{"@context":"https://schema.org","@type":"NewsArticle","generatedAt":"2026-10-23T07:00:45.542Z","headline":"LMSYS 发布开源聊天模型 Vicuna-13B，用 ShareGPT 对话微调 LLaMA，训练成本约 $300","description":"LMSYS 团队发布 Vicuna-13B，通过约 70K ShareGPT 用户共享对话微调 LLaMA，训练成本约 $300，代码、权重和在线 demo 以非商业许可公开。GPT-4 作为评审的初步评估显示其达到 ChatGPT/Bard 90% 以上质量，在 45% 问题上不逊于 ChatGPT；团队同时提出基于 GPT-4 的自动评测框架，并说明其尚非严谨方法。","url":"https://www.aioga.com/news/eu8pag1qg6pf93gn9k1ql0m0z/","mainEntityOfPage":"https://www.aioga.com/news/eu8pag1qg6pf93gn9k1ql0m0z/","datePublished":"2026-10-03T15:50:46.000Z","dateModified":"2026-10-03T15:50:46.000Z","inLanguage":"zh-CN","publisher":{"@type":"NewsMediaOrganization","name":"Aioga","url":"https://www.aioga.com"},"citation":["https://www.lmsys.org/blog/2023-03-30-vicuna","https://aihot.news/items/eu8pag1qg6pf93gn9k1ql0m0z"],"canonicalUrl":"https://www.aioga.com/news/eu8pag1qg6pf93gn9k1ql0m0z/","directAnswer":{"@type":"Answer","text":"LMSYS 发布开源聊天模型 Vicuna-13B，基于 LLaMA 并使用约 7 万条 ShareGPT 用户共享对话进行微调。团队称其训练成本约 300 美元，代码、权重和在线演示以非商业许可公开。","url":"https://www.aioga.com/news/eu8pag1qg6pf93gn9k1ql0m0z/","dateCreated":"2026-10-03T15:50:46.000Z","author":{"@type":"Organization","@id":"https://www.aioga.com/authors/aioga-editorial/#editorial-team","name":"Aioga Editorial Team","url":"https://www.aioga.com/authors/aioga-editorial/"}},"evidence":[{"@type":"CreativeWork","name":"LMSYS：Blog source article","url":"https://www.lmsys.org/blog/2023-03-30-vicuna","datePublished":"2026-10-03T15:50:46.000Z","provider":{"@type":"Organization","name":"LMSYS：Blog","url":"https://www.lmsys.org/blog/2023-03-30-vicuna"}},{"@type":"CreativeWork","name":"AIHot archive record","url":"https://aihot.news/items/eu8pag1qg6pf93gn9k1ql0m0z","datePublished":"2026-10-03T15:50:46.000Z","provider":{"@type":"Organization","name":"AIHot","url":"https://aihot.news/items/eu8pag1qg6pf93gn9k1ql0m0z"}}],"aggregationSource":"LMSYS：Blog","originalPublisher":{"name":"LMSYS：Blog","url":"https://www.lmsys.org/blog/2023-03-30-vicuna"},"geoDeepAnswer":null,"article":{"id":"eu8pag1qg6pf93gn9k1ql0m0z","slug":"eu8pag1qg6pf93gn9k1ql0m0z","url":"https://www.aioga.com/news/eu8pag1qg6pf93gn9k1ql0m0z/","title":"LMSYS 发布开源聊天模型 Vicuna-13B，用 ShareGPT 对话微调 LLaMA，训练成本约 $300","title_en":"","summary":"LMSYS 团队发布 Vicuna-13B，通过约 70K ShareGPT 用户共享对话微调 LLaMA，训练成本约 $300，代码、权重和在线 demo 以非商业许可公开。GPT-4 作为评审的初步评估显示其达到 ChatGPT/Bard 90% 以上质量，在 45% 问题上不逊于 ChatGPT；团队同时提出基于 GPT-4 的自动评测框架，并说明其尚非严谨方法。","source":"LMSYS：Blog","sourceUrl":"https://www.lmsys.org/blog/2023-03-30-vicuna","aiHotUrl":"https://aihot.news/items/eu8pag1qg6pf93gn9k1ql0m0z","publishedAt":"2026-10-03T15:50:46.000Z","category":"行业动态","score":72,"selected":true,"articleBody":["We introduce Vicuna-13B, an open-source chatbot trained by fine-tuning LLaMA on user-shared conversations collected from ShareGPT. Preliminary evaluation using GPT-4 as a judge shows Vicuna-13B achieves more than 90%* quality of OpenAI ChatGPT and Google Bard while outperforming other models like LLaMA and Stanford Alpaca in more than 90% * of cases. The cost of training Vicuna-13B is around $300. The code：https://github.com/lm-sys/FastChat and weights：https://github.com/lm-sys/FastChat#vicuna-weights, along with an online demo：https://chat.lmsys.org, are publicly available for non-commercial use.","Vicuna (generated by stable diffusion 2.1)","*According to a fun and non-scientific evaluation with GPT-4. Further rigorous evaluation is needed.","After fine-tuning Vicuna with 70K user-shared ChatGPT conversations, we discover that Vicuna becomes capable of generating more detailed and well-structured answers compared to Alpaca (see examples below), with the quality on par with ChatGPT.","However, evaluating chatbots is never a simple task. With recent advancements in GPT-4, we are curious whether its capabilities have reached a human-like level that could enable an automated evaluation framework for benchmark generation and performance assessments. Our initial finding indicates that GPT-4 can produce highly consistent ranks and detailed assessment when comparing chatbots’ answers (see above example of GPT-4 judgment). Preliminary evaluations based on GPT-4, summarized in Figure 1, show that Vicuna achieves 90% * capability of Bard/ChatGPT. While this proposed framework shows a potential to automate chatbot assessment, it is not yet a rigorous approach . Building an evaluation system for chatbots remains an open question requiring further research. More details are provided in the evaluation section.","Figure 1. Relative Response Quality Assessed by GPT-4*","Try the Vicuna-13B demo here：https://chat.lmsys.org!","The rapid advancement of large language models (LLMs) has revolutionized chatbot systems, resulting in unprecedented levels of intelligence as seen in OpenAI's ChatGPT. However, despite its impressive performance, the training and architecture details of ChatGPT remain unclear, hindering research and open-source innovation in this field. Inspired by the Meta LLaMA and Stanford Alpaca project, we introduce Vicuna-13B, an open-source chatbot backed by an enhanced dataset and an easy-to-use, scalable infrastructure. By fine-tuning a LLaMA base model on user-shared conversations collected from ShareGPT.com, Vicuna-13B has demonstrated competitive performance compared to other open-source models like Stanford Alpaca. This blog post provides a preliminary evaluation of Vicuna-13B's performance and describes its training and serving infrastructure. We also invite the community to interact with our online demo to test the capabilities of this chatbot.","Figure 2 provides an overview of our work. To begin, we collected around 70K conversations from ShareGPT.com, a website where users can share their ChatGPT conversations. Next, we enhanced the training scripts provided by Alpaca to better handle multi-turn conversations and long sequences. The training was done with PyTorch FSDP on 8 A100 GPUs in one day. For serving the demo, we implemented a