Use case · ML evaluation

Realistic, labeled data for the cases your model barely sees.

The events that make or break a model - rare anomalies, edge cases, failure modes - are exactly the ones your data has too few of, or none at all. Rockfish generates realistic, labeled, domain-specific time-series that amplify those rare classes and fill the gaps, so you can stress-test your model honestly and improve its accuracy where it counts.

Book a demo → Get a trial key
The problem

Your model is only as good as the data it has for the cases that matter.

Time-series data is imbalanced by nature - the events that matter most are the rarest, and some scenarios haven't happened yet at all. That caps both how well you can test a model and how accurate you can make it:

Rare by nature

The critical events barely appear in your data.

A handful of anomalies across months of traffic isn't enough to measure recall on - let alone enough for the model to learn the pattern well. Overall accuracy stays high while the rare class quietly fails.

Not yet seen

Some cases haven't happened - yet.

The outage signature, the new drift, the correlated sensor dropout you'll hit next quarter isn't in any dataset you own. You can't test for it, and the model has never had a chance to handle it.

› "we have 6 real examples of this failure mode"
When 1 in 10,000 records is the one that matters, overall accuracy tells you almost nothing.

The fix for a rare class is more of it - realistic, labeled, and true to how the event actually behaves. That's what Rockfish generates. See how it works →

How it works

Turn scarce data - or none at all - into a rich, labeled dataset.

Rockfish works from whatever you have - a sample of your real time-series, or just a schema - and generates realistic, labeled data that amplifies your rare classes and covers the scenarios you don't have.

1

Start with what you have

  • A sample of real time-series - or just a schema
  • Fields, types, ranges, and relationships
  • No full data export required
2

Amplify + generate

  • Turn a few rare examples into many realistic, labeled ones
  • Add edge cases and scenarios you've never recorded
  • Balanced across the classes that matter
3

Test and improve

  • Score recall, false-positive rate, per-scenario coverage
  • Close the gaps with the labeled data you generated
  • Raise accuracy on the rare events, not just the average
Starting from zero

No data yet? Generate it from a schema.

Cold-start projects, brand-new products, privacy-locked domains - sometimes there's no historical data at all. Describe your fields, types, ranges, and relationships, and Rockfish generates realistic, labeled time-series that behave like the real thing - rare events included.

schema:
  timestamp    datetime   @ 1min
  region       category   [us-east, us-west, eu]
  latency_ms   float      ~ corr(traffic)
  error_rate   float      0.0 - 1.0
  event        label      [normal, spike, drift, dropout]

-> 30 days of labeled, realistic time-series
What you get

Stop letting scarce data cap your model.

Amplify rare classes

Turn a handful of real examples into a balanced, labeled set that reflects how the event actually behaves.

Cover missing scenarios

Fill the edge cases and failure modes your data has never recorded with realistic, domain-specific series.

Improve accuracy where it counts

Raise recall on the rare events that matter instead of chasing an average that hides them.

Give your model the data it's missing.

Realistic, labeled, domain-specific time-series that amplify your rare classes and cover the gaps - so your model performs on the events that matter, not just the average.

Book a demo → Get a trial key