opensite-ai/opensite-skills / aws-ai-ml
aws-ai-ml
VerifiedSelects, deploys, and customizes AI models on Amazon SageMaker. Training or Processing jobs, fine-tuning (SFT/DPO/RLVR/RLAIF), model selection, dataset preparation, evaluation, SageMaker or Bedrock deployment, inference optimization and endpoint diagnostics. Covers the full lifecycle from planning through production. Use when fine-tuning models on SageMaker, choosing which base model to customize, fine-tune, or deploy from SageMaker JumpStart or Hub, SageMakerPublicHub, or the SageMaker public model catalog, transforming or validating training data, evaluating model quality, deploying or optimizing endpoints, configuring IAM/S3 for training, or managing SageMaker Managed MLflow. Use for endpoint health, failures, latency, logs, metrics, errors. Covers Serverless Model Customization, Nova and OSS deployment paths, and PySDK v3. NOT for Ground Truth labeling, Feature Store, or general-purpose AWS infrastructure.
references/finetuning/overview.md
View on GitHub# Finetuning
## Prerequisites
Before starting this workflow, verify:
1. A `use_case_spec.md` file exists
- If missing: Load the `use-case-specification` reference first, then resume
- DON'T EVER offer to create a use case spec without loading the use-case-specification reference.
2. A fine-tuning technique (SFT, DPO, RLVR, RLAIF, or CPT/RFT (for Nova)) and base model have already been selected
- If missing: Load the `model-selection` and/or `finetuning-technique` references to collect what's missing, then resume
- Don't make recommendations on the spot. You MUST load the appropriate reference.
3. A base model name available on SageMakerHub has been identified
- If missing: Load the `model-selection` reference to get it
- **Important:** Only use the model name that `model-selection` retrieves, as it may differ from other commonly used names for the same model
4. The SDK environment has been verified (SDK version, region, execution role)
- If not done: Load the `sdk-getting-started` reference first, then resume
5. The region supports serverless model customization
- Check the environment's region against `../planning/references/region-availability.md`. If it is NOT supported, STOP and give the user the blocking message from that file — do not proceed with fine-tuning. (This is normally enforced by the planning skill, but is repeated here in case fine-tuning is reached without it.)
6. A training dataset uploaded to a bucket in the environment's default region.
- If not met: Help the user upload the dataset to the correct S3
---
### Out of Scope
The following are supported by SageMaker and AWS but do not have a validated workflow in this reference. If the user's request matches one of these, let them know and proceed with best-effort guidance using general AWS knowledge:
- Pre-training or creating a model from scratch
- Bring-your-own container/algorithm training
- Server-ful (non-serverless) SageMaker training jobs
- HyperPod training
- Traditional ML training (XGBoost, scikit-learn, custom algorithms)
---
## Critical Rules
### Code Generation Rules
- ✅ Use EXACTLY the imports shown in each code template
- ❌ Do NOT add additional imports even if they seem helpful
- ❌ Do NOT create variables before they're needed in that section
- 📋 Copy the code structure precisely - no improvisation
- 🎯 Follow the minimal code principle strictly
- ✅ When writing code, make sure the indentation and f strings are correct
- ✅ ALWAYS call `list_hyperparameters()` to validate HP names before using them in generated code. This is MANDATORY -- never use hyperparameter names from the template without first confirming they exist for the selected model via the discovery API.
### User Communication Rules
- ❌ NEVER offer to move on to a downstream reference while training is in progress (logically impossible)
- ❌ NEVER set ACCEPT_EULA to True without explicit user confirmation in the conversation
- ✅ Always mention both the number AND title of sections you reference
- ✅ If user asks how to run (notebook): If `run_cell` is available, offer to run it. Otherwise, tell them to run cells one by one (mention ipykernel requirement).
- ✅ If user asks how to run (script): Tell them to run with `python3 <script>.py`
---
## Workflow
### 1. Code Generation Setup
#### 1.1 Directory Setup
1. Identify project directory from conversation context
- If unclear (multiple relevant directories exist) → Ask user which folder to use
- If no project directory exists → load the **directory-management** reference to set one up
⏸ Wait for user.
#### 1.2 Select Code Template
Read `references/code_output_guide.md` for output format rules, then read the code template matching the finetuning strategy:
- SFT → `code_templates/sft.py`
- DPO → `code_templates/dpo.py`
- RLVR → `code_templates/rlvr.py`
- RLAIF with built-in rewards → `code_templates/rlaif_builtin.py`
- RLAIF with custom prompt → `code_templates/rlaif_custom_prompt.py`
The template is a Python file where each `# Cell N: Label` comment marks the start of a new section. Split on these markers — everything between one marker and the next becomes one unit of output.
