/automated-interpretability

Primary LanguagePython

Automated interpretability

Code and tools

This repository contains code and tools associated with the Language models can explain neurons in language models paper, specifically:

  • Code for automatically generating, simulating, and scoring explanations of neuron behavior using the methodology described in the paper. See the neuron-explainer README for more information.

Note: if you run into errors of the form "Error: Could not find any credentials that grant access to storage account: 'openaipublic' and container: 'neuron-explainer'"." you might be able to fix this by signing up for an azure account and specifying the credentials as described in the error message.

  • A tool for viewing neuron activations and explanations, accessible here. See the neuron-viewer README for more information.

Public datasets

Together with this code, we're also releasing public datasets of GPT-2 XL neurons and explanations. Here's an overview of those datasets.

  • Neuron activations: az://openaipublic/neuron-explainer/data/collated-activations/{layer_index}/{neuron_index}.json
    • Tokenized text sequences and their activations for the neuron. We provide multiple sets of tokens and activations: top-activating ones, random samples from several quantiles; and a completely random sample. We also provide some basic statistics for the activations.
    • Each file contains a JSON-formatted NeuronRecord dataclass.
  • Neuron explanations: az://openaipublic/neuron-explainer/data/explanations/{layer_index}/{neuron_index}.jsonl
    • Scored model-generated explanations of the behavior of the neuron, including simulation results.
    • Each file contains a JSON-formatted NeuronSimulationResults dataclass.
  • Related neurons: az://openaipublic/neuron-explainer/data/related-neurons/weight-based/{layer_index}/{neuron_index}.json
    • Lists of the upstream and downstream neurons with the most positive and negative connections (see below for definition).
    • Each file contains a JSON-formatted dataclass whose definition is not included in this repo.
  • Tokens with high average activations: az://openaipublic/neuron-explainer/data/related-tokens/activation-based/{layer_index}/{neuron_index}.json
    • Lists of tokens with the highest average activations for individual neurons, and their average activations.
    • Each file contains a JSON-formatted TokenLookupTableSummaryOfNeuron dataclass.
  • Tokens with large inbound and outbound weights: az://openaipublic/neuron-explainer/data/related-tokens/weight-based/{layer_index}/{neuron_index}.json
    • List of the most-positive and most-negative input and output tokens for individual neurons, as well as the associated weight (see below for definition).
    • Each file contains a JSON-formatted WeightBasedSummaryOfNeuron dataclass.

Update (July 5, 2023): We also released a set of explanations for GPT-2 Small. The methodology is slightly different from the methodology used for GPT-2 XL so the results aren't directly comparable.

  • Neuron activations: az://openaipublic/neuron-explainer/gpt2_small_data/collated-activations/{layer_index}/{neuron_index}.json
  • Neuron explanations: az://openaipublic/neuron-explainer/gpt2_small_data/explanations/{layer_index}/{neuron_index}.jsonl

Update (August 30, 2023): We recently discovered a bug in how we performed inference on the GPT-2 series models used for the paper and for these datasets. Specifically, we used an optimized GELU implementation rather than the original GELU implementation associated with GPT-2. While the model’s behavior is very similar across these two configurations, the post-MLP activation values we used to generate and simulate explanations differ from the correct values by the following amounts for GPT-2 small:

  • Median: 0.0090
  • 90th percentile: 0.0252
  • 99th percentile: 0.0839
  • 99.9th percentile: 0.1736

Definition of connection weights

Refer to GPT-2 model code for understanding of model weight conventions.

Neuron-neuron: For two neurons (l1, n1) and (l2, n2) with l1 < l2, the connection strength is defined as h{l1}.mlp.c_proj.w[:, n1, :] @ diag(h{l2}.ln_2.g) @ h{l2}.mlp.c_fc.w[:, :, n2].

Neuron-token: For token t and neuron (l, n), the input weight is computed as wte[t, :] @ diag(h{l}.ln_2.g) @ h{l}.mlp.c_fc.w[:, :, n] and the output weight is computed as h{l}.mlp.c_proj.w[:, n, :] @ diag(ln_f.g) @ wte[t, :].

Misc Lists of Interesting Neurons

Lists of neurons we thought were interesting according to different criteria, with some preliminary descriptions.