Datasets
Training data, open by default.
High-quality synthetic data, generated and validated by our pipeline — released openly with the dataset card, the quality breakdowns, and the harness to reproduce every verdict.
The pipeline
Every sample survives a gauntlet.
Four containerized pass/fail stages sit between raw generation and release — no model in the loop, so every verdict is reproducible.
Parse
Malformed generations are rejected at the door — nothing unparseable gets in.
Compile
Every sample builds in its language's real toolchain, not a linter approximation.
Execute
Samples run against generated tests and have to pass them to survive.
Mutate
A mutation check catches tests that pass vacuously — green suites that prove nothing.
~40%
of raw generations fail somewhere in the gauntlet and never reach a release — and that rejected slice is exactly the data you don't want in a training run.
In the box
Tokens are the smallest part of a release.
A dataset you can't interrogate is a liability — so everything needed to trust or rebuild ours ships alongside it.
The dataset card
Per-sample provenance metadata and the verdicts from every pipeline stage.
The validation harness
The same containers we ran — regenerate a slice, extend the pipeline, or re-verify our verdicts.
Quality breakdowns
Per-language and per-slice quality reports, so you can filter to what matters for your run.
Releases
Released so far.
Dataset releases land in the research registry first — check back soon.
Custom data
Need data built for your domain?
The same pipeline that produces our open releases can generate and validate data for your training runs — commissioned, filtered, and delivered with its harness.