task_id stringlengths 8 12 | spec stringlengths 142 246 | observation stringlengths 124 750 | reasoning stringclasses 32
values | actions listlengths 1 14 | verified bool 1
class | source stringclasses 1
value | messages listlengths 3 3 |
|---|---|---|---|---|---|---|---|
gen-mlp-0 | Design a dense classifier for a 162-feature input with 154 output classes. Use 4 hidden linear layers with ReLU and a linear head. Keep it under 10M params. | Model: gen-mlp-0
Components:
- input (input) params={"shape":[1,162]}
- output (output)
Connections:
- input -> output | [
{
"type": "add_component",
"componentType": "linear",
"name": "linear1",
"afterName": "input",
"params": {
"inFeatures": 162,
"outFeatures": 512
}
},
{
"type": "add_component",
"componentType": "relu",
"name": "relu1",
"afterName": "linear1",
"params": {}
... | true | grok:verified | [
{
"role": "system",
"content": "You are a neural-architecture design agent. You edit a structured model graph by emitting actions.\nRespond with ONE JSON object and nothing else: { \"actions\": [ <action> ... ] }\n\nAction types:\n- { \"type\": \"add_component\", \"componentType\": \"<layer>\", \"name\": \"... | |
gen-ae-1 | Design a dense autoencoder for 1486-dim input: encode down to a 57-dim bottleneck and decode back to 1486. Keep it valid and connected, under 20M params. | Model: gen-ae-1
Components:
- input (input) params={"shape":[1,1486]}
- output (output)
Connections:
- input -> output | [
{
"type": "add_component",
"componentType": "linear",
"name": "enc1",
"afterName": "input",
"params": {
"inFeatures": 1486,
"outFeatures": 512
}
},
{
"type": "add_component",
"componentType": "linear",
"name": "enc2",
"afterName": "enc1",
"params": {
... | true | grok:verified | [
{
"role": "system",
"content": "You are a neural-architecture design agent. You edit a structured model graph by emitting actions.\nRespond with ONE JSON object and nothing else: { \"actions\": [ <action> ... ] }\n\nAction types:\n- { \"type\": \"add_component\", \"componentType\": \"<layer>\", \"name\": \"... | |
gen-gqa-4 | Design a grouped-query-attention encoder for 101-token sequences with 56 output classes: token embedding, 1 GQA block(s) each followed by layer normalization, then a linear head. numHeads must be divisible by numKVHeads. Keep it under 60M params. | Model: gen-gqa-4
Components:
- input (input) params={"shape":[1,101]}
- output (output)
Connections:
- input -> output | [
{
"type": "add_component",
"componentType": "embedding",
"name": "embed",
"afterName": "input",
"params": {
"numEmbeddings": 10000,
"embeddingDim": 256
}
},
{
"type": "add_component",
"componentType": "groupedQueryAttention",
"name": "gqa",
"afterName": "embed... | true | grok:verified | [
{
"role": "system",
"content": "You are a neural-architecture design agent. You edit a structured model graph by emitting actions.\nRespond with ONE JSON object and nothing else: { \"actions\": [ <action> ... ] }\n\nAction types:\n- { \"type\": \"add_component\", \"componentType\": \"<layer>\", \"name\": \"... | |
gen-fix-5 | This encoder fails validation: numHeads (8) is not divisible by numKVHeads (3). Repair the attention configuration in place with at most 2 actions. Do not rebuild the model from scratch. | Model: gen-fix-5
Components:
- input (input) params={"shape":[1,232]}
- embed (embedding) params={"numEmbeddings":14650,"embeddingDim":504}
- attn (groupedQueryAttention) params={"embedDim":504,"numHeads":8,"numKVHeads":3}
- head (linear) params={"inFeatures":504,"outFeatures":37}
- output (output)
Connection... | [
{
"type": "update_params",
"name": "attn",
"params": {
"numKVHeads": 2
}
}
] | true | grok:verified | [
{
"role": "system",
"content": "You are a neural-architecture design agent. You edit a structured model graph by emitting actions.\nRespond with ONE JSON object and nothing else: { \"actions\": [ <action> ... ] }\n\nAction types:\n- { \"type\": \"add_component\", \"componentType\": \"<layer>\", \"name\": \"... | |
gen-trim-6 | This 3-hidden-layer MLP uses 2371-wide layers and blows a 2M parameter budget. Shrink the widths in place so total params fit the budget, keeping all 3 hidden layers and consistent in/out features. Do not rebuild the model from scratch. | Model: gen-trim-6
Components:
- input (input) params={"shape":[1,95]}
- fc1 (linear) params={"inFeatures":95,"outFeatures":2371}
- act1 (relu)
- fc2 (linear) params={"inFeatures":2371,"outFeatures":2371}
- act2 (relu)
- fc3 (linear) params={"inFeatures":2371,"outFeatures":2371}
- act3 (relu)
- head (lin... | [
{
"type": "update_params",
"name": "fc1",
"params": {
"outFeatures": 966
}
},
{
"type": "update_params",
"name": "fc2",
"params": {
"inFeatures": 966,
"outFeatures": 966
}
},
{
"type": "update_params",
"name": "fc3",
"params": {
"inFeatures... | true | grok:verified | [
{
"role": "system",
"content": "You are a neural-architecture design agent. You edit a structured model graph by emitting actions.\nRespond with ONE JSON object and nothing else: { \"actions\": [ <action> ... ] }\n\nAction types:\n- { \"type\": \"add_component\", \"componentType\": \"<layer>\", \"name\": \"... | |
gen-norm-7 | This 146-feature, 72-class MLP trains unstably. Insert a batchNorm1d after each of the 3 hidden linear layers (between the linear and its activation), with numFeatures matching each layer's output width. Do not rebuild the model from scratch. | Model: gen-norm-7
Components:
- input (input) params={"shape":[1,146]}
- fc1 (linear) params={"inFeatures":146,"outFeatures":472}
- act1 (relu)
- fc2 (linear) params={"inFeatures":472,"outFeatures":343}
- act2 (relu)
- fc3 (linear) params={"inFeatures":343,"outFeatures":78}
- act3 (relu)
- head (linear)... | [
{
"type": "add_component",
"componentType": "batchNorm1d",
"name": "bn1",
"afterName": "fc1",
"params": {
"numFeatures": 472
}
},
{
"type": "add_component",
"componentType": "batchNorm1d",
"name": "bn2",
"afterName": "fc2",
"params": {
"numFeatures": 343
... | true | grok:verified | [
{
"role": "system",
"content": "You are a neural-architecture design agent. You edit a structured model graph by emitting actions.\nRespond with ONE JSON object and nothing else: { \"actions\": [ <action> ... ] }\n\nAction types:\n- { \"type\": \"add_component\", \"componentType\": \"<layer>\", \"name\": \"... | |
