← eval history

Qwen/Qwen3-Embedding-8B + deepseek-chat · k=20

run #28 · 2026-06-25 18:01 · git 16da4bb · deploy:q8-sf-xdocfix

Retrieval hit@k
98%
Answer accuracy
98%
Cases
45 (40 grounded)

qwen8b + Family3 + per-sub-query rerank for cross-doc

Δ vs baseline

Compared to Qwen/Qwen3-Embedding-8B + deepseek-chat · k=20 (2026-06-25 17:44). A retrieval delta with the same generator isolates the embedder/retrieval change.

Failures in this run (2)

CaseQuestionRetrievalAnswerSymptom
q024How much larger were Berkshire Hathaway's 2023 net earnings than its 2023 operating earnings?PASSFAILretrieved but generated wrong (generation/arithmetic)
q043Which had the highest net income in 2023: JPMorgan Chase, Microsoft, or Costco?FAILPASSanswered without the gold chunk (parametric leak?)