---
license: other
license_name: dinov3-license
license_link: https://ai.meta.com/resources/models-and-libraries/dinov3-license
library_name: block-sparse-featurizer
tags:
- block-sparse-featurizer
- sparse-autoencoder
- dinov3
---

# Vanilla BSF — starter notebook

This is a trained block-sparse featurizer reproduced from
[Goodfire's original repository](https://github.com/goodfire-ai/block-sparse-featurizer) at immutable commit
`0bf2d9a6ae959452d57bc169374c8902135e0f02`. It is part of the
[Block-Sparse Featurizers on DINOv3 Rabbits collection](https://huggingface.co/collections/BurnyCoder/block-sparse-featurizers-on-dinov3-rabbits-6a629047facccb1d34e808c2).

## Exact recipe

- Source: `starters/03_vanilla.ipynb` (Git blob `b132cde470b0716b7796566f0b06a871fab5bfd7`)
- Constructor arguments: `{"d": "activation_width", "group_size": 3, "l0": 8, "n_groups": 256}`
- `bsf.train` arguments: `{"epochs": 300, "lr": 0.003}`
- Effective defaults: batch size 2048, SNR 0.1,
  learning rate 0.003
- DINO backbone: `facebook/dinov3-vitb16-pretrain-lvd1689m` at `5931719e67bbdb9737e363e781fb0c67687896bc`
- Reconstruction R²: 0.821171
- Mean active blocks: 8.0
- Dead groups: 2

The full machine-readable provenance, input hashes, environment, and metrics are
in `manifest.json`.

## Loading

The hardened loader and immutable catalog live in
[`block-sparse-featurizer-experiments`](https://github.com/BurnyCoder/block-sparse-featurizer-experiments). Install that
application from a revision containing this model's Hub commit:

```bash
git clone --recurse-submodules https://github.com/BurnyCoder/block-sparse-featurizer-experiments.git
cd block-sparse-featurizer-experiments
uv sync --frozen
```

```python
from bsf_experiments.artifacts import restore_checkpoint
from bsf_experiments.hub_phase import (
    download_hub_checkpoint,
    get_hub_checkpoint_spec,
)
from bsf_experiments.types import PretrainedRecipe

spec = get_hub_checkpoint_spec(PretrainedRecipe.VANILLA_NOTEBOOK)
path = download_hub_checkpoint(spec)
model, model_config = restore_checkpoint(path)
```

The application catalog pins a full Hub commit, file size, input width, and
SHA-256 before this code restores the checkpoint. Loading avoids BSF retraining;
new images still require the same DINOv3 patch-token extraction,
positional-mean subtraction, and RMS scaling.

## Limitations and licenses

The checkpoint was trained only on the bundled 300-image rabbit dataset and
should not be assumed to generalize to other data. Training intentionally
preserved the original unseeded stochastic behavior, so an independent
reproduction need not be bit-identical.

Goodfire's BSF software terms are in `LICENSE-goodfire.txt`. DINOv3's license,
which governs the backbone and its materials, is in `LICENSE-dinov3.md`; review
those terms before redistribution or use.