Synthetic Data

References

A synthesis of selected conference talks and technical references. Citations link to the source material; they do not imply that every talk on this subject is included.

1. MITRE: Synthea Synthetic Data Overview

   Synthetic data is artificially constructed rather than collected from actual events. Synthea illustrates construction without a teacher model: it creates fictional individuals from regional demographic distributions and simulates their histories using disease modules informed by incidence, prevalence and care guidelines. The resulting records describe simulated events, not observations of actual patients.

2. Pseudo-Label: The Simple and Efficient Semi-Supervised Learning Method for Deep Neural Networks

   Lee's pseudo-labeling method assigns an unlabeled input the class with the highest model-predicted probability and uses that prediction as a training target alongside genuine labeled examples. The input itself need not be generated. In this method, predictions are refreshed during training rather than supplied by a necessarily separate teacher.

3. Improving Neural Machine Translation Models with Monolingual Data

   Back-translation turns existing target-language sentences into additional translation pairs by automatically generating their source-language counterparts. Only the source side is synthetic; the desired output remains text from the monolingual corpus. Sennrich, Haddow and Birch used this to exploit monolingual data without adding a separate language model or changing the translation architecture. Their English-to-German WMT experiments improved translation scores over parallel-data baselines. Continuing the baseline for another week produced no BLEU improvement, providing a control against training duration alone explaining gains.

4. Foundations of Computer Vision: Training for Robustness and Generality

   Label-preserving augmentation assumes y(T(x))=y(x). This is task-dependent: mirroring can preserve a scene category while changing a character's identity. When targets have spatial structure, transform image and target together so y(Tx)=T_y(y(x)); a segmentation crop must crop the label map at the same coordinates. Augmentation teaches selected invariances or equivariances by adding transformed training examples. It cannot repair an invalid label-preservation assumption or supply an independent evaluation sample.

5. Distilling the Knowledge in a Neural Network

   Distillation transfers behavior from a trained model or ensemble to another model using the source model’s predictions as supervision. The original method uses class-probability distributions as soft targets; its transfer inputs can be existing training examples or unlabeled data. Thus, the model supplying supervision need not generate the inputs, and supervision need not consist of generated text.

6. The Monte Carlo Method

   In September 1949, Nicholas Metropolis and S. Ulam described Monte Carlo computation as a statistical alternative for mathematical-physics problems whose integro-differential equations were laborious or unavailable in closed form. Their procedure samples chains of stochastic and deterministic events from specified distributions, repeats them to approximate evolving distributions, and uses computing machinery to produce the required pseudorandom values and calculations.

7. Unsupervised Word Sense Disambiguation Rivaling Supervised Methods

   David Yarowsky's June 1995 method addressed costly hand-tagging by beginning with small sense-specific seed sets, training a classifier, adding high-confidence predictions from the unlabeled residual, and repeating until the residual stabilized. Added cases could return to the residual if their confidence later fell. The study reported 96.1% average supervised accuracy and up to 98.6% for its strongest tested bootstrapped configuration across twelve ambiguous words.

8. SMOTE: Synthetic Minority Over-sampling Technique

   Published in June 2002, SMOTE addressed classification datasets with relatively few examples of an important class. Instead of repeatedly copying minority examples, its basic method constructs feature vectors between a minority example and selected minority-class neighbors, assigning the minority label. The motivation was to encourage broader decision regions rather than increasingly specific regions around duplicates. The paper supplies a numerical construction using the points (6,4) and (4,3), with a random fraction of their difference added to the first point.

9. Domain Randomization for Transferring Deep Neural Networks from Simulation to the Real World

   On March 20, 2017, Josh Tobin and colleagues reported domain randomization as an alternative to precisely matching a simulator to reality. They generated labeled images while varying camera position, lighting, object placement, distractors, and unrealistic textures. A detector trained only on simulation localized tested real geometric objects to about 1.5 centimeters and supported grasping in clutter. Ablations showed real-image error depended on the chosen randomizations: fewer than 1,000 textures degraded performance, and omitting distractors harmed robustness to real distractors.

