Dataset Viewer
Auto-converted to Parquet Duplicate
key
stringlengths
3
17
name
stringlengths
4
30
category
stringclasses
9 values
description
stringlengths
38
145
whenToUse
stringlengths
116
229
layers
int64
4
434
params
int64
2.72k
666B
paramsHuman
stringlengths
4
7
inputShape
listlengths
1
3
outputShape
listlengths
1
3
verdict
stringclasses
2 values
findings
listlengths
0
8
page
stringlengths
31
45
markdown
stringlengths
29
43
graph
stringlengths
45
59
openInApp
stringlengths
34
48
eegnet
EEGNet
Biosignal
Compact CNN for EEG/BCI — depthwise + separable convs make it 10× lighter than standard CNNs (Lawhern 2018)
Pick when channel count is low (4–8) and labelled trials are scarce (<400). Strong default for motor imagery / P300 BCI on consumer headsets.
16
2,724
2.7K
[ 1, 22, 1000 ]
[ 4 ]
pass
[]
https://neurarch.com/a/eegnet.html
https://neurarch.com/a/eegnet.md
https://neurarch.com/templates/eegnet/model.json
https://neurarch.com/?template=eegnet
eeg-conformer
EEG Conformer
Biosignal
Conv stem + Transformer encoder — SOTA for high-channel motor imagery EEG (Song 2023)
Pick when you have ≥16 channels and ~400+ trials per subject. Best published accuracy on BCI IV-2a/2b; expect tricky regularization.
23
78,564
78.6K
[ 1, 22, 1000 ]
[ 4 ]
warn
[ { "rule": "attention-no-pe", "severity": "warn" } ]
https://neurarch.com/a/eeg-conformer.html
https://neurarch.com/a/eeg-conformer.md
https://neurarch.com/templates/eeg-conformer/model.json
https://neurarch.com/?template=eeg-conformer
patch-tst
PatchTST
Time-series
Channel-independent patching + Transformer for multivariate time-series (Nie 2023)
Pick for multivariate time-series with long windows where channels can be treated independently. Same backbone serves classification and forecasting.
19
395,396
395.4K
[ 1, 22, 1000 ]
[ 4 ]
pass
[]
https://neurarch.com/a/patch-tst.html
https://neurarch.com/a/patch-tst.md
https://neurarch.com/templates/patch-tst/model.json
https://neurarch.com/?template=patch-tst
cnn-lstm-1d
1D CNN + LSTM
Time-series
Conv1D + LSTM baseline for ECG/PPG/IMU and other long-form physio signals
Pick for long-form 1D physiological signals (ECG, PPG, IMU) where local morphology + temporal context both matter. Solid baseline before reaching for transformers.
14
311,749
311.7K
[ 12, 5000 ]
[ 5 ]
pass
[]
https://neurarch.com/a/cnn-lstm-1d.html
https://neurarch.com/a/cnn-lstm-1d.md
https://neurarch.com/templates/cnn-lstm-1d/model.json
https://neurarch.com/?template=cnn-lstm-1d
simple-cnn
Simple CNN
Computer Vision
Simple Convolutional Neural Network for image classification
Pick as a fast baseline for small images (≤64px, e.g. CIFAR). Trains in minutes, easy to debug — use before reaching for ResNet.
9
804,554
804.6K
[ 1, 28, 28 ]
[ 10 ]
pass
[ { "rule": "deep-no-norm", "severity": "info" } ]
https://neurarch.com/a/simple-cnn.html
https://neurarch.com/a/simple-cnn.md
https://neurarch.com/templates/simple-cnn/model.json
https://neurarch.com/?template=simple-cnn
resnet-block
ResNet Block
Computer Vision
ResNet residual block with skip connections
Pick when you need a depth-friendly CV backbone. Stack 2–4 blocks for CIFAR, or use as the building block of ResNet-18/50 for ImageNet-scale.
9
73,984
74.0K
[ 64, 32, 32 ]
[ 64, 32, 32 ]
pass
[]
https://neurarch.com/a/resnet-block.html
https://neurarch.com/a/resnet-block.md
https://neurarch.com/templates/resnet-block/model.json
https://neurarch.com/?template=resnet-block
unet
U-Net
Computer Vision
Encoder-decoder with skip connections — Ronneberger et al. 2015. The standard for biomedical and small-data image segmentation.
Pick for image segmentation when training data is limited (<10k images). Skip connections preserve fine spatial detail that pure encoder-decoders lose. Default for medical imaging, satellite, and any pixel-level binary mask task.
24
720,705
720.7K
[ 3, 256, 256 ]
[ 1, 256, 768 ]
warn
[ { "rule": "deep-no-residual", "severity": "warn" } ]
https://neurarch.com/a/unet.html
https://neurarch.com/a/unet.md
https://neurarch.com/templates/unet/model.json
https://neurarch.com/?template=unet
vit-b16
ViT-B/16
Computer Vision
Vision Transformer — patch embedding stem + 1 encoder block (768D, 12 heads)
