H-D-T/Select-Stack

Buzz: Advancing Efficiency through Iterative Fine-Tuning

Introduction

• Alignment Lab AI is pleased to introduce our latest research efforts with:

Buzz, a highly curated pretraining scale assistant dataset, unifying RL and SFT, developed in collaboration with Hive Digital Technologies.

The Buzz model, Dataset, and Code are to be released to build a toolkit that aims to demonstrate the potential for reuse and optimization of existing pretrained language models to continuously refine the heights of performance that can be achieved with optimal use of FlOps. Alongside Buzz-5b-Medium, we release

• The Buzz Dataset
• Buzz-2.5b-Small
• Buzz-8B-Large

Features

Buzz contains over 500 datasets, deduplicated, with formatting built to maintain and extend compatibility between training types and the current local ecosystem.

the datasets within are comprised of various high quality instruction following, conversational, storytelling, and coding datasets, as well as over 5 million new rows of data, in addition to several million reaugmented rows of data, comprising the totality of the learned techniques since our release of Open-Orca

Iterative Fine-Tuning Methodology

Our research builds upon the concepts introduced in several key papers, including:

• Simple and Scalable Strategies to Continually Pre-train Large Language Models
• NEFTune: Noisy Embeddings Improve Instruction Finetuning
• An Optimistic Acceleration of AMSGrad for Nonconvex Optimization
• Improving Generalization Performance by Switching from Adam to SGD
• Orca: Progressive Learning from Complex Explanation Traces of GPT-4

By combining high quality data, iterative fine-tuning with carefully selected "grounding" distributions from previous epochs, we have developed a cost-effective approach that pushes the boundaries of model reuse and optimization.

notably, we observe that training on a single epoch of high quality in domain data can still achieve remarkably low loss values before overfitting.

Data structure and formatting

buzz should be out of the box compatible with the sharegpt type in Axolotl and lmsys' FastChat during training it containsthe following structure

{
  "source": "string containing the source dataset",
  "stack": "chosen/rejected for RL techniques",
  "question_index": optional row, only contained in DPO specific dataset to match dpo pairs - int64
  "conversations": [
    {
      "from": "system",
      "value": "an initial system prompt or user query, may or may not be present depending on the row"
    },
    {
      "from": "human or system",
      "value": "an initial 'human' query"
    },
    {
      "from": "gpt",
      "value": "a response to the previous turn, may be followed by additional human/gpt alternations"
    }
  ]
}

Conclusion

We intend to focus on updating and improving the dataset, tools to construct it, and other surrounding open sourced infrastructure. Our next effort will focus on context and implementing the research currently being conducted by Wing-Lian, the lead developer of the Axolotl training framework that underpins these experiments. We encourage the community to explore Wing-Lian's work, such as the Llama-3-8b-64k-PoSE and llama-3-8b-256k-PoSE models, which showcase the potential for further advancements in language modeling.

Buzz hopes to be a proof of concept, and a toolkit to demonstrate and enable the community in the pursuit of efficient and effective locally run, personally owned, language models. Through collaboration with Hive Digital Technologies who have enabled us to perform this research, we have demonstrated the immense potential for model reuse and optimization. The Buzz models and dataset are open sourced with [////////].

Credits

to the many researchers who have open sourced their knowledge and tools to allow us to pursue this,

to Hive Digital Technologies for providing compute, advice, and meaningful research insight.

to Meta for developing the Llama models, and maintaining a philosophy of supporting open research and open source.

To wing et al. with Open Access AI Collective for developing axolotl, assisting with research, and generally being geniuses.

to Thomas Capelle et al. working on LLM_Surgery

as well as many, many others who are too numerous to name.

Citations

@misc{ibrahim2024simple,
      title={Simple and Scalable Strategies to Continually Pre-train Large Language Models}, 
      author={Adam Ibrahim and Benjamin Thérien and Kshitij Gupta and Mats L. Richter and Quentin Anthony and Timothée Lesort and Eugene Belilovsky and Irina Rish},
      year={2024},
      eprint={2403.08763},
      archivePrefix={arXiv},
      primaryClass={cs.LG}
}

@misc{jain2023neftune,
      title={NEFTune: Noisy Embeddings Improve Instruction Finetuning}, 
      author={Neel Jain and Ping-yeh Chiang and Yuxin Wen and John Kirchenbauer and Hong-Min Chu and Gowthami Somepalli and Brian R. Bartoldson and Bhavya Kailkhura and Avi Schwarzschild and Aniruddha Saha and Micah Goldblum and Jonas Geiping and Tom Goldstein},
      year={2023},
      eprint={2310.05914},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}

@misc{wang2020optimistic,
      title={An Optimistic Acceleration of AMSGrad for Nonconvex Optimization}, 
      author={Jun-Kun Wang and Xiaoyun Li and Belhal Karimi and Ping Li},
      year={2020},
      eprint={1903.01435},
      archivePrefix={arXiv},
      primaryClass={stat.ML}
}

@misc{keskar2017improving,
      title={Improving Generalization Performance by Switching from Adam to SGD}, 
      author={Nitish Shirish Keskar and Richard Socher},
      year={2017},
      eprint={1712.07628},
      archivePrefix={arXiv},
      primaryClass={cs.LG}
}

@misc{mukherjee2023orca,
      title={Orca: Progressive Learning from Complex Explanation Traces of GPT-4}, 
      author={Subhabrata Mukherjee and Arindam Mitra and Ganesh Jawahar and Sahaj Agarwal and Hamid Palangi and Ahmed Awadallah},
      year={2023},
      eprint={2306.02707},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}