Upload zimg-fdloss/RECIPE.md with huggingface_hub

zimg-fdloss/RECIPE.md

@@ -1,4 +1,12 @@
-# Flow Matching Mid/Post-training — Final Recipe Card (frozen 2026-08-04)
+# Flow Matching Mid/Post-training — Final Recipe Card (rev 2026-08-05)
+
+> **2026-08-05 champion promotion (pre-registered 4-seed test)**: the b64 +
+> 250-step-finishing stack ("I1combo") beat the frozen b16 champion at BOTH
+> operating points on 4 seeds (553×4 = 2212 paired items): GenEval @8
+> z=+2.91, @1 z=+2.42 (rule: z≥+2 at one endpoint AND z>0 at the other;
+> per-seed @1 deltas all positive). 4-seed macro: **0.910 @8 / 0.858 @1**
+> (old champion 0.899/0.842). Decomposition: b64 stage +2.82 @8 / +1.94 @1;
+> finishing stage +0.14 @8 / +0.87 @1. The recipe below reflects this.
 
 Model: Z-Image 6B S3-DiT @1024². Reference data: BLIP3o-60k (GPT-4o set),
 holdout = sorted-enumeration idx % 12 == 0.
@@ -18,12 +26,15 @@ holdout = sorted-enumeration idx % 12 == 0.
    standalone cross-covariance term on a factor-probe basis; conditional
    coupling via top-32 diag-whitened CCA text coordinates (K=32 optimal;
    K=64 does not transfer); EMA population moments (β=0.999). **Global
-   batch 64 (4-node) if available** — the sole surviving upgrade.
-3. **Optional TFC finishing dial** — short (250–500 step) pass tying the
-   1-jump endpoint's SigLIP2 features to sg(own multi-step endpoint):
-   redistributes quality between NFE operating points along a measured
-   exchange curve (~1.9 pts gained @1 per pt lost @8, pooled over 5 runs).
-   Not net creation. **Mechanism caveat (measured)**: the dial as shipped
+   batch 64 (4-node): promoted component — on 4 seeds this stage alone is
+   z=+2.82 @8 / +1.94 @1 over b16.**
+3. **TFC finishing pass (promoted @ b64, 250 steps, lr 3e-6)** — ties the
+   1-jump endpoint's SigLIP2 features to sg(own multi-step endpoint). On
+   the b64 base it adds @1 without measurable @8 cost (4-seed +0.14 @8 /
+   +0.87 @1); on b16 bases it acts as a redistribution dial along a
+   measured exchange curve (~1.9 pts gained @1 per pt lost @8, pooled
+   over 5 runs — not net creation). **Mechanism caveat (measured)**: the
+   pass as shipped
    uses the standard stats root, whose warm-start moments come from the
    iMF-stage model — a stale anchor relative to the checkpoint being
    finished. Rebuilding the warm-start from the finished model's own
@@ -32,18 +43,20 @@ holdout = sorted-enumeration idx % 12 == 0.
    champion): part of the dial's @1 gain is stale-anchor replay, not TFC
    alone. Keep the standard (stale) stats root when using the dial.
 
-## Headline numbers (GenEval macro, 553×2 VLM-judged; FDr held-out judge mean)
+## Headline numbers (GenEval macro; FDr held-out judge mean)
 
-| model | GenEval@8 | GenEval@1 | FDr@8 | FDr@1 |
-|---|---|---|---|---|
-| Z-Image-Turbo (distilled baseline) | 0.811 | 0.719 | 2.174 | 1.700 |
-| **champion (this recipe, b16)** | **0.900** | **0.842** | 1.612 | 1.519 |
-| champion @ batch 64 | 0.910 | 0.846 | 1.621 | 1.447 |
-| 1-NFE specialist (TFC full-run) | 0.843 | 0.879 | 1.909 | 1.596 |
+| model | seeds | GenEval@8 | GenEval@1 | FDr@8 | FDr@1 |
+|---|---|---|---|---|---|
+| Z-Image-Turbo (distilled baseline) | 2 | 0.811 | 0.719 | 2.174 | 1.700 |
+| b16 recipe (old champion) | 4 | 0.899 | 0.842 | 1.612 | 1.519 |
+| + batch 64 | 4 | 0.910 | 0.853 | 1.621 | 1.447 |
+| **+ 250-step finishing (CHAMPION)** | **4** | **0.910** | **0.858** | — | — |
+| 1-NFE specialist (TFC full-run, b16) | 2 | 0.843 | 0.879 | 1.909 | 1.596 |
 
-McNemar vs turbo: z=+7.0/+8.2 (Holm-robust). No teacher, no discriminator,
-no fake-score network, no CFG at train or inference; ~4.5 h on one 8×H200
-node (b16).
+McNemar: champion vs b16 recipe +2.91 @8 / +2.42 @1 (4 seeds, pre-registered
+promotion rule); b16 recipe vs turbo z=+7.0/+8.2 (Holm-robust). No teacher,
+no discriminator, no fake-score network, no CFG at train or inference;
+~4.5 h on one 8×H200 node (b16) / ~1.5 h wall on 4 nodes (b64 + finishing).
 
 ## The three laws (measured)
 
@@ -65,8 +78,9 @@ node (b16).
 
 ## Checkpoints in this folder
 
-- `champion-D1c-clean/` — the b16 champion (FD ckpt + iMF init)
-- `champion-D1c-b64/` — batch-64 variant (best both-points nominal)
+- `champion-I1combo/` — **the promoted champion** (b64 FD + 250s finishing)
+- `champion-D1c-clean/` — the b16 recipe (FD ckpt + iMF init)
+- `champion-D1c-b64/` — batch-64 stage (the promoted champion's base)
 - `T1-1nfe-record/` — best pure-1-NFE model (0.879@1)
 - `T3-balanced-finisher/`, `T4a-near-pareto-finisher/` — TFC dial points
 - `V1g0-clean-rerun-g0c/`, `champion-V1g0/` — recipe-ancestor reference