stationhouse

stationhouse

Holds sparse autoencoder checkpoints, logs and feature census files for Marin-32B, with a TopK SAE of width 65536 for each of its 64 layers. A research archive for interpretability work.

Made by alxcspr on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: stationhouse is a research archive of sparse autoencoders, or SAEs, trained on the activations of Marin-32B for a neuron atlas. The main set is a TopK SAE with width 65536 and k set to 32, trained for each of the 64 layers. It also holds width probes at layers 45, 57 and 63, plus logs, statistics and census files.

Who it suits: It suits interpretability researchers who study Marin-32B and want ready-made SAE checkpoints, with the training records beside them. Read the input convention warning in Good to know first.

Industry | Who uses it | How they use it | Example
AI interpretability | Researchers studying Marin-32B | Load a layer's SAE to inspect the features it learned | Examine the features of layer 45
Model analysis | Teams comparing SAE runs | Compare validation logs inside one width family | Check val_fvu across layers of the 64k sweep
Research tooling | Engineers building feature browsers | Read the census files for each layer | Build a viewer from the uniform census data

You can skip it if you want a model to chat with, since this archive holds research files and not a finished assistant.

Where to use it:

• Python on Windows, macOS and Linux, to download the files
• Interpretability pipelines that can feed Marin-32B activations to an SAE
• Analysis notebooks that read parquet, JSON and JSONL files

Use case | Who | What they get back
Feature inspection | Interpretability researchers | A trained SAE for each layer of the model
Training review | Methodology reviewers | Train and validation logs for each layer
Feature census | Tool builders | Census tables describing the learned features

How to set it up: You need these first:

• Python, to use Hugging Face's download client
• Marin-32B activations that you can feed to the SAEs
• Enough disk space for a full sweep, so check the repository size on its page first

1. Install or update the Hugging Face download client.
2. Download this repository, or only the layer folders you need.
3. Pick a layer folder under sae-64k and check its files against the sha256 list.
4. Feed the SAE raw activations if you use the 64k sweep.

pip install -U huggingface_hub
hf download alxcspr/stationhouse

How to use it: You load a layer's final checkpoint and run Marin-32B activations through it. The visible card text gives no code, so use your own SAE loading code. The 64k sweep expects raw activations, as the warning explains.

Tool | What it does
stats.json | Training statistics for a layer
val_log.jsonl | Validation log, using a trailing holdout of about 1.92M rows
train_log.jsonl | Training log for a layer
sha256.txt | Checksums for a layer's files
census_uniform.parquet | Feature census with the row groups spread evenly
census_prefix.parquet | The original census from the start of the corpus, kept for provenance
hf download | Fetches the repository's files to your computer

Example: A researcher downloads the layer 45 folder of the 64k sweep and feeds raw Marin-32B activations to its final checkpoint. They then read the validation log, where the trainer's val_fvu is the reference number. The card gives no sample code.

For newcomers and experts: A sparse autoencoder is a tool for seeing which features a model's neurons carry. A good first step is to read the whole card text, especially the input convention warning, before you download anything.

For readers who know the field, each layer has a TopK SAE trained with plateau-early stopping and best-validation checkpoints, taking about 44 to 75 minutes per layer. In the 64k sweep, a bug in reading the layer label meant the per-neuron scale was never applied, so training and validation used raw activations. The 32k family used per-neuron scaling.

Choose a layer folder
Check the sha256 list
Prepare raw activations for 64k
Run the TopK SAE
Read features and val_fvu

Good to know:

• Input convention: 32k and 64k checkpoints expect different inputs, so do not compare fvu across widths until they are unified.
• Valid comparisons: inside one width family, validation to validation and census to census still hold.
• Artifact files: files named with scaled_ARTIFACT measure the normalization mismatch, not model quality.
• Width choice: the 128k probes were stopped, and 64k was adopted for the full sweep.

Title: Median validation gain of the 128k probe against the bar
128k probe median gain: 0.0044
Bar to continue: 0.02

• License: Apache 2.0.
• Cost to run: none named.
• Data it reaches: the files you download. The text available here names no network use.
• Not for: cross-width fvu comparisons, which the card calls invalid until the conventions are unified.

On Hugging Face, 13 November 2026

• Downloads in the last 30 days: 0
• Likes: 3
• License: apache-2.0

Links

• Model: https://huggingface.co/alxcspr/stationhouse
• Files: https://huggingface.co/alxcspr/stationhouse/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download alxcspr/stationhouse

Then follow the usage in its model card. Nothing in this kit runs by itself.