gen-grow-9 | This 2-hidden-layer MLP (109-feature input, 5 classes) is far too small for its workload. Widen the hidden layers in place so total params land between 400k and 4M, keeping consistent in/out features. Do not rebuild the model from scratch. | Model: gen-grow-9
Components:
- input (input) params={"shape":[1,109]}
- fc1 (linear) params={"inFeatures":109,"outFeatures":20}
- act1 (relu)
- fc2 (linear) params={"inFeatures":20,"outFeatures":20}
- act2 (relu)
- head (linear) params={"inFeatures":20,"outFeatures":5}
- output (output)
Connections:
- ... | [
{
"type": "update_params",
"name": "fc1",
"params": {
"outFeatures": 580
}
},
{
"type": "update_params",
"name": "fc2",
"params": {
"inFeatures": 580,
"outFeatures": 580
}
},
{
"type": "update_params",
"name": "head",
"params": {
"inFeature... | true | grok:verified | [
{
"role": "system",
"content": "You are a neural-architecture design agent. You edit a structured model graph by emitting actions.\nRespond with ONE JSON object and nothing else: { \"actions\": [ <action> ... ] }\n\nAction types:\n- { \"type\": \"add_component\", \"componentType\": \"<layer>\", \"name\": \"... | |
gen-mlp-10 | Design a dense classifier for a 214-feature input with 100 output classes. Use 2 hidden linear layers with ReLU and a linear head. Keep it under 10M params. | Model: gen-mlp-10
Components:
- input (input) params={"shape":[1,214]}
- output (output)
Connections:
- input -> output | [
{
"type": "add_component",
"componentType": "linear",
"name": "hidden1",
"afterName": "input",
"params": {
"inFeatures": 214,
"outFeatures": 512
}
},
{
"type": "add_component",
"componentType": "relu",
"name": "relu1",
"afterName": "hidden1",
"params": {}
... | true | grok:verified | [
{
"role": "system",
"content": "You are a neural-architecture design agent. You edit a structured model graph by emitting actions.\nRespond with ONE JSON object and nothing else: { \"actions\": [ <action> ... ] }\n\nAction types:\n- { \"type\": \"add_component\", \"componentType\": \"<layer>\", \"name\": \"... | |
gen-ae-11 | Design a dense autoencoder for 1905-dim input: encode down to a 134-dim bottleneck and decode back to 1905. Keep it valid and connected, under 20M params. | Model: gen-ae-11
Components:
- input (input) params={"shape":[1,1905]}
- output (output)
Connections:
- input -> output | [
{
"type": "add_component",
"componentType": "linear",
"name": "enc_fc1",
"afterName": "input",
"params": {
"inFeatures": 1905,
"outFeatures": 512
}
},
{
"type": "add_component",
"componentType": "linear",
"name": "enc_fc2",
"afterName": "enc_fc1",
"params"... | true | grok:verified | [
{
"role": "system",
"content": "You are a neural-architecture design agent. You edit a structured model graph by emitting actions.\nRespond with ONE JSON object and nothing else: { \"actions\": [ <action> ... ] }\n\nAction types:\n- { \"type\": \"add_component\", \"componentType\": \"<layer>\", \"name\": \"... | |
gen-gqa-14 | Design a grouped-query-attention encoder for 138-token sequences with 35 output classes: token embedding, 1 GQA block(s) each followed by layer normalization, then a linear head. numHeads must be divisible by numKVHeads. Keep it under 60M params. | Model: gen-gqa-14
Components:
- input (input) params={"shape":[1,138]}
- output (output)
Connections:
- input -> output | [
{
"type": "add_component",
"componentType": "embedding",
"name": "embed",
"afterName": "input",
"params": {
"numEmbeddings": 50000,
"embeddingDim": 256
}
},
{
"type": "add_component",
"componentType": "groupedQueryAttention",
"name": "gqa",
"afterName": "embed... | true | grok:verified | [
{
"role": "system",
"content": "You are a neural-architecture design agent. You edit a structured model graph by emitting actions.\nRespond with ONE JSON object and nothing else: { \"actions\": [ <action> ... ] }\n\nAction types:\n- { \"type\": \"add_component\", \"componentType\": \"<layer>\", \"name\": \"... | |
gen-fix-15 | This encoder fails validation: embedDim (400) is not divisible by numHeads (7). Repair the attention configuration in place with at most 2 actions. Do not rebuild the model from scratch. | Model: gen-fix-15
Components:
- input (input) params={"shape":[1,327]}
- embed (embedding) params={"numEmbeddings":8331,"embeddingDim":400}
- attn (multiHeadAttention) params={"embedDim":400,"numHeads":7}
- head (linear) params={"inFeatures":400,"outFeatures":29}
- output (output)
Connections:
- input -> em... | [
{
"type": "update_params",
"name": "attn",
"params": {
"numHeads": 8
}
}
] | true | grok:verified | [
{
"role": "system",
"content": "You are a neural-architecture design agent. You edit a structured model graph by emitting actions.\nRespond with ONE JSON object and nothing else: { \"actions\": [ <action> ... ] }\n\nAction types:\n- { \"type\": \"add_component\", \"componentType\": \"<layer>\", \"name\": \"... | |
gen-trim-16 | This 4-hidden-layer MLP uses 3938-wide layers and blows a 2M parameter budget. Shrink the widths in place so total params fit the budget, keeping all 4 hidden layers and consistent in/out features. Do not rebuild the model from scratch. | Model: gen-trim-16
Components:
- input (input) params={"shape":[1,104]}
- fc1 (linear) params={"inFeatures":104,"outFeatures":3938}
- act1 (relu)
- fc2 (linear) params={"inFeatures":3938,"outFeatures":3938}
- act2 (relu)
- fc3 (linear) params={"inFeatures":3938,"outFeatures":3938}
- act3 (relu)
- fc4 (l... | [
{
"type": "update_params",
"name": "fc1",
"params": {
"outFeatures": 798
}
},
{
"type": "update_params",
"name": "fc2",
"params": {
"inFeatures": 798,
"outFeatures": 798
}
},
{
"type": "update_params",
"name": "fc3",
"params": {
"inFeatures... | true | grok:verified | [
{
"role": "system",
"content": "You are a neural-architecture design agent. You edit a structured model graph by emitting actions.\nRespond with ONE JSON object and nothing else: { \"actions\": [ <action> ... ] }\n\nAction types:\n- { \"type\": \"add_component\", \"componentType\": \"<layer>\", \"name\": \"... | |
gen-norm-17 | This 196-feature, 28-class MLP trains unstably. Insert a batchNorm1d after each of the 3 hidden linear layers (between the linear and its activation), with numFeatures matching each layer's output width. Do not rebuild the model from scratch. | Model: gen-norm-17
Components:
- input (input) params={"shape":[1,196]}
- fc1 (linear) params={"inFeatures":196,"outFeatures":68}
- act1 (relu)
- fc2 (linear) params={"inFeatures":68,"outFeatures":480}
- act2 (relu)
