This is the submission-facing ARMOR-DP checkpoint for the BabyLM 2026 multilingual track. It uses UTF-8 byte tokenization, language adapters, and a curriculum-dynamic exposure controller over English, Dutch, and Chinese.

Repository: liaocccccc/armor-dp-babylm-2026-d768-currdyn
Source checkpoint: final_100m10_l512_b4_d768l12_curriculum_dynamic
Target budget: 100000000
Predicted non-padding tokens: 86991960
Context length: 512
Architecture: d=768, layers=12, heads=12
Schedule: curriculum_dynamic
Mechanism: language_adapters
Max-epoch cap: 10
Consumed epochs: EN=6.6105, NL=6.5892, ZH=6.3699

Use with trust_remote_code=True:

python

from transformers import AutoTokenizer, AutoModelForCausalLM

tok = AutoTokenizer.from_pretrained("liaocccccc/armor-dp-babylm-2026-d768-currdyn", trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained("liaocccccc/armor-dp-babylm-2026-d768-currdyn", trust_remote_code=True)