mistral-7b-sae-layer16

Mistral-7B Layer 16 Sparse Autoencoder (JumpReLU)

A Sparse Autoencoder (SAE) trained on layer 16 residual stream activations of Mistral-7B-Instruct-v0.1, using the JumpReLU architecture from Anthropic's Scaling Monosemanticity.

Weights: lmxxf/mistral-7b-sae-layer16

Motivation

LEACE-based linear erasure experiments on Mistral-7B failed to separate JSON structural features from natural language semantic features — all 20+ configurations exhibited sharp phase transitions where both capabilities collapsed simultaneously (subspace entanglement EUO = 1.59). A linear knife cannot cut entangled subspaces. This SAE expands the 4096-dim residual stream to 16384 dimensions, disentangling structure and semantics into distinct sparse features suitable for targeted intervention.

Training Configuration

Validation Results

Quantitative Metrics

Feature Interpretability

The SAE cleanly separates JSON structural features from natural language semantic features. Top-20 features for each domain show near-zero overlap.

JSON structure features (examples):

Natural language features (examples):

JSON punctuation–preferring features (top-10) all have activation ratio > 1e8 — they fire exclusively on JSON punctuation with zero activation on non-punctuation tokens.

Usage

from sae_lens import SAE

sae = SAE.from_pretrained(
    release="lmxxf/mistral-7b-sae-layer16",
    sae_id=".",
)

# sae.encode(activations) -> sparse features
# sae.decode(features) -> reconstructed activations

For a full pipeline with TransformerLens:

from transformer_lens import HookedTransformer

model = HookedTransformer.from_pretrained("mistralai/Mistral-7B-Instruct-v0.1")
_, cache = model.run_with_cache("Hello world", prepend_bos=True)

layer_16_act = cache["blocks.16.hook_resid_post"]
features = sae.encode(layer_16_act)  # (batch, seq, 16384)
reconstructed = sae.decode(features)

Environment

License

This SAE is released for research purposes. The base model (Mistral-7B-Instruct-v0.1) is subject to its own license.