Update v2 numbers (retrained head + fuser)

README.md

@@ -40,14 +40,15 @@ in-house model and a cloud frontier VLM reference.
 <!-- RESULTS-TABLE:START — filled from leaderboard/leaderboard.json by the scoring run -->
 | Benchmark (macro-F1) | `document-classification-v2` | best generalist | best cloud VLM |
 | --- | ---: | ---: | ---: |
-| **DocLayNet** (page types) | **0.88** | 0.63 | 0.83 |
+| **DocLayNet** (page types) | **0.97** | 0.63 | 0.83 |
 | **Forms** | **1.00** | 0.11 | 1.00 |
-| **Tobacco** (doc types) | 0.69 | 0.43 | **0.85** |
-| **OOD** (unseen doc types) | **0.97** | — | — |
-| **OOV** (synonym wording) | **0.80** | — | — |
+| **Tobacco** (doc types) | 0.74 | 0.43 | **0.85** |
+| **OOD** (unseen doc types) | **0.95** | — | — |
+| **OOV** (synonym wording) | **0.83** | — | — |
 
-Latency: **~8 docs/s on an A40** (p50 106 ms / p95 194 ms). The open-weight
-[`v1`](https://huggingface.co/nutrientdocs/document-classification-v1) runs ~3–7 docs/s on the same GPU.
+Latency (fused image+text, A40): **~4.9 pages/s** on text-heavy pages; the image branch alone runs ~8 docs/s
+when no OCR is present. The open-weight
+[`v1`](https://huggingface.co/nutrientdocs/document-classification-v1) runs ~5.7 pages/s fused on the same GPU.
 <!-- RESULTS-TABLE:END -->
 
 Every model — ours and cloud — is scored by the same open macro-F1 scorer; full per-model ranking on the
@@ -55,10 +56,10 @@ Every model — ours and cloud — is scored by the same open macro-F1 scorer; f
 
 - **Matches/leads the cloud on visual document-type tracks** (DocLayNet, Forms) at **zero per-request API
   cost** — and runs on a single GPU.
-- **Trails on Tobacco** (0.69 vs 0.85): that track rewards *reading* fine header text (memo vs letter vs
+- **Trails on Tobacco** (0.74 vs 0.85): that track rewards *reading* fine header text (memo vs letter vs
   email) — a large VLM reads it; an embedding model can't.
-- **OOD robustness:** on document types absent from training (invoices, handwriting, charts, tables), 0.97
-  macro-F1. **Open-vocab (OOV):** under never-seen synonym label wording, 0.80 — matches the *concept*, not
+- **OOD robustness:** on document types absent from training (invoices, handwriting, charts, tables), 0.95
+  macro-F1. **Open-vocab (OOV):** under never-seen synonym label wording, 0.83 — matches the *concept*, not
   your exact string. (Cloud VLMs train on ~all data, so these held-out axes aren't reported for them.)
 
 ## Intended use & limits