For AI labs
Data you can’t scrape.
Private datasets from operating companies, prepared for post-training and evals.
Private datasets from operating companies, prepared for post-training and evals.
None of it has been on the public web. Exclusive terms are available.
Enough to judge it before any contract. Shared under NDA.
A start state, the tools allowed, a verifier and a held-out split.
# task spec, example { "task": "take PR to merge", "start": "repo@pre-fix", "tools": ["shell", "git", "tests"], "verifier": "test suite passes", "split": "held-out" }
Signed before anything is shared.
Anonymous data cards by category.
Test a sample. No training on it.
Signed directly with the data owner.
Tell us the capability you are working on.