lightweight distributed serving system. We conducted a preliminary evaluation of the model quality by creating a set of 80 diverse questions and utilizing GPT-4 to judge the model outputs. To compare two different models, we combine the outputs from each model into a single prompt for each question. The prompts are then sent to GPT-4, which assesses which model provides better responses. A detailed comparison of LLaMA, Alpaca, ChatGPT, and Vicuna is shown in Table 1 below.","Table 1. Comparison between several notable models","Vicuna is created by fine-tuning a LLaMA base model using approximately 70K user-shared conversations gathered from ShareGPT.com with public APIs. To ensure data quality, we convert the HTML back to markdown and filter out some inappropriate or low-quality samples. Additionally, we divide lengthy conversations into smaller segments that fit the model's maximum context length.","Our training recipe builds on top of Stanford’s alpaca：https://crfm.stanford.edu/2023/03/13/alpaca.html with the following improvements.","We build a serving system that is capable of serving multiple models with distributed workers. It supports flexible plug-in of GPU workers from both on-premise clusters and the cloud. By utilizing a fault-tolerant controller and managed spot feature in SkyPilot, this serving system can work well with cheaper spot instances from multiple clouds to reduce the serving costs. It is currently a lightweight implementation and we are working on integrating more of our latest research：https://arxiv.org/abs/2302.11665 into it.","Evaluating AI chatbots is a challenging task, as it requires examining language understanding, reasoning, and context awareness. With AI chatbots becoming more advanced, current open benchmarks may no longer suffice. For instance, the evaluation dataset used in Stanford’s Alpaca, self-instruct：https://github.com/yizhongw/self-instruct/tree/main/human_eval, can be effectively answered by SOTA chatbots, making it difficult for humans to discern differences in performance. More limitations include training/test data contamination and the potentially high cost of creating new benchmarks. To tackle these issues, we propose an evaluation framework based on GPT-4 to automate chatbot performance assessment.","First, we devised eight question categories, such as Fermi problems, roleplay scenarios, and coding/math tasks, to test various aspects of a chatbot's performance. Through careful prompt engineering, GPT-4 is able to generate diverse, challenging questions that baseline models struggle with. We select ten questions per category and collect answers from five chatbots: LLaMA, Alpaca, ChatGPT, Bard, and Vicuna. We then ask GPT-4 to rate the quality of their answers based on helpfulness, relevance, accuracy, and detail. We discover that GPT-4 can produce not only relatively consistent scores but also detailed explanations on why such scores are given (detailed examples link：https://lmsys.org/vicuna_eval/). However, we also notice that GPT-4 is not very good at judging coding/math tasks.","Figure 3. Response Comparison Assessed by GPT-4","Figure 3 displays the comparison results between all baselines and Vicuna. GPT-4 prefers Vicuna over state-of-the-art open-source models (LLaMA, Alpaca) in more than 90% of the questions, and it achieves competitive performance against proprietary models (ChatGPT, Bard). In 45% of the questions, GPT-4 rates Vicuna's response as better or equal to ChatGPT's. As GPT-4 assigns a quantitative score to each response on a scale of 10, we calculate the total score for each (baseline, Vicuna) comparison pair by adding up the scores obtained by each model on 80 questions. As shown in Table 2, Vicuna’s total score is 92% of ChatGPT’s. Despite recent advancements, these chatbots still face limitations, such as struggling with basic math problems or having limited coding ability.","Table 2. Total Scores Assessed by GPT-4.","While this proposed evaluation framework demonstrates the potential for assessing chatbots, it is not yet a rigorous or mature approach, as large language models are prone to hallucinate. Developing a comprehensive, standardized evaluation system for chatbots remains an open question requiring further research.","Edited : After this blog post, we conducted a deeper study on this GPT4-based evaluation approach. You are welcome to read our new Judging LLM-as-a-judge paper：https://arxiv.org/abs/2306.05685 and try the new evaluation tool：https://github.com/lm-sys/FastChat/tree/main/fastchat/llm_judge.","We have noticed that, similar to other large language models, Vicuna has certain limitations. For instance, it is not good at tasks involving reasoning or mathematics, and it may have limitations in accurately identifying itself or ensuring the factual accuracy of its outputs. Additionally, it has not been sufficiently optimized to guarantee safety or mitigate potential toxicity or bias. To address the safety concerns, we use the OpenAI moderation：https://platform.openai.com/docs/guides/moderation/overview API to filter out inappropriate user inputs in our online demo. Nonetheless, we anticipate that Vicuna can serve as an open starting point for future research to tackle these limitations.","In our first release, we will share the training, serving, and evaluation code on a GitHub repo: https://github.com/lm-sys/FastChat：https://github.com/lm-sys/FastChat. We also released the Vicuna-13B model weights：https://github.com/lm-sys/FastChat#vicuna-weights. There is no plan to release the dataset. Join our Discord：https://discord.gg/HSWAKCrnFx server and follow our Twitter：https://twitter.com/lmsysorg to get the latest updates.","We would like to thank Xinyang Geng, Hao Liu, and Eric Wallace from BAIR; Xuecheng Li, and Tianyi Zhang from Stanford Alpaca team for their insightful discussion and feedback; Qirong Ho from MBZUAI for providing support on the serving cluster. Please check out a blog post from BAIR about