Pick for 224px+ images when pretrained weights are available, or when dataset is large enough (>1M images) to train from scratch.
13
8,449,000
8.45M
[ 3, 224, 224 ]
[ 196, 1000 ]
pass
[ { "rule": "dropout-before-bn", "severity": "info" } ]
https://neurarch.com/a/vit-b16.html
https://neurarch.com/a/vit-b16.md
https://neurarch.com/templates/vit-b16/model.json
https://neurarch.com/?template=vit-b16
transformer-block
Transformer Block
NLP/LLM
Transformer encoder block (simplified)
Pick as a generic encoder building block for sequence models when you don't need the modern LLM stack (RoPE/GQA/RMSNorm). Good teaching baseline.
8
7,087,872
7.09M
[ 512, 768 ]
[ 512, 768 ]
warn
[ { "rule": "bn-at-output", "severity": "warn" }, { "rule": "attention-no-pe", "severity": "warn" } ]
https://neurarch.com/a/transformer-block.html
https://neurarch.com/a/transformer-block.md
https://neurarch.com/templates/transformer-block/model.json
https://neurarch.com/?template=transformer-block
bert-base
BERT Base
NLP/LLM
BERT-Base encoder — bidirectional MHA (12 heads, 768D), no causal mask, 30K vocab
Pick for text classification / NLU with limited labels — pretrained encoder + small head fine-tunes reliably. 512-token cap.
11
31,120,896
31.12M
[ 1, 512 ]
[ 1, 512, 768 ]
pass
[]
https://neurarch.com/a/bert-base.html
https://neurarch.com/a/bert-base.md
https://neurarch.com/templates/bert-base/model.json
https://neurarch.com/?template=bert-base
gpt2
GPT-2
NLP/LLM
GPT-2 Small — causal transformer block (768D, 12 heads, 4× FFN)
Pick for single-GPU language modeling experiments and as a teaching reference for the canonical decoder-only stack. Modern LLMs prefer LLaMA-3 / Phi-3 blocks.
12
84,331,345
84.33M
[ 1, 1024 ]
[ 1, 1024, 50257 ]
pass
[]
https://neurarch.com/a/gpt2.html
https://neurarch.com/a/gpt2.md
https://neurarch.com/templates/gpt2/model.json
https://neurarch.com/?template=gpt2
llama3-block
LLaMA-3 Block
NLP/LLM
LLaMA-3 decoder block — GQA (32H/8KV), SwiGLU FFN, RMSNorm, residual streams
Pick when building a modern dense LLM (7B–70B class). Strongest open recipe per param; GQA cuts KV cache, RoPE handles long context.
10
702,554,112
702.55M
[ 1, 2048 ]
[ 1, 2048, 4096 ]
pass
[]
https://neurarch.com/a/llama3-block.html
https://neurarch.com/a/llama3-block.md
https://neurarch.com/templates/llama3-block/model.json
https://neurarch.com/?template=llama3-block
mixtral-block
Mixtral MoE Block
NLP/LLM
Mixtral decoder block — GQA + Sparse MoE (8 experts, top-2) + RMSNorm + RoPE
Pick when you have MoE training infra and want best quality per active param. Routing instability and memory cost (still scales with total params) are the trade-off.
9
1,451,270,144
1.45B
[ 1, 4096, 4096 ]
[ 1, 4096, 4096 ]
pass
[ { "rule": "moe-no-aux-loss", "severity": "info" } ]
https://neurarch.com/a/mixtral-block.html
https://neurarch.com/a/mixtral-block.md
https://neurarch.com/templates/mixtral-block/model.json
https://neurarch.com/?template=mixtral-block
t5-small
T5 Small
NLP/LLM
T5 encoder-decoder — bidirectional encoder + masked decoder with cross-attention (512D, 8 heads)
Pick for seq2seq tasks (summarization, translation, QA) where you need both bidirectional understanding and generation in one model.
23
56,734,592
56.73M
[ 1, 512 ]
[ 1, 128, 32128 ]
warn
[ { "rule": "attention-no-pe", "severity": "warn" } ]
https://neurarch.com/a/t5-small.html
https://neurarch.com/a/t5-small.md
https://neurarch.com/templates/t5-small/model.json
https://neurarch.com/?template=t5-small
whisper-small
Whisper Small
Audio
Whisper speech encoder-decoder — conv1d audio stem + transformer encoder/decoder (384D)
Pick for ASR or as a pretrained audio encoder — drop the decoder + add a head for audio classification (UrbanSound, ESC-50).
18
46,986,649
46.99M
[ 1, 80, 3000 ]
[ 1, 448, 51865 ]
pass
[]
https://neurarch.com/a/whisper-small.html
https://neurarch.com/a/whisper-small.md
https://neurarch.com/templates/whisper-small/model.json
https://neurarch.com/?template=whisper-small
simple-rnn
Simple RNN
NLP
Simple Recurrent Neural Network for sequence processing
Pick as a teaching reference. Real workloads should reach for LSTM/GRU or transformer — vanilla RNN suffers from vanishing gradients on anything beyond ~50 steps.
4
1,100,298
1.10M
[ 128, 300 ]
[ 128, 10 ]
pass
[]
