🦠 DeepDecipher

A web page and API that provides interpretability information from many sources on various transformer models.

GitHub API documentation arXiv

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

Available models

Model Activation Function Dataset Layers Neurons per Layer Total Neurons Total Parameters Available Services
solu-1l solu 80% C4 (Web Text) and 20% Python Code 1 2,048 2,048 3,145,728 neuron2graph,neuron2graph-search,neuroscope
gelu-1l gelu 80% C4 (Web Text) and 20% Python Code 1 2,048 2,048 3,145,728 neuron2graph,neuron2graph-search,neuroscope
solu-2l solu 80% C4 (Web Text) and 20% Python Code 2 2,048 4,096 6,291,456 neuron2graph,neuron2graph-search,neuroscope
gelu-2l gelu 80% C4 (Web Text) and 20% Python Code 2 2,048 4,096 6,291,456 neuron2graph,neuron2graph-search,neuroscope
solu-3l solu 80% C4 (Web Text) and 20% Python Code 3 2,048 6,144 9,437,184 neuron2graph,neuron2graph-search,neuroscope
gelu-3l gelu 80% C4 (Web Text) and 20% Python Code 3 2,048 6,144 9,437,184 neuron2graph,neuron2graph-search,neuroscope
solu-4l solu 80% C4 (Web Text) and 20% Python Code 4 2,048 8,192 12,582,912 neuron2graph,neuron2graph-search,neuroscope
gelu-4l gelu 80% C4 (Web Text) and 20% Python Code 4 2,048 8,192 12,582,912 neuron2graph,neuron2graph-search,neuroscope
solu-6l solu 80% C4 (Web Text) and 20% Python Code 6 3,072 18,432 42,467,328 neuroscope
solu-8l solu 80% C4 (Web Text) and 20% Python Code 8 4,096 32,768 100,663,296 neuron2graph,neuron2graph-search,neuroscope
solu-10l solu 80% C4 (Web Text) and 20% Python Code 10 5,120 51,200 196,608,000 neuron2graph,neuron2graph-search,neuroscope
solu-12l solu 80% C4 (Web Text) and 20% Python Code 12 6,144 73,728 339,738,624 neuroscope
gpt2-small gelu Open Web Text 12 3,072 36,864 84,934,656 neuron2graph,neuron2graph-search,neuron_explainer,neuroscope
gpt2-medium gelu Open Web Text 24 4,096 98,304 301,989,888 neuroscope
gpt2-large gelu Open Web Text 36 5,120 184,320 707,788,800 neuron2graph,neuron2graph-search,neuroscope
gpt2-xl gelu Open Web Text 48 6,400 307,200 1,474,560,000 neuron_explainer,neuroscope
solu-1l-pile solu The Pile 1 4,096 4,096 12,582,912 neuroscope
solu-4l-pile solu The Pile 4 2,048 8,192 12,582,912 neuron2graph,neuron2graph-search,neuroscope
solu-6l-pile solu The Pile 6 3,072 18,432 42,467,328 neuron2graph,neuron2graph-search,neuroscope
solu-2l-pile solu The Pile 2 2,944 5,888 12,812,288 neuron2graph,neuron2graph-search,neuroscope
solu-8l-pile solu The Pile 8 4,096 32,768 100,663,296 neuron2graph,neuron2graph-search,neuroscope
solu-10l-pile solu The Pile 10 5,120 51,200 196,608,000 neuron2graph,neuron2graph-search,neuroscope
pythia-70m gelu The Pile 6 2,048 12,288 18,874,368 neuron2graph,neuron2graph-search,neuroscope
pythia-160m gelu The Pile 12 3,072 36,864 84,934,656 neuron2graph,neuron2graph-search,neuroscope
pythia-350m gelu The Pile 24 4,096 98,304 301,989,888 neuron2graph,neuron2graph-search,neuroscope