eval-harness

metric

judge_won

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 share of the item the arm won, by judges who picked the best output (1, or 1/k when k share the best) or compared outputs two at a time (the share of its pairs won): evh judge-score, mode pick or pairwise. The judges were validated in that mode against human preferences.

evh judge-score in pick mode shows every arm's output for an item at once, blind and shuffled, and each judge names the best; in pairwise mode it shows two at a time and the judge says which is better, for every pair. An arm's judge_won for an item is its share of that item (1 alone best, 1/k shared; its share of the pairs won), the mean over judges. Its mean over items is the arm's win rate; the summary adds Bradley-Terry strengths with intervals, and the position and noise checks.

02on cnn_dailymail_200

judge judge-summ-pick-v1cnn_anthropic_l_n200_v1 · judged/judge-summ-pick-v1: 0.402cnn_anthropic_s_n200_v1 · judged/judge-summ-pick-v1: 0.226cnn_anthropic_m_n200_v1 · judged/judge-summ-pick-v1: 0.144cnn_deepseek_m_n200_v1 · judged/judge-summ-pick-v1: 0.083cnn_qwen_l_n200_v1 · judged/judge-summ-pick-v1: 0.059cnn_qwen_m_n200_v1 · judged/judge-summ-pick-v1: 0.02cnn_deepseek_l_n200_v1 · judged/judge-summ-pick-v1: 0.014cnn_openai_s_n200_v1 · judged/judge-summ-pick-v1: 0.0078cnn_glm_l_n200_v1 · judged/judge-summ-pick-v1: 0.0063cnn_gemma_m_n200_v1 · judged/judge-summ-pick-v1: 0.0054cnn_glm_m_n200_v1 · judged/judge-summ-pick-v1: 0.0053cnn_openai_l_n200_v1 · judged/judge-summ-pick-v1: 0.0044cnn_mistral_s_n200_v1 · judged/judge-summ-pick-v1: 0.0038cnn_deepseek_s_n200_v1 · judged/judge-summ-pick-v1: 0.0036cnn_openai_m_n200_v1 · judged/judge-summ-pick-v1: 0.0035cnn_llama_l_n200_v1 · judged/judge-summ-pick-v1: 0.0029cnn_mistral_l_n200_v1 · judged/judge-summ-pick-v1: 0.0025cnn_gemma_s_n200_v1 · judged/judge-summ-pick-v1: 0.0015cnn_llama_s_n200_v1 · judged/judge-summ-pick-v1: 0.0013cnn_qwen_s_n200_v1 · judged/judge-summ-pick-v1: 0.0011cnn_glm_s_n200_v1 · judged/judge-summ-pick-v1: 0.001cnn_gemma_l_n200_v1 · judged/judge-summ-pick-v1: 0.00064cnn_llama_m_n200_v1 · judged/judge-summ-pick-v1: 0.00063cnn_bart_l_n200_v1 · judged/judge-summ-pick-v1: 0cnn_bart_l_xsum_n200_v1 · judged/judge-summ-pick-v1: 0cnn_bart_m_n200_v1 · judged/judge-summ-pick-v1: 0cnn_bart_s_n200_v1 · judged/judge-summ-pick-v1: 0cnn_lead3_n200_v1 · judged/judge-summ-pick-v1: 0cnn_mistral_m_n200_v1 · judged/judge-summ-pick-v1: 00better · 0.402

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_wonitemsscored against
cnn_anthropic_l_n200_v1 · judged/judge-summ-pick-v10.402198judge judge-summ-pick-v1
cnn_anthropic_s_n200_v1 · judged/judge-summ-pick-v10.226198judge judge-summ-pick-v1
cnn_anthropic_m_n200_v1 · judged/judge-summ-pick-v10.144198judge judge-summ-pick-v1
cnn_deepseek_m_n200_v1 · judged/judge-summ-pick-v10.083198judge judge-summ-pick-v1
cnn_qwen_l_n200_v1 · judged/judge-summ-pick-v10.059198judge judge-summ-pick-v1
cnn_qwen_m_n200_v1 · judged/judge-summ-pick-v10.020198judge judge-summ-pick-v1
cnn_deepseek_l_n200_v1 · judged/judge-summ-pick-v10.014198judge judge-summ-pick-v1
cnn_openai_s_n200_v1 · judged/judge-summ-pick-v10.008198judge judge-summ-pick-v1
cnn_glm_l_n200_v1 · judged/judge-summ-pick-v10.006198judge judge-summ-pick-v1
cnn_gemma_m_n200_v1 · judged/judge-summ-pick-v10.005198judge judge-summ-pick-v1
cnn_glm_m_n200_v1 · judged/judge-summ-pick-v10.005198judge judge-summ-pick-v1
cnn_openai_l_n200_v1 · judged/judge-summ-pick-v10.004198judge judge-summ-pick-v1
cnn_mistral_s_n200_v1 · judged/judge-summ-pick-v10.004198judge judge-summ-pick-v1
cnn_deepseek_s_n200_v1 · judged/judge-summ-pick-v10.004198judge judge-summ-pick-v1
cnn_openai_m_n200_v1 · judged/judge-summ-pick-v10.004198judge judge-summ-pick-v1
cnn_llama_l_n200_v1 · judged/judge-summ-pick-v10.003198judge judge-summ-pick-v1
cnn_mistral_l_n200_v1 · judged/judge-summ-pick-v10.003198judge judge-summ-pick-v1
cnn_gemma_s_n200_v1 · judged/judge-summ-pick-v10.002198judge judge-summ-pick-v1
cnn_llama_s_n200_v1 · judged/judge-summ-pick-v10.001198judge judge-summ-pick-v1
cnn_qwen_s_n200_v1 · judged/judge-summ-pick-v10.001198judge judge-summ-pick-v1
cnn_glm_s_n200_v1 · judged/judge-summ-pick-v10.001198judge judge-summ-pick-v1
cnn_gemma_l_n200_v1 · judged/judge-summ-pick-v10.001198judge judge-summ-pick-v1
cnn_llama_m_n200_v1 · judged/judge-summ-pick-v10.001198judge judge-summ-pick-v1
cnn_bart_l_n200_v1 · judged/judge-summ-pick-v10.000198judge judge-summ-pick-v1
cnn_bart_l_xsum_n200_v1 · judged/judge-summ-pick-v10.000198judge judge-summ-pick-v1
cnn_bart_m_n200_v1 · judged/judge-summ-pick-v10.000198judge judge-summ-pick-v1
cnn_bart_s_n200_v1 · judged/judge-summ-pick-v10.000198judge judge-summ-pick-v1
cnn_lead3_n200_v1 · judged/judge-summ-pick-v10.000198judge judge-summ-pick-v1
cnn_mistral_m_n200_v1 · judged/judge-summ-pick-v10.000198judge judge-summ-pick-v1