https://neurarch.com/a/simple-rnn.html
https://neurarch.com/a/simple-rnn.md
https://neurarch.com/templates/simple-rnn/model.json
https://neurarch.com/?template=simple-rnn
mamba-block
Mamba SSM Block
NLP/LLM
Mamba State Space Model — selective SSM + causal conv gating, no attention (O(T) complexity)
Pick for very long sequences where attention's O(T²) cost is the bottleneck (DNA, audio, long-context LM). Tuning is trickier than transformers.
18
168,271,976
168.27M
[ 1, 1024 ]
[ 1, 1024, 50280 ]
pass
[]
https://neurarch.com/a/mamba-block.html
https://neurarch.com/a/mamba-block.md
https://neurarch.com/templates/mamba-block/model.json
https://neurarch.com/?template=mamba-block
phi3-mini
Phi-3 Mini Block
NLP/LLM
Phi-3 Mini 3.8B decoder block — full MHA (32H, 3072D), SwiGLU FFN (8192), RMSNorm, RoPE
Pick for compact LLMs (≤4B params) when you want modern ingredients (RoPE, SwiGLU) without the full LLaMA-3 footprint.
12
310,288,704
310.29M
[ 1, 2048 ]
[ 1, 2048, 32064 ]
pass
[]
https://neurarch.com/a/phi3-mini.html
https://neurarch.com/a/phi3-mini.md
https://neurarch.com/templates/phi3-mini/model.json
https://neurarch.com/?template=phi3-mini
two-tower
Two-Tower
Recommendation
User+Item dual encoder for retrieval — embeddings → MLP per side → dot product score
Pick for retrieval at scale (billions of items) where item embeddings can be precomputed and indexed. Not suitable for re-ranking — no cross-features.
12
70,433,152
70.43M
[ 1 ]
[ 1, 64 ]
pass
[ { "rule": "deep-no-norm", "severity": "info" } ]
https://neurarch.com/a/two-tower.html
https://neurarch.com/a/two-tower.md
https://neurarch.com/templates/two-tower/model.json
https://neurarch.com/?template=two-tower
wide-and-deep
Wide & Deep
Recommendation
Memorization (wide linear) + generalization (deep MLP) joint trained — Cheng et al. 2016
Pick as a simpler, more interpretable alternative to DLRM for CTR. Linear part captures memorized rules, deep part generalizes.
13
3,652,882
3.65M
[ 10000 ]
[ 1 ]
pass
[ { "rule": "output-activation", "severity": "info" }, { "rule": "vanishing-gradient", "severity": "info" }, { "rule": "deep-no-norm", "severity": "info" } ]
https://neurarch.com/a/wide-and-deep.html
https://neurarch.com/a/wide-and-deep.md
https://neurarch.com/templates/wide-and-deep/model.json
https://neurarch.com/?template=wide-and-deep
dlrm
DLRM
Recommendation
Meta's Deep Learning Recommendation Model — bottom MLP for dense, embedding for sparse, feature interaction, top MLP
Pick when you have substantial dense + sparse features and need a production-validated CTR baseline. Bare-bones — needs feature engineering.
14
32,347,553
32.35M
[ 13 ]
[ 32, 1 ]
pass
[ { "rule": "output-activation", "severity": "info" }, { "rule": "vanishing-gradient", "severity": "info" }, { "rule": "deep-no-norm", "severity": "info" }, { "rule": "init-activation-mismatch", "severity": "info" } ]
https://neurarch.com/a/dlrm.html
https://neurarch.com/a/dlrm.md
https://neurarch.com/templates/dlrm/model.json
https://neurarch.com/?template=dlrm
neumf
Neural Collaborative Filtering
Recommendation
He et al. 2017 NeuMF — GMF + MLP fused for recommendation
Pick for pure user/item collaborative filtering when you have no side features. Fused GMF+MLP variant; concat-only variant is `ncf`.
16
105,610,401
105.61M
[ 1 ]
[ 1, 1 ]
pass
[ { "rule": "output-activation", "severity": "info" }, { "rule": "vanishing-gradient", "severity": "info" }, { "rule": "deep-no-norm", "severity": "info" }, { "rule": "init-activation-mismatch", "severity": "info" } ]
https://neurarch.com/a/neumf.html
https://neurarch.com/a/neumf.md
https://neurarch.com/templates/neumf/model.json
https://neurarch.com/?template=neumf
graph-sage-rec
GraphSAGE Recommender
Recommendation
Inductive node embeddings via neighbor sampling + aggregation — for graph-based recommenders (PinSage style)
Pick when your recsys has a rich item-item or user-item graph and cold-start items must generalize via neighbors (PinSage-style production setup).
10
263,040
263.0K
[ 128 ]
[ 128, 64 ]
warn
[ { "rule": "bn-at-output", "severity": "warn" } ]
https://neurarch.com/a/graph-sage-rec.html
https://neurarch.com/a/graph-sage-rec.md
https://neurarch.com/templates/graph-sage-rec/model.json
