DeepSeek-Coder-V2: Breaking the Barrier of Closed-Source Models in Code Intelligence
DeepSeek-Coder-V2：打破闭源模型在代码智能领域的壁垒

In summary, our main contributions are:
总结而言，我们的主要贡献如下：

• •

  We introduce DeepSeek-Coder-V2 with 16B and 236B parameters based on the DeepSeekMoE framework, which has activation parameters of only 2.4B and 21B, efficiently supporting diverse computational and application needs. Additionally, DeepSeek-Coder-V2 supports 338 programming languages and a maximum context length of 128K tokens.

  Report issue for preceding element
  
  • 我们推出了基于 DeepSeekMoE 框架的 DeepSeek-Coder-V2，拥有 16B 和 236B 参数，而激活参数仅为 2.4B 和 21B，有效满足多样化的计算与应用需求。此外，DeepSeek-Coder-V2 支持 338 种编程语言，并具备最高 128K 令牌的上下文长度。
• •

  We make the first attempt to develop an open-source hundred-billion-parameter code model to advance the field of code intelligence. Experimental results indicate that DeepSeek-Coder-V2 236B outperforms state-of-the-art closed-source models, such as GPT4-Turbo, Claude 3 Opus, and Gemini 1.5 Pro, in both coding and mathematics tasks.

  Report issue for preceding element
  
  • 我们首次尝试开发一个开源的千亿参数代码模型，以推动代码智能领域的发展。实验结果显示，DeepSeek-Coder-V2 236B 在编码和数学任务上均超越了当前最先进的闭源模型，如 GPT4-Turbo、Claude 3 Opus 以及Gemini 1.5 Pro。
• •

  DeepSeek-Coder-V2 models are released publicly under a permissive license, allowing for both research and unrestricted commercial use.

  Report issue for preceding element
  
  • DeepSeek-Coder-V2 模型在宽松许可下公开发布，允许进行研究和无限制的商业使用。

• •

  Code: Regarding code generation benchmark evaluation, DeepSeek-Coder-V2 demonstrates remarkable superiority over all open source models while exhibiting performance on par with the leading closed-source models, such as GPT4-Turbo, Claude 3 Opus, and Gemini 1.5 Pro. Notably, we achieve a 90.2% score on HumanEval (Chen et al., 2021), a 76.2% score on MBPP (Austin et al., 2021a) (establishing a new state-of-the-art result with EvalPlus evaluation pipeline), and a 43.4% score on LiveCodeBench (Jain et al., 2024) (questions from Dec. 2023 to June. 2024). Additionally, DeepSeek-Coder-V2 is the first open-source model that surpasses a score of 10% on SWEBench (Jimenez et al., 2023).

  Report issue for preceding element
  
  • 代码：在代码生成基准评估方面，DeepSeek-Coder-V2 展现出对所有开源模型的显著优势，同时与 GPT4-Turbo、Claude 3 Opus 以及Gemini 1.5 Pro 等领先的闭源模型性能相当。特别地，我们在 HumanEval（Chen et al., 2021）上获得了 90.2%的分数，在 MBPP（Austin et al., 2021a）上获得了 76.2%的分数（通过 EvalPlus 评估管道创造了新的最先进结果），以及在 LiveCodeBench（Jain et al., 2024）上获得了 43.4%的分数（问题涵盖 2023 年 12 月至 2024 年 6 月）。此外，DeepSeek-Coder-V2 是首个在 SWEBench（Jimenez et al., 2023）上突破 10%分数的开源模型。
• •

  Math: DeepSeek-Coder-V2 exhibits strong mathematical reasoning abilities, rivaling top closed-source models such as GPT-4o, Gemini 1.5 Pro, and Claude 3 Opus on both elementary benchmarks like GSM8K (Cobbe et al., 2021) and advanced competition-level benchmarks including MATH (Hendrycks et al., 2021), AIME (MAA, 2024), and Math Odyssey (Netmind.AI, 2024). Notably, DeepSeek-Coder-V2 attains an accuracy of 75.7% on the MATH benchmark, nearly matching the state-of-the-art accuracy of 76.6% achieved by GPT-4o. Furthermore, it surpasses the performance of these closed-source models in the AIME 2024 competition.