- fc3 (linear) params={"inFeatures":480,"outFeatures":50}
- act3 (relu)
- head (linear) ... | [
{
"type": "add_component",
"componentType": "batchNorm1d",
"name": "bn1",
"afterName": "fc1",
"params": {
"numFeatures": 68
}
},
{
"type": "add_component",
"componentType": "batchNorm1d",
"name": "bn2",
"afterName": "fc2",
"params": {
"numFeatures": 480
... | true | grok:verified | [
{
"role": "system",
"content": "You are a neural-architecture design agent. You edit a structured model graph by emitting actions.\nRespond with ONE JSON object and nothing else: { \"actions\": [ <action> ... ] }\n\nAction types:\n- { \"type\": \"add_component\", \"componentType\": \"<layer>\", \"name\": \"... | |
gen-grow-19 | This 2-hidden-layer MLP (76-feature input, 17 classes) is far too small for its workload. Widen the hidden layers in place so total params land between 400k and 4M, keeping consistent in/out features. Do not rebuild the model from scratch. | Model: gen-grow-19
Components:
- input (input) params={"shape":[1,76]}
- fc1 (linear) params={"inFeatures":76,"outFeatures":6}
- act1 (relu)
- fc2 (linear) params={"inFeatures":6,"outFeatures":6}
- act2 (relu)
- head (linear) params={"inFeatures":6,"outFeatures":17}
- output (output)
Connections:
- inpu... | [
{
"type": "update_params",
"name": "fc1",
"params": {
"outFeatures": 640
}
},
{
"type": "update_params",
"name": "fc2",
"params": {
"inFeatures": 640,
"outFeatures": 640
}
},
{
"type": "update_params",
"name": "head",
"params": {
"inFeature... | true | grok:verified | [
{
"role": "system",
"content": "You are a neural-architecture design agent. You edit a structured model graph by emitting actions.\nRespond with ONE JSON object and nothing else: { \"actions\": [ <action> ... ] }\n\nAction types:\n- { \"type\": \"add_component\", \"componentType\": \"<layer>\", \"name\": \"... | |
gen-mlp-20 | Design a dense classifier for a 11-feature input with 53 output classes. Use 2 hidden linear layers with ReLU and a linear head. Keep it under 10M params. | Model: gen-mlp-20
Components:
- input (input) params={"shape":[1,11]}
- output (output)
Connections:
- input -> output | [
{
"type": "add_component",
"componentType": "linear",
"name": "fc1",
"afterName": "input",
"params": {
"inFeatures": 11,
"outFeatures": 256
}
},
{
"type": "add_component",
"componentType": "relu",
"name": "relu1",
"afterName": "fc1",
"params": {}
},
{
... | true | grok:verified | [
{
"role": "system",
"content": "You are a neural-architecture design agent. You edit a structured model graph by emitting actions.\nRespond with ONE JSON object and nothing else: { \"actions\": [ <action> ... ] }\n\nAction types:\n- { \"type\": \"add_component\", \"componentType\": \"<layer>\", \"name\": \"... | |
gen-ae-21 | Design a dense autoencoder for 1525-dim input: encode down to a 129-dim bottleneck and decode back to 1525. Keep it valid and connected, under 20M params. | Model: gen-ae-21
Components:
- input (input) params={"shape":[1,1525]}
- output (output)
Connections:
- input -> output | [
{
"type": "add_component",
"componentType": "linear",
"name": "enc1",
"afterName": "input",
"params": {
"inFeatures": 1525,
"outFeatures": 512
}
},
{
"type": "add_component",
"componentType": "linear",
"name": "enc2",
"afterName": "enc1",
"params": {
... | true | grok:verified | [
{
"role": "system",
"content": "You are a neural-architecture design agent. You edit a structured model graph by emitting actions.\nRespond with ONE JSON object and nothing else: { \"actions\": [ <action> ... ] }\n\nAction types:\n- { \"type\": \"add_component\", \"componentType\": \"<layer>\", \"name\": \"... | |
gen-cnn-22 | Design a convolutional classifier for a 3x43x43 image with 80 classes. Use 4 conv layers with ReLU and a linear head. Keep it under 50M params. | Model: gen-cnn-22
Components:
- input (input) params={"shape":[1,3,43,43]}
- output (output)
Connections:
- input -> output | [
{
"type": "add_component",
"componentType": "conv2d",
"name": "conv1",
"afterName": "input",
"params": {
"inChannels": 3,
"outChannels": 32,
"kernelSize": 3
}
},
{
"type": "add_component",
"componentType": "relu",
"name": "relu1",
"afterName": "conv1",
... | true | grok:verified | [
{
"role": "system",
"content": "You are a neural-architecture design agent. You edit a structured model graph by emitting actions.\nRespond with ONE JSON object and nothing else: { \"actions\": [ <action> ... ] }\n\nAction types:\n- { \"type\": \"add_component\", \"componentType\": \"<layer>\", \"name\": \"... | |
gen-txf-23 | Design a transformer encoder for 157-token sequences with 37 output classes. Token embedding, 2 multi-head attention block(s) with a valid head configuration, then a linear head. Keep it under 60M params. | Model: gen-txf-23
Components:
- input (input) params={"shape":[1,157]}
- output (output)
Connections:
- input -> output | [
{
"type": "add_component",
"componentType": "embedding",
"name": "token_emb",
"afterName": "input",
"params": {
"numEmbeddings": 10000,
"embeddingDim": 512
}
},
{
"type": "add_component",
"componentType": "multiHeadAttention",
"name": "mha1",
"afterName": "tok... | true | grok:verified | [
{
"role": "system",
"content": "You are a neural-architecture design agent. You edit a structured model graph by emitting actions.\nRespond with ONE JSON object and nothing else: { \"actions\": [ <action> ... ] }\n\nAction types:\n- { \"type\": \"add_component\", \"componentType\": \"<layer>\", \"name\": \"... | |
gen-gqa-24 | Design a grouped-query-attention encoder for 462-token sequences with 25 output classes: token embedding, 1 GQA block(s) each followed by layer normalization, then a linear head. numHeads must be divisible by numKVHeads. Keep it under 60M params. | Model: gen-gqa-24
Components:
- input (input) params={"shape":[1,462]}
- output (output)
Connections:
- input -> output | [
{
"type": "add_component",
"componentType": "embedding",
"name": "embed",
"afterName": "input",
"params": {
"numEmbeddings": 10000,
"embeddingDim": 512
}
},
{
"type": "add_component",
"componentType": "groupedQueryAttention",
"name": "gqa",
"afterName": "embed... | true | grok:verified | [
{
"role": "system",
"content": "You are a neural-architecture design agent. You edit a structured model graph by emitting actions.\nRespond with ONE JSON object and nothing else: { \"actions\": [ <action> ... ] }\n\nAction types:\n- { \"type\": \"add_component\", \"componentType\": \"<layer>\", \"name\": \"... | |
gen-fix-25 | This encoder fails validation: numHeads (8) is not divisible by numKVHeads (7). Repair the attention configuration in place with at most 2 actions. Do not rebuild the model from scratch. | Model: gen-fix-25
Components:
- input (input) params={"shape":[1,198]}
- embed (embedding) params={"numEmbeddings":41296,"embeddingDim":400}
- attn (groupedQueryAttention) params={"embedDim":400,"numHeads":8,"numKVHeads":7}
- head (linear) params={"inFeatures":400,"outFeatures":19}
- output (output)
Connectio... | [
{
"type": "update_params",
"name": "attn",
"params": {
"numKVHeads": 4
}
}
] | true | grok:verified | [
{
"role": "system",
"content": "You are a neural-architecture design agent. You edit a structured model graph by emitting actions.\nRespond with ONE JSON object and nothing else: { \"actions\": [ <action> ... ] }\n\nAction types:\n- { \"type\": \"add_component\", \"componentType\": \"<layer>\", \"name\": \"... | |
gen-trim-26 | This 2-hidden-layer MLP uses 3140-wide layers and blows a 2M parameter budget. Shrink the widths in place so total params fit the budget, keeping all 2 hidden layers and consistent in/out features. Do not rebuild the model from scratch. | Model: gen-trim-26
Components:
- input (input) params={"shape":[1,208]}
- fc1 (linear) params={"inFeatures":208,"outFeatures":3140}
- act1 (relu)
- fc2 (linear) params={"inFeatures":3140,"outFeatures":3140}
- act2 (relu)
- head (linear) params={"inFeatures":3140,"outFeatures":20}
- output (output)
Connect... | [
{
"type": "update_params",
"name": "fc1",
"params": {
"outFeatures": 1300
}
},
{
"type": "update_params",
"name": "fc2",
"params": {
"inFeatures": 1300,
"outFeatures": 1300
}
},
{
"type": "update_params",
"name": "head",
"params": {
"inFeat... | true | grok:verified | [
{
"role": "system",
"content": "You are a neural-architecture design agent. You edit a structured model graph by emitting actions.\nRespond with ONE JSON object and nothing else: { \"actions\": [ <action> ... ] }\n\nAction types:\n- { \"type\": \"add_component\", \"componentType\": \"<layer>\", \"name\": \"... | |
gen-norm-27 | This 249-feature, 50-class MLP trains unstably. Insert a batchNorm1d after each of the 2 hidden linear layers (between the linear and its activation), with numFeatures matching each layer's output width. Do not rebuild the model from scratch. | Model: gen-norm-27
Components:
- input (input) params={"shape":[1,249]}
- fc1 (linear) params={"inFeatures":249,"outFeatures":206}
- act1 (relu)
- fc2 (linear) params={"inFeatures":206,"outFeatures":150}
- act2 (relu)
- head (linear) params={"inFeatures":150,"outFeatures":50}
- output (output)
Connections... | [
{
"type": "add_component",
"componentType": "batchNorm1d",
"name": "bn1",
"afterName": "fc1",
"params": {
"numFeatures": 206
}
},
{
"type": "add_component",
"componentType": "batchNorm1d",
"name": "bn2",
"afterName": "fc2",
"params": {
"numFeatures": 150
... | true | grok:verified | [
{
"role": "system",
"content": "You are a neural-architecture design agent. You edit a structured model graph by emitting actions.\nRespond with ONE JSON object and nothing else: { \"actions\": [ <action> ... ] }\n\nAction types:\n- { \"type\": \"add_component\", \"componentType\": \"<layer>\", \"name\": \"... | |
gen-grow-29 | This 2-hidden-layer MLP (30-feature input, 8 classes) is far too small for its workload. Widen the hidden layers in place so total params land between 400k and 4M, keeping consistent in/out features. Do not rebuild the model from scratch. | Model: gen-grow-29
Components:
- input (input) params={"shape":[1,30]}
- fc1 (linear) params={"inFeatures":30,"outFeatures":20}
- act1 (relu)
- fc2 (linear) params={"inFeatures":20,"outFeatures":20}
- act2 (relu)
- head (linear) params={"inFeatures":20,"outFeatures":8}
- output (output)
Connections:
- i... | [
{
"type": "update_params",
"name": "fc1",
"params": {
"outFeatures": 1000
}
},
{
"type": "update_params",
"name": "fc2",
"params": {
"inFeatures": 1000,
"outFeatures": 1000
}
},
{
"type": "update_params",
"name": "head",
"params": {
"inFeat... | true | grok:verified | [
{
"role": "system",
"content": "You are a neural-architecture design agent. You edit a structured model graph by emitting actions.\nRespond with ONE JSON object and nothing else: { \"actions\": [ <action> ... ] }\n\nAction types:\n- { \"type\": \"add_component\", \"componentType\": \"<layer>\", \"name\": \"... | |
gen-mlp-30 | Design a dense classifier for a 115-feature input with 50 output classes. Use 3 hidden linear layers with ReLU and a linear head. Keep it under 10M params. | Model: gen-mlp-30
Components:
- input (input) params={"shape":[1,115]}
- output (output)
Connections:
- input -> output | [
{
"type": "add_component",
"componentType": "linear",
"name": "fc1",
"afterName": "input",
"params": {
"inFeatures": 115,
"outFeatures": 512
}
},
{
"type": "add_component",
"componentType": "relu",
"name": "relu1",
"afterName": "fc1",
"params": {}
},
{... | true | grok:verified | [
{
"role": "system",
"content": "You are a neural-architecture design agent. You edit a structured model graph by emitting actions.\nRespond with ONE JSON object and nothing else: { \"actions\": [ <action> ... ] }\n\nAction types:\n- { \"type\": \"add_component\", \"componentType\": \"<layer>\", \"name\": \"... | |
gen-ae-31 | Design a dense autoencoder for 1107-dim input: encode down to a 194-dim bottleneck and decode back to 1107. Keep it valid and connected, under 20M params. | Model: gen-ae-31
Components:
- input (input) params={"shape":[1,1107]}
- output (output)
Connections:
- input -> output | [
{
"type": "add_component",
"componentType": "linear",
"name": "enc1",
"afterName": "input",
"params": {
"inFeatures": 1107,
"outFeatures": 512
}
},
{
"type": "add_component",
"componentType": "linear",
"name": "enc2",
"afterName": "enc1",
"params": {
... | true | grok:verified | [
{
"role": "system",
"content": "You are a neural-architecture design agent. You edit a structured model graph by emitting actions.\nRespond with ONE JSON object and nothing else: { \"actions\": [ <action> ... ] }\n\nAction types:\n- { \"type\": \"add_component\", \"componentType\": \"<layer>\", \"name\": \"... | |
gen-cnn-32 | Design a convolutional classifier for a 3x60x60 image with 7 classes. Use 2 conv layers with ReLU and a linear head. Keep it under 50M params. | Model: gen-cnn-32
