ceselder/maemm-uplift-acts_sae

MAEMM cross-uplift arm acts_sae: 100k real activations + 100k SAE-feature directions

Midtrain (1 epoch, lr 1e-4) on a 200k bank of 100k real activations + 100k SAE-feature directions, from the 23M real-activation SFT init, then 100 RL steps (CISPO / ScaleRL, 128 directions × 16 samples per step, lr 1e-5, 10 warmup steps) on the same bank. Report: http://5.78.192.0/reports/view/maemm-uplift-matrix/report.html

All checkpoints are LoRA adapters (r 64, α 16, rsLoRA, all linear layers) of the MAEMM activation→text inverter for Qwen3.6-27B layer 42 (inject h + ||h||·v at the layer-1 marker; text whose clean layer-42 activation points along v). Code: https://github.com/ceselder/maemm. Eval = 512 held-out directions/family, best-of-4 at T=1, cosine of the clean base L42 activation (max over last 5 tokens). Subfolders are PEFT adapters: PeftModel.from_pretrained(base, repo, subfolder="<name>").

Held-out evals

checkpoint	mean_all	realact	SAE norm_act	SAE rank-1	BSF	probes	MLP fire-back
init (23M realact SFT)	0.368	0.477	0.416	0.189	0.296	0.226	0.121
sft_final (after midtrain)	0.258	0.325	0.404	0.176	0.242	0.184	0.073
rl_step_25	0.339	0.440	0.576	0.244	0.281	0.218	0.157
rl_step_50	0.376	0.494	0.685	0.297	0.300	0.237	0.248
rl_step_100	0.398	0.518	0.805	0.336	0.312	0.251	0.407