05-dataset.md

title	Workshop: Instructor Notes for Datasets
description	Facilitator notes for framing datasets as product scope, seeding curated examples into Langfuse, and connecting items back to monitoring.

05 Dataset

Learner guide: 05 Dataset

Instructor notes

• Frame the dataset as product scope written down: expected user inputs, expected behavior, and metadata for slicing.
• The code path is intentionally simple: seed curated JSON into Langfuse, then inspect the hosted dataset.
• Call out why expectedOutput has two fields: idealAnswer feeds the Correctness judge, and expectedKeywords feeds the deterministic keyword_overlap check in the next chapter.
• Connect this chapter back to monitoring: good dataset items often come from surprising production traces.

Demo rhythm

1. Open data/seed-dataset.json and point out input, expectedOutput, and metadata.
2. Run npm run dataset:seed.
3. Open the dataset list and item table in Langfuse.

Watch for

• Learners expecting exact-answer matching. The ideal answer is reference material for evaluators, not a string equality target.
• Duplicate dataset items if the seed script is run repeatedly against the same project.