Models

solu-1l

Model name

neuron2graph, neuron2graph-search, neuroscope

Available Services

solu

Activation Function

1

Layers

2,048

Neurons per Layer

2,048

Total Neurons

3,145,728

Total Parameters

80% C4 (Web Text) and 20% Python Code

Dataset

gelu-1l

Model name

neuron2graph, neuron2graph-search, neuroscope

Available Services

gelu

Activation Function

1

Layers

2,048

Neurons per Layer

2,048

Total Neurons

3,145,728

Total Parameters

80% C4 (Web Text) and 20% Python Code

Dataset

solu-2l

Model name

neuron2graph, neuron2graph-search, neuroscope

Available Services

solu

Activation Function

2

Layers

2,048

Neurons per Layer

4,096

Total Neurons

6,291,456

Total Parameters

80% C4 (Web Text) and 20% Python Code

Dataset

gelu-2l

Model name

neuron2graph, neuron2graph-search, neuroscope

Available Services

gelu

Activation Function

2

Layers

2,048

Neurons per Layer

4,096

Total Neurons

6,291,456

Total Parameters

80% C4 (Web Text) and 20% Python Code

Dataset

solu-3l

Model name

neuron2graph, neuron2graph-search, neuroscope

Available Services

solu

Activation Function

3

Layers

2,048

Neurons per Layer

6,144

Total Neurons

9,437,184

Total Parameters

80% C4 (Web Text) and 20% Python Code

Dataset

gelu-3l

Model name

neuron2graph, neuron2graph-search, neuroscope

Available Services

gelu

Activation Function

3

Layers

2,048

Neurons per Layer

6,144

Total Neurons

9,437,184

Total Parameters

80% C4 (Web Text) and 20% Python Code

Dataset

solu-4l

Model name

neuron2graph, neuron2graph-search, neuroscope

Available Services

solu

Activation Function

4

Layers

2,048

Neurons per Layer

8,192

Total Neurons

12,582,912

Total Parameters

80% C4 (Web Text) and 20% Python Code

Dataset

gelu-4l

Model name

neuron2graph, neuron2graph-search, neuroscope

Available Services

gelu

Activation Function

4

Layers

2,048

Neurons per Layer

8,192

Total Neurons

12,582,912

Total Parameters

80% C4 (Web Text) and 20% Python Code

Dataset

solu-6l

Model name

neuroscope

Available Services

solu

Activation Function

6

Layers

3,072

Neurons per Layer

18,432

Total Neurons

42,467,328

Total Parameters

80% C4 (Web Text) and 20% Python Code

Dataset

solu-8l

Model name

neuron2graph, neuron2graph-search, neuroscope

Available Services

solu

Activation Function

8

Layers

4,096

Neurons per Layer

32,768

Total Neurons

100,663,296

Total Parameters

80% C4 (Web Text) and 20% Python Code

Dataset

solu-10l

Model name

neuron2graph, neuron2graph-search, neuroscope

Available Services

solu

Activation Function

10

Layers

5,120

Neurons per Layer

51,200

Total Neurons

196,608,000

Total Parameters

80% C4 (Web Text) and 20% Python Code

Dataset

solu-12l

Model name

neuroscope

Available Services

solu

Activation Function

12

Layers

6,144

Neurons per Layer

73,728

Total Neurons

339,738,624

Total Parameters

80% C4 (Web Text) and 20% Python Code

Dataset

gpt2-small

Model name

neuron2graph, neuron2graph-search, neuron_explainer, neuroscope

Available Services

gelu

Activation Function

12

Layers

3,072

Neurons per Layer

36,864

Total Neurons

84,934,656

Total Parameters

Open Web Text

Dataset

gpt2-medium

Model name

neuroscope

Available Services

gelu

Activation Function

24

Layers

4,096

Neurons per Layer

98,304

Total Neurons

301,989,888

Total Parameters

Open Web Text

Dataset

gpt2-large

Model name

neuron2graph, neuron2graph-search, neuroscope

Available Services

gelu

Activation Function

36

Layers

5,120

Neurons per Layer

184,320

Total Neurons

707,788,800

Total Parameters

Open Web Text

Dataset

gpt2-xl

Model name

neuron_explainer, neuroscope

Available Services

gelu

Activation Function

48

Layers

6,400

Neurons per Layer

307,200

Total Neurons

1,474,560,000

Total Parameters

Open Web Text

Dataset

solu-1l-pile

Model name

neuroscope

Available Services

solu

Activation Function

1

Layers

4,096

Neurons per Layer

4,096

Total Neurons

12,582,912

Total Parameters

The Pile

Dataset

solu-4l-pile

Model name

neuron2graph, neuron2graph-search, neuroscope

Available Services

solu

Activation Function

4

Layers

2,048

Neurons per Layer

8,192

Total Neurons

12,582,912

Total Parameters

The Pile

Dataset

solu-6l-pile

Model name

neuron2graph, neuron2graph-search, neuroscope

Available Services

solu

Activation Function

6

Layers

3,072

Neurons per Layer

18,432

Total Neurons

42,467,328

Total Parameters

The Pile

Dataset

solu-2l-pile

Model name

neuron2graph, neuron2graph-search, neuroscope

Available Services

solu

Activation Function

2

Layers

2,944

Neurons per Layer

5,888

Total Neurons

12,812,288

Total Parameters

The Pile

Dataset

solu-8l-pile

Model name

neuron2graph, neuron2graph-search, neuroscope

Available Services

solu

Activation Function

8

Layers

4,096

Neurons per Layer

32,768

Total Neurons

100,663,296

Total Parameters

The Pile

Dataset

solu-10l-pile

Model name

neuron2graph, neuron2graph-search, neuroscope

Available Services

solu

Activation Function

10

Layers

5,120

Neurons per Layer

51,200

Total Neurons

196,608,000

Total Parameters

The Pile

Dataset

pythia-70m

Model name

neuron2graph, neuron2graph-search, neuroscope

Available Services

gelu

Activation Function

6

Layers

2,048

Neurons per Layer

12,288

Total Neurons

18,874,368

Total Parameters

The Pile

Dataset

pythia-160m

Model name

neuron2graph, neuron2graph-search, neuroscope

Available Services

gelu

Activation Function

12

Layers

3,072

Neurons per Layer

36,864

Total Neurons

84,934,656

Total Parameters

The Pile

Dataset

pythia-350m

Model name

neuron2graph, neuron2graph-search, neuroscope

Available Services

gelu

Activation Function

24

Layers

4,096

Neurons per Layer

98,304

Total Neurons

301,989,888

Total Parameters

The Pile

Dataset

Features

Feature Description Source
Neuron2Graph Vizualize the activation patterns of neurons as a graph. Each path through the graph is a n-gram which activates the neuron. From this we also derive a set of similar neurons, which are neurons whose graphs are sufficiently similar. Paper
Neuroscope Shows how much the neuron activates to each token in a series of text examples. The examples chosen are the examples with the highest activations for that neuron. Website
Neuron explanation An attempt by GPT-4 to explain what concept the neuron activates on. Only available for models gpt2-small and gpt2-xl. Website