Upload README.md with huggingface_hub

README.md

@@ -1,16 +1,27 @@
----
-tags: [gemma-4, 31b, style-mask, fisher, j-line, merge]
----
-# Esobold Style_Masks
+# Esobold Style Masks — J-line (2026-05-29)
 
-Per-weight **contrastive style-Fisher ratios** for style-aware 31B merges (J-line / target-side style mask).
+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**:
 
-Each subfolder holds a `style_ratio.pt` (bf16, target LM-linear keys only: q/k/v/o_proj + mlp
-gate/up/down in `language_model.layers.*`) + `ratio_layer_summary.json`.
+```
+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)
+```
 
-`r_w = F_A / (F_A + F_B)`, where `F_M = relu(Fisher(M under S+ style prompt) - Fisher(M under S-))`
-over matched generations. `r_w > 0.5` => weight carries more good-style signal in model A;
-neutral 0.5 below a combined-Fisher floor. Consumed by `04_merge_style.py` (trim-then-weight TIES).
+- **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**.
 
-Pipeline: `instruct_mask/style_mask_j1/`. Public so multiple ~35GB masks fit without private-storage limits.
-Initialized 2026-05-29T01:01:19Z.
+## `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.