https://neurarch.com/?template=graph-sage-rec
ncf
Neural Collaborative Filtering
Recommendation
He et al. 2017 NCF — user/item embeddings → concat → MLP → score (concat-then-MLP variant)
Pick as the simplest deep CF baseline before reaching for fused (NeuMF) or graph-based variants. No feature engineering required.
12
35,206,273
35.21M
[ 1 ]
[ 1, 1 ]
pass
[ { "rule": "output-activation", "severity": "info" }, { "rule": "vanishing-gradient", "severity": "info" }, { "rule": "deep-no-norm", "severity": "info" }, { "rule": "init-activation-mismatch", "severity": "info" } ]
https://neurarch.com/a/ncf.html
https://neurarch.com/a/ncf.md
https://neurarch.com/templates/ncf/model.json
https://neurarch.com/?template=ncf
lightgcn
LightGCN
Recommendation
He et al. 2020 — user/item embeddings propagated through 3 light graph-conv layers, layer combination via mean (no transforms, no nonlinearities)
Pick for collaborative filtering with implicit feedback (clicks, plays). Strips graph-conv to its essentials — often beats heavier GCN variants on rec benchmarks.
14
70,424,960
70.42M
[ 1 ]
[ 64 ]
pass
[ { "rule": "deep-no-norm", "severity": "info" } ]
https://neurarch.com/a/lightgcn.html
https://neurarch.com/a/lightgcn.md
https://neurarch.com/templates/lightgcn/model.json
https://neurarch.com/?template=lightgcn
bst
Behavior Sequence Transformer
Recommendation
Alibaba 2019 BST — user behavior sequence + target → Transformer encoder + concat with user features → MLP → CTR
Pick for sequential CTR when user behavior history matters (session intent). Production-deployed at Alibaba scale; needs sequence-aware features.
20
128,756,259
128.76M
[ 50 ]
[ 1 ]
pass
[ { "rule": "output-activation", "severity": "info" }, { "rule": "vanishing-gradient", "severity": "info" }, { "rule": "deep-no-norm", "severity": "info" }, { "rule": "consecutive-linear-no-activation", "severity": "info" }, { "rule": "init-activation-mismatch", ...
https://neurarch.com/a/bst.html
https://neurarch.com/a/bst.md
https://neurarch.com/templates/bst/model.json
https://neurarch.com/?template=bst
sli-rec
SLi-Rec
Recommendation
Yu et al. 2019 — Short and Long-term Interest Recommender. Time-LSTM + ASVD attention fused via gate
Pick when both short-session intent and long-term preferences matter and you want each modelled separately, then fused via a learned gate.
20
64,049,986
64.05M
[ 50 ]
[ 1 ]
warn
[ { "rule": "output-activation", "severity": "info" }, { "rule": "attention-no-pe", "severity": "warn" }, { "rule": "vanishing-gradient", "severity": "info" }, { "rule": "vanishing-gradient", "severity": "info" }, { "rule": "deep-no-norm", "severity": "info" }...
https://neurarch.com/a/sli-rec.html
https://neurarch.com/a/sli-rec.md
https://neurarch.com/templates/sli-rec/model.json
https://neurarch.com/?template=sli-rec
diffusion-unet
Diffusion UNet
Generative
Stable-Diffusion-style noise predictor — latent UNet with cross-attention to a text embedding
Pick when you want to generate images from a text prompt. The full pipeline also needs a VAE encoder/decoder and a text encoder (e.g. CLIP); this template is the denoiser core.
19
6,687,044
6.69M
[ 4, 64, 64 ]
[ 4, 64, 64 ]
pass
[]
https://neurarch.com/a/diffusion-unet.html
https://neurarch.com/a/diffusion-unet.md
https://neurarch.com/templates/diffusion-unet/model.json
https://neurarch.com/?template=diffusion-unet
deepseek-v3
DeepSeek-V3
NLP/LLM
671B MoE LLM — Multi-head Latent Attention (MLA) compresses the KV cache; fine-grained MoE with shared + routed experts (DeepSeek 2024)
Study a frontier MoE design: MLA for cheap long-context inference and shared-expert routing. Fold collapses the 61 repeated decoder layers to one block; expand to see the full depth.
372
666,329,115,904
666.33B
[ 1, 163840 ]
[ 1, 163840, 129280 ]
pass
[ { "rule": "deep-attention-default-init", "severity": "info" } ]
https://neurarch.com/a/deepseek-v3.html
https://neurarch.com/a/deepseek-v3.md
https://neurarch.com/templates/deepseek-v3/model.json
https://neurarch.com/?template=deepseek-v3
llama-4-scout
Llama-4 Scout
NLP/LLM
109B natively-multimodal MoE LLM — interleaved dense/MoE layers (16 experts) with iRoPE (interleaved no-RoPE) for long context (Meta 2025)
