sjgerstner/OLMo-7B-0424-hf_neuron-activations

---
dataset_info:
features:
- name: layer
dtype: int64
- name: neuron
dtype: int64
- name: gate+_in+_freq
dtype: float32
- name: gate+_in+_hook_post_max_values
list: float32
- name: gate+_in+_hook_post_max_indices
list: int64
- name: gate+_in+_hook_post_mean
dtype: float32
- name: gate+_in+_hook_pre_linear_max_values
list: float32
- name: gate+_in+_hook_pre_linear_max_indices
list: int64
- name: gate+_in+_hook_pre_linear_mean
dtype: float32
- name: gate+_in+_hook_pre_max_values
list: float32
- name: gate+_in+_hook_pre_max_indices
list: int64
- name: gate+_in+_hook_pre_mean
dtype: float32
- name: gate+_in+_swish_mean
dtype: float32
- name: gate+_in-_freq
dtype: float32
- name: gate+_in-_hook_post_max_values
list: float32
- name: gate+_in-_hook_post_max_indices
list: int64
- name: gate+_in-_hook_post_mean
dtype: float32
- name: gate+_in-_hook_pre_linear_max_values
list: float32
- name: gate+_in-_hook_pre_linear_max_indices
list: int64
- name: gate+_in-_hook_pre_linear_mean
dtype: float32
- name: gate+_in-_hook_pre_max_values
list: float32
- name: gate+_in-_hook_pre_max_indices
list: int64
- name: gate+_in-_hook_pre_mean
dtype: float32
- name: gate+_in-_swish_mean
dtype: float32
- name: gate-_in+_freq
dtype: float32
- name: gate-_in+_hook_post_max_values
list: float32
- name: gate-_in+_hook_post_max_indices
list: int64
- name: gate-_in+_hook_post_mean
dtype: float32
- name: gate-_in+_hook_pre_linear_max_values
list: float32
- name: gate-_in+_hook_pre_linear_max_indices
list: int64
- name: gate-_in+_hook_pre_linear_mean
dtype: float32
- name: gate-_in+_hook_pre_max_values
list: float32
- name: gate-_in+_hook_pre_max_indices
list: int64
- name: gate-_in+_hook_pre_mean
dtype: float32
- name: gate-_in+_swish_max_values
list: float32
- name: gate-_in+_swish_max_indices
list: int64
- name: gate-_in+_swish_mean
dtype: float32
- name: gate-_in-_freq
dtype: float32
- name: gate-_in-_hook_post_max_values
list: float32
- name: gate-_in-_hook_post_max_indices
list: int64
- name: gate-_in-_hook_post_mean
dtype: float32
- name: gate-_in-_hook_pre_linear_max_values
list: float32
- name: gate-_in-_hook_pre_linear_max_indices
list: int64
- name: gate-_in-_hook_pre_linear_mean
dtype: float32
- name: gate-_in-_hook_pre_max_values
list: float32
- name: gate-_in-_hook_pre_max_indices
list: int64
- name: gate-_in-_hook_pre_mean
dtype: float32
- name: gate-_in-_swish_max_values
list: float32
- name: gate-_in-_swish_max_indices
list: int64
- name: gate-_in-_swish_mean
dtype: float32
splits:
- name: train
num_bytes: 1020133376
num_examples: 352256
download_size: 835972919
dataset_size: 1020133376
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
license: mit
task_categories:
- tabular-regression
language:
- en
tags:
- interpretability
- neuron
pretty_name: Neuron activations of OLMo-7B-0424-hf
size_categories:
- 100K<n<1M
---

<!-- Provide a quick summary of the dataset. -->

This dataset contains activation data of neurons in [OLMo-7B-0424](allenai/OLMo-7B-0424-hf).
(We define a neuron as a hidden dimension in a MLP sublayer.)
To create the dataset, the model was run on [20M tokens from Dolma](sjgerstner/dolma-small).

## Dataset Description

<!-- Provide a longer summary of what this dataset is. -->

Each row corresponds to a neuron, identified by the columns "layer" and "neuron".
(We use zero-based indexing).

The other columns are as follows:

* The first two elements of the name (e.g. "gate+_in+") indicate a sign combination of the activations. For example, "gate+_in+" means the cases in which both the "gate" ("hook_pre") and "in" ("hook_pre_linear") activations of the neuron were positive.
* For each of these sign combinations, we store:
* The relative frequency of the combination (e.g. "gate+_in+_freq")
* Summary statistics about activation values **conditional on this sign combination**, for each combination of:
* intermediate activation ("hook_pre_linear", "hook_pre", "swish", "hook_post") and
* type of summary statistics ("mean", "max_values", "max_indices"). Each of "max_values" and "max_indices" is a list of 16 elements. The indices are with respect to [dolma-small](sjgerstner/dolma-small). We use "max" as a shorthand for "max absolute values", so in many cases the values will actually be negative.

For example, "gate+_in-_hook_post_max_indices" means:
"the indices of the 16 [dolma-small](dolma-small) entries for which this neuron had the smallest (strongest negative) hook_post activations, given that the gate value was positive and the in value was negative".
Why "smallest / strongest negative"?
Because when hook_pre>0 and hook_pre_linear<0, then hook_post=Swish(hook_pre)*hook_pre_linear is automatically negative.

## Uses

<!-- Address questions around how the dataset is intended to be used. -->

This dataset is designed for interpretability research.

We created it for our [GLUScope](https://sjgerstner.github.io/neuroscope) tool.
The tool only shows visualizations for a small number of neurons,
but with this dataset users can quickly create their own visualizations for neurons they are interested in.

It is also possible to use the dataset for more high-level exploration,
such as looking for correlations between layers and certain summary statistics.

## Citation

<!-- If there is a paper or blog post introducing the dataset, the APA and Bibtex information for that should go in this section. -->

**BibTeX:**

[More Information Needed]

## Contact

[More Information Needed]