liveBook · Manning

Edge case

As explained in LLM Customization and Fine-Tuning

An uncommon or difficult input situation that tests whether a model behaves correctly beyond routine scenarios.

Edge cases are uncommon or difficult inputs where a model’s reliability matters most. In financial compliance, routine documentation lookups may dominate queries, while unusual instruments, conflicting guidance, and reporting-threshold boundaries carry greater risk if answered incorrectly. section 3.3.4

Plan for difficult variants before collecting adaptation data:

  • For every ten common-scenario examples, add at least three examples of the hardest variant. section 3.3.4
  • At roughly 13 examples per scenario, a 300-example pilot dataset can cover about 20–25 scenarios; spreading it across 50 risks superficial coverage. section 3.3.4
  • Include boundary conditions and error cases, such as truncated incident threads, descriptions that contradict logs, and genuinely unknown answers. section 3.3.4

How LLM Customization and Fine-Tuning covers it

How deep it goes
Quantifies a hard-variant sampling rule and dataset coverage, but does not specify a collection workflow.
Examples
Uses financial compliance queries, reporting thresholds, and incident threads with missing or contradictory information.
What you can do after
Apply a sampling ratio to plan scenario coverage in an adaptation dataset.
How it connects
Links fine-tuning data to model evaluation, boundary conditions, and error handling.

Also explained in

Compare how 13 books explain edge case →

Related concepts

This explainer was generated by Manning's knowledge graph from the text of LLM Customization and Fine-Tuning. The section links point to the parts of the book it draws on.