ceselder/maemm-uplift-acts_all

MAEMM cross-uplift arm acts_all: 100k real activations + 100k spread over the 5 other families (act-matched 'everything mixed')

Same protocol as the single-family arms: midtrain (1 epoch, lr 1e-4) on a 200k bank of 100k real activations + 20k each of SAE-feature, BSF, cluster-probe, long-context-activation and layer-42 MLP-neuron directions, from the 23M real-activation SFT init, then 100 RL steps (CISPO / ScaleRL, 128 directions × 16 samples per step, lr 1e-5). Report: http://5.78.192.0/reports/view/maemm-uplift-matrix/report.html

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). Code: https://github.com/ceselder/maemm. Eval = 512 held-out directions/family, best-of-4 at T=1. 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)	0.325	0.427	0.363	0.164	0.252	0.196	0.099
rl_step_25	0.379	0.499	0.502	0.227	0.299	0.228	0.196
rl_step_50	0.395	0.518	0.625	0.289	0.312	0.246	0.337
rl_step_100	0.407	0.529	0.728	0.309	0.325	0.257	0.499