a concurrent effort on their chatbot, Koala：https://bair.berkeley.edu/blog/2023/04/03/koala/.","This is a joint effort with collaborators from multiple institutions, including UC Berkeley, CMU, Stanford, UC San Diego, and MBZUAI.","✉ Correspondence to: Lianmin Zheng (lianminzheng@gmail.com：mailto:lianminzheng@gmail.com), Hao Zhang (sjtu.haozhang@gmail.com：mailto:sjtu.haozhang@gmail.com), or LMSYS (lmsys.org@gmail.com：mailto:lmsys.org@gmail.com).","After this blog post, we extended our idea of GPT-4 based evaluation and wrote a more formal paper that systematically studies this \"LLM-as-a-judge\" approach. You are welcome to read and cite this paper: Judging LLM-as-a-judge with MT-Bench and Chatbot Arena：https://arxiv.org/abs/2306.05685."],"articleImages":[{"sourceUrl":"https://www.lmsys.org/images/blog/vicuna/vicuna.jpeg","alt":"","afterParagraph":0,"url":"/media/articles/eu8pag1qg6pf93gn9k1ql0m0z/49f86534df357923.jpg"},{"sourceUrl":"https://www.lmsys.org/images/blog/vicuna/chart.svg","alt":"","afterParagraph":4,"url":"/media/articles/eu8pag1qg6pf93gn9k1ql0m0z/bfc06a03072ab6c0.jpg"},{"sourceUrl":"https://www.lmsys.org/images/blog/vicuna/overview.png","alt":"","afterParagraph":7,"url":"/media/articles/eu8pag1qg6pf93gn9k1ql0m0z/2a305d2776fe0036.png"},{"sourceUrl":"https://www.lmsys.org/images/blog/vicuna/response-compare.png","alt":"","afterParagraph":14,"url":"/media/articles/eu8pag1qg6pf93gn9k1ql0m0z/e16fc69b439de776.png"}],"mediaStatus":"ok","articleBodyZh":["我们介绍 Vicuna-13B，这是一个开源聊天机器人，通过在从 ShareGPT 收集的用户共享对话数据上微调 LLaMA 训练而成。使用 GPT-4 作为评判者的初步评估显示，Vicuna-13B 在超过 90% 的情况下达到 OpenAI ChatGPT 和 Google Bard 的*质量水平，同时在超过 90% 的情况下*优于 LLaMA 和 Stanford Alpaca 等其他模型。训练 Vicuna-13B 的成本约为 300 美元。代码：https://github.com/lm-sys/FastChat 和权重：https://github.com/lm-sys/FastChat#vicuna-weights，以及在线演示：https://chat.lmsys.org，均可公开用于非商业用途。","维库纳（由稳定扩散2.1生成）","*根据使用 GPT-4 进行的一项趣味性且非科学的评估。需要进一步严格的评估。","在使用 7 万条用户共享的 ChatGPT 对话对 Vicuna 进行微调后，我们发现 Vicuna 能够生成比 Alpaca 更加详细和结构良好的回答（见下方示例），其质量与 ChatGPT 相当。","然而，评估聊天机器人从来不是一件简单的事情。随着 GPT-4 的最新进展，我们好奇其能力是否已达到类人水平，从而可能实现基于自动化的基准生成和性能评估框架。我们的初步发现表明，在比较聊天机器人回答时，GPT-4 可以产生高度一致的排序和详细评估（见上方 GPT-4 判定示例）。基于 GPT-4 的初步评估，总结如图 1 所示，显示 Vicuna 达到 Bard/ChatGPT 90%* 的能力水平。虽然该提议的框架显示了自动化聊天机器人评估的潜力，但它尚不是严格的方法。建立针对聊天机器人的评估系统仍然是一个需要进一步研究的开放问题。更多细节请参见评估部分。","图1. GPT-4*评估的相对响应质量","在这里尝试 Vicuna-13B 演示：https://chat.lmsys.org！","大型语言模型（LLM）的快速发展革新了聊天机器人系统，在OpenAI的ChatGPT中展现了前所未有的智能水平。然而，尽管其表现令人印象深刻，ChatGPT的训练和架构细节仍不清楚，这阻碍了该领域的研究和开源创新。受到Meta LLaMA和斯坦福Alpaca项目的启发，我们推出了Vicuna-13B，这是一款基于增强数据集和易于使用、可扩展基础设施的开源聊天机器人。通过在收集自ShareGPT.com的用户共享对话上对LLaMA基础模型进行微调，Vicuna-13B展示了与其他开源模型（如斯坦福Alpaca）相比具有竞争力的性能。本文博客提供了Vicuna-13B性能的初步评估，并描述了其训练与部署基础设施。我们还邀请社区与我们的在线演示进行互动，以测试该聊天机器人的能力。","图2概述了我们的工作。首先，我们从ShareGPT.com收集了约7万条对话，该网站允许用户分享他们的ChatGPT对话。接下来，我们增强了Alpaca提供的训练脚本，以更好地处理多轮对话和长序列。训练使用PyTorch FSDP在8块A100 GPU上一天完成。为了演示的部署，我们实现了一个轻量级的分布式服务系统。我们通过创建一套80个多样化问题，并利用GPT-4来评估模型输出，对模型质量进行了初步评估。为了比较两种不同模型，我们将每个模型的输出合并为每个问题的单一提示，然后将提示发送给GPT-4，由其评估哪种模型提供了更好的回答。LLaMA、Alpaca、ChatGPT和Vicuna的详细比较见下表1。","表1. 几个著名模型的比较","Vicuna是通过在收集自ShareGPT.com的约7万条用户共享对话的基础上，利用公共API对LLaMA基础模型进行微调而创建的。为了保证数据质量，我们将HTML转换回Markdown，并过滤掉一些不适当或低质量的样本。此外，我们将冗长的对话划分为适合模型最大上下文长度的小段。","我们的训练方案基于斯坦福的 Alpaca：https://crfm.stanford.edu/2023/03/13/alpaca.html，并进行了以下改进。","我们构建了一个能够通过分布式工作节点服务多模型的系统。它支持灵活地插入来自本地集群和云端的 GPU 工作节点。通过利用 SkyPilot 的容错控制器和托管抢占实例功能，这个服务系统可以很好地利用来自多个云的廉价抢占实例，从而降低服务成本。目前，它是一个轻量级实现，我们正在努力将更多最新的研究成果：https://arxiv.org/abs/2302.11665 集成到其中。","评估 AI 聊天机器人是一项具有挑战性的任务，因为它需要考察语言理解、推理能力和上下文意识。随着 AI 聊天机器人变得越来越先进，现有的开放基准可能不再足够。例如，斯坦福 Alpaca 中使用的评估数据集 self-instruct：https://github.com/yizhongw/self-instruct/tree/main/human_eval，可以被最先进的聊天机器人有效回答，这使得人类难以判断性能差异。更多的限制还包括训练/测试数据的污染，以及创建新基准的潜在高成本。为了解决这些问题，我们提出了一个基于 GPT-4 的评估框架，以实现聊天机器人性能的自动评估。","首先，我们设计了八类问题，例如费米问题、角色扮演场景以及编码/数学任务，以测试聊天机器人的各方面性能。通过精心的提示工程，GPT-4 能够生成多样且具有挑战性的问题，基线模型难以回答。我们从每个类别中选择十个问题，并收集五个聊天机器人的回答：LLaMA、Alpaca、ChatGPT、Bard 和 Vicuna。然后，我们请 GPT-4 根据有用性、相关性、准确性和细节来评估它们的回答质量。我们发现 GPT-4 不仅可以给出相对一致的分数，还能提供分数背后的详细解释（详细示例链接：https://lmsys.org/vicuna_eval/）。然而，我们也注意到 GPT-4 在判断编码/数学任务方面表现不佳。","图 3. GPT-4 评估的响应比较","图 3 显示了所有基线模型与 Vicuna 的比较结果。GPT-4 在超过 90% 的问题中更倾向于 Vicuna，而不是最先进的开源模型（LLaMA、Alpaca），并且其表现与专有模型（ChatGPT、Bard）相当。在 45% 的问题中，GPT-4 评定 Vicuna 的回答优于或等同于 ChatGPT 的回答。由于 GPT-4 会对每条回答按 10 分制给出量化分数，我们通过将每个模型在 80 个问题中获得的分数相加，计算每对（基线模型，Vicuna）比较的总分。如表 2 所示，Vicuna 的总分为 ChatGPT 的 92%。尽管近期取得了进展，这些聊天机器人仍存在局限，例如在基础数学问题上表现不佳或编码能力有限。","表 2. GPT-4 评估的总分。","虽然该评估框架展示了评估聊天机器人的潜力，但它仍不是严格或成熟的方法，因为大语言模型容易产生幻觉。开发一个全面、标准化的聊天机器人评估系统仍是一个开放问题，需要进一步研究。","编辑：在这篇博客文章之后，我们对基于 GPT-4 的评估方法进行了更深入的研究。欢迎阅读我们的新论文《Judging LLM-as-a-judge》：https://arxiv.org/abs/2306.05685 并尝试新的评估工具：https://github.com/lm-sys/FastChat/tree/main/fastchat/llm_judge。","我们注意到，类似于其他大语言模型，Vicuna 也存在一定的局限性。例如，它在涉及推理或数学的任务上表现不佳，并且在准确识别自身或确保输出事实准确性方面可能存在限制。此外，它尚未得到充分优化以保证安全性或减轻潜在的有害性或偏见。为应对安全问题，我们在在线演示中使用 OpenAI moderation：https://platform.openai.com/docs/guides/moderation/overview API 来过滤不适当的用户输入。尽管如此，我们预计 Vicuna 可以作为未来研究的开放起点，以解决这些局限性。","在我们的首次发布中，我们将在 GitHub 仓库上分享训练、服务和评估代码：https://github.com/lm-sys/FastChat：https://github.com/lm-sys/FastChat。我们还发布了 Vicuna-13B 模型权重：https://github.com/lm-sys/FastChat#vicuna-weights。目前没有发布数据集的计划。加入我们的 Discord 服务器：https://discord.gg/HSWAKCrnFx 并关注我们的 Twitter：https://twitter.com/lmsysorg 以获取最新更新。","我们要感谢 BAIR 的 Xinyang Geng、Hao Liu 和 Eric Wallace；Stanford Alpaca 团队的 Xuecheng Li 和 Tianyi Zhang，感谢他们的深刻讨论和反馈；MBZUAI 的 Qirong Ho，感谢他在服务集群上的支持。请查看 BAIR 对其聊天机器人 Koala 的相关博客文章：https://bair.berkeley.edu/blog/2023/04/03/koala/。","这是与多所机构的合作者共同完成的工作，包括 UC Berkeley、CMU、Stanford、UC San Diego 和 MBZUAI。","✉ 通讯作者：郑连民 (lianminzheng@gmail.com：mailto:lianminzheng@gmail.com)、张浩 (sjtu.haozhang@gmail.com：mailto:sjtu.haozhang@gmail.com) 或 LMSYS (lmsys.org@gmail.com：mailto:lmsys.org@gmail.com)。","在这篇博客文章之后，我们扩展了基于 GPT-4 的评估思路，并撰写了一篇更正式的论文，系统地研究了这种“LLM 作为裁判”的方法。欢迎阅读并引用此论文：Judging LLM-as-a-judge with MT-Bench and Chatbot Arena：https://arxiv.org/abs/2306.05685。"],"translationStatus":"translated","bodyOrigin":"source-page","editorial":{"summary":"LMSYS 发布开源聊天模型 Vicuna-13B，基于 LLaMA 并使用约 7 万条 ShareGPT 用户共享对话进行微调。