e12-ft42-sae-l42

E12 R02 fine-tune of ceselder/qwen36-27b-sae-l42-1b (trainer_0).

• Architecture: BatchTopK, k = 64, 131,072 features
• Site: block 42 of Qwen3.6-27B @ 6a9e13bd
• Fine-tuning data: 8.0M parent block-42 tokens (chat 60% / story 25% / web 15%) + header copies
• Training: 1,990 steps, lr 2e-5
• Header EV: 0.867 (original: 0.280)

Encoding

relu, then per-token top-64.

Files

• ae.pt: SAE weights
• config.json: SAE config
• sha256.json: sha256 of ae.pt and config.json