#### 1.2.1 Validate Hyperparameter Names
Before generating code, call `list_hyperparameters()` (requires `sagemaker>=3.22.0`, enforced by the `sdk-getting-started` prerequisite) to validate that the hyperparameter names in the template are correct for this model:
```python
from sagemaker.train import list_hyperparameters
# list_hyperparameters(model_name, technique, training_type)
# technique: "SFT", "DPO", "RLVR", "RLAIF"
# training_type: "LORA"
hp_options = list_hyperparameters("meta-textgeneration-llama-3-2-1b-instruct", "SFT", "LORA")
```
This is an agent-side validation step -- the agent calls this internally before writing code. The generated user-facing code uses `trainer.hyperparameters.get_info()` instead (available on any constructed trainer object, no extra import needed).
Use the model name, technique, and training_type from the user's request. The returned object has attribute access for each valid parameter name.
When generating the hyperparameters display and override sections:
- Only include `trainer.hyperparameters.<name>` references for parameter names that exist in `hp_options`
- If a template references a parameter name that does NOT appear in `hp_options`, omit that line from the generated code
- The template's HP names are guidelines for which params to highlight, not a fixed contract
This prevents breakage when recipe schemas rename parameters upstream.
#### 1.3 Generate Code
1. Write the code from the template following the rules in `code_output_guide.md`
2. Use same order, dependencies, and imports as the template
3. DO NOT improvise or add extra code
4. If the model is **NOT** a Meta/Llama model (model ID does NOT start with `meta-`):
- Omit the `ACCEPT_EULA = False` line from the config cell
- Omit the `accept_eula=ACCEPT_EULA,` line from the trainer call
5. If the model is from the Nova family, omit any code containing `max_epochs` or `lr_warmup_steps_ratio` from the Configure Trainer section and the Hyperparameter Overrides section
#### 1.4 Auto-Generate Configuration Values
**In the 'Setup & Credentials' cell, populate:**
1. **BASE_MODEL**
- Use the exact SageMakerHub model name from context
2. **MODEL_PACKAGE_GROUP_NAME**
- Generate from use case (read `use_case_spec.md` if needed)
- Format rules:
- Lowercase, alphanumeric with hyphens only
- 1-63 characters
- Pattern: `[a-zA-Z0-9](-*[a-zA-Z0-9]){0,62}`
- Example: "Customer Support Chatbot" → `customer-support-chatbot-v1`
3. Save notebook
### 2. RLVR Reward Function (for RLVR only, skip this section if technique is SFT or DPO)
#### 2.1 Check Reward Function Status
- Ask if user has a reward function already, or would like help creating one.
- If user says they have one → Ask for the SageMaker Hub Evaluator ARN. Only proceed to Section 2.3 once the user provides a valid Evaluator ARN. If they don't have it registered as a SageMaker Hub Evaluator, continue to 2.2.
- If user says they do not have one → Continue to 2.2
#### 2.2 Generate Reward Function From Template
1. Follow workflow in `references/rlvr_reward_function.md` section "Helping Users Create Custom Reward Functions"
#### 2.3 Set CUSTOM_REWARD_FUNCTION value
1. Set the value for `CUSTOM_REWARD_FUNCTION` in the Notebook with the ARN of the reward function (either given directly by the user, or from the function generation code as `evaluator.arn`).
### 3. RLAIF (for RLAIF only, skip this section if technique is not RLAIF)
Read `references/rlaif_guide.md` and follow its instructions.
### 4. EULA review and acceptance
1. Look up the official license link for the selected base model from references/eula_links.md
2. Display the license to the user following the phrasing in references/eula_links.md. For OSS models: "This model is licensed under **{License}**. Please review the license terms here: {URL}." For Nova models: "This model is subject to the AWS Service Terms: {URL}."
3. Check if the selected base model is a Meta/Llama model (model ID starts with `meta-`)
- **If Meta/Llama**: Tell the user they must read and agree to the EULA before using this model. Ask: "Do you accept the license terms? (yes/no)". If the user confirms, set `ACCEPT_EULA = True` and uncomment `accept_eula=ACCEPT_EULA` in the generated notebook. If the user declines, leave `ACCEPT_EULA = False` and warn that training will fail without acceptance.
