SAE Probes Benchmark

This repository contains the code for the paper Are Sparse Autoencoders Useful? A Case Study in Sparse Probing, but has been reformatted into a Python package that will work with any SAE that can be loaded in SAELens. This makes it easy to use the sparse probing tasks from the paper as a standalone SAE benchmark.

Installation

pip install sae-probes

Running evaluations

You can run benchmarks directly; any missing model activations are generated on demand. If you don't pass a model_cache_path, a temporary directory is used and cleaned up when the function completes. To persist activations across runs (recommended for repeated experiments), provide a model_cache_path.

Training Probes

Probes can be trained directly on the model activations (baselines) or on SAE activations. In both cases, the following test data-balance settings are available: "normal", "scarcity", and "imbalance". For more details about these settings, see the original paper. For the most standard sparse-probing benchmark, use the normal setting.

SAE Probes

The most standard use of this library is as a sparse probing benchmark for SAEs using the normal setting. This is demonstrated below:

from sae_probes import run_sae_evals
from sae_lens import SAE

# run the benchmark on a Gemma Scope SAE
release = "gemma-scope-2b-pt-res-canonical"
sae_id = "layer_12/width_16k/canonical"
sae = SAE.from_pretrained(release, sae_id)

run_sae_evals(
  sae=sae,
  model_name="gemma-2-2b",
  hook_name="blocks.12.hook_resid_post",
  reg_type="l1",
  setting="normal",
  results_path="/results/output/path",
  # model_cache_path is optional; if omitted, a temp dir is used and cleared after
  model_cache_path="/path/to/saved/activations",
  ks=[1, 16],
)

The sparse probing results for each dataset will be saved to results_path as a JSON file per dataset.

Baseline Probes

You can now run baseline probes using a unified API that matches the SAE evaluation interface:

from sae_probes import run_baseline_evals

# Run baseline probes with consistent API
run_baseline_evals(
  model_name="gemma-2-2b",
  hook_name="blocks.12.hook_resid_post",
  setting="normal",  # or "scarcity", "imbalance"
  results_path="/results/output/path",
  # model_cache_path is optional; if omitted, a temp dir is used and cleared after
  model_cache_path="/path/to/saved/activations",
)

Output Format

Both SAE and baseline probes now save results as JSON files with consistent structure:

• SAE results: sae_probes_{model_name}/{setting}_setting/{dataset}_{hook_name}_{reg_type}.json
• Baseline results: baseline_results_{model_name}/{setting}_setting/{dataset}_{hook_name}_{method}.json

Each JSON file contains a list with metrics and metadata for easy comparison between SAE and baseline approaches.

Optional: Pre-generating model activations

Pre-generating can speed up repeated runs and lets you inspect the saved tensors. It's optional because benchmarks will auto-generate missing activations on their first run if missing.

from sae_probes import generate_dataset_activations

generate_dataset_activations(
  model_name="gemma-2-2b", # the TransformerLens name of the model
  hook_names=["blocks.12.hook_resid_post"], # Any TLens hook names
  batch_size=64,
  device="cuda",
  model_cache_path="/path/to/save/activations",
)

If you skip pre-generation, the benchmarks will create any missing activations automatically. Passing a model_cache_path persists them; if omitted, activations will be written to a temporary directory that is deleted after the run.

Citation

If you use this code in your research, please cite:

@inproceedings{kantamnenisparse,
  title={Are Sparse Autoencoders Useful? A Case Study in Sparse Probing},
  author={Kantamneni, Subhash and Engels, Joshua and Rajamanoharan, Senthooran and Tegmark, Max and Nanda, Neel},
  booktitle={Forty-second International Conference on Machine Learning}
}