Components:
- input (input) params={"shape":[1,3,60,60]}
- output (output)
Connections:
- input -> output | [
{
"type": "add_component",
"componentType": "conv2d",
"name": "conv1",
"afterName": "input",
"params": {
"inChannels": 3,
"outChannels": 32,
"kernelSize": 3
}
},
{
"type": "add_component",
"componentType": "relu",
"name": "relu1",
"afterName": "conv1",
... | true | grok:verified | [
{
"role": "system",
"content": "You are a neural-architecture design agent. You edit a structured model graph by emitting actions.\nRespond with ONE JSON object and nothing else: { \"actions\": [ <action> ... ] }\n\nAction types:\n- { \"type\": \"add_component\", \"componentType\": \"<layer>\", \"name\": \"... | |
gen-gqa-34 | Design a grouped-query-attention encoder for 350-token sequences with 64 output classes: token embedding, 3 GQA block(s) each followed by layer normalization, then a linear head. numHeads must be divisible by numKVHeads. Keep it under 60M params. | Model: gen-gqa-34
Components:
- input (input) params={"shape":[1,350]}
- output (output)
Connections:
- input -> output | [
{
"type": "add_component",
"componentType": "embedding",
"name": "embed",
"afterName": "input",
"params": {
"numEmbeddings": 10000,
"embeddingDim": 256
}
},
{
"type": "add_component",
"componentType": "groupedQueryAttention",
"name": "gqa1",
"afterName": "embe... | true | grok:verified | [
{
"role": "system",
"content": "You are a neural-architecture design agent. You edit a structured model graph by emitting actions.\nRespond with ONE JSON object and nothing else: { \"actions\": [ <action> ... ] }\n\nAction types:\n- { \"type\": \"add_component\", \"componentType\": \"<layer>\", \"name\": \"... | |
gen-fix-35 | This encoder fails validation: embedDim (272) is not divisible by numHeads (5). Repair the attention configuration in place with at most 2 actions. Do not rebuild the model from scratch. | Model: gen-fix-35
Components:
- input (input) params={"shape":[1,223]}
- embed (embedding) params={"numEmbeddings":22641,"embeddingDim":272}
- attn (multiHeadAttention) params={"embedDim":272,"numHeads":5}
- head (linear) params={"inFeatures":272,"outFeatures":40}
- output (output)
Connections:
- input -> e... | [
{
"type": "update_params",
"name": "attn",
"params": {
"numHeads": 8
}
}
] | true | grok:verified | [
{
"role": "system",
"content": "You are a neural-architecture design agent. You edit a structured model graph by emitting actions.\nRespond with ONE JSON object and nothing else: { \"actions\": [ <action> ... ] }\n\nAction types:\n- { \"type\": \"add_component\", \"componentType\": \"<layer>\", \"name\": \"... | |
gen-trim-36 | This 2-hidden-layer MLP uses 2812-wide layers and blows a 2M parameter budget. Shrink the widths in place so total params fit the budget, keeping all 2 hidden layers and consistent in/out features. Do not rebuild the model from scratch. | Model: gen-trim-36
Components:
- input (input) params={"shape":[1,139]}
- fc1 (linear) params={"inFeatures":139,"outFeatures":2812}
- act1 (relu)
- fc2 (linear) params={"inFeatures":2812,"outFeatures":2812}
- act2 (relu)
- head (linear) params={"inFeatures":2812,"outFeatures":26}
- output (output)
Connect... | [
{
"type": "update_params",
"name": "fc1",
"params": {
"outFeatures": 1334
}
},
{
"type": "update_params",
"name": "fc2",
"params": {
"inFeatures": 1334,
"outFeatures": 1334
}
},
{
"type": "update_params",
"name": "head",
"params": {
"inFeat... | true | grok:verified | [
{
"role": "system",
"content": "You are a neural-architecture design agent. You edit a structured model graph by emitting actions.\nRespond with ONE JSON object and nothing else: { \"actions\": [ <action> ... ] }\n\nAction types:\n- { \"type\": \"add_component\", \"componentType\": \"<layer>\", \"name\": \"... | |
gen-norm-37 | This 239-feature, 17-class MLP trains unstably. Insert a batchNorm1d after each of the 3 hidden linear layers (between the linear and its activation), with numFeatures matching each layer's output width. Do not rebuild the model from scratch. | Model: gen-norm-37
Components:
- input (input) params={"shape":[1,239]}
- fc1 (linear) params={"inFeatures":239,"outFeatures":356}
- act1 (relu)
- fc2 (linear) params={"inFeatures":356,"outFeatures":462}
- act2 (relu)
- fc3 (linear) params={"inFeatures":462,"outFeatures":197}
- act3 (relu)
- head (linea... | [
{
"type": "add_component",
"componentType": "batchNorm1d",
"name": "bn1",
"afterName": "fc1",
"params": {
"numFeatures": 356
}
},
{
"type": "add_component",
"componentType": "batchNorm1d",
"name": "bn2",
"afterName": "fc2",
"params": {
"numFeatures": 462
... | true | grok:verified | [
{
"role": "system",
"content": "You are a neural-architecture design agent. You edit a structured model graph by emitting actions.\nRespond with ONE JSON object and nothing else: { \"actions\": [ <action> ... ] }\n\nAction types:\n- { \"type\": \"add_component\", \"componentType\": \"<layer>\", \"name\": \"... | |
gen-grow-39 | This 2-hidden-layer MLP (46-feature input, 43 classes) is far too small for its workload. Widen the hidden layers in place so total params land between 400k and 4M, keeping consistent in/out features. Do not rebuild the model from scratch. | Model: gen-grow-39
Components:
- input (input) params={"shape":[1,46]}
- fc1 (linear) params={"inFeatures":46,"outFeatures":11}
- act1 (relu)
- fc2 (linear) params={"inFeatures":11,"outFeatures":11}
- act2 (relu)
- head (linear) params={"inFeatures":11,"outFeatures":43}
- output (output)
Connections:
- ... | [
{
"type": "update_params",
"name": "fc1",
"params": {
"outFeatures": 590
}
},
{
"type": "update_params",
"name": "fc2",
"params": {
"inFeatures": 590,
"outFeatures": 590
}
},
{
"type": "update_params",
"name": "head",
"params": {
"inFeature... | true | grok:verified | [
{
"role": "system",
"content": "You are a neural-architecture design agent. You edit a structured model graph by emitting actions.\nRespond with ONE JSON object and nothing else: { \"actions\": [ <action> ... ] }\n\nAction types:\n- { \"type\": \"add_component\", \"componentType\": \"<layer>\", \"name\": \"... | |
gen-mlp-40 | Design a dense classifier for a 157-feature input with 15 output classes. Use 3 hidden linear layers with ReLU and a linear head. Keep it under 10M params. | Model: gen-mlp-40
Components:
- input (input) params={"shape":[1,157]}
- output (output)
Connections:
- input -> output | [
