N Neurarch Architectures Checks Data Docs Open the app

Resources / Datasets

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. Every row pairs the static verifier verdict with what actually happened at runtime: whether the module constructed, whether the forward pass survived, whether training made progress, and the initial and final loss. 264 graphs, two seeds, torch 2.8.

264 graphs MIT Free to use What the checks intercepted Source repo On Hugging Face

Get it

Per-graph resultsraw.githubusercontent.com/neurarch-ai/neurarch-arch-bench/main/grounding_results.csv
text/csv
Hugging Face mirror (dataset viewer, load_dataset)huggingface.co/datasets/neurarch-ai/arch-verifier-grounding-264
text/html
 
curl -sO https://raw.githubusercontent.com/neurarch-ai/neurarch-arch-bench/main/grounding_results.csv
python -c "import pandas as pd; d=pd.read_csv('grounding_results.csv'); print(d.groupby('verifierBlocked').forward_ok.mean())"

What is in a row

verifier score verifier blocked constructs forward pass ok makes training progress initial loss final loss

What this dataset is not

The corruptions are injected, not sampled from real mistakes, so the failure modes in here are the ones we chose to plant. This settles that a blocked graph does not run. It says nothing about how often an engineer writes one, and a paper quoting it as a base rate would be quoting it wrong.

Licence and citation

Released under MIT License. Cite it as:

Neurarch. Verifier grounding study (264 graphs). https://neurarch.com/d/grounding-264.html

The rest of the set

Neurarch architecture corpus
36 neural network architectures kept as typed graphs rather than diagrams.
36 architectures · CC0-1.0
Neurarch structural check catalogue
The 41 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.
41 checks · CC-BY-4.0
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-Bench task set
The task definitions behind the benchmark: design-from-spec and repair-and-extend instances for agents that build neural network architectures.
12 curated tasks, 8 fixtures · MIT
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
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