Doge 120M MoE

---

Doge uses Dynamic Mask Attention as sequence transformation and can use Multi-Layer Perceptron or Cross Domain Mixture of Experts as state transformation. Dynamic Mask Attention allows the Transformer to use self-attention during training and state space during inference, and Cross Domain Mixture of Experts can directly inherit the weights of Multi-Layer Perceptron for further training. This model is trained by SmallDoge community, for detailed algorithm and model architecture, paper coming soon, all training details and code are available in the small-doge repository.

Uses

>>> from transformers import AutoTokenizer, AutoModelForCausalLM

>>> tokenizer = AutoTokenizer.from_pretrained("SmallDoge/Doge-120M-MoE")
>>> model = AutoModelForCausalLM.from_pretrained("SmallDoge/Doge-120M-MoE", trust_remote_code=True)
>>> inputs = tokenizer("Hey how are you doing?", return_tensors="pt")

>>> out = model.generate(**inputs, max_new_tokens=100)
>>> print(tokenizer.batch_decode(out))

Model Details

We build the Doge by doing Per-Training on Smollm-Corpus. If you want to continue pre-training this model, you can find the unconverged checkpoint here. These models has not been fine-tuned for instruction, the instruction model is here.

Pre-Training:

Evaluation:

Procedure:

Environment:

• Image: nvcr.io/nvidia/pytorch:24.12-py3
• Hardware: 1x NVIDIA RTX 4090
• Software: Transformers

Citation

@misc{smalldoges,
  title={SmallDoges: A Family of Dynamic UltraFast Small Language Models}, 
  author={Jingze, Shi and Yifan, Wu and Bingheng, Wu and Yuyu, Luo},
  year={2025},
  month={March},
  url={https://github.com/SmallDoges/small-doge}
}