10. Self-Instruct: Aligning Language Models with Self-Generated Instructions

    Self-Instruct starts from 175 human-written seed tasks, generates instructions and input–output instances, filters candidates, and fine-tunes the original model. Input-first classification generation favored particular labels, motivating generation conditioned on chosen labels. Filtering removed similar instructions, conflicting duplicate inputs and malformed instances. Nevertheless, an author’s audit of 200 sampled instructions with one instance each judged 92% of instructions valid, 58% of outputs correct and acceptable, and 54% of complete records valid. Construction, apparent task validity and correct supervision were therefore distinct in this dataset.

11. DataDreamer: A Tool for Synthetic Data Generation and Reproducible LLM Workflows

    DataDreamer’s synthetic data cards trace chained generation steps and record dataset and model identities, execution dates, known licenses, configuration fingerprints and environment information. The authors recommend sharing exact prompts and intermediate outputs. They explicitly acknowledge that reproducibility concerns cannot be fully eliminated with closed-source models, and that automatically collected metadata requires human supplementation.

12. Statistics Canada Quality Guidelines: Coverage and frames

    The target population comprises the units about which information is wanted. Practical exclusions can narrow this to a survey population. A sampling frame identifies and provides access to units; omissions, duplicates, erroneous inclusions and misclassifications create coverage errors. Probability sampling from a restricted survey population does not automatically justify claims about the broader target population. The guidance recommends documenting these differences and periodically checking frame coverage against other sources.

13. The Oracle Problem in Software Testing: A Survey

    A test oracle is a procedure for distinguishing correct from incorrect behavior of the system under test. Generating inputs does not solve the separate problem of deciding whether their outputs are right. Complete specifications can support automated oracles, but specifications may be unavailable or incomplete. Metamorphic relations, documentation and execution evidence can support partial oracles; implicit checks detect only selected classes of failure. Where suitable artifacts are absent, humans must supply or assess expected behavior, incurring construction and review costs.

14. DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning

    DeepSeek-R1’s reported rejection-sampling stage generates multiple responses per reasoning prompt and retains responses judged correct. Some judgments use rules; others use a model supplied with a reference answer. Additional filters remove mixed-language responses, long paragraphs and code blocks. The resulting reasoning examples are combined with non-reasoning data for supervised fine-tuning. The authors also fine-tune smaller Qwen and Llama models on the curated examples as distillation.

15. OpenThoughts: Data Recipes for Reasoning Models

    OpenThoughts compares data-source, filtering, answer-generation and teacher choices through downstream student performance. In its answer-filtering experiments, random selection outperformed tested quality filters for math. The unfiltered arm retained twice as many examples and was explicitly not compute-controlled. Teacher choice also mattered: QwQ-32B-generated supervision outperformed DeepSeek-R1-generated supervision in the tested settings despite the teachers’ opposite standalone benchmark ordering. Selecting strategies by average performance did not yield the strongest result in every domain.

16. PROV-DM: The PROV Data Model

    Provenance records the entities, activities and responsible agents involved in producing or delivering information. Entities include files and document versions; activities use and generate entities; derivation connects an output to an earlier entity through transformation, update or construction. These relationships provide a concrete representation for source-to-derived-data lineage. PROV distinguishes a particular document version from a resource identifying its changing latest version and warns that provenance descriptions must remain valid as resource state changes.

17. Synthetic Data – Anonymisation Groundhog Day

    Experiments on synthetic tabular releases show that replacing original records with generated records does not automatically prevent inference about individuals in the source data. Privacy gains vary across records, generator types and attacker features; selected outlier records remain vulnerable even when randomly selected records receive much stronger protection.

18. NIST AI RMF Playbook: Measure

    Construct validity asks whether an indicator measures the concept it claims to measure; external validity concerns generalization beyond development conditions. NIST calls for documented operating conditions, measurement assumptions, limitations and variance. Evaluations using human-subject data should reflect the population in the context of use. Applied to agent evaluation, define the deployment population and scenario dimensions before sampling, document exclusions, and compare sampled conditions with intended users, tasks and operating environments. A split within an unrepresentative dataset does not establish deployment coverage.

19. Assessing Generative Models via Precision and Recall

    Sajjadi and colleagues' 2018 distributional precision–recall framework separates generated-sample quality from coverage of a reference distribution. In controlled image and text experiments, dropping reference modes reduced recall, while adding unsupported modes reduced precision; models with similar one-number FID scores could therefore exhibit materially different failures. The computation depends on samples represented through a meaningful embedding.

20. How Faithful is your Synthetic Data? Sample-level Metrics for Evaluating and Auditing Generative Models

    Alaa and colleagues separate fidelity, diversity, and authenticity in synthetic-data assessment. Their analysis notes that a generator can attain perfect support-based fidelity and diversity by resampling its training records, without producing novel examples. Their density-sensitive precision and recall curves also distinguish shared support from matching probability mass. Consequently, plausible records and broad aggregate support do not establish novelty or useful independent information.

21. Computer Use at the Edge of the Statistical Precipice

    The speaker reports that most tested frontier models do not maintain equivalent performance when the starting screen or application theme changes.

22. The Art & Science of Benchmarking Agents

    Choose task distributions intentionally: production representativeness and coverage of rare consequential failures are different evaluation goals.

23. Computer Use at the Edge of the Statistical Precipice

    Scale environment generation through a verification pipeline, not just through generating more application code.

24. AmbigQA: Answering Ambiguous Open-domain Questions

    AmbigQA represents an ambiguous question through multiple plausible answers, each paired with a minimally rewritten question specifying its interpretation. Its annotated examples distinguish entity references, events, answer types and time-dependent meanings. Annotators search and read supporting Wikipedia pages rather than assuming that an existing single answer exhausts the question's interpretations.

25. The Oracle Problem in Software Testing: A Survey

    A test oracle decides whether observed behavior is acceptable. The survey formalizes a deterministic oracle as a partial function from test activity sequences to true or false, distinguishing it from conceptual ground truth. An oracle can therefore be missing for some cases or disagree with intended behavior. Engineering inference: independently executing tests does not make their expected answers independent of the implementation's assumptions. If both encode the same mistaken interpretation of a requirement, execution can faithfully confirm their agreement while the delivered behavior remains wrong.

26. Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena

    The study uses anonymous candidate pairs, shared evaluation instructions, expert human votes, reference solutions for some questions, and human–human as well as human–model agreement. It examines order swaps, verbosity and reference-guided judging. In reasoning examples, a judge can reproduce an error present in the candidate answer; independently generated references reduce failures in the tested setting. Human reconsideration after seeing GPT-4's explanation is a separate, influenced judgment. The authors say their limited data cannot determine self-enhancement bias.

27. Correlated Errors in Large Language Models

    The researchers measure whether pairs of models choose the same wrong answer, conditional on both being wrong. Across the studied multiple-choice datasets, wrong-answer agreement exceeded uniform-choice baselines for almost all model pairs. Shared provider and architecture were associated with greater agreement, but distinct providers did not eliminate it. When one model's answers were used as reference labels for another, shared mistakes could inflate measured accuracy. This directly demonstrates why agreement between different models need not supply independent confirmation.

28. Synthetic QA Corpora Generation with Roundtrip Consistency

    The 2019 method selects an answer span from a passage, generates a question conditioned on both, and retains the example when a separately trained answering model recovers that span from the passage and question. Thus, the question generator receives the intended answer while the answering model does not. Filtering improved downstream question-answering performance in the reported ablations. Nevertheless, manual inspection judged only 39% of 46 accepted triples correct, versus 16% of 44 discarded triples. At fixed synthetic-example counts, combining two generation sources improved SQuAD2 development accuracy over one source.

29. Let's Verify Step by Step

    The study distinguishes labels for final outcomes from labels for individual reasoning steps. Its human annotators marked generated steps positive, negative or neutral, retaining ambiguity rather than forcing every judgment into correct or incorrect. The authors explicitly identify false-positive outcome labels when incorrect reasoning reaches the correct final answer. They also selected apparently convincing wrong-answer solutions for human review because those cases exposed mistakes in the current verifier.

30. Documenting Large Webtext Corpora: A Case Study on the Colossal Clean Crawled Corpus

    The authors examined documents excluded by C4's word blocklist, finding scientific, medical and other non-offensive content among removals. Identity-mention analysis and manual inspection showed exclusion of benign material discussing sexual orientations. A dialect-model analysis found disproportionate removal of text assigned African American English and Hispanic-aligned English categories. Inspecting rejected records therefore revealed composition effects that examining only the retained corpus would miss.

