metadata

library_name: transformers
license: apache-2.0
datasets:
  - HuggingFaceTB/smoltalk
  - HuggingFaceH4/ultrafeedback_binarized
base_model:
  - JingzeShi/Doge-60M
language:
  - en
pipeline_tag: question-answering

Doge 60M Instruct

Doge is an ongoing research project where we aim to train a series of small language models to further explore whether the Transformer framework allows for more complex feedforward network structures, enabling the model to have fewer cache states and larger knowledge capacity.

In addition, 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 Jingze Shi, it only allows text input and text generation, for detailed algorithm and model architecture, please refer to Wonderful Matrices, all training details are in here.

Uses

from transformers import AutoTokenizer, AutoModelForCausalLM, GenerationConfig, TextStreamer

tokenizer = AutoTokenizer.from_pretrained("JingzeShi/Doge-60M-Instruct")
model = AutoModelForCausalLM.from_pretrained("JingzeShi/Doge-60M-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

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

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}, 
}