团队称其训练成本约 300 美元，代码、权重和在线演示以非商业许可公开。","background":"来源材料称，GPT-4 作为评审的初步结果显示，Vicuna-13B 的质量超过 ChatGPT 和 Bard 的 90%，并在超过 90% 的案例中优于 LLaMA、Stanford Alpaca 等模型。该评测被明确标注为非科学性方法，仍需更严格验证。","viewpoint":"Aioga 判断：Vicuna 将低成本训练、用户共享对话数据和开源发布结合起来，展示了开放模型迭代的路径。但现有性能结论主要来自 GPT-4 评审，且团队已说明评测框架尚不严谨，不宜据此作出确定性比较。","implications":"可能影响：开源模型研究者或可获得代码、权重和演示进行复现与测试，但非商业许可会限制部分使用场景。GPT-4 自动评测可能提高比较效率，却不足以替代严谨、可复核的模型评估。","nextStep":"后续观察：需要关注 Vicuna 在更严格评测中的结果、其 GPT-4 评审框架的复核情况，以及代码和权重的实际可用范围。对训练成本和模型质量的判断，应以更多公开证据为依据。","evidenceRefs":["title","summary","articleBody","source"],"status":"published","aiGenerated":true,"autoApproved":true,"generatedBy":"aioga-editorial:gpt-5.6-sol","reviewedBy":"aioga-editorial-review:gpt-5.6-sol","generatedAt":"2026-10-03T19:07:33.620Z","sourceHash":"9c06267809eff13d","review":{"approved":true,"groundedness":92,"clarity":91,"duplicationRisk":8,"blockingIssues":[],"notes":["建议将“质量超过 ChatGPT 和 Bard 的 90%”改为“达到 ChatGPT 和 Bard 90%以上的质量”，以避免被理解为超过 ChatGPT 和 Bard 本身。","“开源发布”建议与“代码、权重和在线演示以非商业许可公开”并列表述，以更清晰体现许可限制。"]},"validation":{"passed":true,"mode":"ai-auto","revisions":0,"checks":["schema","length","source-attribution","editorial-labels","inference-boundary","low-source-overlap","no-html","independent-ai-review"]}},"tags":["行业动态","LMSYS：Blog"],"translations":{"zh-CN":{"title":"LMSYS 发布开源聊天模型 Vicuna-13B，用 ShareGPT 对话微调 LLaMA，训练成本约 $300","summary":"LMSYS 团队发布 Vicuna-13B，通过约 70K ShareGPT 用户共享对话微调 LLaMA，训练成本约 $300，代码、权重和在线 demo 以非商业许可公开。GPT-4 作为评审的初步评估显示其达到 ChatGPT/Bard 90% 以上质量，在 45% 问题上不逊于 ChatGPT；团队同时提出基于 GPT-4 的自动评测框架，并说明其尚非严谨方法。","category":"行业动态","source":"LMSYS：Blog","aggregationSource":"LMSYS：Blog","pageTitle":"LMSYS 发布开源聊天模型 Vicuna-13B，用 ShareGPT 对话微调 LLaMA，训练成本约 $300 - Aioga AI资讯","description":"LMSYS 团队发布 Vicuna-13B，通过约 70K ShareGPT 用户共享对话微调 LLaMA，训练成本约 $300，代码、权重和在线 demo 以非商业许可公开。GPT-4 作为评审的初步评估显示其达到 ChatGPT/Bard 90% 以上质量，在 45% 问题上不逊于 ChatGPT；团队同时提出基于 GPT-4 的自动评测框架，并说明其尚...","url":"https://www.aioga.com/news/eu8pag1qg6pf93gn9k1ql0m0z/","articleBody":["我们介绍 Vicuna-13B，这是一个开源聊天机器人，通过在从 ShareGPT 收集的用户共享对话数据上微调 LLaMA 训练而成。使用 GPT-4 作为评判者的初步评估显示，Vicuna-13B 在超过 90% 的情况下达到 OpenAI ChatGPT 和 Google Bard 的*质量水平，同时在超过 90% 的情况下*优于 LLaMA 和 Stanford Alpaca 等其他模型。训练 Vicuna-13B 的成本约为 300 美元。代码：https://github.com/lm-sys/FastChat 和权重：https://github.com/lm-sys/FastChat#vicuna-weights，以及在线演示：https://chat.lmsys.org，均可公开用于非商业用途。","维库纳（由稳定扩散2.1生成）","*根据使用 GPT-4 进行的一项趣味性且非科学的评估。需要进一步严格的评估。","在使用 7 万条用户共享的 ChatGPT 对话对 Vicuna 进行微调后，我们发现 Vicuna 能够生成比 Alpaca 更加详细和结构良好的回答（见下方示例），其质量与 ChatGPT 相当。","然而，评估聊天机器人从来不是一件简单的事情。随着 GPT-4 的最新进展，我们好奇其能力是否已达到类人水平，从而可能实现基于自动化的基准生成和性能评估框架。我们的初步发现表明，在比较聊天机器人回答时，GPT-4 可以产生高度一致的排序和详细评估（见上方 GPT-4 判定示例）。基于 GPT-4 的初步评估，总结如图 1 所示，显示 Vicuna 达到 Bard/ChatGPT 90%* 的能力水平。虽然该提议的框架显示了自动化聊天机器人评估的潜力，但它尚不是严格的方法。建立针对聊天机器人的评估系统仍然是一个需要进一步研究的开放问题。更多细节请参见评估部分。","图1. GPT-4*评估的相对响应质量","在这里尝试 Vicuna-13B 演示：https://chat.lmsys.org！","大型语言模型（LLM）的快速发展革新了聊天机器人系统，在OpenAI的ChatGPT中展现了前所未有的智能水平。然而，尽管其表现令人印象深刻，ChatGPT的训练和架构细节仍不清楚，这阻碍了该领域的研究和开源创新。受到Meta LLaMA和斯坦福Alpaca项目的启发，我们推出了Vicuna-13B，这是一款基于增强数据集和易于使用、可扩展基础设施的开源聊天机器人。通过在收集自ShareGPT.com的用户共享对话上对LLaMA基础模型进行微调，Vicuna-13B展示了与其他开源模型（如斯坦福Alpaca）相比具有竞争力的性能。本文博客提供了Vicuna-13B性能的初步评估，并描述了其训练与部署基础设施。我们还邀请社区与我们的在线演示进行互动，以测试该聊天机器人的能力。","图2概述了我们的工作。首先，我们从ShareGPT.com收集了约7万条对话，该网站允许用户分享他们的ChatGPT对话。接下来，我们增强了Alpaca提供的训练脚本，以更好地处理多轮对话和长序列。训练使用PyTorch FSDP在8块A100 GPU上一天完成。为了演示的部署，我们实现了一个轻量级的分布式服务系统。我们通过创建一套80个多样化问题，并利用GPT-4来评估模型输出，对模型质量进行了初步评估。为了比较两种不同模型，我们将每个模型的输出合并为每个问题的单一提示，然后将提示发送给GPT-4，由其评估哪种模型提供了更好的回答。LLaMA、Alpaca、ChatGPT和Vicuna的详细比较见下表1。","表1. 几个著名模型的比较","Vicuna是通过在收集自ShareGPT.com的约7万条用户共享对话的基础上，利用公共API对LLaMA基础模型进行微调而创建的。为了保证数据质量，我们将HTML转换回Markdown，并过滤掉一些不适当或低质量的样本。此外，我们将冗长的对话划分为适合模型最大上下文长度的小段。","我们的训练方案基于斯坦福的 Alpaca：https://crfm.stanford.edu/2023/03/13/alpaca.html，并进行了以下改进。","我们构建了一个能够通过分布式工作节点服务多模型的系统。它支持灵活地插入来自本地集群和云端的 GPU 工作节点。通过利用 SkyPilot 的容错控制器和托管抢占实例功能，这个服务系统可以很好地利用来自多个云的廉价抢占实例，从而降低服务成本。目前，它是一个轻量级实现，我们正在努力将更多最新的研究成果：https://arxiv.org/abs/2302.11665 集成到其中。","评估 AI 聊天机器人是一项具有挑战性的任务，因为它需要考察语言理解、推理能力和上下文意识。随着 AI 聊天机器人变得越来越先进，现有的开放基准可能不再足够。例如，斯坦福 Alpaca 中使用的评估数据集 self-instruct：https://github.com/yizhongw/self-instruct/tree/main/human_eval，可以被最先进的聊天机器人有效回答，这使得人类难以判断性能差异。更多的限制还包括训练/测试数据的污染，以及创建新基准的潜在高成本。为了解决这些问题，我们提出了一个基于 GPT-4 的评估框架，以实现聊天机器人性能的自动评估。","首先，我们设计了八类问题，例如费米问题、角色扮演场景以及编码/数学任务，以测试聊天机器人的各方面性能。通过精心的提示工程，GPT-4 能够生成多样且具有挑战性的问题，基线模型难以回答。我们从每个类别中选择十个问题，并收集五个聊天机器人的回答：LLaMA、Alpaca、ChatGPT、Bard 和 Vicuna。然后，我们请 GPT-4 根据有用性、相关性、准确性和细节来评估它们的回答质量。我们发现 GPT-4 不仅可以给出相对一致的分数，还能提供分数背后的详细解释（详细示例链接：https://lmsys.org/vicuna_eval/）。然而，我们也注意到 GPT-4 在判断编码/数学任务方面表现不佳。","图 3. GPT-4 评估的响应比较","图 3 显示了所有基线模型与 Vicuna 的比较结果。GPT-4 在超过 90% 的问题中更倾向于 Vicuna，而不是最先进的开源模型（LLaMA、Alpaca），并且其表现与专有模型（ChatGPT、Bard）相当。在 45% 的问题中，GPT-4 评定 Vicuna 的回答优于或等同于 ChatGPT 的回答。由于 GPT-4 会对每条回答按 10 分制给出量化分数，我们通过将每个模型在 80 个问题中获得的分数相加，计算每对（基线模型，Vicuna）比较的总分。如表 2 所示，Vicuna 的总分为 ChatGPT 的 92%。尽管近期取得了进展，这些聊天机器人仍存在局限，例如在基础数学问题上表现不佳或编码能力有限。","表 2. GPT-4 评估的总分。","虽然该评估框架展示了评估聊天机器人的潜力，但它仍不是严格或成熟的方法，因为大语言模型容易产生幻觉。开发一个全面、标准化的聊天机器人评估系统仍是一个开放问题，需要进一步研究。","编辑：在这篇博客文章之后，我们对基于 GPT-4 的评估方法进行了更深入的研究。欢迎阅读我们的新论文《Judging LLM-as-a-judge》：https://arxiv.org/abs/2306.05685 并尝试新的评估工具：https://github.com/lm-sys/FastChat/tree/main/fastchat/llm_judge。","我们注意到，类似于其他大语言模型，Vicuna 也存在一定的局限性。例如，它在涉及推理或数学的任务上表现不佳，并且在准确识别自身或确保输出事实准确性方面可能存在限制。此外，它尚未得到充分优化以保证安全性或减轻潜在的有害性或偏见。为应对安全问题，我们在在线演示中使用 OpenAI moderation：https://platform.openai.com/docs/guides/moderation/overview API 来过滤不适当的用户输入。尽管如此，我们预计 Vicuna 可以作为未来研究的开放起点，以解决这些局限性。","在我们的首次发布中，我们将在 GitHub 仓库上分享训练、服务和评估代码：https://github.com/lm-sys/FastChat：https://github.com/lm-sys/FastChat。我们还发布了 Vicuna-13B 模型权重：https://github.com/lm-sys/FastChat#vicuna-weights。目前没有发布数据集的计划。加入我们的 Discord 服务器：https://discord.gg/HSWAKCrnFx 并关注我们的 Twitter：https://twitter.com/lmsysorg 以获取最新更新。","我们要感谢 BAIR 的 Xinyang Geng、Hao Liu 和 Eric Wallace；Stanford Alpaca 团队的 Xuecheng Li 和 Tianyi Zhang，感谢他们的深刻讨论和反馈；MBZUAI 的 Qirong Ho，感谢他在服务集群上的支持。