Nawah-BERT-6M-bilingual — pretrained from scratch on 5B Arabic + 5B English tokens

Same architecture as Nawah-BERT-6M-v2 (hidden 128, 8 layers, 2 heads, 5,993,600 params) but pretrained on a genuinely balanced bilingual corpus with a tokenizer built for both languages, rather than an Arabic-only backbone.

Why this exists

A router-head experiment on Nawah-BERT-6M-v2 (Arabic-only pretraining) found that training an English routing head on it alone produced a model that learned nothing — loss stayed flat, every eval landed at random chance. Mixing in Arabic router-training data rescued English performance, but that only fixes the downstream task; the backbone itself was still never taught English. This model tests the more direct fix: pretrain the backbone itself on both languages, from scratch.

What's different from Nawah-BERT-6M-v2

	Nawah-BERT-6M-v2	this model
pretraining tokens	15B (Arabic only, 2 stages)	10B (5B Arabic + 5B English, one pass, shuffled)
tokenizer	custom_llama_tokenizer (32K, fit on Arabic only)	bilingual32k_tokenizer (32K, fit on interleaved Arabic+English documents)
English tokenization cost	~2.02 tokens/word (byte-fragmented)	~1.15 tokens/word
Arabic tokenization cost	~1.20 tokens/word	~1.37 tokens/word (the real, accepted trade-off)
architecture	hidden 128, 8 layers	identical

The tokenizer swap is not free for Arabic — giving English a fair share of the 32K vocabulary budget costs Arabic some merge efficiency. That trade-off is deliberate and documented in train_tokenizer_bilingual.py.

Training

• Two 5B-token bins (kaust-generative-ai/fineweb-edu-ar, HuggingFaceFW/fineweb-edu), each tokenized fresh with bilingual32k_tokenizer, sampled 50/50 per chunk with a fixed permutation (DualMemmapDataset, forked from the dialect-encoder run's dual-bin design) — a genuine single pass over both languages with no repetition on either side.
• 2,048 chunks held out from training entirely (1,019 Arabic / 1,029 English), never seen during pretraining.
• Same recipe as the rest of the BERT ladder: 30%→15% MLM mask anneal, LR 6e-4 from scratch, 2,048-token context, sparse MLM head (masked positions only).
• 10B tokens, 38,130 steps, 3h 36m on a single consumer GPU.

Held-out results (2,048 chunks never trained on, measured after training)

	MLM loss	perplexity	masked tokens
Arabic	3.4300	30.88	311,982
English	3.1132	22.49	315,240

Both languages learned; English scoring slightly better than Arabic here despite the tokenizer trade-off running the other way, plausibly reflecting fineweb-edu's more uniform filtering versus fineweb-edu-ar's — not something this run isolates further.

Usage

from transformers import AutoTokenizer, AutoModelForMaskedLM

M = "oddadmix/Nawah-BERT-6M-bilingual"
tok = AutoTokenizer.from_pretrained(M)
model = AutoModelForMaskedLM.from_pretrained(M)

Ships as a plain BertForMaskedLM (the sparse-head training trick used to speed up pretraining is not part of the saved architecture). Downstream fine-tuning (classification, routing) needs mean pooling over non-pad positions, not [CLS] — this is a packed-corpus BERT with no meaningful CLS position, the same trap documented on every other model in this family.

© KAND CA 2026 — PROJECT NAWAH