N Neurarch Architectures Models Checks Data Docs Open the app

Resources / Datasets

Arch-Bench task set

The task definitions behind the benchmark: design-from-spec and repair-and-extend instances for agents that build neural network architectures. Each task pairs a natural-language design brief with a starting graph the agent edits and a set of programmatic pass criteria (structural blockers, parameter budgets and bands, required layer families on an input-to-output path, KV cache and decode-latency ceilings). Twelve curated tasks with eight starting fixtures, plus a deterministic generator that mints a larger split from a seed.

12 curated tasks, 8 fixtures MIT Free to use The environment Source repo On Hugging Face

Get it

Task definitionsraw.githubusercontent.com/neurarch-ai/neurarch-arch-bench/main/tasks.json
application/json
Hugging Face mirror (dataset viewer, load_dataset)huggingface.co/datasets/neurarch-ai/arch-bench-tasks
text/html
 
curl -s https://raw.githubusercontent.com/neurarch-ai/neurarch-arch-bench/main/tasks.json | jq '.tasks[] | {id, spec, constraints}'

What is in a row

task id design spec starting graph pass constraints difficulty

What this dataset is not

Twelve curated tasks is a seed, not a set you can rank models on with confidence. The generated split exists because of that, and the numbers worth comparing are the generated-split ones. The curated tasks are best read as the worked examples that show what a task is.

Licence and citation

Released under MIT License. Cite it as:

Neurarch. Arch-Bench task set. https://neurarch.com/d/bench-tasks.html

The rest of the set

Neurarch architecture corpus
48 neural network architectures kept as typed graphs rather than diagrams.
48 architectures · CC0-1.0
Neurarch structural check catalogue
The 50 structural checks Neurarch runs on a model graph, as data: id, severity, category, the condition that triggers it, why it costs something, and the fix.
50 checks · CC-BY-4.0
Verifier grounding study (264 graphs)
Clean reference architectures plus systematically corrupted variants (broken attention head divisibility, linear width mismatches, severed connections), each built as a real PyTorch model and run on a GPU.
264 graphs · MIT
Arch-Bench arena results
Frontier language models scored on design-from-spec tasks by a deterministic verifier rather than a human or an LLM judge.
18 model-split results · CC-BY-4.0
arch-design-sft: verified architecture-design SFT data
Supervised fine-tuning data for neural architecture design treated as structured graph editing.
3,010 verified examples · MIT
Structure, verdict and trained outcome triples
The corpus that pairs what a design looks like with what it did.
80 trained graphs · MIT
Design selection replay: two trained campaigns and every selector tried
The records needed to score a design selector offline, in the shape of xRouteBench: every candidate design was trained once and the outcome kept, so a rule that picks which unexecuted design gets the GPU is evaluated by replay against the same rows as every previous rule, with no GPU and no model call.
39 trained designs · CC-BY-4.0
Who wrote the design: verifier refusal rate by author
The same pre-flight verifier applied to model designs written by three kinds of author, with the refusal rate per author kept apart.
7 rows · CC-BY-4.0
Verified architecture-design reasoning traces (Claude)
Spec to reasoning to design triples where the design is re-graded by the same deterministic verifier the benchmark uses, and only passing traces are kept.
306 verified traces · MIT
Verified architecture-design reasoning traces (Grok)
The same verified spec to reasoning to design triples as the Claude split, rejection-sampled from a different frontier model, so the two can be pooled for volume or held apart to see how much of the reasoning style is model-specific.
376 verified traces · MIT