  Report issue for preceding element
  
  • 数学：DeepSeek-Coder-V2 展现出强大的数学推理能力，与顶尖的闭源模型如 GPT-4o、Gemini 1.5 Pro 以及Claude 3 Opus 在基础基准测试如 GSM8K（Cobbe et al., 2021）和高级竞赛级别基准测试包括 MATH（Hendrycks et al., 2021）、AIME（MAA, 2024）及 Math Odyssey（Netmind.AI, 2024）上不相上下。值得注意的是，DeepSeek-Coder-V2 在 MATH 基准测试中达到 75.7%的准确率，几乎与 GPT-4o 创下的 76.6%的最新准确率持平。此外，在 2024 年的 AIME 竞赛中，它超越了这些闭源模型的表现。
• •

  Natural Language: DeepSeek-Coder-V2 maintains comparable general language performance to DeepSeek-V2. For example, DeepSeek-Coder-V2 achieves 79.2% on MMLU with OpenAI simple-eval pipeline. Regarding subjective evaluation with GPT-4 as a judger, DeepSeek-Coder-V2 achieves 65.0 on arena-hard (Li et al., 2024), 8.77 on MT-bench (Zheng et al., 2023) and 7.84 on alignbench (Liu et al., 2023c). These scores are significantly better than other code-specific models, even comparable with general open source models.

  Report issue for preceding element
  
  • 自然语言能力：DeepSeek-Coder-V2 在通用语言表现上与 DeepSeek-V2 相当。例如，通过 OpenAI 的 simple-eval 管道，DeepSeek-Coder-V2 在 MMLU 上达到 79.2%。在以 GPT-4 为评判的主观评估中，DeepSeek-Coder-V2 在 arena-hard（李等，2024）上获得 65.0 分，在 MT-bench（郑等，2023）上得 8.77 分，在 alignbench（刘等，2023c）上得 7.84 分。这些成绩显著优于其他专注于代码的模型，甚至可与通用开源模型相媲美。

We use two training objectives for DeepSeek-Coder-v2 16B: Next-Token-Prediction and Fill-In-Middle (FIM)  (Li et al., 2023b; Bavarian et al., 2022; Guo et al., 2024). For DeepSeek-Coder-v2 236B, we only utilize the Next-Token-Prediction objective. Here we give a brief introduction of the FIM training policy. We adopt the FIM training approach for the development of DeepSeek-Coder-v2-16B, leveraging the PSM (Prefix, Suffix, Middle) mode. This method structures the content reconstruction in the sequence: Prefix, Suffix, and Middle, as illustrated below:
我们为 DeepSeek-Coder-v2 16B 设定了两个训练目标：下一词预测（Next-Token-Prediction）和填空中间（Fill-In-Middle, FIM）（Li 等人，2023b；Bavarian 等人，2022；Guo 等人，2024）。对于 DeepSeek-Coder-v2 236B，我们仅采用下一词预测目标。在此，简要介绍 FIM 训练策略。我们采用 FIM 训练方法来开发 DeepSeek-Coder-v2-16B，利用 PSM（前缀、后缀、中间）模式。该方法按以下顺序构建内容重构：前缀、后缀和中间，具体如下所示：

This structure is applied at the document level as part of the pre-packing process. The FIM is utilized at a rate of 0.5, consistent with the PSM framework, to enhance the training efficacy and model performance.
该结构在文档级别作为预打包过程的一部分应用。FIM 以 0.5 的比率使用，与 PSM 框架保持一致，以提升训练效果和模型性能。

Our architecture aligns with that of DeepSeekV2 (DeepSeek-AI, 2024). The hyperparameters settings, 16B and 236B, correspond to those used in DeepSeek-V2-Lite and DeepSeek-V2, respectively. Notably, we encountered instability during training and spikes in gradient values, which we attributed to the exponential normalization technique. To address this, we reverted to the conventional normalization method.
我们的架构与 DeepSeekV2 (DeepSeek-AI, 2024)保持一致。超参数设置 16B 和 236B 分别对应于 DeepSeek-V2-Lite 和 DeepSeek-V2 所采用的参数。值得注意的是，我们在训练过程中遇到了不稳定性和梯度值的尖峰，我们认为这归因于指数归一化技术。为解决这一问题，我们转而采用了传统的归一化方法。

Consistent with the DeepSeek V2 methodology (DeepSeek-AI, 2024), we utilize the AdamW optimizer (Loshchilov and Hutter, 2019), configured with $\beta_{1}=0.9$, $\beta_{2}=0.95$, and a weight decay of 0.1. Batch sizes and learning rates are adjusted according to DeepSeek-V2 specifications. For learning rate scheduling, we employ a cosine decay strategy, starting with 2000 warm-up steps and gradually reducing the learning rate to 10% of its initial value.
与 DeepSeek V2 方法论一致(DeepSeek-AI, 2024)，我们采用 AdamW 优化器(Loshchilov 和 Hutter, 2019)，配置为 ${\beta }_1=0.9\backslash beta\_\{1\}=0.9italic\_\beta \; start\_POSTSUBSCRIPT\; 1\; end\_POSTSUBSCRIPT\; =\; 0.9$ ， ${\beta }_2=0.95\backslash beta\_\{2\}=0.95italic\_\beta \; start\_POSTSUBSCRIPT\; 2\; end\_POSTSUBSCRIPT\; =\; 0.95$ ，以及 0.1 的权重衰减。批次大小和学习率根据 DeepSeek-V2 规范进行调整。对于学习率调度，我们采用余弦衰减策略，起始为 2000 步的预热阶段，并逐渐将学习率降至初始值的 10%。

Both DeepSeek-Coder-V2 and DeepSeek-Coder-V2-Lite are trained using the same methodology. To maintain robust natural language understanding capabilities in DeepSeek-Coder-V2, we continue the pre-training process from an intermediate checkpoint of DeepSeek-V2. The intermediate checkpoint was initially trained on 4.2T tokens. Consequently, DeepSeek-Coder-V2 has been exposed to a total of 10.2T high-quality tokens during the pre-training phase.
DeepSeek-Coder-V2 与 DeepSeek-Coder-V2-Lite 均采用相同的训练方法。为保持 DeepSeek-Coder-V2 强大的自然语言理解能力，我们从中途检查点继续对 DeepSeek-V2 进行预训练。该中途检查点最初在 4.2T 个词元上进行了训练。因此，DeepSeek-Coder-V2 在预训练阶段总共接触了 10.2T 个高质量词元。

Following DeepSeek-V2, we extend the context length of DeepSeek-Coder-V2 to 128K using Yarn (Peng et al., 2023). The hyper-parameters of YARN are the same as DeepSeek-V2: the scale $s$ to 40, $\alpha$ to 1, $\beta$ to 32. We further continue training the model using two stages to enhance its capability for handling long contexts. In the first stage, we utilize a sequence length of 32K and a batch size of 1152 for 1000 steps. In the second stage, we train the model for an additional 1000 steps, employing a sequence length of 128K and a batch size of 288 sequences. It should be noted here we upsample long context data ratio during long context extension. As shown in Figure 2, the results on the “Needle In A Haystack” (NIAH) tests indicate that DeepSeek-Coder-V2 performs well across all context window lengths up to 128K.
继 DeepSeek-V2 之后，我们利用 Yarn(Peng et al., 2023)将 DeepSeek-Coder-V2 的上下文长度扩展至 128K。YARN 的超参数与 DeepSeek-V2 相同：比例 $ssitalic\_s$ 设为 40， $\alpha \backslash alphaitalic\_\alpha$ 设为 1， $\beta \backslash betaitalic\_\beta$ 设为 32。我们进一步通过两个阶段的持续训练来提升模型处理长上下文的能力。在第一阶段，我们采用 32K 的序列长度和 1152 的批次大小进行 1000 步训练。在第二阶段，我们使用 128K 的序列长度和 288 序列的批次大小，再训练 1000 步。需要注意的是，在长上下文扩展过程中，我们增加了长上下文数据的比例。如图2所示，在“大海捞针”(NIAH)测试中，DeepSeek-Coder-V2 在所有上下文窗口长度达到 128K 的情况下表现良好。

