Research infrastructure should preserve knowledge, not just metrics.
Synthetic-data research increasingly depends on complex chains of generators, assets, datasets, training jobs and evaluation environments. The useful information is often distributed across scripts, logs, files, machines and experiment trackers.
SynthGraph is being built to make that history explicit.
The observation behind it is narrow and specific. In synthetic-data work, the configuration that produced a dataset is often more fragile than the dataset itself. A folder name records one parameter out of ten. A notebook cell is re-run. An asset is replaced on disk. Six weeks later the result is still there and the reason for it is not.
Experiment trackers solved this for training. They log metrics, hyperparameters and checkpoints extremely well. What they were not built to capture is what happened before the training run — the generator version, the scene parameters, the seed, the asset versions, the dataset version that came out the other side. That is the part SynthGraph is trying to hold onto.
What SynthGraph is
- A provenance and experiment-management system for synthetic-data research.
- A Python SDK that records what your existing tools already produce.
- A record of generations, dataset versions, training runs and evaluation results, and the relationships between them.
What SynthGraph is not
- It does not generate synthetic data.
- It does not replace W&B or MLflow, your simulator, or your training framework.
- It does not require datasets to be uploaded to it.
- It does not guarantee exact reproduction when external dependencies are missing.
- It does not optimise experiments on your behalf.
Current status
SynthGraph is in a private research pilot. It is being developed with a small number of research workflows before broader access. There are no customers, published benchmarks or shipped integrations to point at yet, and this site does not claim any. When that changes, this page will say so.
Your next experiment should leave a trail.
SynthGraph is currently available through a limited research pilot. Request a demo to see how provenance fits into your existing workflow.
SynthGraph is being developed with a small number of research workflows before broader access.