README.md

GPT2 Multilingual 20% AutoInterp Features

This dataset contains feature interpretations for GPT2 Multilingual model with 20% sparsity.

Structure

• data/layer0.parquet - Features for layer 0
• data/layer1.parquet - Features for layer 1
• ...
• data/layer11.parquet - Features for layer 11

Each parquet file contains feature interpretation data including:

• Feature activations
• Top examples
• Interpretations
• And other feature analysis data

Usage

from datasets import load_dataset

# Load specific layer
dataset = load_dataset("fdraye/autointerp-gpt2-multilingual-20", data_files="data/layer0.parquet")

# Load all layers
dataset = load_dataset("fdraye/autointerp-gpt2-multilingual-20", data_files="data/*.parquet")