{
"type": "add_component",
"componentType": "linear",
"name": "fc1",
"afterName": "input",
"params": {
"inFeatures": 157,
"outFeatures": 512
}
},
{
"type": "add_component",
"componentType": "relu",
"name": "relu1",
"afterName": "fc1",
"params": {}
},
{... | true | grok:verified | [
{
"role": "system",
"content": "You are a neural-architecture design agent. You edit a structured model graph by emitting actions.\nRespond with ONE JSON object and nothing else: { \"actions\": [ <action> ... ] }\n\nAction types:\n- { \"type\": \"add_component\", \"componentType\": \"<layer>\", \"name\": \"... | |
gen-ae-41 | Design a dense autoencoder for 625-dim input: encode down to a 113-dim bottleneck and decode back to 625. Keep it valid and connected, under 20M params. | Model: gen-ae-41
Components:
- input (input) params={"shape":[1,625]}
- output (output)
Connections:
- input -> output | [
{
"type": "add_component",
"componentType": "linear",
"name": "enc1",
"afterName": "input",
"params": {
"inFeatures": 625,
"outFeatures": 256
}
},
{
"type": "add_component",
"componentType": "linear",
"name": "enc2",
"afterName": "enc1",
"params": {
... | true | grok:verified | [
{
"role": "system",
"content": "You are a neural-architecture design agent. You edit a structured model graph by emitting actions.\nRespond with ONE JSON object and nothing else: { \"actions\": [ <action> ... ] }\n\nAction types:\n- { \"type\": \"add_component\", \"componentType\": \"<layer>\", \"name\": \"... | |
gen-gqa-44 | Design a grouped-query-attention encoder for 112-token sequences with 47 output classes: token embedding, 2 GQA block(s) each followed by layer normalization, then a linear head. numHeads must be divisible by numKVHeads. Keep it under 60M params. | Model: gen-gqa-44
Components:
- input (input) params={"shape":[1,112]}
- output (output)
Connections:
- input -> output | [
{
"type": "add_component",
"componentType": "embedding",
"name": "embed",
"afterName": "input",
"params": {
"numEmbeddings": 10000,
"embeddingDim": 256
}
},
{
"type": "add_component",
"componentType": "groupedQueryAttention",
"name": "gqa1",
"afterName": "embe... | true | grok:verified | [
{
"role": "system",
"content": "You are a neural-architecture design agent. You edit a structured model graph by emitting actions.\nRespond with ONE JSON object and nothing else: { \"actions\": [ <action> ... ] }\n\nAction types:\n- { \"type\": \"add_component\", \"componentType\": \"<layer>\", \"name\": \"... | |
gen-fix-45 | This encoder fails validation: embedDim (456) is not divisible by numHeads (5). Repair the attention configuration in place with at most 2 actions. Do not rebuild the model from scratch. | Model: gen-fix-45
Components:
- input (input) params={"shape":[1,400]}
- embed (embedding) params={"numEmbeddings":40489,"embeddingDim":456}
- attn (multiHeadAttention) params={"embedDim":456,"numHeads":5}
- head (linear) params={"inFeatures":456,"outFeatures":3}
- output (output)
Connections:
- input -> em... | [
{
"type": "update_params",
"name": "attn",
"params": {
"numHeads": 8
}
}
] | true | grok:verified | [
{
"role": "system",
"content": "You are a neural-architecture design agent. You edit a structured model graph by emitting actions.\nRespond with ONE JSON object and nothing else: { \"actions\": [ <action> ... ] }\n\nAction types:\n- { \"type\": \"add_component\", \"componentType\": \"<layer>\", \"name\": \"... | |
gen-trim-46 | This 2-hidden-layer MLP uses 3187-wide layers and blows a 2M parameter budget. Shrink the widths in place so total params fit the budget, keeping all 2 hidden layers and consistent in/out features. Do not rebuild the model from scratch. | Model: gen-trim-46
Components:
- input (input) params={"shape":[1,83]}
- fc1 (linear) params={"inFeatures":83,"outFeatures":3187}
- act1 (relu)
- fc2 (linear) params={"inFeatures":3187,"outFeatures":3187}
- act2 (relu)
- head (linear) params={"inFeatures":3187,"outFeatures":48}
- output (output)
Connectio... | [
{
"type": "update_params",
"name": "fc1",
"params": {
"outFeatures": 1350
}
},
{
"type": "update_params",
"name": "fc2",
"params": {
"inFeatures": 1350,
"outFeatures": 1350
}
},
{
"type": "update_params",
"name": "head",
"params": {
"inFeat... | true | grok:verified | [
{
"role": "system",
"content": "You are a neural-architecture design agent. You edit a structured model graph by emitting actions.\nRespond with ONE JSON object and nothing else: { \"actions\": [ <action> ... ] }\n\nAction types:\n- { \"type\": \"add_component\", \"componentType\": \"<layer>\", \"name\": \"... | |
gen-norm-47 | This 240-feature, 52-class MLP trains unstably. Insert a batchNorm1d after each of the 4 hidden linear layers (between the linear and its activation), with numFeatures matching each layer's output width. Do not rebuild the model from scratch. | Model: gen-norm-47
Components:
- input (input) params={"shape":[1,240]}
- fc1 (linear) params={"inFeatures":240,"outFeatures":420}
- act1 (relu)
- fc2 (linear) params={"inFeatures":420,"outFeatures":484}
- act2 (relu)
- fc3 (linear) params={"inFeatures":484,"outFeatures":132}
- act3 (relu)
- fc4 (linear... | [
{
"type": "add_component",
"componentType": "batchNorm1d",
"name": "bn1",
"afterName": "fc1",
"params": {
"numFeatures": 420
}
},
{
"type": "add_component",
"componentType": "batchNorm1d",
"name": "bn2",
"afterName": "fc2",
"params": {
"numFeatures": 484
... | true | grok:verified | [
{
"role": "system",
"content": "You are a neural-architecture design agent. You edit a structured model graph by emitting actions.\nRespond with ONE JSON object and nothing else: { \"actions\": [ <action> ... ] }\n\nAction types:\n- { \"type\": \"add_component\", \"componentType\": \"<layer>\", \"name\": \"... | |
gen-grow-49 | This 2-hidden-layer MLP (36-feature input, 50 classes) is far too small for its workload. Widen the hidden layers in place so total params land between 400k and 4M, keeping consistent in/out features. Do not rebuild the model from scratch. | Model: gen-grow-49
Components:
- input (input) params={"shape":[1,36]}
- fc1 (linear) params={"inFeatures":36,"outFeatures":15}
- act1 (relu)
- fc2 (linear) params={"inFeatures":15,"outFeatures":15}
- act2 (relu)
- head (linear) params={"inFeatures":15,"outFeatures":50}
- output (output)
Connections:
- ... | [
{
"type": "update_params",
"name": "fc1",
"params": {
"inFeatures": 36,
"outFeatures": 1024
}
},
