Dataset · james-ra-henry
Activations of 46 language models for interpretability
james-ra-henry/Rosetta-Activations
Tabular and text data with extracted activations of 46 LLMs across 17 semantic concepts.
What's inside
A mechanistic interpretability research dataset containing contrastive activations extracted from 46 different language models across 17 English semantic concepts. It includes text and tabular data, as well as calibration arrays in .npy format.
What you can do with it
- Researching cross-architectural mechanistic interpretability of language models
- Analyzing and comparing internal representations of identical concepts across different LLMs
- Languages
- English
- Size
- up to 1K rows
- Downloads this month
- 118.7K
- Likes
- 15
How to load it
from datasets import load_dataset
ds = load_dataset("james-ra-henry/Rosetta-Activations")