TheBeagle-v2beta-32B-MGS

This model is an experimental version of our latest innovation: MGS. Its up to you to figure out what does it means, but its very explicit. We didn't applied our known UNA algorithm to the forward pass, but they are entirely compatible and operates in different parts of the neural network and in different ways, tho they both can be seen as a regularization technique.

CHANGELOG

UPDATE: 26/Oct

• Updated tokenizer_config.json (from the base_model)
• Regenerated Quants (being uploaded)
• Re-submitted Leaderboard Evaluation, MATH & IFeval have relevant updates
• Aligned LICENSE with Qwen terms.

MGS

MGS stands for... Many-Geeks-Searching... and thats it. Hint: 1+1 is 2, and 1+1 is not 3

We still believe on 1-Epoch should be enough, so we just did 1 Epoch only.

Dataset

Used here the first decent (corpora & size) dataset on the hub: Magpie-Align/Magpie-Pro-300K-Filtered Kudos to the Magpie team to contribute with some decent stuff that I personally think is very good to ablate.

It achieves the following results on the evaluation set:

• Loss: 0.5378 (1 Epoch), outperforming the baseline model.

Quants

All versions available

Licensing terms:

On top of the Qwen LICENSE, we add an extra term for derivatives to include "Beagle" or "MGS" on the model name, this will help us to track better the study. Thank you

Training

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 8e-05
• train_batch_size: 2
• eval_batch_size: 2
• seed: 42
• distributed_type: multi-GPU
• num_devices: 8
• gradient_accumulation_steps: 4
• total_train_batch_size: 64
• total_eval_batch_size: 16
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 25
• num_epochs: 1

Training results

Training Loss	Epoch	Step	Validation Loss
9.8642	0.0012	1	0.7195
2.077	0.0507	42	0.6161
1.0325	0.1014	84	0.6093
0.8945	0.1520	126	0.5962
0.8532	0.2027	168	0.5869
0.8185	0.2534	210	0.5805
0.81	0.3041	252	0.5719
0.7901	0.3548	294	0.5663
0.7766	0.4054	336	0.5618
0.7687	0.4561	378	0.5590
0.7443	0.5068	420	0.5564
0.7494	0.5575	462	0.5525
0.7787	0.6081	504	0.5485
0.7381	0.6588	546	0.5466
0.7359	0.7095	588	0.5444
0.7447	0.7602	630	0.5435
0.7378	0.8109	672	0.5415
0.7302	0.8615	714	0.5398
0.7476	0.9122	756	0.5391
0.715	0.9629	798	0.5378

Open LLM Leaderboard Evaluation Results

Detailed results can be found here

Metric	Value
Avg.	40.29
IFEval (0-Shot)	45.03
BBH (3-Shot)	58.07
MATH Lvl 5 (4-Shot)	39.43
GPQA (0-shot)	20.13
MuSR (0-shot)	24.50
MMLU-PRO (5-shot)	54.57

Thanks

• Qwen Team for their outstanding model
• MagPie Team for contributing plenty of datasets
• Cybertron Cloud Compute

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{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}
}