eval-harness

metric

coverage

better
higher
kind
quality: measured against a reference or by a judge
decides
Short news summaries (CNN)
weighed by
no goal
recorded on
cnn_dailymail_20, cnn_dailymail_200, cnn_dailymail_dm200

01what it measures, and how

ROUGE-Lsum recall with the summary cut to the reference's length: did it carry the story.

The summary is clipped to as many words as the gold summary has, then scored for ROUGE-Lsum recall against it (sentences split here, stemming on). Writing more cannot buy coverage; only putting the important facts first can. Read beside concision, which rewards the opposite.

02on cnn_dailymail_20

gold/cnn_dailymail_20cnn_bart_l_v1: 0.366cnn_lead3_v1: 0.2620.262better · 0.366

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 coverageitemsscored against
cnn_bart_l_v10.36620gold/cnn_dailymail_20 gold
cnn_lead3_v10.26220gold/cnn_dailymail_20 gold

03on cnn_dailymail_200

gold/cnn_dailymail_200cnn_bart_l_n200_v1: 0.346cnn_bart_l_n200_v1 · linux: 0.346cnn_deepseek_m_n200_v1: 0.346cnn_llama_l_n200_v1: 0.339cnn_llama_m_n200_v1: 0.338cnn_bart_m_n200_v1: 0.337cnn_bart_m_n200_v1 · linux: 0.337cnn_llama_s_n200_v1: 0.333cnn_deepseek_l_n200_v1: 0.332cnn_glm_m_n200_v1: 0.332cnn_deepseek_s_n200_v1: 0.328cnn_bart_s_n200_v1: 0.328cnn_bart_s_n200_v1 · linux: 0.328cnn_bart_s_n200_v1 · repeats: 0.328cnn_openai_m_n200_v1: 0.327cnn_openai_l_n200_v1: 0.327cnn_mistral_s_n200_v1: 0.326cnn_anthropic_s_n200_v1: 0.324cnn_anthropic_m_n200_v1: 0.324cnn_gemma_l_n200_v1: 0.32cnn_openai_s_n200_v1: 0.317cnn_anthropic_l_n200_v1: 0.315cnn_glm_s_n200_v1: 0.314cnn_gemma_m_n200_v1: 0.314cnn_glm_l_n200_v1: 0.312cnn_gemma_s_n200_v1: 0.312cnn_qwen_l_n200_v1: 0.311cnn_mistral_l_n200_v1: 0.308cnn_qwen_s_n200_v1: 0.307cnn_mistral_m_n200_v1: 0.304cnn_qwen_m_n200_v1: 0.303cnn_lead3_n200_v1: 0.272cnn_lead3_n200_v1 · linux: 0.272cnn_bart_l_xsum_n200_v1: 0.207cnn_bart_l_xsum_n200_v1 · linux: 0.207silver/cnn_dailymail_200cnn_anthropic_m_n200_v1 · silver: 0.491cnn_deepseek_m_n200_v1 · silver: 0.471cnn_anthropic_s_n200_v1 · silver: 0.469cnn_deepseek_l_n200_v1 · silver: 0.399cnn_qwen_l_n200_v1 · silver: 0.398cnn_glm_m_n200_v1 · silver: 0.389cnn_llama_s_n200_v1 · silver: 0.388cnn_deepseek_s_n200_v1 · silver: 0.388cnn_openai_s_n200_v1 · silver: 0.384cnn_mistral_s_n200_v1 · silver: 0.382cnn_qwen_m_n200_v1 · silver: 0.381cnn_glm_l_n200_v1 · silver: 0.375cnn_openai_l_n200_v1 · silver: 0.371cnn_llama_l_n200_v1 · silver: 0.37cnn_openai_m_n200_v1 · silver: 0.359cnn_llama_m_n200_v1 · silver: 0.358cnn_gemma_s_n200_v1 · silver: 0.35cnn_gemma_m_n200_v1 · silver: 0.342cnn_gemma_l_n200_v1 · silver: 0.342cnn_mistral_l_n200_v1 · silver: 0.337cnn_qwen_s_n200_v1 · silver: 0.334cnn_mistral_m_n200_v1 · silver: 0.32cnn_bart_m_n200_v1 · silver: 0.32cnn_bart_l_n200_v1 · silver: 0.317cnn_bart_s_n200_v1 · silver: 0.306cnn_glm_s_n200_v1 · silver: 0.303cnn_lead3_n200_v1 · silver: 0.3cnn_bart_l_xsum_n200_v1 · silver: 0.1370.137better · 0.491

Not drawn: cnn_anthropic_l_n200_v1 · silver, scored against the reference it wrote (its row is in the table).

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 coverageitemsscored against
cnn_bart_l_n200_v10.346200gold/cnn_dailymail_200 gold
cnn_bart_l_n200_v1 · linux0.346200gold/cnn_dailymail_200 gold
cnn_deepseek_m_n200_v10.346200gold/cnn_dailymail_200 gold
cnn_llama_l_n200_v10.339200gold/cnn_dailymail_200 gold
cnn_llama_m_n200_v10.338200gold/cnn_dailymail_200 gold
cnn_bart_m_n200_v10.337200gold/cnn_dailymail_200 gold
cnn_bart_m_n200_v1 · linux0.337200gold/cnn_dailymail_200 gold
cnn_llama_s_n200_v10.333200gold/cnn_dailymail_200 gold
cnn_deepseek_l_n200_v10.332200gold/cnn_dailymail_200 gold
cnn_glm_m_n200_v10.332200gold/cnn_dailymail_200 gold
cnn_deepseek_s_n200_v10.328200gold/cnn_dailymail_200 gold
cnn_bart_s_n200_v10.328200gold/cnn_dailymail_200 gold
cnn_bart_s_n200_v1 · linux0.328200gold/cnn_dailymail_200 gold
cnn_bart_s_n200_v1 · repeats0.328200gold/cnn_dailymail_200 gold
cnn_openai_m_n200_v10.327200gold/cnn_dailymail_200 gold
cnn_openai_l_n200_v10.327200gold/cnn_dailymail_200 gold
cnn_mistral_s_n200_v10.326200gold/cnn_dailymail_200 gold
cnn_anthropic_s_n200_v10.324200gold/cnn_dailymail_200 gold
cnn_anthropic_m_n200_v10.324200gold/cnn_dailymail_200 gold
cnn_gemma_l_n200_v10.320200gold/cnn_dailymail_200 gold
cnn_openai_s_n200_v10.317200gold/cnn_dailymail_200 gold
cnn_anthropic_l_n200_v10.315200gold/cnn_dailymail_200 gold
cnn_glm_s_n200_v10.314200gold/cnn_dailymail_200 gold
cnn_gemma_m_n200_v10.314200gold/cnn_dailymail_200 gold
cnn_glm_l_n200_v10.312200gold/cnn_dailymail_200 gold
cnn_gemma_s_n200_v10.312200gold/cnn_dailymail_200 gold
cnn_qwen_l_n200_v10.311200gold/cnn_dailymail_200 gold
cnn_mistral_l_n200_v10.308200gold/cnn_dailymail_200 gold
cnn_qwen_s_n200_v10.307200gold/cnn_dailymail_200 gold
cnn_mistral_m_n200_v10.304200gold/cnn_dailymail_200 gold
cnn_qwen_m_n200_v10.303200gold/cnn_dailymail_200 gold
cnn_lead3_n200_v10.272200gold/cnn_dailymail_200 gold
cnn_lead3_n200_v1 · linux0.272200gold/cnn_dailymail_200 gold
cnn_bart_l_xsum_n200_v10.207200gold/cnn_dailymail_200 gold
cnn_bart_l_xsum_n200_v1 · linux0.207200gold/cnn_dailymail_200 gold
cnn_anthropic_l_n200_v1 · silver1.000200silver/cnn_dailymail_200 silver
cnn_anthropic_m_n200_v1 · silver0.491200silver/cnn_dailymail_200 silver
cnn_deepseek_m_n200_v1 · silver0.471200silver/cnn_dailymail_200 silver
cnn_anthropic_s_n200_v1 · silver0.469200silver/cnn_dailymail_200 silver
cnn_deepseek_l_n200_v1 · silver0.399200silver/cnn_dailymail_200 silver
cnn_qwen_l_n200_v1 · silver0.398200silver/cnn_dailymail_200 silver
cnn_glm_m_n200_v1 · silver0.389200silver/cnn_dailymail_200 silver
cnn_llama_s_n200_v1 · silver0.388200silver/cnn_dailymail_200 silver
cnn_deepseek_s_n200_v1 · silver0.388200silver/cnn_dailymail_200 silver
cnn_openai_s_n200_v1 · silver0.384200silver/cnn_dailymail_200 silver
cnn_mistral_s_n200_v1 · silver0.382200silver/cnn_dailymail_200 silver
cnn_qwen_m_n200_v1 · silver0.381200silver/cnn_dailymail_200 silver
cnn_glm_l_n200_v1 · silver0.375200silver/cnn_dailymail_200 silver
cnn_openai_l_n200_v1 · silver0.371200silver/cnn_dailymail_200 silver
cnn_llama_l_n200_v1 · silver0.370200silver/cnn_dailymail_200 silver
cnn_openai_m_n200_v1 · silver0.359200silver/cnn_dailymail_200 silver
cnn_llama_m_n200_v1 · silver0.358200silver/cnn_dailymail_200 silver
cnn_gemma_s_n200_v1 · silver0.350200silver/cnn_dailymail_200 silver
cnn_gemma_m_n200_v1 · silver0.342200silver/cnn_dailymail_200 silver
cnn_gemma_l_n200_v1 · silver0.342200silver/cnn_dailymail_200 silver
cnn_mistral_l_n200_v1 · silver0.337200silver/cnn_dailymail_200 silver
cnn_qwen_s_n200_v1 · silver0.334200silver/cnn_dailymail_200 silver
cnn_mistral_m_n200_v1 · silver0.320200silver/cnn_dailymail_200 silver
cnn_bart_m_n200_v1 · silver0.320200silver/cnn_dailymail_200 silver
cnn_bart_l_n200_v1 · silver0.317200silver/cnn_dailymail_200 silver
cnn_bart_s_n200_v1 · silver0.306200silver/cnn_dailymail_200 silver
cnn_glm_s_n200_v1 · silver0.303200silver/cnn_dailymail_200 silver
cnn_lead3_n200_v1 · silver0.300200silver/cnn_dailymail_200 silver
cnn_bart_l_xsum_n200_v1 · silver0.137200silver/cnn_dailymail_200 silver

04on cnn_dailymail_dm200

gold/cnn_dailymail_dm200cnn_bart_l_dm200_v1: 0.443cnn_bart_m_dm200_v1: 0.443cnn_bart_s_dm200_v1: 0.405cnn_lead3_dm200_v1: 0.37cnn_bart_l_xsum_dm200_v1: 0.1520.152better · 0.443

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