亚马逊 Nova 食谱 - 亚马逊 SageMaker AI
Amazon Web Services 文档中描述的 Amazon Web Services 服务或功能可能因区域而异。要查看适用于中国区域的差异,请参阅 中国的 Amazon Web Services 服务入门 (PDF)

本文属于机器翻译版本。若本译文内容与英语原文存在差异,则一律以英文原文为准。

亚马逊 Nova 食谱

你可以从食谱库中获取 Amazon Nova SageMaker HyperPod 食谱。Nova 配方是一个 YAML 配置文件,它向 SageMaker AI 提供了有关如何运行模型自定义作业的详细信息。它提供基础模型名称,设置训练超参数,定义优化设置,并包括成功微调或训练模型所需的任何其他选项。

你也可以通过 Amazon SageMaker Studio 和 Amazon Uni SageMaker fied Studio 访问 Nova 食谱,方法是导航到 JumpStart 模型中心 Amazon,选择并浏览 Amazon Nova 模型以查找其相关食谱。Amazon SageMaker Studio 和 Amazon SageMaker Unified Studio 都为每个食谱提供了示例笔记本,其中包括修改配方和使用 SageMaker 人工智能训练作业或亚马逊 SageMaker HyperPod 环境运行自定义任务的所有必要步骤。

要访问 Amazon SageMaker Studio 中的食谱页面,执行角色必须具有以下权限。

{ "Version": "2012-10-17", "Statement": [ { "Effect": "Allow", "Action": [ "s3:GetObject" ], "Resource": [ "arn:aws:s3:::*model-customization-recipes*" ] } ] }

要在 SageMaker 训练作业上执行示例笔记本 SageMaker HyperPod,请使用以下 SageMaker 分发映像版本之一:2.7.1+2.8.0+3.2.1+3.3.0+。这适用于亚马逊 SageMaker Studio 和亚马逊 SageMaker 联合工作室。

获取亚马逊 Nova 食谱

要获取基本的 Amazon Nova 配方,请运行以下命令克隆SageMaker HyperPod 配方存储库。

git clone https://github.com/aws/sagemaker-hyperpod-recipes.git

基本食谱可在以下网址获得recipes_collection/recipes/

cd recipes_collection/recipes/

Amazon Nova 定制食谱位于以下文件夹中。

食谱类型 文件夹
SFT(全等级和 PEFT)、PPO、DPO(满等级和 PEFT) 微调/新星
评估 评估/新星
CPT 训练/新星

可用模型和算法

下表汇总了 Amazon Nova 机型的自定义功能以及支持的 SageMaker AI 算法。

模型名称

模型 ID

微调

备注

Amazon Nova Micro

亚马逊。 nova-micro-v1:0:128 k

对于 SFT 和 DPO,此模型接受文本作为输入,仅生成文本作为输出。

Amazon Nova Lite

亚马逊。 nova-lite-v1:0300 k

  • SFT-接受文本 and/or 图像或文本 and/or 视频作为输入并生成文本作为输出。单个作业不能在同一次运行中合并文本、图像和视频。

  • DPO-接受文本和图像作为输入并生成文本作为输出。

Amazon Nova Pro

亚马逊。 nova-pro-v1:0300 k

  • SFT-接受文本 and/or 图像或文本 and/or 视频作为输入并生成文本作为输出。单个作业不能在同一次运行中合并文本、图像和视频。

  • DPO-接受文本和图像作为输入并生成文本作为输出。

亚马逊 Nova 食谱参考

下表列出了 Amazon Nova 食谱参考的详细信息。

模型 类别/子类别 方法 食谱名称 图片 URI(SageMaker 训练作业) 图片 URI (SageMaker HyperPod) 计算实例
Nova Lite 训练/微调

监督微调 (LoRa)

nova_lite_p5_gpu_lora_sft.yaml 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-fine-tune-repo:SM-TJ-SFT-latest 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-fine-tune-repo:SM-HP-SFT-latest ml.p5.48xlarge
Nova Lite 训练/微调