请查看 BAIR 对其聊天机器人 Koala 的相关博客文章：https://bair.berkeley.edu/blog/2023/04/03/koala/。","这是与多所机构的合作者共同完成的工作，包括 UC Berkeley、CMU、Stanford、UC San Diego 和 MBZUAI。","✉ 通讯作者：郑连民 (lianminzheng@gmail.com：mailto:lianminzheng@gmail.com)、张浩 (sjtu.haozhang@gmail.com：mailto:sjtu.haozhang@gmail.com) 或 LMSYS (lmsys.org@gmail.com：mailto:lmsys.org@gmail.com)。","在这篇博客文章之后，我们扩展了基于 GPT-4 的评估思路，并撰写了一篇更正式的论文，系统地研究了这种“LLM 作为裁判”的方法。欢迎阅读并引用此论文：Judging LLM-as-a-judge with MT-Bench and Chatbot Arena：https://arxiv.org/abs/2306.05685。"]},"en":{"title":"LMSYS releases the open-source chat model Vicuna-13B, fine-tuned from LLaMA using ShareGPT conversations, with a training cost of about $300","summary":"The LMSYS team released Vicuna-13B, fine-tuned on LLaMA with approximately 70K ShareGPT user-shared conversations, with a training cost of about $300. The code, weights, and online demo are publicly available under a non-commercial license. Preliminary evaluations using GPT-4 as a reviewer show that it reaches over 90% of ChatGPT/Bard quality, and performs no worse than ChatGPT on 45% of questions; the team also proposed a GPT-4-based automatic evaluation framework, noting that it is not yet a rigorous method.","category":"Industry","source":"LMSYS：Blog","aggregationSource":"LMSYS：Blog","pageTitle":"LMSYS releases the open-source chat model Vicuna-13B, fine-tuned from LLaMA using ShareGPT conversations, with a training cost of about $300 - Aioga AI News","description":"The LMSYS team released Vicuna-13B, fine-tuned on LLaMA with approximately 70K ShareGPT user-shared conversations, with a training cost of about $300. The code, weights, and online...","url":"https://www.aioga.com/en/news/eu8pag1qg6pf93gn9k1ql0m0z/","contentTranslated":true,"translationStatus":"translated","translationRetryAt":"","translationError":"","sourceHash":"6f8ada70b07698c4","translatedAt":"2026-10-03T18:08:09.366Z"},"ja":{"title":"LMSYS、オープンソースのチャットモデルVicuna-13Bをリリース。ShareGPTの対話データでLLaMAをファインチューニングし、訓練コストは約$300","summary":"LMSYSチームはVicuna-13Bを発表しました。これは約7万件のShareGPTユーザー共有の対話を用いてLLaMAを微調整したもので、トレーニングコストは約300ドルです。コード、重み、オンラインデモは非商用ライセンスで公開されています。GPT-4を評価者として行った初期評価では、ChatGPTやBardの品質の90%以上に達し、45%の問題ではChatGPTに劣らないことが示されました。チームは同時にGPT-4ベースの自動評価フレームワークも提案しており、これはまだ厳密な方法ではないことを説明しています。","category":"業界動向","source":"LMSYS：Blog","aggregationSource":"LMSYS：Blog","pageTitle":"LMSYS、オープンソースのチャットモデルVicuna-13Bをリリース。ShareGPTの対話データでLLaMAをファインチューニングし、訓練コストは約$300 - Aioga AIニュース","description":"LMSYSチームはVicuna-13Bを発表しました。これは約7万件のShareGPTユーザー共有の対話を用いてLLaMAを微調整したもので、トレーニングコストは約300ドルです。コード、重み、オンラインデモは非商用ライセンスで公開されています。GPT-4を評価者として行った初期評価では、ChatGPTやBardの品質の90%以上に達し、45%の問題ではCh...","url":"https://www.aioga.com/ja/news/eu8pag1qg6pf93gn9k1ql0m0z/","contentTranslated":true,"translationStatus":"translated","translationRetryAt":"","translationError":"","sourceHash":"6f8ada70b07698c4","translatedAt":"2026-10-03T18:09:18.715Z"},"ko":{"title":"LMSYS가 오픈소스 챗 모델 Vicuna-13B를 출시했으며, ShareGPT로 LLaMA를 대화 미세 조정했으며, 훈련 비용은 약 $300입니다.","summary":"LMSYS 팀이 Vicuna-13B를 발표했으며, 약 7만 명의 ShareGPT 사용자 공유 대화를 통해 LLaMA를 미세 조정했으며, 훈련 비용은 약 300달러입니다. 코드, 가중치 및 온라인 데모는 비상업적 라이선스로 공개됩니다. GPT-4를 평가자로 한 초기 평가에서는 ChatGPT/Bard의 90% 이상의 품질에 도달했으며, 45%의 문제에서는 ChatGPT에 뒤지지 않는 것으로 나타났습니다. 팀은 동시에 GPT-4 기반 자동 평가 프레임워크를 제안하고 이것이 아직 엄격한 방법이 아님을 설명했습니다.","category":"업계 동향","source":"LMSYS：Blog","aggregationSource":"LMSYS：Blog","pageTitle":"LMSYS가 오픈소스 챗 모델 Vicuna-13B를 출시했으며, ShareGPT로 LLaMA를 대화 미세 조정했으며, 훈련 비용은 약 $300입니다. - Aioga AI 뉴스","description":"LMSYS 팀이 Vicuna-13B를 발표했으며, 약 7만 명의 ShareGPT 사용자 공유 대화를 통해 LLaMA를 미세 조정했으며, 훈련 비용은 약 300달러입니다. 코드, 가중치 및 온라인 데모는 비상업적 라이선스로 공개됩니다. GPT-4를 평가자로 한 초기 평가에서는 ChatGPT/Bard의 90% 이상의 품질에...","url":"https://www.aioga.com/ko/news/eu8pag1qg6pf93gn9k1ql0m0z/","contentTranslated":true,"translationStatus":"translated","translationRetryAt":"","translationError":"","sourceHash":"6f8ada70b07698c4","translatedAt":"2026-10-03T18:08:57.064Z"},"es":{"title":"LMSYS lanza el modelo de chat de código abierto Vicuna-13B, ajusta LLaMA con conversaciones de ShareGPT; el costo de entrenamiento es de aproximadamente $300.","summary":"El equipo de LMSYS lanzó Vicuna-13B, ajustando LLaMA mediante conversaciones compartidas por aproximadamente 70,000 usuarios de ShareGPT, con un costo de entrenamiento de alrededor de $300. El código, los pesos y la demostración en línea se han publicado bajo una licencia no comercial. La evaluación preliminar realizada por GPT-4 como revisor muestra que alcanza más del 90% de la calidad de ChatGPT/Bard, y en el 45% de las preguntas no es inferior a ChatGPT; al mismo tiempo, el equipo propuso un marco de evaluación automática basado en GPT-4, indicando que aún no es un método riguroso.","category":"Industria","source":"LMSYS：Blog","aggregationSource":"LMSYS：Blog","pageTitle":"LMSYS lanza el modelo de chat de código abierto Vicuna-13B, ajusta LLaMA con conversaciones de ShareGPT; el costo de entrenamiento es de aproximadamente $300. - Aioga Noticias de IA","description":"El equipo de LMSYS lanzó Vicuna-13B, ajustando LLaMA mediante conversaciones compartidas por aproximadamente 70,000 usuarios de ShareGPT, con un costo de entrenamiento de alrededor...","url":"https://www.aioga.com/es/news/eu8pag1qg6pf93gn9k1ql0m0z/","contentTranslated":true,"translationStatus":"translated","translationRetryAt":"","translationError":"","sourceHash":"6f8ada70b07698c4","translatedAt":"2026-10-03T18:14:20.786Z"},"fr":{"title":"LMSYS publie le modèle de chat open source Vicuna-13B, affiné par dialogue avec ShareGPT sur LLaMA, le coût d'entraînement est d'environ 300 $","summary":"L'équipe LMSYS a publié Vicuna-13B, ajusté finement sur LLaMA à partir d'environ 70 000 conversations partagées par les utilisateurs de ShareGPT, avec un coût d'entraînement d'environ 300 $, le code, les poids et la démonstration en ligne étant rendus publics sous licence non commerciale. Une évaluation préliminaire utilisant GPT-4 comme examinateur montre que sa qualité atteint plus de 90 % de celle de ChatGPT/Bard, et qu'il