SmallDoge/Doge-160M-Instruct

Doge 160M Instruct

---

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, please refer to Wonderful Matrices, all training details and code are publicly available on the small-doge repository.

Uses

from transformers import AutoTokenizer, AutoModelForCausalLM, GenerationConfig, TextStreamer

tokenizer = AutoTokenizer.from_pretrained("SmallDoge/Doge-160M-Instruct")
model = AutoModelForCausalLM.from_pretrained("SmallDoge/Doge-160M-Instruct", trust_remote_code=True)

generation_config = GenerationConfig(
      max_new_tokens=100, 
      use_cache=True, 
      do_sample=True, 
      temperature=0.8, 
      top_p=0.9,
      repetition_penalty=1.0
)
steamer = TextStreamer(
      tokenizer=tokenizer, 
      skip_prompt=True
)

prompt = "Hi, how are you doing today?"
conversation = [
      {"role": "user", "content": prompt}
]
inputs = tokenizer.apply_chat_template(
    conversation=conversation,
    tokenize=True,
    return_tensors="pt",
)

outputs = model.generate(
    inputs, 
    tokenizer=tokenizer,
    generation_config=generation_config, 
    streamer=steamer
)

Model Details

We build the Doge-Instruct by first SFT on SmolTalk and then DPO on UltraFeedback Binarized.

TODO: The larger model is under training and will be uploaded soon.

SFT:

Model	Training Data	Epochs	Content Length	LR	Batch Size	Precision
Doge-20M-Instruct-SFT	HuggingFaceTB/smoltalk	2	2048	8e-4	0.25M	bfloat16
Doge-60M-Instruct-SFT	HuggingFaceTB/smoltalk	2	2048	6e-4	0.25M	bfloat16
Doge-160M-Instruct-SFT	HuggingFaceTB/smoltalk	2	2048	4e-4	0.25M	bfloat16

DPO:

Model	Training Data	Epochs	Content Length	LR	Batch Size	Precision
Doge-20M-Instruct	HuggingFaceH4/ultrafeedback_binarized	2	1024	8e-5	0.125M	bfloat16
Doge-60M-Instruct	HuggingFaceH4/ultrafeedback_binarized	2	1024	6e-5	0.125M	bfloat16
Doge-160M-Instruct	HuggingFaceH4/ultrafeedback_binarized	2	1024	4e-5	0.125M	bfloat16

Evaluation:

Model	IFEval (Prompt Strict Acc)	MMLU	BBH	ARC	PIQA	HellaSwag	tokens / s on i7-11 CPU
Doge-20M-Instruct	7.3	26.3	18.3	29.2	57.8	27.8	142
Doge-60M-Instruct	7.4	27.5	27.7	37.5	61.4	32.1	62
SmolLM-135M-Instruct	12.2	28.3	25.2	33.9	64.0	38.9	30
Doge-160M-Instruct	16.8	29.7	29.1	42.8	64.1	37.1	28

Procedure:

SFT:

DPO:

Environment:

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

Citation

@misc{shi2024wonderfulmatrices,
      title={Wonderful Matrices: Combining for a More Efficient and Effective Foundation Model Architecture}, 
      author={Jingze Shi and Bingheng Wu},
      year={2024},
      eprint={2412.11834},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2412.11834}, 
}