Top-K SAE on gpt2-medium layer-12 residual stream (blocks.12.hook_resid_pre, d_model=1024). Trained on 1000 documents from the Pile to demonstrate that the platform can train a usable SAE at non-toy scale on a single 128 GB Apple Silicon MBP. 2048 features with k=32 gives a ~64x expansion factor and roughly 3% feature density, which is typical for mid-network SAEs on GPT-2 class models.

Or via the CLI: