eval-harness

metric

judge_score

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 validated against human ratings.

evh judge-score shows each item and the outputs of several arms to two or more judges from different vendor families, blinded and in a seeded order, in blocks. Each judge scores every output on each of the scorer's dimensions; an arm's judge_score for an item is the mean over the dimensions and the judges. The scorer's validation names the committed runs and the agreement with people they reached; the pass also re-asks a fraction of items in another order (position) and again (noise), and reports how much the scores moved.

02on cnn_dailymail_200

judge judge-summ-v1cnn_anthropic_l_n200_v1 · judged/judge-summ-v1: 4.976cnn_anthropic_m_n200_v1 · judged/judge-summ-v1: 4.963cnn_anthropic_s_n200_v1 · judged/judge-summ-v1: 4.952cnn_deepseek_m_n200_v1 · judged/judge-summ-v1: 4.923cnn_deepseek_l_n200_v1 · judged/judge-summ-v1: 4.877cnn_qwen_m_n200_v1 · judged/judge-summ-v1: 4.876cnn_glm_m_n200_v1 · judged/judge-summ-v1: 4.861cnn_openai_s_n200_v1 · judged/judge-summ-v1: 4.825cnn_deepseek_s_n200_v1 · judged/judge-summ-v1: 4.818cnn_qwen_l_n200_v1 · judged/judge-summ-v1: 4.812cnn_gemma_s_n200_v1 · judged/judge-summ-v1: 4.801cnn_mistral_s_n200_v1 · judged/judge-summ-v1: 4.798cnn_openai_m_n200_v1 · judged/judge-summ-v1: 4.795cnn_openai_l_n200_v1 · judged/judge-summ-v1: 4.773cnn_gemma_l_n200_v1 · judged/judge-summ-v1: 4.715cnn_mistral_m_n200_v1 · judged/judge-summ-v1: 4.708cnn_gemma_m_n200_v1 · judged/judge-summ-v1: 4.707cnn_llama_s_n200_v1 · judged/judge-summ-v1: 4.704cnn_llama_l_n200_v1 · judged/judge-summ-v1: 4.7cnn_mistral_l_n200_v1 · judged/judge-summ-v1: 4.698cnn_llama_m_n200_v1 · judged/judge-summ-v1: 4.674cnn_qwen_s_n200_v1 · judged/judge-summ-v1: 4.609cnn_glm_l_n200_v1 · judged/judge-summ-v1: 4.607cnn_glm_s_n200_v1 · judged/judge-summ-v1: 4.53cnn_bart_l_n200_v1 · judged/judge-summ-v1: 3.868cnn_bart_m_n200_v1 · judged/judge-summ-v1: 3.564cnn_bart_l_xsum_n200_v1 · judged/judge-summ-v1: 3.516cnn_lead3_n200_v1 · judged/judge-summ-v1: 3.421cnn_bart_s_n200_v1 · judged/judge-summ-v1: 3.185judge judge-summ-v2cnn_anthropic_l_n200_v1 · judged/judge-summ-v2: 4.888cnn_deepseek_m_n200_v1 · judged/judge-summ-v2: 4.854cnn_openai_l_n200_v1 · judged/judge-summ-v2: 4.847cnn_openai_m_n200_v1 · judged/judge-summ-v2: 4.811cnn_anthropic_m_n200_v1 · judged/judge-summ-v2: 4.809cnn_deepseek_l_n200_v1 · judged/judge-summ-v2: 4.763cnn_openai_s_n200_v1 · judged/judge-summ-v2: 4.747cnn_anthropic_s_n200_v1 · judged/judge-summ-v2: 4.728cnn_glm_m_n200_v1 · judged/judge-summ-v2: 4.718cnn_deepseek_s_n200_v1 · judged/judge-summ-v2: 4.644cnn_gemma_l_n200_v1 · judged/judge-summ-v2: 4.603cnn_qwen_m_n200_v1 · judged/judge-summ-v2: 4.6cnn_gemma_m_n200_v1 · judged/judge-summ-v2: 4.585cnn_qwen_l_n200_v1 · judged/judge-summ-v2: 4.573cnn_llama_l_n200_v1 · judged/judge-summ-v2: 4.529cnn_llama_m_n200_v1 · judged/judge-summ-v2: 4.488cnn_llama_s_n200_v1 · judged/judge-summ-v2: 4.469cnn_qwen_s_n200_v1 · judged/judge-summ-v2: 4.461cnn_gemma_s_n200_v1 · judged/judge-summ-v2: 4.429cnn_glm_s_n200_v1 · judged/judge-summ-v2: 4.42cnn_mistral_s_n200_v1 · judged/judge-summ-v2: 4.408cnn_glm_l_n200_v1 · judged/judge-summ-v2: 4.384cnn_mistral_m_n200_v1 · judged/judge-summ-v2: 4.309cnn_mistral_l_n200_v1 · judged/judge-summ-v2: 4.037cnn_bart_l_n200_v1 · judged/judge-summ-v2: 3.821cnn_bart_l_xsum_n200_v1 · judged/judge-summ-v2: 3.366cnn_lead3_n200_v1 · judged/judge-summ-v2: 3.339cnn_bart_m_n200_v1 · judged/judge-summ-v2: 3.235cnn_bart_s_n200_v1 · judged/judge-summ-v2: 2.9462.946better · 4.976

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_scoreitemsscored against
cnn_anthropic_l_n200_v1 · judged/judge-summ-v14.976197judge judge-summ-v1
cnn_anthropic_m_n200_v1 · judged/judge-summ-v14.963197judge judge-summ-v1
cnn_anthropic_s_n200_v1 · judged/judge-summ-v14.952197judge judge-summ-v1
cnn_deepseek_m_n200_v1 · judged/judge-summ-v14.923197judge judge-summ-v1
cnn_anthropic_l_n200_v1 · judged/judge-summ-v24.888200judge judge-summ-v2
cnn_deepseek_l_n200_v1 · judged/judge-summ-v14.877197judge judge-summ-v1
cnn_qwen_m_n200_v1 · judged/judge-summ-v14.876197judge judge-summ-v1
cnn_glm_m_n200_v1 · judged/judge-summ-v14.861197judge judge-summ-v1
cnn_deepseek_m_n200_v1 · judged/judge-summ-v24.854200judge judge-summ-v2
cnn_openai_l_n200_v1 · judged/judge-summ-v24.847200judge judge-summ-v2
cnn_openai_s_n200_v1 · judged/judge-summ-v14.825197judge judge-summ-v1
cnn_deepseek_s_n200_v1 · judged/judge-summ-v14.818197judge judge-summ-v1
cnn_qwen_l_n200_v1 · judged/judge-summ-v14.812197judge judge-summ-v1
cnn_openai_m_n200_v1 · judged/judge-summ-v24.811200judge judge-summ-v2
cnn_anthropic_m_n200_v1 · judged/judge-summ-v24.809200judge judge-summ-v2
cnn_gemma_s_n200_v1 · judged/judge-summ-v14.801197judge judge-summ-v1
cnn_mistral_s_n200_v1 · judged/judge-summ-v14.798197judge judge-summ-v1
cnn_openai_m_n200_v1 · judged/judge-summ-v14.795197judge judge-summ-v1
cnn_openai_l_n200_v1 · judged/judge-summ-v14.773197judge judge-summ-v1
cnn_deepseek_l_n200_v1 · judged/judge-summ-v24.763200judge judge-summ-v2
cnn_openai_s_n200_v1 · judged/judge-summ-v24.747200judge judge-summ-v2
cnn_anthropic_s_n200_v1 · judged/judge-summ-v24.728200judge judge-summ-v2
cnn_glm_m_n200_v1 · judged/judge-summ-v24.718200judge judge-summ-v2
cnn_gemma_l_n200_v1 · judged/judge-summ-v14.715197judge judge-summ-v1
cnn_mistral_m_n200_v1 · judged/judge-summ-v14.708197judge judge-summ-v1
cnn_gemma_m_n200_v1 · judged/judge-summ-v14.707197judge judge-summ-v1
cnn_llama_s_n200_v1 · judged/judge-summ-v14.704197judge judge-summ-v1
cnn_llama_l_n200_v1 · judged/judge-summ-v14.700197judge judge-summ-v1
cnn_mistral_l_n200_v1 · judged/judge-summ-v14.698197judge judge-summ-v1
cnn_llama_m_n200_v1 · judged/judge-summ-v14.674197judge judge-summ-v1
cnn_deepseek_s_n200_v1 · judged/judge-summ-v24.644200judge judge-summ-v2
cnn_qwen_s_n200_v1 · judged/judge-summ-v14.609197judge judge-summ-v1
cnn_glm_l_n200_v1 · judged/judge-summ-v14.607197judge judge-summ-v1
cnn_gemma_l_n200_v1 · judged/judge-summ-v24.603200judge judge-summ-v2
cnn_qwen_m_n200_v1 · judged/judge-summ-v24.600200judge judge-summ-v2
cnn_gemma_m_n200_v1 · judged/judge-summ-v24.585200judge judge-summ-v2
cnn_qwen_l_n200_v1 · judged/judge-summ-v24.573200judge judge-summ-v2
cnn_glm_s_n200_v1 · judged/judge-summ-v14.530197judge judge-summ-v1
cnn_llama_l_n200_v1 · judged/judge-summ-v24.529200judge judge-summ-v2
cnn_llama_m_n200_v1 · judged/judge-summ-v24.488200judge judge-summ-v2
cnn_llama_s_n200_v1 · judged/judge-summ-v24.469200judge judge-summ-v2
cnn_qwen_s_n200_v1 · judged/judge-summ-v24.461200judge judge-summ-v2
cnn_gemma_s_n200_v1 · judged/judge-summ-v24.429200judge judge-summ-v2
cnn_glm_s_n200_v1 · judged/judge-summ-v24.420200judge judge-summ-v2
cnn_mistral_s_n200_v1 · judged/judge-summ-v24.408200judge judge-summ-v2
cnn_glm_l_n200_v1 · judged/judge-summ-v24.384200judge judge-summ-v2
cnn_mistral_m_n200_v1 · judged/judge-summ-v24.309200judge judge-summ-v2
cnn_mistral_l_n200_v1 · judged/judge-summ-v24.037200judge judge-summ-v2
cnn_bart_l_n200_v1 · judged/judge-summ-v13.868197judge judge-summ-v1
cnn_bart_l_n200_v1 · judged/judge-summ-v23.821200judge judge-summ-v2
cnn_bart_m_n200_v1 · judged/judge-summ-v13.564197judge judge-summ-v1
cnn_bart_l_xsum_n200_v1 · judged/judge-summ-v13.516197judge judge-summ-v1
cnn_lead3_n200_v1 · judged/judge-summ-v13.421197judge judge-summ-v1
cnn_bart_l_xsum_n200_v1 · judged/judge-summ-v23.366200judge judge-summ-v2
cnn_lead3_n200_v1 · judged/judge-summ-v23.339200judge judge-summ-v2
cnn_bart_m_n200_v1 · judged/judge-summ-v23.235200judge judge-summ-v2
cnn_bart_s_n200_v1 · judged/judge-summ-v13.185197judge judge-summ-v1
cnn_bart_s_n200_v1 · judged/judge-summ-v22.946200judge judge-summ-v2