Training and testing large deep learning models often require substantial amounts of data. Even a large dataset might be missing important examples a model needs. Rare conditions may be poorly represented and observations may have incomplete labels. Synthetic data can fill these gaps with generated examples and improve the labels on existing observations.

Ample real-world evidence demonstrates the potential gains from adding synthetic training examples and improved labels to real datasets, but published results also highlight limitations in replacing real data entirely. Generated tests can expose failures, but they can also overstate performance or produce misleading comparisons between models. Prior to joining HASH, at Helsing I led the team generating images and video to train and test AI systems. In this post, I examine what these results suggest for generating process histories to train and evaluate HASH’s process foundation models.

What is the data gap?

A model needs data that covers the conditions it will meet. Scaling studies in language modeling and image recognition show that performance keeps improving with more training data when capacity and compute are sufficient. Each additional training example costs money and time to collect and annotate.

Collected datasets often contain many common cases and few rare ones. A driving dataset can have many hours of routine traffic but few examples of unusual hazards, such as an animal crossing the road. Obtaining enough examples to train and evaluate a model on those hazards can require collecting far more footage.

Some data exists but cannot be used (or accessed at all) because of confidentiality. Other data can be used but is labeled too coarsely for the task. For example, a single class label per image carries less information than object masks.

Generating data and labels

Building a dataset involves deciding which examples to include and what labels the task requires. Generative models can produce new images, video, or text from patterns learned during training. The generation process can be guided with prompts, spatial inputs and output constraints. Models can also predict or refine labels for existing examples.

Text prompts can describe objects, their attributes, and their spatial relationships but they provide limited control over exact scene structure:

1. ControlNet adds spatial conditioning to text-to-image diffusion models through inputs such as edge maps, depth maps, segmentation masks, and human poses.
2. Cosmos-Transfer1 provides similar control for video, using spatial inputs across frames.

Using these inputs in combination let us guide scene structure while varying the generated appearance.

For dataset generation, existing labels can become inputs to a generator. For example, a segmentation mask can guide the generation of images with similar layout but different appearances. This can reduce annotation work by reusing labels from an existing dataset.

By choosing the prompts and spatial inputs used for generation, we can target rare conditions and increase their proportions in a dataset. We still need to verify that the requested conditions appear in the generated samples and that the labels remain accurate.

Generative models can also produce the labels alongside the samples by using internal representations during inference:

1. MosaicFusion derives instance masks from cross-attention maps associated with object prompts.
2. InstaGen adds a trained grounding head that predicts object categories and bounding boxes from diffusion features.

These methods produce image-label pairs without requiring a predefined mask or bounding box for each new image.

Language models can generate particular kinds of training examples. AgentInstruct generates instruction-response pairs from documents and code using defined skills and task types. Its tool-use examples include tasks requiring multiple APIs and requests with missing parameters or unavailable tools.

Text generation can also be constrained to a specified output format:

1. Grammar-constrained decoding restricts which tokens a model can produce so that completed outputs follow a defined grammar.
2. Tools such as XGrammar support constraints derived from JSON schema. This enables specifying required fields and value types for generated records.

Factual accuracy and requirements outside those constraints still need separate checks.

Using synthetic data for training

Improving coverage where real data is scarce

Several vision studies report larger gains from generated training data when real examples are scarce:

1. InstaGen demonstrates this by varying the amount of real training data while retaining all COCO categories. With Faster R-CNN and a ResNet-50 backbone, adding generated images improved average precision (AP) by 5.2 points when 10% of the real training set was used. The gain fell to 1.7 points at 50% and 1.0 point with the full set.
2. MosaicFusion shows a similar pattern across categories. On LVIS, the authors added generated images and instance masks to the real training data for a Mask R-CNN with a ResNet-50-FPN backbone. Mask AP increased by 5.6 points on rare categories and 1.4 points overall. In the same setup, Copy-Paste augmentation improved rare-category mask AP by 1.0 point.
3. DiffuLT presents similar findings in image classification. It trains a diffusion model from scratch on a long-tailed dataset, without external data or pretrained weights. It generates examples for underrepresented classes and discards those too far from the real class centers in feature space. A classifier then trains on the combined dataset, with generated examples given less weight than real ones. On ImageNet-LT with a ResNet-50, this procedure improved accuracy over a standard cross-entropy baseline by 21.8 percentage points on medium-frequency classes and 33.6 points on rare classes. Accuracy on frequent classes fell by 0.7 points.

Generated images can also help when no real examples are available for adapting a pretrained classifier. He et al. used GLIDE to generate images from prompts describing the target classes. They filtered the generated images and trained a classifier on frozen CLIP-RN50 features. Across 17 datasets, their procedure improved top-1 accuracy by an average of 4.31 percentage points over zero-shot CLIP, without using real training images from those datasets for adaptation. In separate few-shot experiments, the gains from synthetic data generally diminished as the number of real training images increased from 1 to 16 per class.

