release: multi-model card, fixed links

README.md

@@ -6,28 +6,56 @@ colorTo: blue
 sdk: docker
 app_port: 7860
 pinned: false
+models:
+  - oddadmix/Nawah-Router-BERT-6M-v2
+  - oddadmix/Nawah-Router-v3
+  - oddadmix/Nawah-BERT-6M-v2
+datasets:
+  - oddadmix/arabic-prompt-routing
+tags:
+  - arabic
+  - zero-shot-classification
+  - prompt-routing
 short_description: توجيه عربي صفري — نص + أي فئات تكتبها → الفئة المناسبة
 ---
 
 # Nawah-Router — Arabic zero-shot prompt routing
 
 Write a text and **any categories in plain Arabic**; the model scores all of them in a single
-forward pass. Categories are free text typed at runtime — the model has no fixed taxonomy.
-Arabic counterpart to [LFM2.5-Encoder-350M-Prompt-Router](https://huggingface.co/LiquidAI/LFM2.5-Encoder-350-Prompt-Router).
+forward pass. Categories are free text typed at runtime — the model has no fixed taxonomy, so
+this is zero-shot classification over a label set it has never seen.
 
 Edit any category inline, add or remove lanes, and the routing updates as you type.
+Arabic counterpart to [LFM2.5-Encoder-350M-Prompt-Router](https://huggingface.co/LiquidAI/LFM2.5-Encoder-350-Prompt-Router).
 
-| eval set | accuracy | random |
+## Two backbones, one head — switch from the picker
+
+| | [`Nawah-Router-BERT-6M-v2`](https://huggingface.co/oddadmix/Nawah-Router-BERT-6M-v2) | [`Nawah-Router-v3`](https://huggingface.co/oddadmix/Nawah-Router-v3) |
 |---|---:|---:|
-| unseen category sets | **0.9305** | 0.2137 |
-| unseen domains | 0.6976 | 0.2521 |
-| **unseen axes** (dimensions absent from training) | 0.6130 | 0.2358 |
-| deliberately adjacent categories | 0.9008 | 0.2109 |
+| parameters | **5,977,985** | 52,049,409 |
+| backbone | BERT encoder | Llama decoder |
+| base model | [`Nawah-BERT-6M-v2`](https://huggingface.co/oddadmix/Nawah-BERT-6M-v2) | `50M-2048-Emhotob` |
+
+Both scored in one session by the same `eval_router_only.py`; Router-v3 reproduced its published
+card exactly.
 
-Model: [`oddadmix/Nawah-Router-v3`](https://huggingface.co/oddadmix/Nawah-Router-v3) — 52M
-parameters, CPU-only, one forward pass regardless of how many categories you give it.
+| eval set | **6M encoder** | 52M decoder | random |
+|---|---:|---:|---:|
+| unseen category sets | **0.9327** | 0.9308 | 0.2137 |
+| unseen domains | **0.7009** | 0.6975 | 0.2521 |
+| deliberately adjacent categories | **0.9101** | 0.9017 | 0.2109 |
+| unseen axes (dimensions absent from training) | 0.6000 | **0.6127** | 0.2358 |
+
+The 6M wins three of four at **1/8.7 the size**. It loses on unseen axes — the column the 52M's
+card calls its strongest claim — and three of the four margins are under one point, which is
+inside what a single run can tell you.
 
 Confidence is not calibrated: clear cases saturate near 100%. Use the ranking, not the number.
 
-`MODEL_HF_TOKEN` is required while the model repo is private (`HF_TOKEN` is reserved by Spaces and
-never reaches the container).
+Data: [`oddadmix/arabic-prompt-routing`](https://huggingface.co/datasets/oddadmix/arabic-prompt-routing)
+— 207,097 training rows, 58,008 distinct category sets, 12 routing axes, with the full generation
+and verification pipeline.
+
+CPU-only, one forward pass regardless of how many categories you give it.
+
+© KAND CA 2026 — PROJECT NAWAH