n'est pas inférieur à ChatGPT pour 45 % des questions ; l'équipe a également proposé un cadre d'évaluation automatique basé sur GPT-4, en précisant qu'il ne s'agit pas encore d'une méthode rigoureuse.","category":"Industrie","source":"LMSYS：Blog","aggregationSource":"LMSYS：Blog","pageTitle":"LMSYS publie le modèle de chat open source Vicuna-13B, affiné par dialogue avec ShareGPT sur LLaMA, le coût d'entraînement est d'environ 300 $ - Aioga Actualités IA","description":"L'équipe LMSYS a publié Vicuna-13B, ajusté finement sur LLaMA à partir d'environ 70 000 conversations partagées par les utilisateurs de ShareGPT, avec un coût d'entraînement d'envi...","url":"https://www.aioga.com/fr/news/eu8pag1qg6pf93gn9k1ql0m0z/","contentTranslated":true,"translationStatus":"translated","translationRetryAt":"","translationError":"","sourceHash":"6f8ada70b07698c4","translatedAt":"2026-10-03T18:14:31.728Z"},"de":{"title":"LMSYS veröffentlicht das Open-Source-Chatmodell Vicuna-13B und stimmt LLaMA mit ShareGPT-Dialogen fein ab. Die Trainingskosten belaufen sich auf etwa $300.","summary":"Das LMSYS-Team hat Vicuna-13B veröffentlicht, das durch Feineinstellung von LLaMA mit etwa 70.000 von ShareGPT geteilten Benutzerdialogen trainiert wurde, die Trainingskosten betrugen etwa 300 USD. Code, Gewichte und Online-Demo sind unter nicht-kommerzieller Lizenz öffentlich zugänglich. Eine vorläufige Bewertung durch GPT-4 zeigt, dass es mehr als 90 % der ChatGPT/Bard-Qualität erreicht und bei 45 % der Fragen nicht schlechter als ChatGPT abschneidet; das Team stellte gleichzeitig einen auf GPT-4 basierenden automatischen Bewertungsrahmen vor und wies darauf hin, dass es sich noch nicht um eine rigorose Methode handelt.","category":"行业动态","source":"LMSYS：Blog","aggregationSource":"LMSYS：Blog","pageTitle":"LMSYS veröffentlicht das Open-Source-Chatmodell Vicuna-13B und stimmt LLaMA mit ShareGPT-Dialogen fein ab. Die Trainingskosten belaufen sich auf etwa $300. - Aioga KI-News","description":"Das LMSYS-Team hat Vicuna-13B veröffentlicht, das durch Feineinstellung von LLaMA mit etwa 70.000 von ShareGPT geteilten Benutzerdialogen trainiert wurde, die Trainingskosten betru...","url":"https://www.aioga.com/de/news/eu8pag1qg6pf93gn9k1ql0m0z/","contentTranslated":true,"translationStatus":"translated","translationRetryAt":"","translationError":"","sourceHash":"6f8ada70b07698c4","translatedAt":"2026-10-03T18:14:40.281Z"},"pt-BR":{"title":"LMSYS lança modelo de chat open source Vicuna-13B, afinado com conversas do ShareGPT no LLaMA, custo de treinamento cerca de $300","summary":"A equipe LMSYS lançou o Vicuna-13B, ajustando finamente o LLaMA com diálogos compartilhados por cerca de 70 mil usuários do ShareGPT, com um custo de treinamento de aproximadamente $300. O código, os pesos e a demonstração online foram disponibilizados com licença não comercial. Uma avaliação preliminar feita pelo GPT-4 como avaliador mostrou que ele atingiu mais de 90% da qualidade do ChatGPT/Bard, não ficando atrás do ChatGPT em 45% das questões; a equipe também propôs uma estrutura de avaliação automática baseada no GPT-4, ressaltando que ainda não é um método rigoroso.","category":"行业动态","source":"LMSYS：Blog","aggregationSource":"LMSYS：Blog","pageTitle":"LMSYS lança modelo de chat open source Vicuna-13B, afinado com conversas do ShareGPT no LLaMA, custo de treinamento cerca de $300 - Aioga Notícias de IA","description":"A equipe LMSYS lançou o Vicuna-13B, ajustando finamente o LLaMA com diálogos compartilhados por cerca de 70 mil usuários do ShareGPT, com um custo de treinamento de aproximadamente...","url":"https://www.aioga.com/pt-BR/news/eu8pag1qg6pf93gn9k1ql0m0z/","contentTranslated":true,"translationStatus":"translated","translationRetryAt":"","translationError":"","sourceHash":"6f8ada70b07698c4","translatedAt":"2026-10-03T18:20:22.881Z"},"ru":{"title":"LMSYS выпустила открытые чат-модели Vicuna-13B, использующие ShareGPT для тонкой настройки LLaMA, стоимость обучения около $300","summary":"Команда LMSYS выпустила Vicuna-13B, дообученный на LLaMA с помощью примерно 70 тысяч диалогов, предоставленных пользователями ShareGPT, при этом стоимость обучения составила около $300; код, веса и онлайн-демо опубликованы с некоммерческой лицензией. Первоначальная оценка GPT-4 в качестве рецензента показывает, что его качество превысило 90% ChatGPT/Bard, а по 45% вопросов он не уступает ChatGPT; команда также предложила автоматическую оценочную систему на основе GPT-4 и указала, что она пока не является строгим методом.","category":"行业动态","source":"LMSYS：Blog","aggregationSource":"LMSYS：Blog","pageTitle":"LMSYS выпустила открытые чат-модели Vicuna-13B, использующие ShareGPT для тонкой настройки LLaMA, стоимость обучения около $300 - Aioga Новости ИИ","description":"Команда LMSYS выпустила Vicuna-13B, дообученный на LLaMA с помощью примерно 70 тысяч диалогов, предоставленных пользователями ShareGPT, при этом стоимость обучения составила около...","url":"https://www.aioga.com/ru/news/eu8pag1qg6pf93gn9k1ql0m0z/","contentTranslated":true,"translationStatus":"translated","translationRetryAt":"","translationError":"","sourceHash":"6f8ada70b07698c4","translatedAt":"2026-10-03T18:20:22.368Z"},"ar":{"title":"أصدرت LMSYS نموذج دردشة مفتوح المصدر Vicuna-13B، حيث تم تعديل LLaMA للحوار باستخدام ShareGPT، وتكلفة التدريب حوالي 300 دولار","summary":"أصدر فريق LMSYS نموذج Vicuna-13B، الذي جرى ضبطه الدقيق باستخدام LLaMA من خلال محادثات شاركها نحو 70K من مستخدمي ShareGPT، وبلغت تكلفة تدريبه نحو $300. وقد أُتيح الكود والأوزان والعرض التجريبي عبر الإنترنت بموجب ترخيص غير تجاري. وأظهر تقييم أولي باستخدام GPT-4 كمحكّم أن جودته تتجاوز 90% من جودة ChatGPT/Bard، وأنه لا يقل عن ChatGPT في 45% من الأسئلة؛ كما اقترح الفريق إطارًا للتقييم الآلي قائمًا على GPT-4، موضحًا في الوقت نفسه أنه ليس بعدُ منهجًا صارمًا.","category":"行业动态","source":"LMSYS：Blog","aggregationSource":"LMSYS：Blog","pageTitle":"أصدرت LMSYS نموذج دردشة مفتوح المصدر Vicuna-13B، حيث تم تعديل LLaMA للحوار باستخدام