Skip-Lens: Multi-Token NLA Lenses (Qwen3.6-27B)
Collection
a multi token J lens approach that may or may not be abandoned, here are a bunch of checkpoints β’ 12 items β’ Updated
agreelens arm D β L62-trained skip-lens variant (DISAGREE/span-miss-filtered)
Twin of ceselder/skip-lens-qwen36-27b-agreelens-d: exact same 244,367
(position, on-policy 12-token span) pairs, but the input activation is the layer-62
residual instead of layer-42. Intended test-time feed: J_42->62 @ h42 (the official released
Jacobian, camilablank/workspace-lenses) β the skip-lens trick. Same recipe otherwise
(bs 64, lr 1e-4, LoRA r64 Ξ±16 rsLoRA scope-all, 1 epoch, seed 0).
Fedlayer workspace agreement (Sonnet vs official J, 353 items): 0.660 J-fed / 0.659 raw-h42-fed (L42 twin: see sibling repo). The J-fed L62 arms are the strongest workspace readers in the agreelens series β skip-lens+Jacobian adds +0.14β0.17 over matched-L42 training.
Usage: PEFT LoRA on Qwen3.6-27B; capture h42 at the read position, apply J_42->62, inject at marker γ (id 158983) via Karvonen norm-matched hook; template in bundled nla_meta.yaml.
- Downloads last month
- 7
Inference Providers NEW
This model isn't deployed by any Inference Provider. π Ask for provider support
Model tree for ceselder/skip-lens-qwen36-27b-agreelens-d-l62
Base model
Qwen/Qwen3.6-27B