Downloads · 30 days
0
maemm-uplift-all7-mixed-rlE
ceselder/maemm-uplift-all7-mixed-rlE
maemm-uplift-all7-mixed-rlE is a machine learning model from ceselder. Use it for the machine learning task on the model card, and read the license before you ship it in a product. It is set up for peft.
The reference row of the cross-uplift matrix: from the 23M real-act SFT init, a 1.1M-example all-families midtrain (sftfinal, lr 1e-4, 1 epoch) then RL (CISPO, 16 samples × 256 directions per step, lr 1e-5) on the 7-f…
Access
Public
Updated Sep 5, 2026
Repo size
9.3 GB
Likes
0
Public
Repo makeup
Click a slice to open those files.
.safetensors9.3 GB · 100%
At a glance
- Library
- peft
- Access
- Public
- Created
- Sep 5, 2026
- Updated
- Sep 5, 2026
- SHA
- 426d934c
Base models
Model card
From the Hugging Face model README
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>").
- Library
- peft
- Created
- Sep 5, 2026
- Updated
- Sep 5, 2026