diff --git a/assets/causebench-recall-at-1.svg b/assets/causebench-recall-at-1.svg deleted file mode 100644 index bc55be9..0000000 --- a/assets/causebench-recall-at-1.svg +++ /dev/null @@ -1,46 +0,0 @@ - - CauseBench root-cause recall@1 — Vestige 60% / 50%, every resemblance retriever 0% - - Root-cause recall@1 — can the retriever rank the true cause #1? - CauseBench · a cause that shares an entity with the failure but none of its words - - - - - - - 0% - 20% - 40% - 60% - - Vestige - - - 60% synthetic - 50% real - - Dense vector (nomic) - - 0% - - Hybrid-RRF (BM25 + dense) - - 0% - - BM25 - - 0% - - TF-IDF - - 0% - - Recency - - 0% - - - Every resemblance retriever ranks the true cause #1 exactly 0% of the time. - Vestige also lands the cause in the top 5 in 100% of cases. Source: benchmarks/causebench/RESULTS.md · reproduce with 3 commands. -