Esobold/Style_Masks

Esobold Style Masks — J-line (2026-05-29)

Target-side per-weight style masks for Gemma-4-31B merges. Each style_ratio.pt is a dict {linear_weight_key: r_w} (bf16), where r_w ∈ [0,1] is the per-element interpolation weight toward model A vs model B, derived from a contrastive teacher-forced Fisher:

F_A = relu( Fisher(A | S+) − Fisher(A | S−) )      # style-salient weights of A
F_B = relu( Fisher(B | S+) − Fisher(B | S−) )      # style-salient weights of B
each normalized by its own global max (models have different Fisher scales)
r_w = F_A / (F_A + F_B + eps)
where (F_A+F_B) below the 50th-pct floor → r_w = 0.5 (neutral)

• S+ = "good style" system prompt, S− = "slop" system prompt (see system_prompts.json).
• Generation: transformers greedy, thinking-OFF, 40 seeded prompts (Eso-5 genre set).
• 410 target LM-linear keys, 60 layers. ratio_layer_summary.json = per-layer mean r_w / mean F_A / mean F_B.
• r_w > 0.5 → keep weight closer to A; r_w < 0.5 → closer to B.

j1_iv_glimmer/ — the J-line style mask

• A = Iv3 = Gv3-rebuilt ⊕ Equinox, TIES density 0.3 / scale 0.8, thinkfisher mask.
• B = glimmer-on-Iv3 = toasty glimmer-rp v0.1 LoRA folded onto Iv3.
• mean r_w 0.380 (leans toward glimmer). Purpose: inject glimmer's prose style into Iv3 while protecting instruct-Fisher directions.

gv3_equinox/ — Mask 2 (#74)

• A = Gv3 (Esobold/g4_31b_Gv3_Stage1), B = Equinox (LatitudeGames/Equinox-31B).
• mean r_w 0.343 (leans toward Equinox).

The stage-04 merge that applies a mask runs separately (uses the IT-base instruct_fisher to protect instruction-following). Built on RunPod H100 NVL, 2026-05-29.