ShareGPT، وتكلفة التدريب حوالي 300 دولار - Aioga أخبار الذكاء الاصطناعي","description":"أصدر فريق LMSYS نموذج Vicuna-13B، الذي جرى ضبطه الدقيق باستخدام LLaMA من خلال محادثات شاركها نحو 70K من مستخدمي ShareGPT، وبلغت تكلفة تدريبه نحو $300. وقد أُتيح الكود والأوزان والع...","url":"https://www.aioga.com/ar/news/eu8pag1qg6pf93gn9k1ql0m0z/","contentTranslated":true,"translationStatus":"translated","translationRetryAt":"","translationError":"","sourceHash":"6f8ada70b07698c4","translatedAt":"2026-10-03T18:20:09.394Z"},"hi":{"title":"LMSYS ने ओपन-सोर्स चैट मॉडल Vicuna-13B जारी किया, ShareGPT संवादों का उपयोग करके LLaMA को फाइन-ट्यून किया; प्रशिक्षण लागत लगभग $300 रही।","summary":"LMSYS टीम ने Vicuna-13B जारी किया, जो लगभग 70K ShareGPT उपयोगकर्ताओं द्वारा साझा की गई बातचीत के माध्यम से LLaMA का सूक्ष्म अनुकूलन करके तैयार किया गया, प्रशिक्षण लागत लगभग $300 थी, कोड, वज़न और ऑनलाइन डेमो गैर-व्यावसायिक लाइसेंस के तहत सार्वजनिक किए गए। GPT-4 द्वारा की गई प्रारंभिक समीक्षा मूल्यांकन में दिखाया गया कि इसकी गुणवत्ता ChatGPT/Bard के 90% से अधिक तक पहुँची है, और 45% सवालों में यह ChatGPT से कमतर नहीं है; टीम ने साथ ही GPT-4 आधारित स्वचालित मूल्यांकन ढांचा प्रस्तुत किया, और बताया कि यह अभी पूरी तरह से सटीक पद्धति नहीं है।","category":"行业动态","source":"LMSYS：Blog","aggregationSource":"LMSYS：Blog","pageTitle":"LMSYS ने ओपन-सोर्स चैट मॉडल Vicuna-13B जारी किया, ShareGPT संवादों का उपयोग करके LLaMA को फाइन-ट्यून किया; प्रशिक्षण लागत लगभग $300 रही। - Aioga AI समाचार","description":"LMSYS टीम ने Vicuna-13B जारी किया, जो लगभग 70K ShareGPT उपयोगकर्ताओं द्वारा साझा की गई बातचीत के माध्यम से LLaMA का सूक्ष्म अनुकूलन करके तैयार किया गया, प्रशिक्षण लागत लगभग $300 थी...","url":"https://www.aioga.com/hi/news/eu8pag1qg6pf93gn9k1ql0m0z/","contentTranslated":true,"translationStatus":"translated","translationRetryAt":"","translationError":"","sourceHash":"6f8ada70b07698c4","translatedAt":"2026-10-03T18:25:39.930Z"},"it":{"title":"LMSYS pubblica il modello di chat open source Vicuna-13B, mette a punto LLaMA con conversazioni di ShareGPT, con un costo di addestramento di circa $300","summary":"Il team LMSYS ha pubblicato Vicuna-13B, ottenuto mediante il fine-tuning di LLaMA su circa 70K conversazioni condivise dagli utenti di ShareGPT, con un costo di addestramento di circa $300; il codice, i pesi e la demo online sono stati resi pubblici con una licenza non commerciale. Una valutazione preliminare, con GPT-4 come giudice, mostra che raggiunge oltre il 90% della qualità di ChatGPT/Bard e che non è inferiore a ChatGPT nel 45% dei problemi; il team ha inoltre proposto un framework di valutazione automatica basato su GPT-4, precisando che non si tratta ancora di un metodo rigoroso.","category":"行业动态","source":"LMSYS：Blog","aggregationSource":"LMSYS：Blog","pageTitle":"LMSYS pubblica il modello di chat open source Vicuna-13B, mette a punto LLaMA con conversazioni di ShareGPT, con un costo di addestramento di circa $300 - Aioga Notizie IA","description":"Il team LMSYS ha pubblicato Vicuna-13B, ottenuto mediante il fine-tuning di LLaMA su circa 70K conversazioni condivise dagli utenti di ShareGPT, con un costo di addestramento di ci...","url":"https://www.aioga.com/it/news/eu8pag1qg6pf93gn9k1ql0m0z/","contentTranslated":true,"translationStatus":"translated","translationRetryAt":"","translationError":"","sourceHash":"6f8ada70b07698c4","translatedAt":"2026-10-03T18:25:21.247Z"},"nl":{"title":"LMSYS brengt het open-sourcechatmodel Vicuna-13B uit, waarbij LLaMA wordt gefinetuned met ShareGPT-dialogen; de trainingskosten bedragen ongeveer $300.","summary":"Het LMSYS-team heeft Vicuna-13B uitgebracht, een op LLaMA gefinetuned model op basis van ongeveer 70K door ShareGPT-gebruikers gedeelde gesprekken, met trainingskosten van ongeveer $300. De code, gewichten en online demo zijn onder een niet-commerciële licentie openbaar gemaakt. Een voorlopige evaluatie waarbij GPT-4 als beoordelaar werd gebruikt, laat zien dat de kwaliteit meer dan 90% van die van ChatGPT/Bard bereikt en op 45% van de vragen niet onderdoet voor ChatGPT. Het team stelde tegelijkertijd een op GPT-4 gebaseerd automatisch evaluatiekader voor en benadrukte dat dit nog geen rigoureuze methode is.","category":"行业动态","source":"LMSYS：Blog","aggregationSource":"LMSYS：Blog","pageTitle":"LMSYS brengt het open-sourcechatmodel Vicuna-13B uit, waarbij LLaMA wordt gefinetuned met ShareGPT-dialogen; de trainingskosten bedragen ongeveer $300. - Aioga AI-nieuws","description":"Het LMSYS-team heeft Vicuna-13B uitgebracht, een op LLaMA gefinetuned model op basis van ongeveer 70K door ShareGPT-gebruikers gedeelde gesprekken, met trainingskosten van ongeveer...","url":"https://www.aioga.com/nl/news/eu8pag1qg6pf93gn9k1ql0m0z/","contentTranslated":true,"translationStatus":"translated","translationRetryAt":"","translationError":"","sourceHash":"6f8ada70b07698c4","translatedAt":"2026-10-03T18:25:53.964Z"},"tr":{"title":"LMSYS, açık kaynaklı sohbet modeli Vicuna-13B'yi yayımladı; LLaMA'yı ShareGPT diyaloglarıyla ince ayarladı, eğitim maliyeti yaklaşık $300.","summary":"LMSYS ekibi, yaklaşık 70K ShareGPT kullanıcısının paylaştığı konuşmalarla LLaMA’yı ince ayardan geçirerek Vicuna-13B’yi yayımladı. Eğitim maliyeti yaklaşık $300 oldu; kod, ağırlıklar ve çevrimiçi demo ticari olmayan bir lisansla paylaşıldı. GPT-4’ün hakem olarak kullanıldığı ilk değerlendirme, modelin ChatGPT/Bard kalitesinin %90’ından fazlasına ulaştığını ve soruların %45’inde ChatGPT’den geri kalmadığını gösterdi. Ekip ayrıca GPT-4 tabanlı otomatik bir değerlendirme çerçevesi önerdi ve bunun henüz titiz bir yöntem olmadığını belirtti.","category":"行业动态","source":"LMSYS：Blog","aggregationSource":"LMSYS：Blog","pageTitle":"LMSYS, açık kaynaklı sohbet modeli Vicuna-13B'yi yayımladı; LLaMA'yı ShareGPT diyaloglarıyla ince ayarladı, eğitim maliyeti yaklaşık $300. - Aioga AI Haberleri","description":"LMSYS ekibi, yaklaşık 70K ShareGPT kullanıcısının paylaştığı konuşmalarla LLaMA’yı ince ayardan geçirerek Vicuna-13B’yi yayımladı. Eğitim