cybertron-v4-qw7B-UNAMGS

UNA IS BACK Cybertron v4 UNA-MGS, Based on the amazing Qwen2.5 7B

SCORING #1 7-8B LLM WITH NO CONTAMINATION 21.11.2024 with avg. 31.82

This special edition went thru UNA at MLP layers just like miniclaus-1.5B

Here we use our novel approach called MGS. Its up to you to figure out what it means. On top of that we used UNA: Uniform Neural Alignment

Cybertron V4 went thru SFT with MGS & UNA over Magpie-Align/Magpie-Qwen2.5-Pro-1M-v0.1 dataset.

Contamination Benchmark

https://gair-nlp.github.io/benbench/

• MATH:

5gram-Qwen2.5-7B-Instruct-orgn-MATH-test.jsonl: 37.52666666666667
5gram-Qwen2.5-7B-Instruct-orgn-MATH-train.jsonl: 46.36666666666667

vs

5gram-UNA-cybertron-v4-qw7B-MGS-orgn-MATH-test.jsonl: 37.42666666666667
5gram-UNA-cybertron-v4-qw7B-MGS-orgn-MATH-train.jsonl: 46.053333333333335

vs

5gram-Homer-v0.5-orgn-MATH-test.jsonl: 38.77333333333333
5gram-Homer-v0.5-orgn-MATH-train.jsonl: 47.16666666666667

Quantz

Available at bartowski repo

https://huggingface.co/bartowski/cybertron-v4-qw7B-UNAMGS-GGUF

Open LLM Leaderboard Evaluation Results

Detailed results can be found here

MGS & UNA & Details

• MGS, 1+1 = 2 and not 3
• UNA, 1+1 = 2 obviously

We also followed https://arxiv.org/pdf/2410.21228 insights.

Training procedure

1 Epoch as usual.

datasets:
  - path: Magpie-Align/Magpie-Qwen2.5-Pro-1M-v0.1
    split: train
    type: chat_template
    field_messages: conversations
    message_field_role: from
    message_field_content: value
    roles:
      user: ["human", "user"]
      assistant: ["gpt", "assistant", "ai"]
      system: ["system"]

Training hyperparameters

The following hyperparameters were used during training:

• seed: 42
• distributed_type: multi-GPU
• num_devices: 8
• total_train_batch_size: 64
• total_eval_batch_size: 16
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• num_epochs: 1

Training results

Framework versions

• PEFT 0.13.2
• Transformers 4.45.2 (UNA & MGS patch)
• Pytorch 2.3.0+cu121
• Datasets 3.0.1
• Tokenizers 0.20.1

Citations

@misc{thebeagle-v2,
  title={TheBeagle v2: MGS}, 
  author={Xavier Murias},
  year={2024},
  publisher = {HuggingFace},
  journal = {HuggingFace repository},
  howpublished = {\url{https://huggingface.co/fblgit/TheBeagle-v2beta-32B-MGS}},
}

@misc{Magpie,
    title={Magpie: Alignment Data Synthesis from Scratch by Prompting Aligned LLMs with Nothing}, 
    author={Zhangchen Xu and Fengqing Jiang and Luyao Niu and Yuntian Deng and Radha Poovendran and Yejin Choi and Bill Yuchen Lin},
    year={2024},
    eprint={2406.08464},
    archivePrefix={arXiv},
    primaryClass={cs.CL}
}

@misc{qwen2.5,
    title = {Qwen2.5: A Party of Foundation Models},
    url = {https://qwenlm.github.io/blog/qwen2.5/},
    author = {Qwen Team},
    month = {September},
    year = {2024}
}

@article{qwen2,
      title={Qwen2 Technical Report}, 
      author={An Yang and Baosong Yang and Binyuan Hui and Bo Zheng and Bowen Yu and Chang Zhou and Chengpeng Li and Chengyuan Li and Dayiheng Liu and Fei Huang and Guanting Dong and Haoran Wei and Huan Lin and Jialong Tang and Jialin Wang and Jian Yang and Jianhong Tu and Jianwei Zhang and Jianxin Ma and Jin Xu and Jingren Zhou and Jinze Bai and Jinzheng He and Junyang Lin and Kai Dang and Keming Lu and Keqin Chen and Kexin Yang and Mei Li and Mingfeng Xue and Na Ni and Pei Zhang and Peng Wang and Ru Peng and Rui Men and Ruize Gao and Runji Lin and Shijie Wang and Shuai Bai and Sinan Tan and Tianhang Zhu and Tianhao Li and Tianyu Liu and Wenbin Ge and Xiaodong Deng and Xiaohuan Zhou and Xingzhang Ren and Xinyu Zhang and Xipin Wei and Xuancheng Ren and Yang Fan and Yang Yao and Yichang Zhang and Yu Wan and Yunfei Chu and Yuqiong Liu and Zeyu Cui and Zhenru Zhang and Zhihao Fan},
      journal={arXiv preprint arXiv:2407.10671},
      year={2024}
}

@article{xu2024benchmarking,
      title={Benchmarking Benchmark Leakage in Large Language Models},
      author={Xu, Ruijie and Wang, Zengzhi and Fan, Run-Ze and Liu, Pengfei},
      year={2024},
      journal={arXiv preprint arXiv:2404.18824},
      url={https://arxiv.org/abs/2404.18824}
}