maemm-27b-rl-last16-lr5e-7

MAEMM 27B activation-to-text inverter — RL, last-16-token reward, lr 5e-7 (step 300)

Full-parameter Qwen3.6-27B checkpoint from the "simple2m" chain (2026-09-17/18). Given a unit direction in the layer-42 residual space (injected at layer 1 on the ? marker as h + coeff * ||h|| * unit(v)), the model writes text whose own layer-42 activation points along that direction.

Recipe

• SFT init (/data/sft_mix/simple2m_sft/final): one full fine-tune from the base model on 8M rows — exact 50/50 of (a) 4M Ultra-FineWeb activations with 8-64 tokens of context whose target is the full context, and (b) 2M encoder-column + 2M decoder-row directions of a 2M-feature layer-42 BatchTopK SAE (max-activating windows; SFT feature split of 1,847,152 features, 100k eval + 150k RL features held out). lr 1e-5 one-cycle, batch 4,096, one epoch.
• RL: full-parameter CISPO/GRPO (ScaleRL recipe), 8 prompts x 2,048 rollouts per step, lr 5e-7 (flat after a 25-step warm-up), 300 steps, reward = max cosine over the last 16 generated tokens (--reward-window-last 16), length penalty 2.5e-4/token beyond 8, pool = even split of full-document activations (contexts 64-2048) and RL-split 2M-SAE rows (941,132 rows).
• Scorer caveat: the RL reward and the training-time evals used the legacy raw-activation cosine cos(unit(h), d); the numbers below were RE-SCORED with the corrected centered cosine cos(unit(h - mu), d) (mu = the layer-42 corpus mean), which is the metric to compare against.

Held-out evaluation (centered scorer; 512 held-out directions per family, best-of-4 at T=1)

SFT init: mean_all 0.420. For comparison the 8x2048 lr 1e-6 whole-span-reward arm ends at mean_all 0.595 with MLP norm_act 0.581 and 2M enc/dec fired 0.312 / 0.363: this low-lr arm reaches the same fidelity with the best out-of-pool generalization of the batch and held-out 2M-SAE firing still rising at step 300.

Files

HF-format bf16 safetensors shards (model-*.safetensors + model-nontext.safetensors = the non-text base tensors), tokenizer, run_meta.json. Load with transformers as a Qwen3.6-27B causal LM; the injection hook lives in the MAEMM repo (mxf/inject.py).