DeepSeek LLM: Scaling Open-Source Language Models with Longtermism

TL;DR

DeepSeek LLM scales open-source models; 67B model outperforms LLaMA-2 70B in code, math, reasoning.

cs.CL 🔴 Advanced 2024-01-06 40 views
DeepSeek-AI : Xiao Bi Deli Chen Guanting Chen Shanhuang Chen Damai Dai Chengqi Deng Honghui Ding Kai Dong Qiushi Du Zhe Fu Huazuo Gao Kaige Gao Wenjun Gao Ruiqi Ge Kang Guan Daya Guo Jianzhong Guo Guangbo Hao Zhewen Hao Ying He Wenjie Hu Panpan Huang Erhang Li Guowei Li Jiashi Li Yao Li Y. K. Li Wenfeng Liang Fangyun Lin A. X. Liu Bo Liu Wen Liu Xiaodong Liu Xin Liu Yiyuan Liu Haoyu Lu Shanghao Lu Fuli Luo Shirong Ma Xiaotao Nie Tian Pei Yishi Piao Junjie Qiu Hui Qu Tongzheng Ren Zehui Ren Chong Ruan Zhangli Sha Zhihong Shao Junxiao Song Xuecheng Su Jingxiang Sun Yaofeng Sun Minghui Tang Bingxuan Wang Peiyi Wang Shiyu Wang Yaohui Wang Yongji Wang Tong Wu Y. Wu Xin Xie Zhenda Xie Ziwei Xie Yiliang Xiong Hanwei Xu R. X. Xu Yanhong Xu Dejian Yang Yuxiang You Shuiping Yu Xingkai Yu B. Zhang Haowei Zhang Lecong Zhang Liyue Zhang Mingchuan Zhang Minghua Zhang Wentao Zhang Yichao Zhang Chenggang Zhao Yao Zhao Shangyan Zhou Shunfeng Zhou Qihao Zhu Yuheng Zou
deep learning language models open-source longtermism scaling laws

Key Findings

Methodology

The study explores scaling laws to develop DeepSeek LLM, focusing on long-term growth. It uses a dataset of 2 trillion tokens for pre-training and applies supervised fine-tuning and direct preference optimization to base models, creating DeepSeek Chat models.

Key Results

  • DeepSeek LLM 67B surpasses LLaMA-2 70B in benchmarks for code, math, and reasoning.
  • In open-ended evaluations, DeepSeek LLM 67B Chat outperforms GPT-3.5.
  • Data quality significantly affects model/data scaling strategies, enabling larger models with high-quality data.

Significance

This research offers a new perspective on scaling open-source language models, particularly enhancing performance in code and math. It provides a theoretical foundation for future model scaling and highlights the critical role of data quality in model performance.

Technical Contribution

Introduces a new model/data scaling allocation strategy using non-embedding FLOPs/token for more accurate scaling and performance prediction. Utilizes a multi-step learning rate scheduler for continuous training.

Novelty

First to systematically study scaling laws across different datasets, revealing the impact of data quality on scaling strategies. Offers a more precise model/data scaling allocation strategy compared to existing methods.

Limitations

  • Scaling laws' applicability across datasets is limited and requires further validation.
  • Model performance in specific scenarios remains suboptimal.

Future Work

Future research will continue to explore the impact of data quality on scaling laws and optimize scaling strategies across different datasets.

AI Executive Summary

The rapid development of open-source large language models has garnered significant attention. However, existing scaling laws offer inconsistent conclusions, challenging model scaling. The DeepSeek LLM project analyzes scaling laws to propose a new scaling strategy focused on long-term growth.

The project developed a dataset of 2 trillion tokens for pre-training models in 7B and 67B configurations. Through supervised fine-tuning and direct preference optimization, DeepSeek Chat models were created. Experimental results show that DeepSeek LLM 67B outperforms LLaMA-2 70B in code, math, and reasoning.

Additionally, open-ended evaluations indicate that DeepSeek LLM 67B Chat excels in both Chinese and English dialogues compared to GPT-3.5. This research provides new directions for the future development of open-source language models, particularly in data quality and model scaling strategies.

