Nawah-Router-BERT-6M-v2 — موجّه عربي صفري

Give it a text and any categories in plain Arabic; it scores all of them in one forward pass. No fixed taxonomy. 5,977,985 parameters against Nawah-Router-v3's 52,049,409 — 8.7× smaller.

Results — it beats the 52M on three of four

All three models scored in one session by the shipped eval_router_only.py; Router-v3 reproduced its published card (0.9305 / 0.6976 / 0.6130 / 0.9008).

eval	Router-v3 (Llama, 52M)	BERT-6M v1	this model, v2	random
unseen category sets	0.9308	0.9137	0.9327	0.2137
unseen domains	0.6975	0.6696	0.7009	0.2521
unseen axes	0.6127	0.6024	0.6000	0.2358
deliberately adjacent categories	0.9017	0.8924	0.9101	0.2109

A 6M encoder outperforms the 52M decoder on unseen category sets (+0.2), unseen domains (+0.3) and adjacent categories (+0.8), at 1/8.7 the size.

It loses on unseen axes — 0.6000 against 0.6127 — and that is the column the original card calls its strongest claim: tools and retrieval appear nowhere in training. Note also that v2 is worse than v1 there (0.6024), the only metric across all three tasks where more pretraining hurt. Three of the four margins are under one point, which is within what a single run can tell you.

Why no pooling change was needed

RouterModel pools each category's token span via bmm(cat_pool, hidden) rather than a fixed position, so it is architecture-agnostic and a BERT backbone drops in unchanged. The sibling Nawah-Guard-BERT-6M-v2 needed a mean-pooling patch; this did not.

Usage

from transformers import AutoTokenizer
from routing_model import RouterModel, route      # ships in this repo

M = "oddadmix/Nawah-Router-BERT-6M-v2"
tok = AutoTokenizer.from_pretrained(M)
model = RouterModel.from_pretrained(M)

route(model, tok, "كم صار سعر صرف الدولار اليوم؟",
      ["بحث في الويب", "حاسبة", "تقويم ومواعيد", "لا يحتاج أداة"])

Trained on oddadmix/arabic-prompt-routing (207,097 rows, 12 axes) with the dataset's own train_router_head.py at its defaults — decoder-tuned.