Downloads · 30 days
12
29% of all-time downloads
ceselder/skip-lens-olens-pi0-100k-qwen36-27b
skip-lens-olens-pi0-100k-qwen36-27b 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.
Inject a raw layer-42 residual activation and this LoRA emits four bullet points describing what the model is about to generate next — a native L42 reader, no cross-layer transport at inference.
Downloads · 30 days
12
29% of all-time downloads
All-time downloads
42
Public
Repo size
1.9 GB
Likes
0
Public
Click a slice to open those files.
.safetensors1.9 GB · 99%
From the Hugging Face model README
Inject a raw layer-42 residual activation and this LoRA emits four bullet points describing what the model is about to generate next — a native L42 reader, no cross-layer transport at inference.
The olens (oracle-lens) is built by best-of-N expert iteration and is a post-train of the
RL-surprisal activation verbalizer (ceselder/skip-lens-qwen36-27b-RL-surprisal). π₀ is the seed round.
J[42]→L62, camilablank/workspace-lenses)
to mine 32 rollouts / activation.ceselder/skip-lens-ar-l42-jacobian-qwen36-27b) against the raw L42 activation; whitened
NNOLS-select 4 bullets (FVE-ordered).Qwen/Qwen3.6-27B.㈜ marker (token id 158983), on
the output of transformer block 1: h'_p = h_p + ‖h_p‖ · v/‖v‖.*-prefixed bullets (what the model is about to say).There is no separate system message — this is the single user turn, wrapped by the Qwen3.6 chat
template with enable_thinking=False (matching training). The ㈜ (id 158983) is the injection marker
where the raw L42 activation is spliced (block-1 norm-match); do not replace it with text.
You are shown an internal activation vector captured from a language model as it reads a passage of text. The vector, enclosed in <concept> tags, is taken at one position and encodes what the model is about to generate next. Output four bullet points, each starting with '*', describing what the model is most likely about to generate.
<concept>㈜</concept>
Build it in code as tokenizer.apply_chat_template([{"role":"user","content": PROMPT}], add_generation_prompt=True, enable_thinking=False). The same template (and injection config: layer, marker, canonical neighbors) is in nla_meta.yaml.
Injection config (layer, marker, neighbors) and the actor prompt live in nla_meta.yaml.
Built via BoN expert iteration; π₀ = seed. Iteration (π₁…) closes the 0.37→0.49 whitened-FVE gap.