id stringlengths 9 23 | spec stringlengths 148 306 | start stringclasses 8
values | budget unknown | constraints unknown |
|---|---|---|---|---|
cnn-cifar | Design a small convolutional image classifier for CIFAR-10 (3x32x32 input, 10 classes). Use at least two conv layers with nonlinearities and a final linear classifier. Keep it under 50M params. | empty-image | {
"maxParams": 50000000
} | {
"forbidBlockers": true,
"minScore": 55,
"maxParams": 50000000,
"mustContainTypes": [
"conv2d",
"linear"
],
"minComponents": 5,
"mustReachOutput": true
} |
mlp-tabular | Build a multilayer perceptron for tabular binary classification with a 32-feature input. Two or three hidden linear layers with ReLU, ending in a linear head. Keep it tiny, under 5M params. | empty-tabular | {
"maxParams": 5000000
} | {
"forbidBlockers": true,
"minScore": 55,
"maxParams": 5000000,
"mustContainTypes": [
"linear",
"relu"
],
"minComponents": 5,
"mustReachOutput": true
} |
text-encoder | Design a small transformer encoder for text classification. Token embedding, then a couple of multi-head attention blocks with a valid head configuration, then a linear classifier. Embedding dim must be divisible by the number of heads. | empty-text | {
"maxParams": 200000000
} | {
"forbidBlockers": true,
"minScore": 55,
"maxParams": 200000000,
"mustContainTypes": [
"embedding",
"multiHeadAttention",
"linear"
],
"minComponents": 5,
"mustReachOutput": true
} |
fix-broken-attention | This model has a broken attention layer: embedDim 100 is not divisible by numHeads 7. Fix the head configuration so the graph is valid, changing as little else as possible. | broken-attention | {} | {
"forbidBlockers": true,
"minScore": 50,
"mustReachOutput": true,
"maxActions": 4
} |
deepen-resnet | Take this tiny CNN and deepen it: add a couple more conv + ReLU stages before the classifier, keeping the graph valid and connected and staying under 80M params. | tiny-cnn | {
"maxParams": 80000000
} | {
"forbidBlockers": true,
"minScore": 55,
"maxParams": 80000000,
"mustContainTypes": [
"conv2d",
"relu"
],
"minComponents": 7,
"mustReachOutput": true
} |
add-normalization | This MLP trains unstably. Add normalization (batch norm or layer norm) between the linear layers without breaking the shapes, and keep it connected. | unstable-mlp | {} | {
"forbidBlockers": true,
"minScore": 55,
"mustContainTypesAny": [
"batchNorm1d",
"layerNorm",
"batchNorm2d"
],
"mustReachOutput": true
} |
scale-under-budget | Make this transformer bigger to increase capacity, but the total parameter count MUST stay under 100M. Widen or deepen as you like. The result must reach at least 15M parameters (the starting graph has ~9M), so leaving it unchanged does not count. | small-transformer | {
"maxParams": 100000000
} | {
"forbidBlockers": true,
"maxParams": 100000000,
"minParams": 15000000,
"minScore": 50,
"mustReachOutput": true
} |
autoencoder | Design a simple dense autoencoder for 784-dim flattened images: an encoder that compresses to a small bottleneck and a decoder that reconstructs back to 784. Keep it valid and connected. | empty-784 | {
"maxParams": 20000000
} | {
"forbidBlockers": true,
"minScore": 50,
"maxParams": 20000000,
"mustContainTypes": [
"linear"
],
"minComponents": 6,
"mustReachOutput": true
} |
two-tower-retrieval | Design a two-tower retrieval encoder for candidate generation: embed sparse ids, then a projection MLP with a nonlinearity producing an embedding that an ANN index can search. This is the YouTube and Pinterest candidate-retrieval pattern (a query/item tower scored by similarity). Keep it under 30M params. | empty-text | {
"maxParams": 30000000
} | {
"forbidBlockers": true,
"minScore": 55,
"maxParams": 30000000,
"mustContainTypes": [
"embedding",
"linear",
"relu"
],
"minComponents": 6,
"mustReachOutput": true
} |
dlrm-ctr-ranking | Design a DLRM-style click-through-rate ranking model: categorical-feature embeddings feeding a bottom-then-top MLP with nonlinearities and a single click-probability head. This is the Meta DLRM and Google Wide-and-Deep ranking pattern. Keep it under 100M params. | empty-text | {
"maxParams": 100000000
} | {
"forbidBlockers": true,
"minScore": 55,
"maxParams": 100000000,
"mustContainTypes": [
"embedding",
"linear",
"relu"
],
"minComponents": 8,
"mustReachOutput": true
} |
bst-sequence-ranking | Design a Behavior Sequence Transformer for ranking: embed the user behavior sequence, apply a self-attention block (embedding dim divisible by the head count), then an MLP head over the result. This is the Alibaba Taobao BST and Pinterest TransAct pattern. Keep it under 50M params. | empty-text | {
"maxParams": 50000000
} | {
"forbidBlockers": true,
"minScore": 55,
"maxParams": 50000000,
"mustContainTypes": [
"embedding",
"multiHeadAttention",
"linear"
],
"minComponents": 7,
"mustReachOutput": true
} |
semantic-search-encoder | Design a semantic-search text encoder for an embedding service: token embedding, a self-attention block, and a projection to a fixed-size sentence vector for ANN retrieval and re-ranking. This is the Spotify Voyager and LinkedIn semantic-search pattern. Keep it under 40M params. | empty-text | {
"maxParams": 40000000
} | {
"forbidBlockers": true,
"minScore": 55,
"maxParams": 40000000,
"mustContainTypes": [
"embedding",
"multiHeadAttention",
"linear"
],
"minComponents": 6,
"mustReachOutput": true
} |
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.
- Size: 12 curated tasks, 8 fixtures
- Licence: MIT License
- Canonical page: https://neurarch.com/d/bench-tasks.html
- On the site: https://neurarch.com/env.html
What the viewer shows
The viewer shows one row per task. tasks.json is the file the harness loads, and it carries the starting fixtures alongside the tasks.
Fields
- 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.
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. Arch-Bench task set. https://neurarch.com/d/bench-tasks.html
- Downloads last month
- 39