- **If non-Meta**: Inform the user of the license for their awareness. No code-level action needed — the `ACCEPT_EULA` variable and `accept_eula` parameter should already be omitted from the notebook (see Step 1.3).
### 5. Post-Generation
After generating the code, offer to run it. Training can take hours depending on your dataset and model.
**Notebook mode:** If `run_cell` is available, offer to run the cells. Otherwise tell the user to run cells themselves.
**Script mode:** Present the user with options:
> "Would you like me to:
>
> 1. Leave it to you — run with `python scripts/[script_name]`
> 2. Run it and wait until it's done
> 3. Start it but don't wait — we can check status later"
- **Option 1:** Done. Wait for user to come back.
- **Option 2:** Execute the script as-is. `trainer.train(wait=True)` blocks until complete. Report final status.
- **Option 3:** Change `wait=True` to `wait=False` in the script, execute, report the training job name.
**Checking status:**
- `describe-training-job --training-job-name NAME` → `TrainingJobStatus`, `FailureReason`, `SecondaryStatusTransitions`
- For model package ARN after completion: `list-model-packages --model-package-group-name GROUP_NAME --sort-by CreationTime --sort-order Descending --max-results 1`
**Showing results after completion:**
- Use `scripts/mlflow_reference.py` as the pattern to query MLflow metrics
- Present loss by epoch as a text table. Metric names depend on the trainer, so list `run.data.metrics.keys()` first and use what was logged (SFT: total_loss/val_eval_total_loss or train/loss/val/loss; DPO: rewards/margins or dpo/reward_margin; RLVR: critic/rewards/mean)
**CRITICAL:**
- DON'T suggest moving to next steps before training completes
- DON'T elaborate on the next steps unless the user specifically asks you about them.
### 6. Continuous Customization
If the user wants to finetune a model they had already customized, follow the instructions in references/continuous_customization.md
---
## References
- `rlvr_reward_function.md` - Lambda reward function creation guide (RLVR only)
- `templates/rlvr_reward_function_source_template.py` - Lambda reward function source template for open-weights models (RLVR only)
- `templates/nova_rlvr_reward_function_source_template.py` - Lambda reward function source template for Nova 2.0 Lite (RLVR only)
- `code_templates/sft.py` - Complete notebook template for Supervised Fine-Tuning (OSS path)
- `code_templates/dpo.py` - Complete notebook template for Direct Preference Optimization (OSS path)
- `code_templates/rlvr.py` - Complete notebook template for Reinforcement Learning from Verifiable Rewards (OSS path)
- `references/continuous_customization.md` - Instructions on fine-tuning an already fine-tuned model.
- `rlaif_guide.md` - instructions on RLAIF finetuning options
- `rlaif_builtin.py` - Code template for RLAIF with built-in judge prompt
- `rlaif_custom_prompt.py` - Code template for RLAIF with custom judge prompt
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| writing-skillsobra/superpowersUse when creating new skills, editing existing skills, or verifying skills work before deployment | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |
| xlsxanthropics/skillsUse this skill any time a spreadsheet file is the primary input or output. This means any task where the user wants to: open, read, edit, or fix an existing .xlsx, .xlsm, .xltx, .csv, or .tsv file (e.g., adding columns, computing formulas, formatting, charting, cleaning messy data); create a new spreadsheet from scratch or from other data sources; or convert between tabular file formats. Trigger especially when the user references a spreadsheet file by name or path — even casually (like "the xlsx in my downloads") — and wants something done to it or produced from it. Also trigger for cleaning or restructuring messy tabular data files (malformed rows, misplaced headers, junk data) into proper spreadsheets. The deliverable must be a spreadsheet file. Do NOT trigger when the primary deliverable is a Word document, HTML report, standalone Python script, database pipeline, or Google Sheets API integration, even if tabular data is involved. | Stars: 198k | Updated | Claude CodeCodex | Verified | Install |
| shipping-and-launchaddyosmani/agent-skillsPrepares production launches. Use when preparing to deploy to production, or when asking what needs to be in place before shipping. Use when you need a pre-launch checklist, when setting up monitoring, when planning a staged rollout, or when you need a rollback strategy. | Stars: 119k | Updated | Claude CodeCodex | Verified | Install |
| childhoodthedotmack/claude-memBabysit a shipped release after merge, publication, or deployment. Verify published packages, production behavior, installed plugins, and late CI or user reports until the observation window is clean; use babysit for PR readiness before release. | Stars: 118k | Updated | Claude CodeCodex | Verified | Install |