It extends the original Qwen3.5 base with 16K new Kazakh tokens and continues pretraining on a large Kazakh + English corpus, to improve Kazakh fluency and knowledge while retaining the original English capabilities.

Training details

Tokenizer extension (+16K). The Qwen3.5 tokenizer (248,320) was extended with 16,000 new, predominantly Cyrillic-Kazakh tokens → 264,192 total. New embedding rows were initialized with FVT (Fast Vocabulary Transfer): each new row is the unweighted mean of its constituent original sub-token embeddings, giving the model a warm start on the new vocabulary. Applied under tied embeddings (FVT applied to the shared input/output matrix).
Continued pretraining. Full-parameter CPT for one epoch at sequence length 32,768, learning rate 1e-5 (cosine → 1e-6, 2% warmup), weight decay 0.1, gradient clip 1.0, on 2× 8×H100.

Data

~15.7B new tokenizer tokens total with a 50/50 Kazakh:English mix. The Kazakh half is a cross-source deduplicated corpus (FineWeb-2, CulturaX, FinePDFs, FineWiki). The English half is a replay mix for retention (FineWeb-Edu web, Nemotron-CC math + wiki-rewrite, Stack-Edu code).

Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer

name = "issai/Qwen3.5-4B-Base-Kazakh"
tokenizer = AutoTokenizer.from_pretrained(name)
model = AutoModelForCausalLM.from_pretrained(name, torch_dtype="bfloat16", device_map="auto")

prompt = "Сұрақ: Үстірт қорығы қай облыста орналасқан?\nЖауап:"

ids = tokenizer(prompt, return_tensors="pt").to(model.device)

print(tokenizer.decode(model.generate(**ids, max_new_tokens=100)[0][ids.input_ids.shape[1]:]))

Requires a recent transformers with Qwen3.5 support (>=5.9).

License & attribution

Continued-pretrained from Qwen/Qwen3.5-4B-Base; the original Qwen3.5 weights and their Apache-2.0 license apply.