Extend dataset bootstrap coverage and improve humming hard-case weighting

Broaden external dataset bootstrap support and replace naive hard-case oversampling with a more targeted weighting signal that measurably helps humming-like queries while preserving the release/eval workflow.

Constraint: Hard-case optimization must be evidence-driven and preserve a record of mixed outcomes across iterations
Rejected: Reuse naive oversampling after regression | it already showed worse overall behavior with no hard-case gain
Confidence: medium
Scope-risk: moderate
Directive: Next iteration should target confused-case negatives explicitly; do not assume humming gains transfer to confusion robustness
Tested: bootstrap generation for MTG-Jamendo and ModelScope placeholders; 2-epoch CPU training for models_v5; index_v5 build; fast eval JSON generation for smoke-v5
Not-tested: real audio ingestion for the new datasets; full melody-aware slow evaluation on models_v5

+# modelscope_music bootstrap
+
+- Fill raw audio files under `raw/`
+- Review license before training
+- Convert to final catalog/query manifests

+[
+  {
+    "song_id": "modelscope_music_track_0000",
+    "audio_path": "raw/modelscope_music_track_0000.wav",
+    "duration": 0.0,
+    "type": "reference",
+    "source_dataset": "modelscope_music",
+    "license_status": "deny_until_whitelisted"
+  },
+  {
+    "song_id": "modelscope_music_track_0001",
+    "audio_path": "raw/modelscope_music_track_0001.wav",
+    "duration": 0.0,
+    "type": "reference",
+    "source_dataset": "modelscope_music",
+    "license_status": "deny_until_whitelisted"
+  },
+  {
+    "song_id": "modelscope_music_track_0002",
+    "audio_path": "raw/modelscope_music_track_0002.wav",
+    "duration": 0.0,
+    "type": "reference",
+    "source_dataset": "modelscope_music",
+    "license_status": "deny_until_whitelisted"
+  }
+]
\ No newline at end of file

+# mtg_jamendo bootstrap
+
+- Fill raw audio files under `raw/`
+- Review license before training
+- Convert to final catalog/query manifests

+[
+  {
+    "song_id": "mtg_jamendo_track_0000",
+    "audio_path": "raw/mtg_jamendo_track_0000.wav",
+    "duration": 0.0,
+    "type": "reference",
+    "source_dataset": "mtg_jamendo",
+    "license_status": "review_required"
+  },
+  {
+    "song_id": "mtg_jamendo_track_0001",
+    "audio_path": "raw/mtg_jamendo_track_0001.wav",
+    "duration": 0.0,
+    "type": "reference",
+    "source_dataset": "mtg_jamendo",
+    "license_status": "review_required"
+  },
+  {
+    "song_id": "mtg_jamendo_track_0002",
+    "audio_path": "raw/mtg_jamendo_track_0002.wav",
+    "duration": 0.0,
+    "type": "reference",
+    "source_dataset": "mtg_jamendo",
+    "license_status": "review_required"
+  }
+]
\ No newline at end of file

+{
+  "song_0000": 0,
+  "song_0001": 1,
+  "song_0002": 2,
+  "song_0003": 3,
+  "song_0004": 4,
+  "song_0005": 5,
+  "song_0006": 6,
+  "song_0007": 7,
+  "song_0008": 8,
+  "song_0009": 9,
+  "song_0010": 10,
+  "song_0011": 11,
+  "song_0012": 12,
+  "song_0013": 13,
+  "song_0014": 14,
+  "song_0015": 15
+}
\ No newline at end of file

+{
+  "split": "test",
+  "num_queries": 20,
+  "top1": 0.6,
+  "topk": 0.9,
+  "by_type": {
+    "clean": {
+      "n": 8,
+      "top1": 1.0,
+      "topk": 1.0
+    },
+    "augmented": {
+      "n": 4,
+      "top1": 0.5,
+      "topk": 1.0
+    },
+    "humming_like": {
+      "n": 4,
+      "top1": 0.5,
+      "topk": 0.75
+    },
+    "confused": {
+      "n": 4,
+      "top1": 0.0,
+      "topk": 0.75
+    }
+  },
+  "hard_case_summary": {
+    "humming_like": {
+      "n": 4,
+      "top1": 0.5,
+      "topk": 0.75
+    },
+    "confused": {
+      "n": 4,
+      "top1": 0.0,
+      "topk": 0.75
+    }
+  },
+  "sample_failures": [
+    {
+      "truth": "song_0020",
+      "query": "segments/song_0020_seg_04_confused.wav",
+      "type": "confused",
+      "preds": [
+        "song_0002",
+        "song_0022",
+        "song_0006",
+        "song_0023",
+        "song_0001"
+      ]
+    },
+    {
+      "truth": "song_0022",
+      "query": "segments/song_0022_seg_03_humming_like.wav",
+      "type": "humming_like",
+      "preds": [
+        "song_0021",
+        "song_0001",
+        "song_0000",
+        "song_0003",
+        "song_0023"
+      ]
+    }
+  ]
+}
\ No newline at end of file

@@ -228,6 +228,9 @@ class SongPairDataset(Dataset):
         else:
         else:
             a, b = random.sample(choices, 2)
             a, b = random.sample(choices, 2)
 
