Skip to content

Evidence-retrieval dev metric recall@10 is degenerate (<=4 candidates per query) #2

Description

@fsy2004

Observed on the 2026-08-17 server run: each dev query has 1 positive + up to 3 hard negatives (30,272 pairs), so ranking top-10 always includes every candidate and development_recall_at_10 is 1.0 for ANY ranker (trained model and TF-IDF both score 1.0). Meaningful alternatives computed by mw-baseline.py: trained MRR=0.824 / P@1=0.698 vs TF-IDF MRR=0.556 / P@1=0.316. Proposed fix: replace the runner metric with positive-vs-candidates rank metrics (MRR, P@1, recall@k over the full dev document corpus), and record both in the execution receipt.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions