juiceb0xc0de/gemma-4-e2b-SAE-sqlite

Gemma 4 E2B SAE SQLite Atlas

An exact, queryable SQLite representation of all 35 residual-stream sparse autoencoders from `juiceb0xc0de/gemma-4-e2b-it-SAE`. The database contains 1,720,320 feature rows. Encoder and decoder vectors preserve the source checkpoints' float32 values exactly.

Files

• —gemma-4-e2b-sae.sqlite3 — SQLite database (20.63 GiB)
• —manifest.json — source revision, dimensions, SHA-256, and integrity result

Database SHA-256: d376882e1f618d0f7a662b6b543247fca62967311af319c0bfbd4d5026820768 SQLite PRAGMA quick_check: ok

Schema

• —atlas_info: format and vector-encoding metadata
• —layers: layer dimensions, decoder bias, source path/hash, and original metadata
• —features: one row per (layer, feature), with encoder/decoder vectors, biases, log-thresholds, and vector norms
• —feature_thresholds: convenience view exposing exp(log_threshold)

Vectors are little-endian IEEE-754 float32 BLOBs. Decode one without copying:

import sqlite3
import numpy as np

con = sqlite3.connect("gemma-4-e2b-sae.sqlite3")
blob = con.execute(
    "SELECT decoder_f32le FROM features WHERE layer=? AND feature=?", (12, 42)
).fetchone()[0]
decoder = np.frombuffer(blob, dtype="<f4")  # shape: (1536,)

Construction

Built in seven bounded five-layer waves on Hugging Face Jobs. Training checkpoints were deliberately excluded; only sae.pt and meta.json were downloaded. Each layer was committed transactionally and its staging files deleted before the next wave, keeping peak scratch below the 50 GiB job limit.