AWS sagemaker documentation change
Summary
Removed detailed iterative training documentation and replaced with a redirect to Nova 1.0/2.0 user guides. Deleted content included technical implementation details, workflows, monitoring instructions, best practices, cost considerations, limitations, and troubleshooting guides.
Security assessment
The changes involve content removal/relocation rather than addressing security vulnerabilities. While the original content mentioned security aspects like escrow buckets and IAM permissions, the removal doesn't indicate resolution of security issues - it simply moves documentation elsewhere. No evidence of security vulnerability fixes or new security features being documented.
Diff
diff --git a/sagemaker/latest/dg/nova-iterative-training.md b/sagemaker/latest/dg/nova-iterative-training.md index 384d6114f..7ab9ff955 100644 --- a//sagemaker/latest/dg/nova-iterative-training.md +++ b//sagemaker/latest/dg/nova-iterative-training.md @@ -5,2 +4,0 @@ -OverviewHow it worksWhen to use iterative trainingExample workflow: SFT → RFTMonitoring progress across iterationsBest practicesCost considerationsLimitationsTroubleshooting - @@ -9,546 +7 @@ OverviewHow it worksWhen to use iterative trainingExample workflow: SFT → RFTM -## Overview - -Iterative training is the process of repeatedly fine-tuning a model through multiple training cycles across different training methods — train, evaluate, analyze errors, adjust data/objectives/hyperparameters — with each round starting from the previous checkpoint. This approach allows you to systematically target model failure modes, incorporate curated examples addressing specific weaknesses, and adapt to changing requirements over time. - -**Benefits over single-pass training:** - - * **Targeted improvement** : Address specific failure patterns discovered through evaluation - - * **Adaptive refinement** : Respond to distribution shifts or evolving product requirements - - * **Risk mitigation** : Validate improvements incrementally rather than committing to a single long training run - - * **Data efficiency** : Focus data collection efforts on areas where the model underperforms - - * **Curriculum Training** : Multiple round of training with increasingly higher quality data - - - - -## How it works - -### Checkpoint location and access - -After each training job completes, a manifest file is generated in the output location specified by the `output_path` parameter in your training configuration. - -**To access your checkpoint** - - * Navigate to your specified `output_path` in S3 - - * Download and extract the `output.tar.gz` file - - * Open the `manifest.json` file inside - - * Locate the `checkpoint_s3_bucket` parameter, which contains the S3 URI of your trained model - - - - -**Example manifest.json structure** - - - { - "checkpoint_s3_bucket": "s3://customer-escrow-<account-number>-smtj-<unique-identifier>/<job-name>/stepID", - ... - } - -### Understanding escrow buckets - -Since Amazon Nova weights are proprietary, trained model checkpoints are stored in **escrow S3 buckets** within AWS-managed accounts rather than being copied to your account. These escrow buckets: - - * Contain your customized model weights securely - - * Can be referenced by other AWS services (Inference, Evaluation, and subsequent training jobs) - - * Are accessible only to your AWS account via IAM permissions - - * Incur standard S3 storage charges in your account (see Cost considerations) - - - - -You can use the escrow bucket path as the `model_name_or_path` in your next training run to continue iterative training. - -### Using checkpoints for iterative training - -Configure your next training job to use the previous checkpoint as the base model: - - - run: - name: "my-iterative-training-job" - model_type: amazon.nova-2-lite-v1:0:256k - model_name_or_path: "s3://customer-escrow-<account-number>-smtj-<unique-identifier>/<previous-job-name>" - data_s3_path: s3://<bucket>/<data-file>.jsonl - replicas: 4 - -## When to use iterative training - -### Ideal use cases - -Use iterative training when you have: - - * **Feedback loops** – Ability to collect real-world failure cases and systematically address them - - * **Dynamic environments** – Evolving documentation, APIs, or support topics requiring periodic model updates - - * **Robust evaluation** – Strong benchmarks and evaluation frameworks (see examples below) to measure improvements confidently - - * **ML operations capability** – Resources to manage multiple training cycles and version control - - - - -**Examples of robust evaluation frameworks** - - * Automated benchmark suites with pass/fail thresholds - - * Human evaluation protocols with inter-rater reliability metrics - - * Red-team testing scenarios covering edge cases and adversarial inputs - - * A/B testing infrastructure to measure production impact - - - - -### Common patterns - -**SFT → RFT Pipeline** : A frequently used iterative pattern involves: - - * **SFT first** – Teach the model how to solve problems through demonstration examples - - * **RFT second** – Optimize performance across the broader problem space using reward signals - - - - -This sequence is essential when models perform poorly initially—RFT on near-zero accuracy models will not improve performance without first establishing basic problem-solving capabilities through SFT. - -### When not to use iterative training - -Avoid iterative training for: - - * **Stable, well-defined tasks** – Stationary data with consistent requirements already achieving near-ceiling performance - - * **Simple classification problems** – Narrow tasks where single-pass training suffices - - * **Resource constraints** – Lacking dedicated ML operations capabilities to manage multiple training cycles - - * **Marginal gains** – When overhead doesn't justify minimal performance improvements - - - - -## Example workflow: SFT → RFT - -This example demonstrates a common iterative training pattern for reasoning models. - -### Step 1: Initial SFT training - -Configure and launch your SFT training job with your dataset: - - - run: - name: "initial-sft-training" - model_type: amazon.nova-2-lite-v1:0:256k - model_name_or_path: "nova-lite-2/prod" - data_s3_path: s3://<bucket>/sft-training-data.jsonl - validation_data_s3_path: s3://<bucket>/sft-validation-data.jsonl - -**Rationale** : SFT provides additional demonstrations that shape model outputs into your desired format and voice, establishing foundational capabilities. - -**After training completes** - - * Note the `output_path` configured in your training job - - * Download `output.tar.gz` from that location - - * Extract and locate `manifest.json` - - * Copy the `checkpoint_s3_bucket` value - - - - -### Step 2: RFT training on SFT checkpoint - -Create a new RFT training job using the SFT checkpoint: - - - run: - name: "rft-on-sft-checkpoint" - model_type: amazon.nova-2-lite-v1:0:256k - model_name_or_path: "s3://customer-escrow-<account-number>-smtj-<unique-identifier>/<initial-sft-training>" - data_s3_path: s3://<bucket>/rft-training-data.jsonl - reward_lambda_arn: <your-reward-function-arn> - -**Rationale** : RFT training builds on the SFT foundation, allowing the model to develop more complex reasoning patterns optimized by your reward function. - -### Step 3: Evaluate and iterate - -Run evaluation on the RFT checkpoint to assess performance: - - - run: - name: "evaluate-rft-checkpoint" - model_type: amazon.nova-2-lite-v1:0:256k - model_name_or_path: "s3://customer-escrow-<account-number>-smtj-<unique-identifier>/<rft-on-sft-checkpoint>" - data_s3_path: s3://<bucket>/evaluation-data.jsonl - -If target metrics are not satisfied, continue iterating with adjusted data or hyperparameters. - -###### Important