Downloads · 30 days
3
8% of all-time downloads
ceselder/skip-lens-qwen36-27b-agreelens-d
skip-lens-qwen36-27b-agreelens-d 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 card lists the license as apache-2.0.
One of four token-matched arms ablating what a futurelens learns from J-lens/answer-agreement-filtered training data. DISAGREE (span-miss) arm — trained ONLY on positions where the model's next token sits at J-lens ra…
Downloads · 30 days
3
8% of all-time downloads
All-time downloads
36
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
One of four token-matched arms ablating what a futurelens learns from J-lens/answer-agreement-filtered training data. DISAGREE (span-miss) arm — trained ONLY on positions where the model's next token sits at J-lens rank > 1000, entropy < 1.0, AND zero of the generated span's content tokens appear in the J-lens top-64 (the workspace provably contains nothing of the span). Weakest lens on BOTH harnesses (fedlayer workspace −0.071 vs G, p=0.001; near-bottom on workspace-bench) yet trained fine — it learned to read NON-J-space information (best held-out NLL on exactly the spans the J-lens misses).
All four arms are identical except the data filter: 244,367 on-policy 12-token spans each
(model's own continuations, temp 1.0 top-p 0.95) from a fresh 200k-doc FineFineWeb slice
(md5-disjoint from prior skip-lens pools, 67 domains), activations = exact full-context layer-42
residuals. bs 64 (no accumulation), lr 1e-4, LoRA r=64 α=16 rsLoRA scope-all (12 module types),
1 epoch = 3,818 steps, seed 0. Arms A/D/P share the entropy<1.0 gate with bin-matched entropy
histograms. Filter reference: the official released J-lens
(camilablank/workspace-lenses qwen3.6-27b/j-lens/lens.pt).
| metric | value |
|---|---|
| fedlayer workspace agreement (Sonnet-judged vs official J, 353 items, fed raw h42) | 0.543 |
| fedlayer surface-answer agreement | 0.414 |
| workspace-bench AO family wins (of 21 decided) | 2 |
| workspace-bench directed-modulation-mt net (opus-judged) | 0.083 |
| workspace-bench order_ops mean net (frozen L56/L60 cells) | -0.066 |
Full experiment + all 12 workspace-bench families:
http://5.78.192.0/reports/view/agreelens-tokmatch/report.html (internal) — key findings: agree-filtering
does NOT add J-lens agreement over entropy-matched random (A−P = +0.000, p=1.0) but produces the best
surfacer of imminent/held content on workspace-bench; disagree/span-miss training degrades workspace
readout on both harnesses while learning non-J-space reading.
PEFT LoRA adapter on Qwen/Qwen3.6-27B. Inject a raw layer-42 residual at the marker token ㈜
(id 158983) via the Karvonen norm-matched hook (h'_p = h_p + ||h_p||·v/||v||), prompt =
prompt_templates.actor in the bundled nla_meta.yaml (chat template, enable_thinking=False),
then generate (temp 1.0, top-p 0.95). Sibling arms:
ceselder/skip-lens-qwen36-27b-agreelens-{a,d,g,p}.