eval-harness

metric

judge_score_unvalidated

better
higher
kind
quality: measured against a reference or by a judge
decides
no goal
weighed by
no goal
recorded on
cnn_dailymail_200

01what it measures, and how

The judges' mean score for the output on the scorer's scale, over its dimensions and judges (evh judge-score). The judges were not validated against human ratings: a trial, not a measurement.

Computed as judge_score is, by judges that were never compared with human ratings for this use. It ranks arms by what the judges prefer, which may not be what people prefer: read it as a trial. A validated scorer turns it into judge_score.

02on cnn_dailymail_200

judge judge-summ-pointwise-v1cnn_deepseek_m_n200_v1 · judged/judge-summ-pointwise-v1: 4.94cnn_bart_l_n200_v1 · judged/judge-summ-pointwise-v1: 4.1814.181better · 4.94

scored against silverscored by a judgewrote the reference it is scored againstscored against a label set, not a reference setunmarked rows are scored against gold

armmean judge_score_unvalidateditemsscored against
cnn_deepseek_m_n200_v1 · judged/judge-summ-pointwise-v14.940200judge judge-summ-pointwise-v1
cnn_bart_l_n200_v1 · judged/judge-summ-pointwise-v14.181200judge judge-summ-pointwise-v1