eval-harness

metric

ndcg_10

better
higher
kind
quality: measured against a reference or by a judge
decides
Evidence retrieval for scientific claims (SciFact)
weighed by
no goal
recorded on
scifact_200

01what it measures, and how

nDCG@10: how high the relevant documents sit in the first 10, against the best possible order.

Gain 2^grade − 1 per relevant document (grade above 0), discounted by log2 of rank + 1, divided by the ideal ordering's DCG. Ids are normalised and duplicates collapsed (first occurrence wins); an id that is not a corpus document keeps its slot and earns nothing. An unreadable reply scores 0.

02on scifact_200

gold/scifact_200sf_glm_s_e5first_n200_v1 · pair: 0.789sf_qwen_s_e5first_n200_v1 · pair: 0.771sf_glm_s_n200_v1: 0.744sf_glm_s_bm25first_n200_v1 · pair: 0.742sf_glm_m_n200_v1: 0.742sf_deepseek_m_n200_v1: 0.74sf_deepseek_s_n200_v1: 0.738sf_gemma_m_n200_v1: 0.736sf_gemma_l_n200_v1: 0.736sf_llama_l_n200_v1: 0.728sf_e5_base_n200_v1: 0.719sf_qwen_s_n200_v1 · repeats: 0.719sf_qwen_s_n200_v1: 0.718sf_llama_s_n200_v1: 0.717sf_llama_l_n200_v1 · repeats: 0.715sf_qwen_s_bm25first_n200_v1 · pair: 0.714sf_llama_s_n200_v1 · repeats: 0.713sf_bge_small_n200_v1: 0.71sf_llama_m_n200_v1: 0.709sf_mistral_s_n200_v1: 0.707sf_gemma_s_n200_v1: 0.702sf_mpnet_n200_v1: 0.664sf_minilm_n200_v1: 0.653sf_bm25_n200_v1: 0.645sf_random_n200_v1: 0.00051sf_first_k_n200_v1: 00better · 0.789

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