| --- |
| base_model: Qwen/Qwen3.6-27B |
| library_name: peft |
| tags: [interpretability, activation-inversion, lora, grpo] |
| --- |
| # MAEMM 'all 7 families mixed' reference run (RL-E, steps 25–100) |
|
|
| The reference row of the cross-uplift matrix: from the 23M real-act SFT init, a 1.1M-example all-families midtrain (`sft_final`, lr 1e-4, 1 epoch) then RL (CISPO, 16 samples × 256 directions per step, lr 1e-5) on the 7-family bank incl. layer-42 MLP neurons. Checkpoints 25/40/63/100 = 102k/164k/258k/410k rollouts. Held-out at step 100: mean 0.409, real acts 0.530, SAE norm_act 0.845, rank-1 0.340, BSF 0.331, probes 0.262, MLP fire-back 0.577. The run's best checkpoint (step 250) is `ceselder/maemm-qwen36-27b-inverter-rlE-step250`. |
| |
| LoRA adapters (r 64, α 16, rsLoRA, all linear layers) of the MAEMM activation→text inverter for Qwen3.6-27B layer 42. Code: https://github.com/ceselder/maemm. Report: http://5.78.192.0/reports/view/maemm-uplift-matrix/report.html. Load a subfolder with `PeftModel.from_pretrained(base, repo, subfolder="<name>")`. |
|
|