31. SelectiveNet: A Deep Neural Network with an Integrated Reject Option

    Selective prediction pairs predictor f with selection function g: accept when g(x)=1, otherwise abstain. A confidence threshold or learned selection score determines acceptance. Coverage is phi = E[g(X)]; selective risk is E[loss(f(X),Y)g(X)]/phi for phi>0. On labeled evaluation data, measure coverage as accepted cases divided by all cases, and risk as average loss among accepted cases. Varying the threshold produces a risk–coverage curve. SelectiveNet jointly learns prediction and selection while optimizing risk subject to a target coverage constraint. This makes abstention measurable instead of treating a confidence statement as sufficient evidence of reliability.

32. Training Krea 2 - What matters in generative model training.

    Curate against the intended aesthetic coverage rather than relying exclusively on conventional aesthetic or image-quality scores.

33. Synthetic Simplicity: Unveiling Bias in Medical Data Augmentation

    The researchers deliberately vary the association between class labels and whether images are real or synthetic. In digit and echocardiogram classification experiments, strong source–label associations produce poor performance on combinations that break that association. A classifier can therefore exploit generation-related features as shortcuts instead of learning the intended class distinction, even when synthetic images appear realistic.

34. AI models collapse when trained on recursively generated data

    The study examines repeated learning from preceding models’ generated samples. It describes early model collapse as loss of low-probability regions and later collapse as substantial departure from the original distribution. Finite resampling can omit rare events; approximation and learning errors can compound across generations. Its mathematical replacement setting uses samples from the immediately preceding model, while numerical experiments also examine mixtures.

35. Is Model Collapse Inevitable? Breaking the Curse of Recursion by Accumulating Real and Synthetic Data

    The authors compare replacing prior datasets with newly generated samples against retaining earlier data and accumulating new samples. Replacement worsens error across the tested recursive settings, whereas accumulation keeps error bounded. Their language-model ablation also grows replacement datasets to match accumulation sizes; deterioration persists, though more slowly.

36. Language Models are Few-Shot Learners

    Few-shot inference supplies demonstrations as input conditioning while keeping model weights fixed; fine-tuning changes pretrained weights through training. Examples consume bounded context and influence subsequent predictions without becoming parameter updates. Separately, benchmark contamination means evaluation material overlaps training data, weakening claims of generalization to unseen examples. GPT-3's study compares original scores with subsets lacking detected n-gram overlap, but acknowledges false positives and possible distribution differences between clean and original subsets. Conceptually, contamination concerns exposure to evaluation data; optimizing a proxy concerns objective mismatch, while biased reviewer labels concern measurement. Those problems can occur independently.

37. Rethinking Benchmark and Contamination for Language Models with Rephrased Samples

    The authors generate paraphrased and translated benchmark cases that evade tested overlap detectors, then train models on those variants and observe inflated benchmark performance. Detection quality varies with subject, language, embedding model and threshold. Their dataset inspection also identifies rephrased benchmark cases in synthetic instruction datasets. Changing wording therefore does not reliably restore independence from an assessment case.

38. Generalization in Adaptive Data Analysis and Holdout Reuse

    Repeatedly choosing hypotheses or configurations using earlier holdout results can overfit the holdout itself. The paper identifies hyperparameter tuning, feature selection, stopping decisions and analyst feedback as routes through which a supposedly held-out sample influences the selected system.

39. imbalanced-learn: Common pitfalls and recommended practices

    Resampling a complete dataset before separating training and assessment can leak information: generating or selecting training examples may depend on records later assigned to the test set. It also changes the test distribution, potentially replacing naturally imbalanced conditions with an artificially balanced assessment. The documentation demonstrates keeping resampling within training folds through a pipeline.

40. When Will The Benchmaxxing Plague End?

    Public benchmark questions and answers can be memorized without deliberate test-set training; use private holdouts and investigate benchmark-specific recall.