maliyeti yaklaşık $300 oldu; kod, ağırlıkl...","url":"https://www.aioga.com/tr/news/eu8pag1qg6pf93gn9k1ql0m0z/","contentTranslated":true,"translationStatus":"translated","translationRetryAt":"","translationError":"","sourceHash":"6f8ada70b07698c4","translatedAt":"2026-10-03T18:31:15.576Z"},"vi":{"title":"LMSYS phát hành mô hình trò chuyện mã nguồn mở Vicuna-13B, tinh chỉnh LLaMA bằng các cuộc hội thoại từ ShareGPT, với chi phí đào tạo khoảng $300","summary":"Đội ngũ LMSYS đã phát hành Vicuna-13B, tinh chỉnh LLaMA bằng các cuộc hội thoại được khoảng 70K người dùng ShareGPT chia sẻ. Chi phí huấn luyện khoảng $300; mã nguồn, trọng số và bản demo trực tuyến được công khai theo giấy phép phi thương mại. Đánh giá sơ bộ với GPT-4 làm giám khảo cho thấy chất lượng của mô hình đạt trên 90% so với ChatGPT/Bard và không thua kém ChatGPT ở 45% số câu hỏi. Đội ngũ đồng thời đề xuất một khung đánh giá tự động dựa trên GPT-4, đồng thời cho biết đây vẫn chưa phải là một phương pháp nghiêm ngặt.","category":"行业动态","source":"LMSYS：Blog","aggregationSource":"LMSYS：Blog","pageTitle":"LMSYS phát hành mô hình trò chuyện mã nguồn mở Vicuna-13B, tinh chỉnh LLaMA bằng các cuộc hội thoại từ ShareGPT, với chi phí đào tạo khoảng $300 - Tin tức AI Aioga","description":"Đội ngũ LMSYS đã phát hành Vicuna-13B, tinh chỉnh LLaMA bằng các cuộc hội thoại được khoảng 70K người dùng ShareGPT chia sẻ. Chi phí huấn luyện khoảng $300; mã nguồn, trọng số và b...","url":"https://www.aioga.com/vi/news/eu8pag1qg6pf93gn9k1ql0m0z/","contentTranslated":true,"translationStatus":"translated","translationRetryAt":"","translationError":"","sourceHash":"6f8ada70b07698c4","translatedAt":"2026-10-03T18:31:01.107Z"},"id":{"title":"LMSYS merilis model percakapan sumber terbuka Vicuna-13B, yang dibuat dengan menyetel halus LLaMA menggunakan percakapan ShareGPT, dengan biaya pelatihan sekitar $300.","summary":"Tim LMSYS merilis Vicuna-13B, yang dilatih dengan melakukan fine-tuning pada LLaMA menggunakan percakapan yang dibagikan oleh sekitar 70K pengguna ShareGPT, dengan biaya pelatihan sekitar $300. Kode, bobot, dan demo online-nya dirilis di bawah lisensi nonkomersial. Evaluasi awal dengan GPT-4 sebagai penilai menunjukkan bahwa kualitasnya mencapai lebih dari 90% kualitas ChatGPT/Bard, dan tidak kalah dari ChatGPT pada 45% pertanyaan. Tim tersebut juga mengusulkan kerangka evaluasi otomatis berbasis GPT-4, serta menjelaskan bahwa metode ini masih belum ketat.","category":"行业动态","source":"LMSYS：Blog","aggregationSource":"LMSYS：Blog","pageTitle":"LMSYS merilis model percakapan sumber terbuka Vicuna-13B, yang dibuat dengan menyetel halus LLaMA menggunakan percakapan ShareGPT, dengan biaya pelatihan sekitar $300. - Berita AI Aioga","description":"Tim LMSYS merilis Vicuna-13B, yang dilatih dengan melakukan fine-tuning pada LLaMA menggunakan percakapan yang dibagikan oleh sekitar 70K pengguna ShareGPT, dengan biaya pelatihan...","url":"https://www.aioga.com/id/news/eu8pag1qg6pf93gn9k1ql0m0z/","contentTranslated":true,"translationStatus":"translated","translationRetryAt":"","translationError":"","sourceHash":"6f8ada70b07698c4","translatedAt":"2026-10-03T18:31:01.313Z"},"th":{"title":"LMSYS เปิดตัวโมเดลแชทโอเพนซอร์ส Vicuna-13B ใช้ ShareGPT ปรับแต่ง LLaMA สำหรับการสนทนา โดยมีค่าใช้จ่ายในการฝึกประมาณ $300","summary":"ทีม LMSYS เปิดตัว Vicuna-13B โดยปรับแต่ง LLaMA ด้วยบทสนทนาที่ผู้ใช้ ShareGPT ประมาณ 70K รายแบ่งปันมา มีค่าใช้จ่ายในการฝึกประมาณ $300 และเผยแพร่โค้ด น้ำหนักโมเดล และ demo ออนไลน์ภายใต้ใบอนุญาตที่ไม่อนุญาตให้ใช้ในเชิงพาณิชย์ การประเมินเบื้องต้นโดยใช้ GPT-4 เป็นผู้ตัดสินแสดงให้เห็นว่าโมเดลมีคุณภาพสูงกว่า 90% ของ ChatGPT/Bard และไม่ด้อยกว่า ChatGPT ใน 45% ของคำถาม ขณะเดียวกัน ทีมยังเสนอกรอบการประเมินอัตโนมัติที่ใช้ GPT-4 และชี้แจงว่าวิธีการดังกล่าวยังไม่ใช่วิธีที่เข้มงวด","category":"行业动态","source":"LMSYS：Blog","aggregationSource":"LMSYS：Blog","pageTitle":"LMSYS เปิดตัวโมเดลแชทโอเพนซอร์ส Vicuna-13B ใช้ ShareGPT ปรับแต่ง LLaMA สำหรับการสนทนา โดยมีค่าใช้จ่ายในการฝึกประมาณ $300 - ข่าว AI Aioga","description":"ทีม LMSYS เปิดตัว Vicuna-13B โดยปรับแต่ง LLaMA ด้วยบทสนทนาที่ผู้ใช้ ShareGPT ประมาณ 70K รายแบ่งปันมา มีค่าใช้จ่ายในการฝึกประมาณ $300 และเผยแพร่โค้ด น้ำหนักโมเดล และ demo ออนไลน์ภาย...","url":"https://www.aioga.com/th/news/eu8pag1qg6pf93gn9k1ql0m0z/","contentTranslated":true,"translationStatus":"translated","translationRetryAt":"","translationError":"","sourceHash":"6f8ada70b07698c4","translatedAt":"2026-10-03T18:36:41.989Z"},"pl":{"title":"LMSYS udostępniło otwartoźródłowy model konwersacyjny Vicuna-13B, dostrajając model LLaMA na podstawie rozmów z ShareGPT; koszt treningu wyniósł około 300 USD.","summary":"Zespół LMSYS opublikował Vicuna-13B, dostrajając model LLaMA na podstawie około 70K udostępnionych rozmów użytkowników ShareGPT. Koszt treningu wyniósł około $300, a kod, wagi i demo online udostępniono na licencji niekomercyjnej. Wstępna ocena przeprowadzona przez GPT-4 jako sędziego wykazała, że jakość modelu przekracza 90% jakości ChatGPT/Bard, a w przypadku 45% pytań nie ustępuje on ChatGPT. Zespół zaproponował również opartą na GPT-4 strukturę automatycznej ewaluacji, zaznaczając, że nie jest to jeszcze rygorystyczna metoda.","category":"行业动态","source":"LMSYS：Blog","aggregationSource":"LMSYS：Blog","pageTitle":"LMSYS udostępniło otwartoźródłowy model konwersacyjny Vicuna-13B, dostrajając model LLaMA na podstawie rozmów z ShareGPT; koszt treningu wyniósł około 300 USD. - Aioga Wiadomości AI","description":"Zespół LMSYS opublikował Vicuna-13B, dostrajając model LLaMA na podstawie około 70K udostępnionych rozmów użytkowników ShareGPT. Koszt treningu wyniósł około $300, a kod, wagi i de...","url":"https://www.aioga.com/pl/news/eu8pag1qg6pf93gn9k1ql0m0z/","contentTranslated":true,"translationStatus":"translated","translationRetryAt":"","translationError":"","sourceHash":"6f8ada70b07698c4","translatedAt":"2026-10-03T18:36:11.451Z"}},"evidenceTier":"verified-news","reviewStatus":"editorial-selected","indexable":true,"editorialCover":""}}