juiceb0xc0de/qwen3-0.6b-SAE-v2.0

Twenty-eight JumpReLU SAEs for Qwen3-0.6B, trained in one hands-off pass in 2h 14m. One command, no tuning, no restarts. The run I started layer 0 at 09:10:23 UTC and layer 27 finished at 11:25:06 UTC on 2026-07-29. 2h 14m 43s, start to stop, for a complete 28-layer SAE stack. I did not touch it. Not once. There is no hyperparameter sweep behind this, no per-layer L0 tuning, no relaunching a layer that blew up. Every one of the 28 runs logged the same k, lr, n_features… See the full description on the dataset page: https://huggingface.co/datasets/juiceb0xc0de/qwen3-0.6b-SAE-v2.0.

<iframe src="https://silytics.si/embed/item/7c681263-d1d5-47a5-be76-90a8638df8d8" width="420" height="190" style="border:0" loading="lazy" title="Silytics entry"></iframe>

Silytics shows public metadata from Hugging Face and links to the original. Information may be incomplete or out of date. Last refreshed 18 October 2026. Licenses and terms of the original apply to the work itself. Not financial, legal or professional advice.