts, filters, or stopping choices after inspecting those results adapts the system to the assessment. Once reused this way, it is development evidence; the original independence claim requires a new assessment. [36]

Synthetic challenge cases remain useful for targeted diagnosis. They establish behavior on constructed conditions, while representative real cases support a different claim about intended use. Combining those results into one unlabeled average hides the distinction. [16]

Recursive generation and retained evidence

Recursive synthetic training makes a learner from one round a generator or teacher for the next. Finite resampling can omit rare cases; approximation and learning errors can then compound. Model collapse names degradation documented under particular recursive conditions, including loss of low-probability regions. It is distinct from repetitive one-shot generation or a single shortcut. [37]

Replacement and accumulation expose later learners to different evidence. One study found deterioration under replacement but bounded error under accumulation in its tested settings. Its language experiments began with TinyStories, itself generated, and accumulation increased training work as datasets grew. Retaining earlier data was consequential, but it did not establish a universally safe mixture. [38]

Fresh observations, authoritative task constraints, and independent checks introduce evidence that does not depend solely on the latest generator. Their value must still be tested. For the routing fixture, track policy interpretation and rare-condition coverage across rounds, retain ancestry, and vary generation count and synthetic proportion as experimental factors rather than treating any ratio as safe. [39][40]

Downstream utility on protected real cases

Downstream utility is improvement in the intended use of the resulting system. Train-on-synthetic/test-on-real evaluates a predictor fitted on generated examples against held-out real cases—cases excluded from fitting. Testing only on generated cases can miss regions the generator omits. This replacement-data experiment does not establish that adding synthetic examples improves a real-data baseline. [28]

An ablation changes an ingredient to investigate its contribution. Hold model initialization, evaluation cases, tuning access, and the stated training budget constant where the comparison requires it. Separate content changes from extra exposure. Controlled offline comparisons supplies the broader method; synthetic-data experiments must identify which part of the recipe changed. [41]

Assess the targeted Spanish mixed-intent cases and ordinary requests separately. Report eligible counts, unresolved judgments, and uncertainty around differences; an aggregate can conceal regressions. Uncertainty in scores and differences explains interpretation. Coverage of the intended users, tasks, and operating conditions remains necessary even when the test set was held out correctly. [43]

Offline routing accuracy is not yet evidence that support work improved. Appropriate handoff and reduced misrouting burden are additional workflow outcomes. The measurement must match the intended improvement rather than substituting fluent explanations or high filter scores. [44]

Repair, expansion and stopping decisions

Failure diagnosis determines the next intervention. Failure analysis separates observations from competing explanations. More candidates address scarcity; corrected targets address bad supervision; broader retention addresses exclusion. None automatically repairs a learner that ignores the intended distinction. [45]

Small, failure-focused experiments support rapid iteration; broader checks reveal regressions outside the repaired slice. Stop expanding a recipe when meaningful gains disappear or harms dominate. Pause when evidence is insufficient. Release and revision decisions should apply predeclared requirements, rather than equating a larger accepted dataset with progress. [48]

Accepted examples influence learning through their sampling and mixture, not membership alone. Training schedules belong to Post-training and Alignment and Pretraining and Midtraining. Here, the decision is which evidence merits additional exposure, and whether the comparison demonstrates the intended benefit. [19]

Open questions

1. 01

   Predicting teacher suitability before expensive student training remains unresolved. Teacher accuracy alone misses transfer differences, while proposed explanations such as trace format and length remain unisolated. Progress would predict student gains across domains while controlling example count and training exposure. [49]

2. 02

   Separating meaningful novelty from extra exposure remains difficult. Rephrasing reinforces selected information without necessarily adding a decision condition. Matched comparisons of repetitions, paraphrases, and new condition combinations would clarify which contribution improves protected real-case performance. [19][14]

3. 03

   Reliable verification without shared errors remains costly. A judge can inherit a candidate's mistaken reasoning, while human agreement does not establish truth. Progress would measure incorrect acceptance and rejection against independently established references across difficult case families. [22][50]

4. 04

   Long-run retention strategies need evidence beyond existing recursive experiments. Accumulation changes available data and training work together, complicating attribution. Progress would compare retention and replacement across rounds under matched resources while tracking rare-case performance. [38]