Ricette Amazon Nova - Amazon SageMaker AI

Ricette Amazon Nova

È possibile ottenere una ricetta Amazon Nova dal repository di ricette di SageMaker HyperPod. Una ricetta Nova è un file di configurazione YAML che fornisce dettagli a SageMaker AI su come eseguire il processo di personalizzazione del modello. Fornisce il nome del modello di base, imposta gli iperparametri di addestramento, definisce le impostazioni di ottimizzazione e include tutte le opzioni aggiuntive necessarie per eseguire il fine-tuning o addestrare correttamente il modello.

È anche possibile accedere alle ricette Nova tramite Amazon SageMaker Studio e Amazon SageMaker Unified Studio passando all’hub del modello JumpStart, selezionando AWS ed esplorando i modelli Amazon Nova per trovare le ricette associate. Amazon SageMaker Studio e Amazon SageMaker Unified Studio forniscono entrambi notebook di esempio per ogni ricetta, che includono tutte le fasi necessarie per modificare le ricette ed eseguire processi di personalizzazione utilizzando job di addestramento di SageMaker AI o ambienti Amazon SageMaker HyperPod.

Per accedere alle pagine delle ricette in Amazon SageMaker Studio, il ruolo di esecuzione deve disporre delle seguenti autorizzazioni.

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

Per eseguire i notebook di esempio sui job di addestramento di SageMaker e SageMaker HyperPod, utilizza una delle seguenti versioni dell’immagine di distribuzione SageMaker: 2.7.1+, 2.8.0+, 3.2.1+, 3.3.0+. Questa condizione si applica sia ad Amazon SageMaker Studio che ad Amazon SageMaker Unified Studio.

Recupero di ricette Amazon Nova

Per ottenere una ricetta di base di Amazon Nova, clona il repository di ricette SageMaker HyperPod eseguendo il comando seguente.

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

Le ricette di base sono disponibili in recipes_collection/recipes/.

cd recipes_collection/recipes/

Le ricette di personalizzazione di Amazon Nova si trovano nelle seguenti cartelle.

Recipe type Cartella
SFT (a dimensionalità completa e PEFT), PPO, DPO (a dimensionalità completa e PEFT) fine-tuning/nova
Valutazione evaluation/nova
CPT training/nova

Modelli e algoritmi disponibili

La tabella seguente riepiloga la disponibilità della personalizzazione per i modelli Amazon Nova e gli algoritmi supportati mediante SageMaker AI.

Nome modello

ID modello

Ottimizzazione

Note

Amazon Nova Micro

amazon.nova-micro-v1:0:128k

For both SFT and DPO, this model accepts text as input and produces text as output only.

Amazon Nova Lite

amazon.nova-lite-v1:0:300k

  • SFT - Accetta testo e/o immagini o testo e/o video come input e produce testo come output. Un singolo job non può combinare testo, immagini e video nella stessa esecuzione.

  • DPO - Accetta testo e immagini come input e produce testo come output.

Amazon Nova Pro

amazon.nova-pro-v1:0:300k

  • SFT - Accetta testo e/o immagini o testo e/o video come input e produce testo come output. Un singolo job non può combinare testo, immagini e video nella stessa esecuzione.

  • DPO - Accetta testo e immagini come input e produce testo come output.

Riferimento per le ricette di Amazon Nova

La tabella seguente elenca informazioni dettagliate sul riferimento per le ricette di Amazon Nova.

Modello Categoria/Sottocategoria Tecnica Nome ricetta URI immagine (job di addestramento SageMaker) URI immagine (SageMaker HyperPod) Istanza di calcolo
Nova Lite Training/Fine-tuning

Fine-tuning supervisionato (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 Training/Fine-tuning

Fine-tuning supervisionato (completo)

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 Training/Fine-tuning

Ottimizzazione diretta delle preferenze (completa)

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 Training/Fine-tuning

Ottimizzazione diretta delle preferenze (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 Training/Reinforcement learning

Apprendimento per rinforzo basato su ricompense (PPO)

nova_lite_p5_gpu_ppo.yaml N/A 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-fine-tune-repo:SMHP-PPO-TRAIN-latest ml.p5.48xlarge
Nova Lite Training/Continued Pre-Training Preaddestramento continuo (modello di base) nova_lite_gpu_p5x16_pretrain.yaml N/A 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-fine-tune-repo:HP-CPT-latest ml.p5.48xlarge
Nova Lite Evaluation/Evaluate Benchmark testuali standard 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 Evaluation/Evaluate

Valutazione dei set di dati personalizzati

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 Evaluation/Evaluate

Benchmark multimodali

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 Evaluation/Evaluate

LLM-as-a-judge

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 Training/Fine-tuning

Fine-tuning supervisionato (LoRA)

nova_micro_g5_gpu_lora_sft.yaml 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-fine-tune-repo:SM-TJ-SFT-latest N/A ml.g5.12xlarge
Nova Micro Training/Fine-tuning

Fine-tuning supervisionato (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 Training/Fine-tuning

Fine-tuning supervisionato (completo)

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 Training/Fine-tuning

Ottimizzazione diretta delle preferenze (completa)

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 Training/Fine-tuning

Ottimizzazione diretta delle preferenze (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 Training/Reinforcement learning

Apprendimento per rinforzo basato su ricompense (PPO)

nova_micro_p5_gpu_ppo.yaml N/A 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-fine-tune-repo:SMHP-PPO-TRAIN-latest ml.p5.48xlarge
Nova Micro Training/Continued Pre-Training Preaddestramento continuo (modello di base) nova_micro_gpu_p5x8_pretrain.yaml N/A 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-fine-tune-repo:HP-CPT-latest ml.p5.48xlarge
Nova Micro Evaluation/Evaluate Benchmark testuale generale 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 Evaluation/Evaluate

Benchmark Bring Your Own Dataset (gen_qa)

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 Evaluation/Evaluate

LLM-as-a-judge

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 Training/Fine-tuning

Fine-tuning supervisionato (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 Training/Fine-tuning

Fine-tuning supervisionato (completo)

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 Training/Fine-tuning

Ottimizzazione diretta delle preferenze (completa)

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 Training/Fine-tuning

Ottimizzazione diretta delle preferenze (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 Training/Reinforcement learning

Apprendimento per rinforzo basato su ricompense (PPO)

nova_pro_p5_gpu_ppo.yaml N/A 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-fine-tune-repo:SMHP-PPO-TRAIN-latest ml.p5.48xlarge
Nova Pro Training/Continued Pre-Training Preaddestramento continuo (modello di base) nova_pro_gpu_p5x24_pretrain.yaml N/A 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-fine-tune-repo:HP-CPT-latest ml.p5.48xlarge
Nova Pro Training/Data augmentation Distillazione di modelli per il post-addestramento nova_pro_r5_cpu_distill.yaml N/A 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-distillation-repo:SM-TJ-DISTILL-LATEST ml.r5.24xlarge
Nova Pro Evaluation/Evaluate Benchmark testuali standard 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 Evaluation/Evaluate Valutazione dei set di dati personalizzati 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 Evaluation/Evaluate Benchmark multimodali 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 Evaluation/Evaluate LLM-as-a-judge 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
Nova Premier Addestramento Distillazione di modelli per il post-addestramento

nova_premier_r5_cpu_distill.yaml

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

N/A

ml.r5.24xlarge