To evaluate the bug-fixing capabilities of the model, we used the Defects4J ⁷⁷7https://github.com/rjust/defects4j, SWE-bench (Jimenez et al., 2023), and Aider ⁸⁸8https://github.com/paul-gauthier/aider datasets for testing. Defects4J is a widely used dataset in the field of software engineering, specifically designed for the purpose of evaluating and testing program repair techniques. It consists of a collection of real-world software bugs from various open-source projects, including but not limited to Apache Commons, JFreeChart, and Closure Compiler. Each bug in the dataset is accompanied by test suites that can be used to validate the effectiveness of program repair tools. Since the original bugs in Defec4J may need modify several files in the repository resulting in a long context, we collect 238 bugs that only need to modify one method from this benchmark.
为了评估模型的错误修复能力，我们使用了 Defects4J⁷⁷7https://github.com/rjust/defects4j、SWE-bench（Jimenez 等人，2023）和 Aider⁸⁸8https://github.com/paul-gauthier/aider数据集进行测试。Defects4J 是软件工程领域广泛使用的数据集，专门设计用于评估和测试程序修复技术。它包含来自多个开源项目的真实软件错误，例如 Apache Commons、JFreeChart 和 Closure Compiler 等。数据集中的每个错误都附带了测试套件，可用于验证程序修复工具的有效性。由于 Defec4J 中的原始错误可能需要修改仓库中的多个文件，导致上下文较长，我们从该基准中收集了仅需修改一个方法的 238 个错误。

SWE-bench is a comprehensive benchmark designed to evaluate the performance of large language models in addressing real-world software issues sourced from GitHub. The benchmark presents a codebase alongside a specific issue, challenging the language model to generate a patch that effectively resolves the described problem. This rigorous evaluation framework ensures that the language model’s ability to understand and fix real-world software issues is thoroughly tested, providing a clear measure of its practical utility and effectiveness in software development tasks.
SWE-bench 是一个全面的基准测试，旨在评估大型语言模型在解决来自 GitHub 的实际软件问题方面的性能。该基准测试展示一个代码库及其特定问题，要求语言模型生成一个能有效解决所述问题的补丁。这一严格的评估框架确保了语言模型理解和修复实际软件问题的能力得到彻底检验，从而清晰地衡量其在软件开发任务中的实用性和有效性。

Aider’s code editing benchmark evaluates the LLM’s ability to modify Python source files, completing 133 distinct coding tasks. This benchmark not only tests the LLM’s coding skills but also checks its consistency in producing code edits according to the specifications in the prompt. For DeepSeek-Coder-V2 models, we use whole format to evaluate.
Aider 的代码编辑基准测试评估了LLM修改 Python 源文件的能力，完成了 133 项不同的编码任务。此基准不仅测试了LLM的编码技能，还检验了其在根据提示中的规范生成代码编辑时的一致性。对于 DeepSeek-Coder-V2 模型，我们采用整体格式进行评估。

Table 7 outlines the performances of different language models on software repair benchmarks, including Defects4J, SWE-Bench, and Aider. Among open-source models, DeepSeek-Coder-Instruct emerges as a standout, achieving the best performance within the open source models. It scores 21% in Defects4J and 12.7% in SWE-Bench, closely approaching the results of leading closed-source models and demonstrating significant capability in handling longer code sequences. Notably, DeepSeek-Coder-V2-Instruct achieves the highest score of 73.7% in Aider, surpassing all other models listed, including closed-source counterparts. This superior performance highlights its efficiency and robustness in automated code repair tasks, positioning DeepSeek-Coder-V2-Instruct as the top open-source model and a formidable competitor to closed-source alternatives in the field.
表 7 概述了不同语言模型在软件修复基准测试中的表现，包括 Defects4J、SWE-Bench 和 Aider。在开源模型中，DeepSeek-Coder-Instruct 脱颖而出，在开源模型中取得了最佳性能。它在 Defects4J 中得分为 21%，在 SWE-Bench 中为 12.7%，紧追领先闭源模型的成绩，展现出处理较长代码序列的显著能力。值得注意的是，DeepSeek-Coder-V2-Instruct 在 Aider 中取得了最高的 73.7% 分数，超越了所有其他列出的模型，包括闭源模型。这一卓越表现凸显了其在自动化代码修复任务中的高效性和鲁棒性，使 DeepSeek-Coder-V2-Instruct 成为顶尖的开源模型，并在该领域成为闭源替代方案的强劲竞争者。