Benefits also appear when the full real training set is available. Azizi et al. fine-tuned Imagen on ImageNet-1K and added 1.2 million generated images to the real training set. Top-1 accuracy increased from 76.39% to 78.17% for ResNet-50 and from 81.79% to 82.84% for DeiT-B. Separate scaling experiments found that some mixtures with larger proportions of generated images performed worse than the real-data baseline.

In these studies, synthetic data helped the most for rare classes and small real datasets. On the other hand, it only added 1 to 2 points to the performance when the full real set is available and some larger mixtures performed even worse than real data alone.

Self-generation of training data

Models can improve by learning from selected responses they generate. Algorithmic checks or learned judgments can help determine which outputs to use and those responses can address existing training problems:

1. STaR uses GPT-J (a model with six billion parameters) to generate reasoning steps and answers for questions whose correct answers are already known. Starting with ten example explanations, it retains outputs whose final answers match the known answers and uses them for fine-tuning. The generation and training process repeats. On the CommonsenseQA development set, this procedure achieved 68.8% accuracy. In comparison, when GPT-J is fine-tuned to predict answers directly, it achieved 60.0%. A variant that supplies the correct answer as a hint for initially failed questions achieved 72.5%. Those hints are removed from the training prompts. The answer check selects training examples without establishing that every reasoning step is correct.
2. ReST-EM applies a similar procedure to PaLM 2 models. It selects generated solutions by matching known answers on MATH and running test cases on APPS programming problems. With PaLM 2-L, the procedure improved test accuracy over the pretrained model by 6.34 percentage points on MATH and 6.4 points on APPS-Introductory. In a separate comparison using roughly 5,000 MATH questions that the model could solve, training on one accepted model-generated solution per question still outperformed training on one human solution per question. Further iterations brought smaller gains on MATH, while APPS performance declined in the second iteration. The authors suggest overfitting as a possible cause.
3. NVIDIA used a related approach to develop Nemotron-4 340B. Over 98% of the data used for supervised and preference fine-tuning was synthetic and only ~20,000 human-annotated examples used across the alignment process. The workflow began with data generated by Mixtral and then used intermediate Nemotron models to generate training data for later versions.

Generating training responses can reduce the need to write them by hand. It also requires deciding which outputs are suitable for training. Known answers and executable tests can guide that choice. The resulting model still needs evaluation on held-out problems to establish whether those examples helped and whether further rounds of generation and training will continue to improve performance.

Improving labels for existing data

Images can contain more useful information than their annotations capture. Models can help create more detailed labels or refine existing ones. This eventually leads to a more informative model training:

1. ReLabel uses an EfficientNet-L2 teacher trained with additional JFT-300M data to predict class probabilities for different image regions. Training uses supervision corresponding to the region selected by each image crop. This increased ResNet-50’s ImageNet top-1 accuracy from 77.5% to 78.9%. In an ablation with the same teacher, whole-image class probabilities reached 78.5% and one hard label per image reached 77.5%.
2. COCO-ReM combines mask refinement using the Segment Anything Model (SAM) with additional human annotations from LVIS. These changes improve mask boundaries and include objects missing from the original annotations. Training Mask R-CNN in the ViTDet-B configuration on the revised annotations increased mask AP from 50.5 to 55.2, a gain of 4.7 points. Both models used the same training images and were evaluated on the COCO-ReM validation set. The 4.7 point gain comes from SAM refinement and the added LVIS annotations together.

How far real data replacement goes

Generated data can supply all the images used in a particular training stage. Whether it can replace a real dataset depends on the performance required. Here, “synthetic-only” describes the training images. The generator, annotations, or pre-trained weights may still depend on real data.

For image classification, Sariyildiz et al. trained a ResNet-50 from scratch on a synthetic ImageNet-1K dataset and reached 42.9% top-1 accuracy on real validation images. Their strongest real-data baseline reached 80.1%. The training sets had the same size but the models used different training recipes, so the comparison does not isolate the effect of replacing real images.

The same synthetic-trained encoder was more competitive when it was used to extract features for other tasks. With a new linear classifier trained on real labeled examples for each task, it achieved 68.4% average accuracy across ten transfer datasets, compared with 66.5% for that real-data baseline. This supports replacing the encoder’s pre-training images with synthetic images in that setting but real samples are still necessary for downstream classifier training.

ODGEN provides a comparison with equal training-set sizes for object detection. The researchers trained YOLOv7 from scratch on either 10,000 real COCO images or 10,000 generated images using the corresponding bounding-box annotations. Evaluated on the same 31,000 held-out real images, the models reached 26.3 and 14.2 AP, respectively, averaged over intersection over union (IoU) thresholds from 0.50 to 0.95. Synthetic images remained substantially less effective in this experiment. The generator itself was trained on real COCO data, and generation used annotations from real images.

Dataset Diffusion shows how label quality can limit the value of generated images. Training DeepLabV3 with a ResNet-101 backbone on 40,000 generated images and the method’s segmentation masks achieved a mean intersection over union (mIoU) of 64.8 on the PASCAL VOC validation set. Replacing those masks with predictions from a segmentation model trained on real VOC data raised the result to 72.6 mIoU, a gain of 7.8 points. The same paper reports 79.9 mIoU for the same architecture trained on approximately 11,500 real VOC images and evaluated on the same validation set. The larger generated dataset remained below this baseline even after receiving labels from a model trained on real data.