{
"type": "update_params",
"name": "fc2",
"params": {
"inFeatures": 1024,
"outFeatures": 1024
}
},
{
"type": "update_params",
"name": "head",
"... | true | grok:verified | [
{
"role": "system",
"content": "You are a neural-architecture design agent. You edit a structured model graph by emitting actions.\nRespond with ONE JSON object and nothing else: { \"actions\": [ <action> ... ] }\n\nAction types:\n- { \"type\": \"add_component\", \"componentType\": \"<layer>\", \"name\": \"... | |
gen-mlp-50 | Design a dense classifier for a 180-feature input with 165 output classes. Use 5 hidden linear layers with ReLU and a linear head. Keep it under 10M params. | Model: gen-mlp-50
Components:
- input (input) params={"shape":[1,180]}
- output (output)
Connections:
- input -> output | [
{
"type": "add_component",
"componentType": "linear",
"name": "linear1",
"afterName": "input",
"params": {
"inFeatures": 180,
"outFeatures": 512
}
},
{
"type": "add_component",
"componentType": "relu",
"name": "relu1",
"afterName": "linear1",
"params": {}
... | true | grok:verified | [
{
"role": "system",
"content": "You are a neural-architecture design agent. You edit a structured model graph by emitting actions.\nRespond with ONE JSON object and nothing else: { \"actions\": [ <action> ... ] }\n\nAction types:\n- { \"type\": \"add_component\", \"componentType\": \"<layer>\", \"name\": \"... | |
gen-ae-51 | Design a dense autoencoder for 1902-dim input: encode down to a 178-dim bottleneck and decode back to 1902. Keep it valid and connected, under 20M params. | Model: gen-ae-51
Components:
- input (input) params={"shape":[1,1902]}
- output (output)
Connections:
- input -> output | [
{
"type": "add_component",
"componentType": "linear",
"name": "enc1",
"afterName": "input",
"params": {
"inFeatures": 1902,
"outFeatures": 512
}
},
{
"type": "add_component",
"componentType": "linear",
"name": "enc2",
"afterName": "enc1",
"params": {
... | true | grok:verified | [
{
"role": "system",
"content": "You are a neural-architecture design agent. You edit a structured model graph by emitting actions.\nRespond with ONE JSON object and nothing else: { \"actions\": [ <action> ... ] }\n\nAction types:\n- { \"type\": \"add_component\", \"componentType\": \"<layer>\", \"name\": \"... | |
gen-cnn-52 | Design a convolutional classifier for a 3x56x56 image with 39 classes. Use 2 conv layers with ReLU and a linear head. Keep it under 50M params. | Model: gen-cnn-52
Components:
- input (input) params={"shape":[1,3,56,56]}
- output (output)
Connections:
- input -> output | [
{
"type": "add_component",
"componentType": "conv2d",
"name": "conv1",
"afterName": "input",
"params": {
"inChannels": 3,
"outChannels": 16,
"kernelSize": 3
}
},
{
"type": "add_component",
"componentType": "relu",
"name": "relu1",
"afterName": "conv1",
... | true | grok:verified | [
{
"role": "system",
"content": "You are a neural-architecture design agent. You edit a structured model graph by emitting actions.\nRespond with ONE JSON object and nothing else: { \"actions\": [ <action> ... ] }\n\nAction types:\n- { \"type\": \"add_component\", \"componentType\": \"<layer>\", \"name\": \"... | |
gen-fix-55 | This encoder fails validation: embedDim (156) is not divisible by numHeads (5). Repair the attention configuration in place with at most 2 actions. Do not rebuild the model from scratch. | Model: gen-fix-55
Components:
- input (input) params={"shape":[1,290]}
- embed (embedding) params={"numEmbeddings":31546,"embeddingDim":156}
- attn (multiHeadAttention) params={"embedDim":156,"numHeads":5}
- head (linear) params={"inFeatures":156,"outFeatures":47}
- output (output)
Connections:
- input -> e... | [
{
"type": "update_params",
"name": "attn",
"params": {
"numHeads": 6
}
}
] | true | grok:verified | [
{
"role": "system",
"content": "You are a neural-architecture design agent. You edit a structured model graph by emitting actions.\nRespond with ONE JSON object and nothing else: { \"actions\": [ <action> ... ] }\n\nAction types:\n- { \"type\": \"add_component\", \"componentType\": \"<layer>\", \"name\": \"... | |
gen-trim-56 | This 4-hidden-layer MLP uses 2565-wide layers and blows a 2M parameter budget. Shrink the widths in place so total params fit the budget, keeping all 4 hidden layers and consistent in/out features. Do not rebuild the model from scratch. | Model: gen-trim-56
Components:
- input (input) params={"shape":[1,87]}
- fc1 (linear) params={"inFeatures":87,"outFeatures":2565}
- act1 (relu)
- fc2 (linear) params={"inFeatures":2565,"outFeatures":2565}
- act2 (relu)
- fc3 (linear) params={"inFeatures":2565,"outFeatures":2565}
- act3 (relu)
- fc4 (lin... | [
{
"type": "update_params",
"name": "fc1",
"params": {
"outFeatures": 796
}
},
{
"type": "update_params",
"name": "fc2",
"params": {
"inFeatures": 796,
"outFeatures": 796
}
},
{
"type": "update_params",
"name": "fc3",
"params": {
"inFeatures... | true | grok:verified | [
{
"role": "system",
"content": "You are a neural-architecture design agent. You edit a structured model graph by emitting actions.\nRespond with ONE JSON object and nothing else: { \"actions\": [ <action> ... ] }\n\nAction types:\n- { \"type\": \"add_component\", \"componentType\": \"<layer>\", \"name\": \"... | |
gen-norm-57 | This 27-feature, 67-class MLP trains unstably. Insert a batchNorm1d after each of the 4 hidden linear layers (between the linear and its activation), with numFeatures matching each layer's output width. Do not rebuild the model from scratch. | Model: gen-norm-57
Components:
- input (input) params={"shape":[1,27]}
- fc1 (linear) params={"inFeatures":27,"outFeatures":179}
- act1 (relu)
- fc2 (linear) params={"inFeatures":179,"outFeatures":80}
- act2 (relu)
- fc3 (linear) params={"inFeatures":80,"outFeatures":41}
- act3 (relu)
- fc4 (linear) par... | [
{
"type": "add_component",
"componentType": "batchNorm1d",
"name": "bn1",
"afterName": "fc1",
"params": {
"numFeatures": 179
}
},
{
"type": "add_component",
"componentType": "batchNorm1d",
"name": "bn2",
"afterName": "fc2",
"params": {
"numFeatures": 80
... | true | grok:verified | [
{
"role": "system",
"content": "You are a neural-architecture design agent. You edit a structured model graph by emitting actions.\nRespond with ONE JSON object and nothing else: { \"actions\": [ <action> ... ] }\n\nAction types:\n- { \"type\": \"add_component\", \"componentType\": \"<layer>\", \"name\": \"... | |
