更新同歌手加同时长判断

@@ -33,7 +33,7 @@ python scripts/aliyun_dna/evaluate_aliyun_dna.py \
       --queries downloads_db/queries.csv \
       --queries downloads_db/queries.csv \
       --out results/dna_eval.csv \
       --out results/dna_eval.csv \
       --db-id f08aee1806c14df782e03cc1ab93f1ea \
       --db-id f08aee1806c14df782e03cc1ab93f1ea \
-      --concurrency 4
+      --concurrency 8
 """
 """
 
 
 import argparse
 import argparse

@@ -15,10 +15,10 @@
     python scripts/download_from_db.py --song-ids 1,2,3 --extra-versions -1
     python scripts/download_from_db.py --song-ids 1,2,3 --extra-versions -1
 
 
     # 本地：导出查询结果到 CSV（不下载）
     # 本地：导出查询结果到 CSV（不下载）
-    python scripts/download_from_db.py --song-limit 500 --extra-versions 3 --export-csv records.csv
+    python scripts/download_from_db.py --song-limit  --extra-versions 3 --export-csv records_test.csv
 
 
     # 服务器：从导出的 CSV 下载（不需要连接数据库）
     # 服务器：从导出的 CSV 下载（不需要连接数据库）
-    python scripts/download_from_db.py --from-csv records.csv --concurrency 8
+    python scripts/download_from_db.py --from-csv records_test.csv --concurrency 8
 
 
 输出目录结构:
 输出目录结构:
     {output_dir}/
     {output_dir}/
@@ -477,6 +477,15 @@ def main() -> None:
     # 生成 reference.csv 和 queries.csv
     # 生成 reference.csv 和 queries.csv
     ok_results = [r for r in results if r.get("status") == "ok" and r.get("local_path")]
     ok_results = [r for r in results if r.get("status") == "ok" and r.get("local_path")]
 
 
+    # 先建 song_id → 主版本时长 映射，用于过滤负样本
+    main_duration: dict[str, float] = {}
+    for r in ok_results:
+        if int(r["is_main_version"]) == 1 and r.get("duration"):
+            try:
+                main_duration[str(r["song_id"])] = float(r["duration"])
+            except (ValueError, TypeError):
+                pass
+
     ref_rows, query_rows = [], []
     ref_rows, query_rows = [], []
     variants_dir = output_dir / "variants"
     variants_dir = output_dir / "variants"
 
 
@@ -528,9 +537,20 @@ def main() -> None:
                             continue
                             continue
                     query_rows.append(_query_row(dst, vname))
                     query_rows.append(_query_row(dst, vname))
         else:
         else:
-            # 负样本：翻唱/其他版本，不应匹配到 DNA 库中的主版本
+            # 负样本：同歌手 + 时长相近 → 很可能是同一母带的不同分发，跳过
+            singer = r["singer_name"] or ""
+            original = r["original_singer"] or ""
+            if singer and original and singer == original:
+                ref_dur = main_duration.get(str(r["song_id"]))
+                try:
+                    cover_dur = float(r["duration"]) if r.get("duration") else None
+                except (ValueError, TypeError):
+                    cover_dur = None
+                if ref_dur and cover_dur and abs(cover_dur - ref_dur) / ref_dur < 0.05:
+                    logger.debug("跳过同歌手且时长相近的版本（非负样本）: song_id=%s singer=%s dur=%.0f/%.0f",
+                                 r["song_id"], singer, cover_dur, ref_dur)
+                    continue
             platform = r["platform_name"] or "unknown"
             platform = r["platform_name"] or "unknown"
-            singer = r["singer_name"] or "unknown"
             query_rows.append({
             query_rows.append({
                 "song_id": r["song_id"],
                 "song_id": r["song_id"],
                 "audio_path": r["local_path"],
                 "audio_path": r["local_path"],