41. Constitutional AI: Harmlessness from AI Feedback

    Constitutional AI constructs preference examples by presenting a feedback model with a prompt, two candidate responses and a written principle for choosing between them. The model's normalized option probabilities become supervision. The researchers supplied the principles and labeling demonstrations; generated preferences therefore reflect those specifications and the feedback model's interpretation. Their experiments replaced human harmlessness comparisons while retaining human helpfulness comparisons.

42. Holistic Evaluation of Language Models

    HELM organizes evaluation into scenarios and desired properties, measures multiple properties under standardized conditions, and documents missing coverage. It separates core scenarios from targeted evaluations that isolate particular skills and risks. This distinction supports reporting challenge-set results as diagnostic evidence rather than silently combining them with a representative-use score. A collection chosen to expose failures does not, merely by being diverse, estimate how often those failures occur in deployment.

43. Metamorphic Testing: A New Approach for Generating Next Test Cases

    The original report constructs follow-up tests from earlier input–output pairs using domain knowledge about expected behavior. Its binary-search example selects an element at a known position in a sorted array and uses that element as the next search key, providing an expected result without scanning the whole array. The authors present this as a complement to existing test-selection strategies, including cases where a general test oracle is unavailable.

44. Real-valued (Medical) Time Series Generation with Recurrent Conditional GANs

    Train on Synthetic, Test on Real evaluates generated data by training a predictor on synthetic labeled examples and testing it on held-out real examples. The published procedure splits real data before fitting the generator and allows generator optimization on a separate validation set. The reverse procedure, training on real and testing on synthetic, can miss a generator’s failure to cover modes because the generated test set omits those cases.

45. How to evaluate a model for your use case

    Evaluate performance on a large dataset independent of training data, and make evaluation part of the development workflow.

46. Ship Real Agents: Hands-On Evals for Agentic Applications

    Use uncertainty around the measured failure rate, rather than the point estimate alone, when deciding whether a reliability target is met.

47. Moondream: how does a tiny vision model slap so hard?

    Match training questions to real user input, including capitalization variation and typos.

48. How We Built Zeta2: Training an Edit Prediction Model in Production — Ben Kunkle, Zed

    Zeta2's distillation pipeline checks frontier-model predictions for task-specific failures and repairs flagged outputs before using them as student targets.

49. Don't be data poor

    Reverse the inference workflow: sample an outcome, select a compatible policy trace, and generate records conditioned on that trace.

50. OpenThoughts: Data Recipes for Reasoning Models

    Model-assessed difficulty and response length worked better than embedding-based or fast text filtering in these experiments, but the best signal depended on domain.

51. Moondream: how does a tiny vision model slap so hard?

    Use distractor questions to teach rejection of unsupported premises, but diversify them so the model cannot learn a topical shortcut.

52. Don't be data poor

    Use coarse-to-fine generation that mirrors how the underlying documents arise.

53. Don't be data poor

    Let domain experts steer generation at intermediate stages to create relevant failure scenarios and deployment-specific datasets.

54. The Messy Reality of Scale: Synthetic Data and Pre-Training

    Scaling the training budget without sufficiently expanding the corpus can repeat high-quality examples enough to cause early saturation; rephrasing was used to reduce that repetition.

55. The Messy Reality of Scale: Synthetic Data and Pre-Training

    A modular synthetic-data pipeline separates source inputs, metadata, generation, and supplementary validation so the same infrastructure can support both cheap transformations and complex workflows.

56. OpenThoughts: Data Recipes for Reasoning Models

    Teacher selection should use student outcomes: the speaker reports Qwen-32B was a stronger teacher than DeepSeek-R1 despite the latter's own benchmark strength.

57. Computer Use at the Edge of the Statistical Precipice

    Confidence intervals should account for both action stochasticity and environment variation while respecting the benchmark's structure.

58. How We Built Zeta2: Training an Edit Prediction Model in Production — Ben Kunkle, Zed

    The experimental 'settled data' approach captures eventual code as a supervision signal, but inactivity does not establish that it is the correct next edit.

59. How We Built Zeta2: Training an Edit Prediction Model in Production — Ben Kunkle, Zed

    Zeta2 evaluation uses held-out inputs, multiple teacher references, text-similarity scoring, and a separate reversal metric.