Loading leaderboard...
{"help-bf443501182decc7":{"title":"HAKARI-Bench","summary":"Every row is one retrieval model, and every score is how well it ranked relevant documents across hundreds of small retrieval tasks.","details":"HAKARI-Bench measures information retrieval: given a query, does the model put the relevant documents at the top? Each task is a Nano-set, a compact rebuild of a public retrieval benchmark with roughly 50 to 200 queries and a corpus capped at about 10K documents, which keeps every positive document plus hard negatives from the original data. Nano-sets are small enough that every model can be re-run on all of them, which is what makes the same-condition comparison on this page possible.\n\nScores are nDCG@10 multiplied by 100 unless you change Metric. Roughly: 100 means the relevant documents were placed perfectly at the top of the first 10 results, and 0 means nothing relevant was found. Borda Score, the default sort, is a ranking-based score instead of an average score; Macro Mean and Micro Mean are two ways of averaging the same per-task numbers. Each of those column headers has its own help icon.\n\nThe controls above the table apply from top to bottom. Evaluation mode picks full-corpus retrieval or reranking of a fixed candidate list. Score and Metric decide how task scores are averaged and which metric is averaged. Benchmark scope picks which benchmark families count. Task facets narrows those tasks by language or category. Table display adds per-task columns, Efficiency variants adds compressed or truncated embedding rows, and Filter results only hides rows from the table you already have.\n\nOnly models that completed every task in the current scope are ranked, so the table always compares like with like; the counts under the controls show how many models and tasks that is. Click a model name for its run metadata, click the book icon beside a benchmark for its documentation, switch to Chart to plot quality against size, and use Download CSV to take the visible table with you.","eyebrow":"Getting started"}}