监督微调(完整)

nova_lite_p5_gpu_sft.yaml 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-fine-tune-repo:SM-TJ-SFT-latest 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-fine-tune-repo:SM-HP-SFT-latest ml.p5.48xlarge
Nova Lite 训练/微调

直接偏好优化(完整)

nova_lite_p5_gpu_dpo.yaml 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-fine-tune-repo:SM-TJ-DPO-latest 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-fine-tune-repo:SM-HP-DPO-latest ml.p5.48xlarge
Nova Lite 训练/微调

直接偏好优化 (LoRa)

nova_lite_p5_gpu_lora_dpo.yaml 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-fine-tune-repo:SM-TJ-DPO-latest 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-fine-tune-repo:SM-HP-DPO-latest ml.p5.48xlarge
Nova Lite 训练/强化学习

基于奖励的强化学习 (PPO)

nova_lite_p5_gpu_ppo.yaml 不适用 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-fine-tune-repo:SMHP-PPO-TRAIN-latest ml.p5.48xlarge
Nova Lite 培训/继续预训练 继续预训练(基础模型) nova_lite_gpu_p5x16_pretrain.yaml 不适用 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-fine-tune-repo:HP-CPT-latest ml.p5.48xlarge
Nova Lite 评估/评估 标准文本基准 nova_lite_p5_48xl_general_text_benchmark_eval.yaml 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-evaluation-repo:SM-TJ-Eval-latest 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-evaluation-repo:SM-HP-Eval-latest ml.p5.48xlarge
Nova Lite 评估/评估

自定义数据集评估

nova_lite_p5_48xl_bring_your_own_dataset_eval.yaml 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-evaluation-repo:SM-TJ-Eval-latest 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-evaluation-repo:SM-HP-Eval-latest ml.p5.48xlarge
Nova Lite 评估/评估

多模式基准

nova_lite_p5_48_general_multi_modal_benchmark_eval.yaml 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-evaluation-repo:SM-TJ-Eval-latest 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-evaluation-repo:SM-HP-Eval-latest ml.p5.48xlarge
Nova Lite 评估/评估

作为评委的法学硕士

nova_lite_p5_48xl_llm_judge_eval.yaml 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-evaluation-repo:SM-TJ-Eval-latest 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-evaluation-repo:SM-HP-Eval-latest ml.p5.48xlarge
Nova Micro 训练/微调

监督微调 (LoRa)

nova_micro_p5_gpu_lora_sft.yaml 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-fine-tune-repo:SM-TJ-SFT-latest 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-fine-tune-repo:SM-HP-SFT-latest ml.p5.48xlarge
Nova Micro 训练/微调

监督微调(完整)

nova_micro_p5_gpu_sft.yaml 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-fine-tune-repo:SM-TJ-SFT-latest 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-fine-tune-repo:SM-HP-SFT-latest ml.p5.48xlarge
Nova Micro 训练/微调

直接偏好优化(完整)

nova_micro_p5_gpu_dpo.yaml 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-fine-tune-repo:SM-TJ-DPO-latest 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-fine-tune-repo:SM-HP-DPO-latest ml.p5.48xlarge
Nova Micro 训练/微调

直接偏好优化 (LoRa)

nova_micro_p5_gpu_lora_dpo.yaml 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-fine-tune-repo:SM-TJ-DPO-latest 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-fine-tune-repo:SM-HP-DPO-latest ml.p5.48xlarge
Nova Micro 训练/强化学习

基于奖励的强化学习 (PPO)

nova_micro_p5_gpu_ppo.yaml 不适用 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-fine-tune-repo:SMHP-PPO-TRAIN-latest ml.p5.48xlarge
Nova Micro 培训/继续预训练 持续预训练(基础模型) nova_micro_gpu_p5x8_pretrain.yaml 不适用 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-fine-tune-repo:HP-CPT-latest ml.p5.48xlarge
Nova Micro 评估/评估 通用文本基准 nova_micro_p5_48xl_general_text_benchmark_eval.yaml 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-evaluation-repo:SM-TJ-Eval-latest 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-evaluation-repo:SM-HP-Eval-latest ml.p5.48xlarge
Nova Micro 评估/评估