Reference the newest open MoE decoder: alternating dense and expert-routed blocks. A strong base for studying sparse-activation LLMs.
434
102,547,680,704
102.55B
[ 1, 10485760 ]
[ 1, 10486336, 202048 ]
warn
[ { "rule": "attention-no-pe", "severity": "warn" }, { "rule": "moe-no-aux-loss", "severity": "info" }, { "rule": "huge-linear-params", "severity": "warn" }, { "rule": "deep-attention-default-init", "severity": "info" } ]
https://neurarch.com/a/llama-4-scout.html
https://neurarch.com/a/llama-4-scout.md
https://neurarch.com/templates/llama-4-scout/model.json
https://neurarch.com/?template=llama-4-scout
jamba
Jamba
NLP/LLM
Hybrid SSM-Transformer-MoE — interleaves Mamba, attention, and MoE blocks in one stack (AI21 2024)
Pick to see three paradigms in one compact graph: state-space, attention, and mixture-of-experts. Good for studying hybrid long-context designs.
53
13,030,961,152
13.03B
[ 1, 4096 ]
[ 1, 4096, 65536 ]
warn
[ { "rule": "attention-no-pe", "severity": "warn" }, { "rule": "moe-no-aux-loss", "severity": "info" } ]
https://neurarch.com/a/jamba.html
https://neurarch.com/a/jamba.md
https://neurarch.com/templates/jamba/model.json
https://neurarch.com/?template=jamba
qwen3-8b
Qwen3-8B
NLP/LLM
Modern dense decoder LLM — GQA with QK-RMSNorm on the query/key projections for training stability (Alibaba 2025)
A clean, current dense-decoder reference. Pick when you want a straightforward modern LLM block without MoE routing.
221
8,190,878,080
8.19B
[ 1, 40960 ]
[ 1, 40960, 151936 ]
warn
[ { "rule": "attention-no-pe", "severity": "warn" }, { "rule": "deep-attention-default-init", "severity": "info" } ]
https://neurarch.com/a/qwen3-8b.html
https://neurarch.com/a/qwen3-8b.md
https://neurarch.com/templates/qwen3-8b/model.json
https://neurarch.com/?template=qwen3-8b
dit-xl2
DiT-XL/2
Generative
Diffusion Transformer — replaces the UNet denoiser with a ViT backbone conditioned on timestep + class via adaLN-Zero (Peebles 2023)
Pick when you want the modern transformer-based image-generation backbone (used by Sora/SD3-era models) instead of a convolutional UNet denoiser.
204
670,684,064
670.68M
[ 4, 32, 32 ]
[ 256, 32 ]
pass
[ { "rule": "deep-attention-default-init", "severity": "info" } ]
https://neurarch.com/a/dit-xl2.html
https://neurarch.com/a/dit-xl2.md
https://neurarch.com/templates/dit-xl2/model.json
https://neurarch.com/?template=dit-xl2
swin-tiny
Swin-Tiny
Computer Vision
Hierarchical vision transformer — shifted-window attention builds a feature pyramid for dense prediction (Liu 2021)
Pick when you need a ViT that produces multi-scale features (detection, segmentation) rather than a single-scale ViT. Windowed attention keeps compute linear in image size.
81
28,260,808
28.26M
[ 3, 224, 224 ]
[ 1000 ]
warn
[ { "rule": "attention-no-pe", "severity": "warn" }, { "rule": "deep-attention-default-init", "severity": "info" } ]
https://neurarch.com/a/swin-tiny.html
https://neurarch.com/a/swin-tiny.md
https://neurarch.com/templates/swin-tiny/model.json
https://neurarch.com/?template=swin-tiny
clip-vit-b32
CLIP ViT-B/32
Multimodal
Dual-encoder contrastive model — a ViT image tower and a Transformer text tower projected into a shared embedding space (OpenAI 2021)
Pick to study cross-modal retrieval / zero-shot classification. Two parallel encoders meet at a contrastive similarity head.
38
151,198,976
151.20M
[ 3, 224, 224 ]
[ 512 ]
pass
[]
https://neurarch.com/a/clip-vit-b32.html
https://neurarch.com/a/clip-vit-b32.md
https://neurarch.com/templates/clip-vit-b32/model.json
https://neurarch.com/?template=clip-vit-b32
llava-1.5-7b
LLaVA-1.5
Multimodal
Vision-language model — CLIP image encoder + MLP projector feed visual tokens into a LLaMA decoder (Liu 2023)
Pick to see the canonical VLM recipe: a frozen vision encoder bridged into an LLM by a small projector. Foundation for image chat / VQA.
229
7,062,339,840
7.06B
[ 3, 336, 336 ]
[ 1, 2624, 32000 ]
warn
[ { "rule": "full-mha-serving-cost", "severity": "info" }, { "rule": "deep-attention-default-init", "severity": "info" }, { "rule": "kv-cache-context-budget", "severity": "warn" } ]
https://neurarch.com/a/llava-1.5-7b.html
https://neurarch.com/a/llava-1.5-7b.md
https://neurarch.com/templates/llava-1.5-7b/model.json
https://neurarch.com/?template=llava-1.5-7b

