376 verified (spec -> reasoning -> design) traces for neural-architecture
design, minted by grok-4 and filtered by a deterministic verifier.
Each row pairs a natural-language architecture spec with grok-4's reasoning and
its final structured design (a batch of typed graph edits). Every row's design
was applied to the task's start graph and re-graded by the deterministic
verifier of neurarch-arch-bench:
no trace enters this set unless its design provably… See the full description on the dataset page:
https://huggingface.co/datasets/neurarch-ai/arch-reasoning-grok.
