Skip to content

Latest commit

 

History

History
57 lines (35 loc) · 3.86 KB

datasets.md

File metadata and controls

57 lines (35 loc) · 3.86 KB
Raw

Collated activation datasets

This document lists the collated activation datasets that are compatible with the Transformer Debugger. These datasets contain some top-activating examples for each MLP neuron, attention head, and autoencoder latent, as well as the corresponding activations for each token (or token pair) in the example. They provide a way to visualize what each neuron, attention head, or autoencoder latent is selective for (obviously in an incomplete way). These activation datasets are used by the neuron viewer to display the top-activating examples for each component, and are also typically used for automated interpretability.

The activations datasets are located on Azure Blob Storage, for example accessible via the blobfile library.

GPT-2 small

Collated activation datasets are available for both the MLP neurons and the attention heads. MLP neuron activations are recorded for each token, while attention head activations are recorded for each token pair.

The datasets are located at the following paths:

  • MLP neurons: az://openaipublic/neuron-explainer/gpt2_small_data/collated-activations/{layer_index}/{neuron_index}.json
  • Attention heads: az://openaipublic/neuron-explainer/gpt2_small/attn_write_norm/collated-activations-by-token-pair/{layer_index}/{head_index}.json

with the following parameters:

  • layer_index is in range(12)
  • neuron_index is in range(3084)
  • head_index is in range(12)

GPT-2 small - MLP autoencoders

MLP autoencoders were trained either on the MLP neurons (after the activation function), or on the MLP-layer output that is written to the residual stream. See Autoencoders for GPT-2 small for more details.

The datasets are located at the following paths:

  • MLP latents: az://openaipublic/neuron-explainer/gpt2-small/autoencoder_latent/{autoencoder_input}{version}/collated-activations/{layer_index}/{latent_index}.pt

with the following parameters:

  • autoencoder_input is in ["mlp_post_act", "resid_delta_mlp"]
  • version is in ["", "_v4"]. (The _v4 versions use slightly different hyperparameters, and should be preferred.)
  • layer_index is in range(12)
  • latent_index is in range(32768)

GPT-2 small - Attention autoencoders

Attention autoencoders were trained on the attention-layer output that is written to the residual stream. See Autoencoders for GPT-2 small for more details. The collated-activations dataset contains autoencoder latent activations for each token, while the collated-activations-by-token-pair dataset contains autoencoder latent attribution to each token pair. To compute the attribution given an autoencoder latent L and a token pair (T1, T2), we multiply the attention pattern A(T1, T2) with the gradient of L with respect to the attention pattern: attribution_L(T1, T2) = A(T1, T2) * ∂L/∂A(T1, T2).

The datasets are located at the following paths:

  • Attention latents (by token): az://openaipublic/neuron-explainer/gpt2-small/autoencoder_latent/resid_delta_attn_v4/collated-activations/{layer_index}/{latent_index}.pt
  • Attention latents (by token pair): az://openaipublic/neuron-explainer/gpt2-small/autoencoder_latent/resid_delta_attn_v4/collated-activations-by-token-pair/{layer_index}/{latent_index}.pt

with the following parameters:

  • layer_index is in range(12)
  • latent_index is in range(10240)

GPT-2 xl

For GPT-2 xl, only the MLP neurons activations are available. The datasets are located at the following paths:

  • MLP neurons: az://openaipublic/neuron-explainer/data/collated-activations/{layer_index}/{neuron_index}.json

with the following parameters:

  • layer_index is in range(48)
  • neuron_index is in range(6400)