latent-sae

Train Sparse Autoencoders on sentence embedding representations. Decomposes dense embedding vectors into sparse, interpretable features.

Fork of EleutherAI/sae, focused on sentence transformer embeddings with fast disk-based data loading, Modal GPU training, and a comprehensive experiment framework.

Published Models

Model	Subfolder	Embedding	Features	k	Best For
sae-all-MiniLM-L6-v2-FineWeb-RedPajama-Pile-150M	128_4	MiniLM (384D)	1,536	128	Fine-grained classification
	128_8	MiniLM (384D)	3,072	128	Retrieval
	64_8	MiniLM (384D)	3,072	64	Maximum feature coverage
sae-nomic-text-v1.5-FineWeb-edu-100BT	64_32	nomic-v1.5 (768D)	24,576	64	Legacy (taxonomy)

Quick Start

# pip install latentsae
from latentsae import Sae
from sentence_transformers import SentenceTransformer
import torch

# Load SAE
sae = Sae.load_from_hub("enjalot/sae-all-MiniLM-L6-v2-FineWeb-RedPajama-Pile-150M", "128_4")

# Embed text
model = SentenceTransformer("sentence-transformers/all-MiniLM-L6-v2")
embeddings = model.encode(["Hello world", "Sparse autoencoders decompose embeddings"],
                          convert_to_tensor=True, normalize_embeddings=True)

# Extract sparse features
features = sae.encode(embeddings)
print(features.top_indices)  # which features activated
print(features.top_acts)     # how strongly

Training

With YAML config (recommended)

# Single run on Modal A10G
modal run train_modal.py --config experiments/configs/minilm_30M_3source.yaml --gpu-type a10g

# Parameter sweep (parallel)
modal run train_modal.py --config experiments/configs/arch_sweep_base.yaml \
  --sweep experiments/configs/arch_sweep_type.yaml --gpu-type a10g

# Local training on M2 Mac
python -m experiments.run_experiment experiments/configs/smoke_test.yaml --device mps

With CLI args (quick experiments)

modal run train_modal.py --batch-size 1024 --k 128 --expansion-factor 4 --gpu-type a10g
python train_local.py --batch_size 512 --k 64 --expansion_factor 8

Experiment Framework

YAML-driven experiment system with config hashing, cartesian sweeps, and WandB integration. See experiments/ for configs and results.

# Dry-run a sweep to see what would execute
python -m experiments.run_experiment experiments/configs/arch_sweep_base.yaml \
  --sweep experiments/configs/arch_sweep_type.yaml --dry-run

# Compare results
python -m experiments.compare_results experiments/results/

# Evaluate a trained SAE
python -m experiments.eval_probes --sae-path checkpoints/sae_topk_128_4.xxx \
  --embedding-model sentence-transformers/all-MiniLM-L6-v2 --suite hard

Evaluation Suite

Task	Type	What it tests
AG News	4-class classification	Coarse topic features
SST-2	2-class classification	Sentiment features
BANKING77	77-class classification	Fine-grained intent features
CLINC150	150-class classification	Very fine-grained (hardest)
STS-B	Similarity (spearman)	Continuous semantic structure
SciFact	Retrieval (nDCG@10)	Information preservation
MMCS	Feature quality	Decoder weight redundancy

Architecture

Supported SAE types: TopK (recommended), Gated, JumpReLU, LISTA.

Training features: auxk dead feature revival, k-annealing, tilted ERM, decoder decorrelation loss, fire rate penalty, mixed precision (AMP), cosine LR schedule.

Key Research Findings

• Expansion factor 4-8x is optimal for embeddings (not 32x like LLM layers)
• k=128 dramatically beats k=64 on fine-grained tasks (CLINC150: 79.6% vs 64.6%)
• 30M diverse samples retains 97% of 150M quality at 6.4x lower cost
• 3-source data mix matters more than any training regularization
• A10G at batch_size=1024 is the cost-optimal GPU config ($2.60/100M samples)

Data Preparation

Training data (pre-computed embeddings) is prepared in latent-data-modal. See the Latent Taxonomy methodology for details.

Part of the latent-* ecosystem

• latent-scope — Interactive dataset exploration
• latent-taxonomy — SAE feature visualization
• latent-basemap — Parametric UMAP for consistent 2D layouts
• latent-data-modal — Data pipelines on Modal