gen-grow-59 | This 2-hidden-layer MLP (113-feature input, 33 classes) is far too small for its workload. Widen the hidden layers in place so total params land between 400k and 4M, keeping consistent in/out features. Do not rebuild the model from scratch. | Model: gen-grow-59
Components:
- input (input) params={"shape":[1,113]}
- fc1 (linear) params={"inFeatures":113,"outFeatures":18}
- act1 (relu)
- fc2 (linear) params={"inFeatures":18,"outFeatures":18}
- act2 (relu)
- head (linear) params={"inFeatures":18,"outFeatures":33}
- output (output)
Connections:
... | [
{
"type": "update_params",
"name": "fc1",
"params": {
"outFeatures": 1024
}
},
{
"type": "update_params",
"name": "fc2",
"params": {
"inFeatures": 1024,
"outFeatures": 1024
}
},
{
"type": "update_params",
"name": "head",
"params": {
"inFeat... | true | grok:verified | [
{
"role": "system",
"content": "You are a neural-architecture design agent. You edit a structured model graph by emitting actions.\nRespond with ONE JSON object and nothing else: { \"actions\": [ <action> ... ] }\n\nAction types:\n- { \"type\": \"add_component\", \"componentType\": \"<layer>\", \"name\": \"... | |
gen-mlp-60 | Design a dense classifier for a 436-feature input with 132 output classes. Use 6 hidden linear layers with ReLU and a linear head. Keep it under 10M params. | Model: gen-mlp-60
Components:
- input (input) params={"shape":[1,436]}
- output (output)
Connections:
- input -> output | [
{
"type": "add_component",
"componentType": "linear",
"name": "l1",
"afterName": "input",
"params": {
"inFeatures": 436,
"outFeatures": 1024
}
},
{
"type": "add_component",
"componentType": "relu",
"name": "r1",
"afterName": "l1",
"params": {}
},
{
... | true | grok:verified | [
{
"role": "system",
"content": "You are a neural-architecture design agent. You edit a structured model graph by emitting actions.\nRespond with ONE JSON object and nothing else: { \"actions\": [ <action> ... ] }\n\nAction types:\n- { \"type\": \"add_component\", \"componentType\": \"<layer>\", \"name\": \"... | |
gen-ae-61 | Design a dense autoencoder for 2039-dim input: encode down to a 186-dim bottleneck and decode back to 2039. Keep it valid and connected, under 20M params. | Model: gen-ae-61
Components:
- input (input) params={"shape":[1,2039]}
- output (output)
Connections:
- input -> output | [
{
"type": "add_component",
"componentType": "linear",
"name": "enc1",
"afterName": "input",
"params": {
"inFeatures": 2039,
"outFeatures": 1024
}
},
{
"type": "add_component",
"componentType": "linear",
"name": "enc2",
"afterName": "enc1",
"params": {
... | true | grok:verified | [
{
"role": "system",
"content": "You are a neural-architecture design agent. You edit a structured model graph by emitting actions.\nRespond with ONE JSON object and nothing else: { \"actions\": [ <action> ... ] }\n\nAction types:\n- { \"type\": \"add_component\", \"componentType\": \"<layer>\", \"name\": \"... | |
gen-fix-65 | This encoder fails validation: numHeads (8) is not divisible by numKVHeads (5). Repair the attention configuration in place with at most 2 actions. Do not rebuild the model from scratch. | Model: gen-fix-65
Components:
- input (input) params={"shape":[1,441]}
- embed (embedding) params={"numEmbeddings":48889,"embeddingDim":488}
- attn (groupedQueryAttention) params={"embedDim":488,"numHeads":8,"numKVHeads":5}
- head (linear) params={"inFeatures":488,"outFeatures":53}
- output (output)
Connectio... | [
{
"type": "update_params",
"name": "attn",
"params": {
"numKVHeads": 4
}
}
] | true | grok:verified | [
{
"role": "system",
"content": "You are a neural-architecture design agent. You edit a structured model graph by emitting actions.\nRespond with ONE JSON object and nothing else: { \"actions\": [ <action> ... ] }\n\nAction types:\n- { \"type\": \"add_component\", \"componentType\": \"<layer>\", \"name\": \"... | |
gen-trim-66 | This 3-hidden-layer MLP uses 2927-wide layers and blows a 2M parameter budget. Shrink the widths in place so total params fit the budget, keeping all 3 hidden layers and consistent in/out features. Do not rebuild the model from scratch. | Model: gen-trim-66
Components:
- input (input) params={"shape":[1,36]}
- fc1 (linear) params={"inFeatures":36,"outFeatures":2927}
- act1 (relu)
- fc2 (linear) params={"inFeatures":2927,"outFeatures":2927}
- act2 (relu)
- fc3 (linear) params={"inFeatures":2927,"outFeatures":2927}
- act3 (relu)
- head (li... | [
{
"type": "update_params",
"name": "fc1",
"params": {
"outFeatures": 979
}
},
{
"type": "update_params",
"name": "fc2",
"params": {
"inFeatures": 979,
"outFeatures": 979
}
},
{
"type": "update_params",
"name": "fc3",
"params": {
"inFeatures... | true | grok:verified | [
{
"role": "system",
"content": "You are a neural-architecture design agent. You edit a structured model graph by emitting actions.\nRespond with ONE JSON object and nothing else: { \"actions\": [ <action> ... ] }\n\nAction types:\n- { \"type\": \"add_component\", \"componentType\": \"<layer>\", \"name\": \"... |
End of preview. Expand in Data Studio
arch-reasoning-grok
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 solves its spec. No human judge, no LLM judge.
Provenance and yield
- Model:
grok-4(xAI API), rejection-sampled at up to 2 tries per task. - Tasks: 500 procedurally generated (generator v2, seed 20260708), spanning ten families (MLP, autoencoder, CNN, transformer, GQA encoder, two-tower, and four edit-in-place repair/scaling families).
- Mint-time yield: 439/500 = 87.8% verified (zero API failures).
- Released rows: 376, each re-verified under the repository's current rubric (v2), which adds the linear-width consistency check of the paper's Algorithm 1. The 63 mint-time-verified rows that fail v2's stricter width check are excluded; they are exactly the blind-spot class the paper's grounding study documents.
Schema
One JSON object per line: task_id, spec, observation (serialized start
graph), reasoning, actions (the verified edit batch), verified: true,
source, and a chat-format messages array ready for SFT.
Reproduce / extend
git clone https://github.com/neurarch-ai/neurarch-arch-bench
XAI_API_KEY=... node training/reasoning_traces.mjs --provider=grok --count=500 --delay=800 --out=my-traces
A private seed produces reasoning data whose designs never appeared on the public web. Companion set: arch-reasoning-claude.
- Downloads last month
- 39