Deep Analysis

Background

In recent years, transformer-based language models have become key to achieving artificial general intelligence. The open-source community, building on models like LLaMA, has focused on training high-quality models but has largely neglected scaling laws research. This study aims to fill this gap.

Core Problem

Existing scaling laws offer inconsistent conclusions and lack complete descriptions of hyperparameter settings, leading to unclear performance optimization under different compute budgets.

Innovation

Proposes a new model/data scaling allocation strategy using non-embedding FLOPs/token for more accurate scaling and performance prediction. Highlights the impact of data quality on scaling strategies.

Methodology

  • �� Pre-train using a dataset of 2 trillion tokens
  • �� Employ a multi-step learning rate scheduler for continuous training
  • �� Create DeepSeek Chat models through supervised fine-tuning and direct preference optimization

Experiments

Evaluate model performance using various benchmarks, particularly in code, math, and reasoning. Conduct open-ended evaluations to compare DeepSeek LLM 67B Chat with GPT-3.5.

Results

DeepSeek LLM 67B surpasses LLaMA-2 70B in multiple benchmarks, especially in code and math. In open-ended evaluations, DeepSeek LLM 67B Chat outperforms GPT-3.5.

Applications

Applicable in scenarios requiring high-performance language understanding and generation, such as code generation, mathematical reasoning, and complex dialogue systems.

Limitations & Outlook

Scaling laws' applicability across datasets is limited and requires further validation. Model performance in specific scenarios remains suboptimal.

Plain Language Accessible to non-experts

Imagine a factory, where DeepSeek LLM is like an intelligent production line. It analyzes vast amounts of data to optimize production processes and improve product quality. Data is like raw materials, and the model learns continuously to enhance production efficiency. Ultimately, DeepSeek LLM excels in complex tasks, much like a factory producing high-quality products.

ELI14 Explained like you're 14

Imagine you're playing a super complex game, and DeepSeek LLM is your ultimate helper. It can quickly find answers and solve tough problems, like a super weapon in the game. By analyzing tons of data, it keeps upgrading and getting stronger. In the end, it helps you tackle all sorts of challenges!

Glossary

Scaling Laws

Describes how model performance changes with compute budget, model scale, and data scale.

Used to guide model and data scaling strategies.

Supervised Fine-Tuning

Refines pre-trained models using labeled data to enhance specific task performance.

Used to create DeepSeek Chat models.

Direct Preference Optimization

A method to optimize model conversational performance by directly adjusting preferences.

Enhances DeepSeek Chat's dialogue capabilities.

Non-embedding FLOPs/token

A new method for representing model scale, considering attention operation overhead.

Used for more accurate model scale description.

Multi-step Learning Rate Scheduler

A strategy for adjusting learning rates in phases to improve training efficiency.

Used in DeepSeek LLM pre-training.

Open Questions Unanswered questions from this research

  • 1 How to optimize scaling laws' applicability across different datasets?
  • 2 How does data quality specifically affect model scaling strategies?

Applications

Immediate Applications

Code Generation

DeepSeek LLM can be used to automatically generate high-quality code, improving development efficiency.

Long-term Vision

General Artificial Intelligence

By continuously optimizing scaling laws, it can drive the development of general artificial intelligence.

Abstract

The rapid development of open-source large language models (LLMs) has been truly remarkable. However, the scaling law described in previous literature presents varying conclusions, which casts a dark cloud over scaling LLMs. We delve into the study of scaling laws and present our distinctive findings that facilitate scaling of large scale models in two commonly used open-source configurations, 7B and 67B. Guided by the scaling laws, we introduce DeepSeek LLM, a project dedicated to advancing open-source language models with a long-term perspective. To support the pre-training phase, we have developed a dataset that currently consists of 2 trillion tokens and is continuously expanding. We further conduct supervised fine-tuning (SFT) and Direct Preference Optimization (DPO) on DeepSeek LLM Base models, resulting in the creation of DeepSeek Chat models. Our evaluation results demonstrate that DeepSeek LLM 67B surpasses LLaMA-2 70B on various benchmarks, particularly in the domains of code, mathematics, and reasoning. Furthermore, open-ended evaluations reveal that DeepSeek LLM 67B Chat exhibits superior performance compared to GPT-3.5.

cs.CL cs.AI cs.LG