Taken together with the training results above, these results show that synthetic images work best alongside real ones. Adding synthetic data to real training data raised accuracy, especially when real examples were scarce. Training on synthetic data alone scored below the real data baselines. The only exception was pre-training where synthetic images replaced the real ones without a loss on transfer tasks. Every setup still depended on real data for the generator, the labels, or the downstream classifier.

Evaluating with synthetic data

Synthetic data can address gaps in evaluation as well as training. It allows developers to construct tests around particular conditions and vary inputs systematically. These uses require checking what the generated examples measure and how closely they represent the conditions of interest.

Controlled variations can reveal weaknesses hidden by an aggregate benchmark score. GSM-Symbolic generated maths problems from templates based on GSM8K questions. Their GSM-NoOp variant adds distracting details that leave the correct calculation unchanged. For example, a question about the total number of harvested fruits mentions that some are smaller than average. The model must recognize that fruit size does not affect the count. At the GPT-4o model tests mean accuracy fell from 94.9% on GSM-Symbolic to 63.1% on GSM-NoOp, a decline of 31.8 percentage points. Generating related test cases made it possible to examine sensitivity to distracting information systematically.

Generated tests can also improve estimates of how a model will perform in a different population. 3S Testing used U.S. patient data and summary information about a U.K. population to generate synthetic test records. First, researchers curated a real held-out test data set from the U.S. patients to approximate the U.K. population’s known characteristics. Then, they trained seven models on U.S. patients data, separate from the held-out test data. Finally, they checked whether synthetic test records or real U.S. test records better predicted how accurately models would perform on U.K. patients. Synthetic records gave closer estimates to U.K. patients for six of the seven models. For one model, the estimate was off by 3.0 percentage points with synthetic records, compared with 9.6 points with the selected U.S. records. This improved the estimates without changing the models themselves. The approach assumes that relationships learned from U.S. data also hold in the U.K. population.

Splitting one generated dataset into training and test sets can preserve errors shared by both. A model may therefore perform well on the synthetic test set while performing substantially worse on real data. Van Breugel et al. trained classifiers on synthetic data and held each classifier fixed while changing its evaluation data. On SEER, mean AUC across 20 runs was 0.877 on test data from the same generator, compared with 0.711 on held-out real data. Their method using separately trained generators for evaluation brought the estimate closer, to 0.743 with five generators.

Human checks on generated questions do not establish that a synthetic benchmark preserves the original comparison between models. Gill et al. generated questions and passage variations for a sample of CondaQA, which tests reasoning about negation. They evaluated examples that passed validity checks and had agreement between human annotators on the answers. Their measure, “full-bundle consistency”, counts a question as successful only when the model answers it correctly across all four passage versions. Averaged over five runs, Llama-3.3-70B-Instruct scored 53.0% on the original sample, compared with GPT-4o’s 32.6%. On the generated version, GPT-4o scored higher: 50.9% versus Llama’s 40.9%. Replacing the real test data with synthetic therefore reversed which model performed better, despite the human checks.

Synthetic evaluation therefore needs to:

1. Define a clear purpose and evidence that its construction supports that purpose.
2. Check generated examples and labels.
3. Report stress tests separately from everyday performance.
4. Check for shared generator errors.
5. Compare scores and model ranking with real data.

Synthetic data at HASH

At HASH, our work on graph-based world models explores learning useful representations from graph structure, as well as the histories of graph states and actions, with a view to better predicting future changes and outcomes. The coverage results reviewed in this post point to one simple hypothesis we’ll be exploring: adding generated histories of underrepresented conditions to the training data could improve performance on those conditions.

Consider a purchasing process in which a request can be returned for missing information. The available data might contain many direct approvals and few examples of requests that need changes before they can be approved. The request, requester, and approver could be represented as related entities. Their attributes and relationships could record the information available and responsibility for the next step.

Language models could help generate histories of these less common cases. They can be guided by observed histories and explicit rules to produce new synthetic data that follow the same structure as the real process. Our text-to-simulation work describes how language models can be leveraged to turn process descriptions into executable models. One possibility is to use such models to simulate variations in the missing information.

To test the value of these examples, we will explore comparing training on real data alone with training that includes synthetic examples, using the same model and training budget. Evaluation on held-out real data will measure performance on both targeted conditions and common cases. Further tests will check whether training on generated histories also improves predictions for related processes.

Generated tests could also examine how models respond when information is missing or a request is resubmitted. We will explore whether those tests expose relevant failures and whether their performance estimates agree with independent real observations.

The studies reviewed here show several ways to address gaps in data: generate additional training examples, improve existing labels, and construct tests for specific conditions. For HASH’s process foundation models, they motivate testing whether generated histories of uncommon events and action sequences improve predictions beyond training on the available real histories alone. Better performance on those cases, while maintaining performance on common ones, will provide evidence for using synthetic histories in training. Independent real observations remain essential for establishing whether those improvements carry over to the processes we want to model.