Downloads · 30 days
15
41% of all-time downloads
qwen35-27b-sae
HumboldtJoker/qwen35-27b-sae
qwen35-27b-sae is a machine learning model from HumboldtJoker. Use it for the machine learning task on the model card, and read the license before you ship it in a product. The card lists the license as other.
Per-layer Sparse Autoencoder (SAE) weights trained on Qwen3.5-27B.
All-time downloads
37
Public
Repo size
208 GB
Likes
0
Public
Repo makeup
Click a slice to open those files.
.pt208 GB · 100%
At a glance
- License
- other
- Model type
- topk_sae
- Access
- Public
- Created
- Aug 29, 2026
- Updated
- Aug 29, 2026
- SHA
- 2997dfed
Model card
From the Hugging Face model README
qwen35-27b-sae
Per-layer Sparse Autoencoder (SAE) weights trained on Qwen3.5-27B.
Contents
62 per-layer SAE weight files (.pt format) covering layers 0-63 (layers 28 and 52 not present). Each file is ~3.1GB. Also includes training config, app.py visualization script, and license.
Purpose
Useful for sparse autoencoder analysis, feature visualization, and mechanistic interpretability research on the Qwen3.5-27B architecture. These are trained SAE weights, not cached activations.
Date Created
May 2026
Why Archived
Regenerable by retraining but compute-costly. Uploaded to preserve the trained weights for future mech interp research without requiring re-training.
- Type
- topk_sae
- License
- other
- Created
- Aug 29, 2026
- Updated
- Aug 29, 2026