Frankenstein v1 — qwen2.5-7B Instruct ⨯ Coder SLERP graft

layer-wise SLERP · observation-informed t profile · 11 triple-confirmed dim anchors
28 layers 3,584 dim 1,697 tensor ops 11 rose-gold anchors
Three cones, same coordinates: Instruct (left), Hybrid (centre), Coder (right). Each point = a (layer, dim, quantile) cell. Hybrid colour = the SLERP t value at that layer. Rose-gold lines = the 11 triple-confirmed anchors preserved across the graft.

Hover any point or line for endpoint info.

layers

Instruct cone
Hybrid cone
Coder cone
11 rose-gold anchorsverticals + bridges
49 cross-model overlapfaint

SLERP t profile (merge-v1)

self_attn t: [0.1, 0.2, 0.4, 0.6, 0.8]
mlp t: [0.1, 0.2, 0.5, 0.7, 0.9]
others: 0.5
base model: Instruct
drag rotate · scroll zoom · hover info · space auto-rotate