自带数据集 (gen_qa) 基准测试

nova_micro_p5_48xl_bring_your_own_dataset_eval.yaml 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-evaluation-repo:SM-TJ-Eval-latest 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-evaluation-repo:SM-HP-Eval-latest ml.p5.48xlarge
Nova Micro 评估/评估

作为评委的法学硕士

nova_micro_p5_48xl_llm_judge_eval.yaml 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-evaluation-repo:SM-TJ-Eval-latest 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-evaluation-repo:SM-HP-Eval-latest ml.p5.48xlarge
Nova Pro 训练/微调

监督微调 (LoRa)

nova_pro_p5_gpu_lora_sft.yaml 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-fine-tune-repo:SM-TJ-SFT-latest 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-fine-tune-repo:SM-HP-SFT-latest ml.p5.48xlarge
Nova Pro 训练/微调

监督微调(完整)

nova_pro_p5_gpu_sft.yaml 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-fine-tune-repo:SM-TJ-SFT-latest 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-fine-tune-repo:SM-HP-SFT-latest ml.p5.48xlarge
Nova Pro 训练/微调

直接偏好优化(完整)

nova_pro_p5_gpu_dpo.yaml 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-fine-tune-repo:SM-TJ-DPO-latest 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-fine-tune-repo:SM-HP-DPO-latest ml.p5.48xlarge
Nova Pro 训练/微调

直接偏好优化 (LoRa)

nova_pro_p5_gpu_lora_dpo.yaml 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-fine-tune-repo:SM-TJ-DPO-latest 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-fine-tune-repo:SM-HP-DPO-latest ml.p5.48xlarge
Nova Pro 训练/强化学习

基于奖励的强化学习 (PPO)

nova_pro_p5_gpu_ppo.yaml 不适用 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-fine-tune-repo:SMHP-PPO-TRAIN-latest ml.p5.48xlarge
Nova Pro 培训/继续预训练 持续预训练(基础模型) nova_pro_gpu_p5x24_pretrain.yaml 不适用 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-fine-tune-repo:HP-CPT-latest ml.p5.48xlarge
Nova Pro 训练/数据增强 模型蒸馏用于后期训练 nova_pro_r5_cpu_distill.yaml 不适用 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-distillation-repo:SM-TJ-DISTILL-LATEST ml.r5.24xlarge
Nova Pro 评估/评估 标准文本基准 nova_pro_p5_48xl_general_text_benchmark_eval.yaml 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-evaluation-repo:SM-TJ-Eval-latest 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-evaluation-repo:SM-HP-Eval-latest ml.p5.48xlarge
Nova Pro 评估/评估 自定义数据集评估 nova_pro_p5_48xl_bring_your_own_dataset_eval.yaml 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-evaluation-repo:SM-TJ-Eval-latest 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-evaluation-repo:SM-HP-Eval-latest ml.p5.48xlarge
Nova Pro 评估/评估 多模式基准 nova_pro_p5_48xl_general_multi_modal_benchmark_eval.yaml 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-evaluation-repo:SM-TJ-Eval-latest 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-evaluation-repo:SM-HP-Eval-latest ml.p5.48xlarge
Nova Pro 评估/评估 作为评委的法学硕士 nova_pro_p5_48xl_llm_judge_eval.yaml 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-evaluation-repo:SM-TJ-Eval-latest 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-evaluation-repo:SM-HP-Eval-latest ml.p5.48xlarge
新星总理 训练 模型蒸馏用于后期训练

nova_premier_r5_cpu_distill.yaml

708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-distillation-repo:SM-TJ-DISTILL-LATEST

不适用

ml.r5.24xlarge