ishangarg183's picture
Add dataset card
f5770d7 verified
metadata
license: other
pretty_name: Crosscoder Multilayer Split Activations
tags:
  - mechanistic-interpretability
  - crosscoder
  - activations
  - sparse-autoencoder

Crosscoder Multilayer Split Activations

Raw split activation artifacts for multilayer SPARC-style crosscoder training.

This dataset stores reusable base-only and aligned-only activation tensors. These are intended to be assembled into matched activations.pt training artifacts before crosscoder training.

Versions

v1

Source local run: interp_utils/crosscoder/results-multi-v1

Layout:

v1/
  base_activations/
    smollm3-union/
    llama32-3b-union/
    qwen3-4b-union/
  aligned_activations/
    smollm3-{dpo,grpo,kto,orpo,ppo,simpo}/
    llama32-3b-{dpo,grpo,kto,orpo,ppo,simpo}/
    qwen3-4b-{dpo,grpo,kto,orpo,ppo,simpo}/

Each run directory contains:

run_meta.json
activations/base_activations.pt      # base-only runs
activations/aligned_activations.pt   # aligned-only runs

The base tensors contain union layer sets. The aligned tensors contain each aligned model's target probe-best layer window. Assembly slices/reorders the base union tensor to the aligned run's layers.

v1 Base Layers

smollm3-union:      [16, 17, 18, 19, 20]
llama32-3b-union:   [10, 11, 12, 13, 14, 23, 24, 25, 26]
qwen3-4b-union:     [19, 20, 21, 22, 23, 24, 25]

Use

Download one base union and one aligned run, then assemble locally with:

.venv/bin/python -m interp_utils.crosscoder.main \
  --stage assemble \
  --crosscoder-kind multilayer_sparc \
  --base-activations-dir path/to/base_union_dir \
  --aligned-activations-dir path/to/aligned_run_dir \
  --output-dir path/to/assembled_run_dir