To assess the code reasoning capabilities of our models, we utilize the CRUXEval benchmark. This benchmark comprises 800 Python functions paired with corresponding input-output examples. It is divided into two distinct tasks: CRUXEval-I, which requires the large language model (LLM) to predict the output based on the given input, and CRUXEval-O, where the model must predict the input from the known output. This structure challenges the model’s ability to understand and reason through Python code in both forward and reverse directions.
为了评估我们模型的代码推理能力，我们采用了 CRUXEval 基准测试。该基准包含 800 个 Python 函数及其相应的输入输出示例。它分为两个不同的任务：CRUXEval-I 要求大型语言模型（LLM）根据给定的输入预测输出，而 CRUXEval-O 则要求模型从已知的输出预测输入。这种结构挑战了模型在正向和反向两个方向上理解和推理 Python 代码的能力。

Table 8 presents the performance of various language models on the CruxEval benchmark, which assesses models on two metrics: CruxEval-I-COT and CruxEval-O-COT. Among the open-source models, DeepSeek-Coder-V2-Instruct stands out significantly. It scores 70.0% on CruxEval-I-COT and 75.1% on CruxEval-O-COT, showcasing its superior capability within the open-source domain. However, when compared to larger closed-source models, there is a performance gap. This performance gap may largely be attributed to the fact that DeepSeek-Coder-V2-Instruct operates with only 21 billion activation parameters, which is considerably fewer than those in larger, more advanced closed-source models like GPT-4o. This limitation in model complexity could restrict its learning and problem-solving capacities.
表 8 展示了各语言模型在 CruxEval 基准测试中的表现，该测试从 CruxEval-I-COT 和 CruxEval-O-COT 两个维度评估模型。在开源模型中，DeepSeek-Coder-V2-Instruct 表现尤为突出，其在 CruxEval-I-COT 上得分为 70.0%，在 CruxEval-O-COT 上为 75.1%，显示出其在开源领域的卓越能力。然而，与规模更大的闭源模型相比，存在一定性能差距。这一差距很大程度上归因于 DeepSeek-Coder-V2-Instruct 仅拥有 210 亿激活参数，远少于 GPT-4 等更先进闭源模型中的参数数量。模型复杂度的这种限制可能制约了其学习和问题解决能力。

To assess the mathematical reasoning capabilities of DeepSeekCoder-V2, we utilized the popular grade-school benchmark GSM8K (Cobbe et al., 2021), along with advanced competition-level benchmarks including MATH (Hendrycks et al., 2021), the American Invitational Mathematics Examination (AIME) 2024 (MAA, 2024), and Math Odyssey (Netmind.AI, 2024)⁹⁹9The performance of DeepSeek-Coder-V2 on the four mathematical benchmarks was obtained with zero-shot chain-of-thought prompting; each test question was concatenated with the instruction: ”$\backslash$nPlease reason step by step, and put your final answer within $\backslash$boxed{}.”.
为了评估 DeepSeekCoder-V2 的数学推理能力，我们采用了广受欢迎的小学基准 GSM8K(Cobbe et al., 2021)，以及包括 MATH(Hendrycks et al., 2021)、2024 年美国邀请数学考试（AIME）(MAA, 2024)和数学奥德赛(Netmind.AI, 2024)⁹⁹9在内的先进竞赛级基准。DeepSeekCoder-V2 在这四个数学基准上的表现是通过零样本思维链提示获得的；每个测试题目都附有以下指令：“$\backslash \backslash backslash\backslash$n 请逐步推理，并将最终答案置于$\backslash \backslash backslash\backslash$boxed{}中。”。

The results, presented in Table 9, were obtained using greedy decoding without the aid of tools or voting techniques, unless otherwise specified. DeepSeek-Coder-V2 achieved an accuracy of 75.7% on the MATH benchmark and 53.7% on Math Odyssey, comparable to the state-of-the-art GPT-4o. Additionally, DeepSeek-Coder-V2 solves more problems from AIME 2024 than the other models, demonstrating its strong mathematical reasoning capabilities.
结果如表9所示，采用贪婪解码方法获得，除非另有说明，未借助任何工具或投票技术。DeepSeek-Coder-V2 在 MATH 基准测试中达到 75.7%的准确率，在 Math Odyssey 上为 53.7%，与顶尖的 GPT-4o 相当。此外，DeepSeek-Coder-V2 解决了更多来自 AIME 2024 的问题，展现了其强大的数学推理能力。

