AlpacaEval

What it measures

Pairwise model judgments compare instruction-following responses with a fixed reference. Original and length-controlled evaluator versions answer different questions.

Reading the measurement

A preference result depends on the opponents, judges and prompts in the comparison pool. Use it to study which outputs people or judges favor, then check correctness and reliability with a separate task-based evaluation. Ratings from different arenas do not share a universal scale.

Published evaluation

Original AlpacaEval win rate against the text-davinci-003 reference, judged with the original GPT-4 evaluator. This is neither AlpacaEval 2 length-controlled win rate nor an evaluation of the LIMA or UltraFeedback training datasets. Standard error is retained.

Where this fits in an evaluation suite

Format compliance, preference, truthfulness, and safety require separate measurements. Keep desirable writing style from masking factual errors.

Start with a capability hypothesis, then choose a second test with a different failure mode. Keep the checkpoint, reasoning effort, agent tools, sampling policy, and scoring procedure fixed when measuring a training intervention. Record failures and missing outputs in the denominator.

A reproducible comparison

Original AlpacaEval win rate against the text-davinci-003 reference, judged with the original GPT-4 evaluator. This is neither AlpacaEval 2 length-controlled win rate nor an evaluation of the LIMA or UltraFeedback training datasets. Standard error is retained.

Publish the exact run configuration, item coverage, score units, and available uncertainty. If a source supplies only an aggregate or a single run, preserve that limit. A new model name or a rounded score alone does not establish a capability improvement.