liveBook · Manning

Format-agnostic evaluation

As explained in Reinforcement Learning from Human Feedback

Format-agnostic evaluation assesses a model's answer without requiring a rigid output syntax unrelated to the answer's substance.

Format-agnostic evaluation scores whether a model produces the correct answer without making success depend on an unrelated output template. Rigid suffixes, answer labels, or regular-expression extraction can turn capable responses into apparent failures when generation does not match the evaluation format. section 16.1.1

Practical evaluation choices include:

  • Use log-likelihood scoring when the task lacks a question-and-answer format, especially in pretraining evaluation. section 16.1.1
  • Treat exact match cautiously in post-training evaluation when answers may omit required phrases such as “The answer is:” or use unexpected wording. section 16.1.1
  • Design extraction and prompting so formatting does not bottleneck the capability being measured; achieving this generally requires substantial iteration. section 16.1.1

Multiple-choice output

A prompt that demands “Correct Answer: (A)” may mark a response incorrect if the model gives the right explanation without the label. A format-agnostic evaluator separates answer correctness from compliance with that template. section 16.1.1

How Reinforcement Learning from Human Feedback covers it

How deep it goes
Contrasts exact match and log-likelihood scoring, identifies extraction failures, and notes implementation effort without prescribing an algorithm.
Examples
Uses rigid answer suffixes, regular-expression extraction, and MMLU-style labeled choices to expose formatting failures.
What you can do after
Evaluate model answers while reducing distortions caused by rigid output syntax and prompt-format noncompliance.
How it connects
Connects exact match, log-likelihood scoring, few-shot prompting, pretraining evaluation, and post-training evaluation.
Cover of Reinforcement Learning from Human Feedback
liveBookmanning.com
Read Reinforcement Learning from Human Feedback on liveBook Get the book at manning.com

Related concepts

This explainer was generated by Manning's knowledge graph from the text of Reinforcement Learning from Human Feedback. The section links point to the parts of the book it draws on.