As DeepSeek-Coder-V2 is built upon DeepSeek-V2, it inherits the strong natural language capability, even surpassing DeepSeek-V2 on reasoning-related benchmarks. We compare DeepSeek-Coder-V2 Instruct with DeepSeek-V2 Chat on standard benchmarks, which covers both English and Chinese benchmarks, including BigBench Hard (BBH) (Suzgun et al., 2022), MMLU (Hendrycks et al., 2020), ARC (Clark et al., 2018), TriviaQA (Joshi et al., 2017), NaturalQuestions (Kwiatkowski et al., 2019), AGIEval (Zhong et al., 2023)， CLUEWSC (Xu et al., 2020), C-Eval (Huang et al., 2023), and CMMLU (Li et al., 2023a). Besides, we also evaluate the open-ended generation ability of models, including Arena-Hard (Li et al., 2024), AlpacaEval2.0 (Dubois et al., 2024), MT-Bench (Zheng et al., 2023), and Alignbench (Liu et al., 2023c). The evaluation pipeline and metrics are the same as in DeepSeek-V2, where the MMLU are evaluated using OpenAI simple-eval package https://github.com/openai/simple-evals.
由于 DeepSeek-Coder-V2 基于 DeepSeek-V2 构建，它继承了强大的自然语言处理能力，甚至在推理相关的基准测试中超越了 DeepSeek-V2。我们在涵盖英汉双语的标准基准上比较了 DeepSeek-Coder-V2 Instruct 与 DeepSeek-V2 Chat 的表现，这些基准包括 BigBench Hard（BBH）（Suzgun et al., 2022）、MMLU （Hendrycks et al., 2020）、ARC （Clark et al., 2018）、TriviaQA （Joshi et al., 2017）、NaturalQuestions （Kwiatkowski et al., 2019）、AGIEval （Zhong et al., 2023）、CLUEWSC （Xu et al., 2020）、C-Eval （Huang et al., 2023）以及 CMMLU （Li et al., 2023a）。此外，我们还评估了模型的开放式生成能力，包括 Arena-Hard （Li et al., 2024）、AlpacaEval2.0 （Dubois et al., 2024)，MT-Bench (Zheng 等，2023)，以及 Alignbench (Liu 等，2023c)。评估流程和指标与 DeepSeek-V2 相同，其中 MMLU 的评估采用 OpenAI 的 simple-eval 包https://github.com/openai/simple-evals。

When comparing the performance of 16B models, it is evident that DeepSeek-Coder-V2-Lite-Instruct outperforms DeepSeek-V2-Lite-Chat in benchmarks like BBH and Arena-Hard. These benchmarks place a high demand on the model’s reasoning ability, which DeepSeek-Coder-V2-Lite-Instruct excels at. However, DeepSeek-Coder-V2-Lite Instruct falls behind in knowledge-intensive benchmarks like TriviaQA, primarily due to the relatively smaller amount of web data used during pre-training.
在比较 16B 模型的性能时，可以明显看出 DeepSeek-Coder-V2-Lite-Instruct 在 BBH 和 Arena-Hard 等基准测试中优于 DeepSeek-V2-Lite-Chat。这些基准测试对模型的推理能力提出了很高的要求，而 DeepSeek-Coder-V2-Lite-Instruct 在这方面表现出色。然而，在知识密集型的基准测试如 TriviaQA 中，DeepSeek-Coder-V2-Lite Instruct 的表现则稍逊一筹，这主要是因为预训练期间使用的网络数据量相对较小。

Moving on to 236B models, DeepSeek-Coder-V2 Instruct exhibits greater strength in reasoning benchmarks, particularly in Arena-Hard, which comprises a substantial proportion of code, math, and reasoning questions. On the other hand, DeepSeek-V2 Chat demonstrates slightly better results in benchmarks such as MT-bench (Zheng et al., 2023), AlpacaEval 2.0 (Dubois et al., 2024), and AlignBench (Liu et al., 2023c). This advantage can be attributed to the general-purpose alignment stage of DeepSeek-V2 Chat.
转向 236B 型号，DeepSeek-Coder-V2 Instruct 在推理基准测试中展现出更强的实力，尤其是在 Arena-Hard 中，该部分包含了大量的代码、数学和推理题目。另一方面，DeepSeek-V2 Chat 在 MT-bench (Zheng et al., 2023)、AlpacaEval 2.0 (Dubois et al., 2024)和 AlignBench (Liu et al., 2023c)等基准测试中表现略优。这一优势可归因于 DeepSeek-V2 Chat 的通用对齐阶段。