Finetuning Technique

Finetuning Technique

Guides the user through selecting a fine-tuning technique based on their use case and validates compatibility with the selected model.

When to Use

• User has decided to finetune and needs to choose a technique
• User wants to change their finetuning technique
• Technique needs to be validated against a selected model

Prerequisites

• A base model has been selected (via model-selection skill). The model name and hub must be known.
• A use_case_spec.md file exists. If not, activate the use-case-specification skill to generate it first.

Workflow

Step 1: Determine Finetuning Technique

Consult references/finetune_technique_selection_guide.md to recommend the best-fit technique based on the use case and the user's needs (SFT, DPO, RLVR, RLAIF).

Present the recommendation and reasoning to the user. Ask if they'd like to go with the recommendation or prefer a different technique.

Step 2: Validate Technique Availability

1. Once the user confirms a technique, retrieve the finetuning techniques available for the selected model by running: python finetuning-technique/scripts/get_recipes.py <model-name> <hub-name>

• This returns only the techniques the model actually supports, filtered to SFT, DPO, RLVR, and RLAIF. Only these four techniques are supported — ignore any other techniques even if the model's recipes include them.

2. If the chosen technique is available for the model, proceed to Step 3. 3. If the chosen technique is not available for the model, explain that the selected model does not support it on SageMaker and offer to go back to model-selection to pick a different model that supports the chosen technique.

Step 3: Confirm Selections

Present a summary to the user:

Here's what we've selected:
- Base model: [model name]
- Fine-tuning technique: [SFT/DPO/RLVR/RLAIF]

References

• references/finetune_technique_selection_guide.md — Technique guidance (SFT/DPO/RLVR/RLAIF)

Finetuning Technique Selection Guide

Not all models support all techniques. Always validate technique availability against the selected model's recipes before recommending. Only SFT, DPO, RLVR, and RLAIF are supported.

Technique Overview

SFT (Supervised Fine-Tuning)

Use when:

• Task has clear right/wrong answers
• Single optimal output per input
• Output represents exemplary responses
• Classification, extraction, structured generation

DPO (Direct Preference Optimization)

Use when:

• Multiple valid outputs, some better than others
• Subjective quality (tone, style, helpfulness)
• Creative tasks with preference judgments

RLVR (Reinforcement Learning from Verifiable Rewards)

Use when:

• Outputs can be verified programmatically
• Want to reward similarity to gold responses
• Code generation (passes tests = reward)
• Math problems (correct answer = reward)
• Constraint satisfaction (meets criteria = reward)

Key difference from SFT:

• SFT: Model learns to imitate gold responses directly
• RLVR: Model learns to maximize rewards (can be gold similarity or verification-based)

RLAIF (Reinforcement Learning from AI Feedback)

Use when:

• Quality is subjective and hard to define with rules (tone, helpfulness, brand voice, safety)
• No human preference data is available and collecting it is too expensive or slow
• You want RLHF-level alignment without human annotators
• Task involves summarization, dialogue, or open-ended generation where "better" is a judgment call
• You need scalable preference signals that can be regenerated as the model improves

Key difference from DPO:

• DPO: Requires a static dataset of preference pairs (chosen/rejected) upfront
• RLAIF: Uses an AI judge model to generate preference signals or reward scores dynamically, enabling iterative improvement

Key difference from RLVR:

• RLVR: Reward is rule-based and programmatic (correct/incorrect, passes tests)
• RLAIF: Reward comes from an AI model evaluating subjective quality (helpfulness, coherence, safety)

When NOT to use RLAIF:

• Task has objectively verifiable answers → use RLVR instead
• You already have high-quality human preference data → use DPO instead
• You have clear gold-standard outputs → use SFT instead
• The AI judge model is weaker than the model being trained (judge quality bounds training quality)