Neurarch architecture corpus

36 neural network architectures kept as typed graphs rather than diagrams. Each entry carries every layer's type and parameters, the tensor shape propagated through it, an estimated parameter count, the verdict of 41 structural checks, and a link to the model.json an agent can fetch, edit and submit back for verification. Families span vision, language, recommendation, diffusion, biosignal and speech.

What the viewer shows

The viewer shows one row per architecture. The full graphs are the 36 files under graphs/, which are deliberately not a table: a conv graph and a linear graph do not share a schema, and globbing them is what breaks the viewer.

Fields

  • layer count
  • estimated parameter count
  • input shape
  • output shape
  • verifier verdict

What this dataset is not

Every graph here passes our own verifier by construction: a template with a blocking finding is held back from the gallery rather than published. That makes this a corpus of designs that are structurally sound, not a sample of designs people actually write, and the wrong set to measure a checker's recall against. The grounding study is the one with failures in it.

Verify a design of your own

These rows describe neural network graphs that were checked by a deterministic verifier, and the same verifier is callable:

curl -X POST https://www.neurarch.com/api/v1/check \
  -H "Authorization: Bearer $NEURARCH_API_KEY" \
  -H "Content-Type: application/json" \
  -d @graph.json

Cite

Neurarch. Neurarch architecture corpus. https://neurarch.com/d/architectures.html
Downloads last month
-