Demonstrations shaped by your model’s use case.
Supervised fine-tuning examples pair a task and its context with an expert response. The response should solve the problem, follow the required format, and account for the constraints in the source material.
We scope the behaviors and situations your model needs to handle, then have specialists write and review examples against those requirements. Evaluation material stays separate from training data.
Define the behavior before collecting examples
Specify the inputs, required output, and what the model should do when information is missing. Vary the task conditions and difficulty, rather than collecting superficial rewrites.
- Cover representative instructions and constraints.
- Preserve source context and expected output formats.
- Include difficult cases and appropriate uncertainty.
Review for correctness and task fit
Check correctness, instruction following, and support for claims. Record recurring errors so the next collection cycle addresses specific gaps.
- Assess correctness and instruction following together.
- Check consistency across reviewers and task variants.
- Use evaluation findings to guide coverage decisions.
What a program can include
Scope can include:
- Expert-authored instruction and response examples
- Supporting context and task specifications
- Review guidance and quality annotations
- An agreed format for your training pipeline
Illustrative task / Not a production record
Make the objective concrete.
- Objective
- Write a grounded summary of a project document.
- Context
- The source document, a request to identify risks, and a required output structure.
- Review
- Verify that the summary captures the relevant facts, cites the supplied context, and follows the requested format.
Common questions
How does SFT data differ from preference data?
SFT demonstrations show a desired completion. Preference data compares alternative completions and records which is better under defined criteria. The two can support different stages of model development.
Can the examples reflect a professional domain?
Professional knowledge is part of the RDL offering. Scope the domain, reviewer expertise, source material, and acceptance criteria with the team before deciding what examples to collect.