 
+        pair_types = {a.get("type", "unknown"), b.get("type", "unknown")}
+        hard_weight = 2.5 if pair_types & {"confused", "humming_like"} else 1.0
+
         wavs = []
         wavs = []
         for sample in (a, b):
         for sample in (a, b):
             y = self._load_clip(sample)
             y = self._load_clip(sample)
@@ -244,4 +247,5 @@ class SongPairDataset(Dataset):
             "mel": torch.stack(wavs, dim=0),
             "mel": torch.stack(wavs, dim=0),
             "song_id": torch.tensor([label, label], dtype=torch.long),
             "song_id": torch.tensor([label, label], dtype=torch.long),
             "song_name": song_id,
             "song_name": song_id,
+            "hard_weight": torch.tensor(hard_weight, dtype=torch.float32),
         }
         }

@@ -48,9 +48,14 @@ class CombinedLoss(nn.Module):
         logits: torch.Tensor,
         logits: torch.Tensor,
         labels: torch.Tensor,
         labels: torch.Tensor,
         supcon_labels: torch.Tensor,
         supcon_labels: torch.Tensor,
+        hard_weight: torch.Tensor | None = None,
     ) -> dict:
     ) -> dict:
         loss_supcon = self.supcon(embedding, supcon_labels)
         loss_supcon = self.supcon(embedding, supcon_labels)
         loss_ce = self.ce(logits, labels)
         loss_ce = self.ce(logits, labels)
+        if hard_weight is not None:
+            weight = hard_weight.float().mean()
+            loss_supcon = loss_supcon * weight
+            loss_ce = loss_ce * weight
 
 
         total = self.supcon_weight * loss_supcon + self.aam_weight * loss_ce
         total = self.supcon_weight * loss_supcon + self.aam_weight * loss_ce
         return {
         return {

@@ -23,17 +23,21 @@ def collate_fn(batch):
     mels = []
     mels = []
     song_ids = []
     song_ids = []
     song_names = []
     song_names = []
+    hard_weights = []
     for b in batch:
     for b in batch:
         mel = b["mel"]
         mel = b["mel"]
+        hw = b.get("hard_weight", torch.tensor(1.0))
         if mel.dim() == 3:
         if mel.dim() == 3:
             for i in range(mel.shape[0]):
             for i in range(mel.shape[0]):
                 mels.append(mel[i])
                 mels.append(mel[i])
                 song_ids.append(b["song_id"][i])
                 song_ids.append(b["song_id"][i])
                 song_names.append(b["song_name"])
                 song_names.append(b["song_name"])
+                hard_weights.append(hw)
         else:
         else:
             mels.append(mel)
             mels.append(mel)
             song_ids.append(b["song_id"])
             song_ids.append(b["song_id"])
             song_names.append(b["song_name"])
             song_names.append(b["song_name"])
+            hard_weights.append(hw)
 
 
     max_t = max(m.shape[1] for m in mels)
     max_t = max(m.shape[1] for m in mels)
     mels_padded = []
     mels_padded = []
@@ -47,6 +51,7 @@ def collate_fn(batch):
         "mel": torch.cat(mels_padded, dim=0),
         "mel": torch.cat(mels_padded, dim=0),
         "song_id": torch.stack(song_ids),
         "song_id": torch.stack(song_ids),
         "song_name": song_names,
         "song_name": song_names,
+        "hard_weight": torch.stack(hard_weights),
     }
     }
 
 
 
 
@@ -60,7 +65,7 @@ def train_epoch(model, loader, optimizer, criterion, scaler, device, epoch, cfg)
 
 
         with torch.amp.autocast("cuda", enabled=cfg["training"]["mixed_precision"] and device.type == "cuda"):
         with torch.amp.autocast("cuda", enabled=cfg["training"]["mixed_precision"] and device.type == "cuda"):
             embedding, logits = model(mel, labels)
             embedding, logits = model(mel, labels)
-            loss_dict = criterion(embedding, logits, labels, labels)
+            loss_dict = criterion(embedding, logits, labels, labels, batch.get("hard_weight", None).to(device) if "hard_weight" in batch else None)
 
 
         optimizer.zero_grad()
         optimizer.zero_grad()
         if scaler:
         if scaler:
@@ -205,7 +210,7 @@ def main():
         mel = batch["mel"].to(device)
         mel = batch["mel"].to(device)
         labels = batch["song_id"].to(device)
         labels = batch["song_id"].to(device)
         embedding, logits = model(mel, labels)
         embedding, logits = model(mel, labels)
-        loss_dict = criterion(embedding, logits, labels, labels)
+        loss_dict = criterion(embedding, logits, labels, labels, batch.get("hard_weight", None).to(device) if "hard_weight" in batch else None)
         loss_dict["loss"].backward()
         loss_dict["loss"].backward()
         print(f"  Forward/backward OK. Loss: {loss_dict['loss']:.4f}")
         print(f"  Forward/backward OK. Loss: {loss_dict['loss']:.4f}")
         print(f"  Embedding shape: {embedding.shape}")
         print(f"  Embedding shape: {embedding.shape}")

@@ -136,3 +136,25 @@
 结论：
 结论：
 - 该轮简单过采样策略无效，且整体精度下降
 - 该轮简单过采样策略无效，且整体精度下降
 - 下一轮应改用更细粒度 hard-negative / melody-aware 正则，而不是继续放大样本重复权重
 - 下一轮应改用更细粒度 hard-negative / melody-aware 正则，而不是继续放大样本重复权重
+
+## 2026-06-02
+
+### Stage: MTG-Jamendo / ModelScope bootstrap + type-aware hard-case weighting
+
+完成项：
+- 补充 `mtg_jamendo` 与 `modelscope_music` 的 bootstrap manifest 生成
+- 在训练链路中加入 type-aware hard-case weighting（针对 `confused` / `humming_like`）
+- 重训 `models_v5`、重建 `index_v5`、重跑 `smoke-v5` 评测
+
+验证结果：
+- `data/external_bootstrap/mtg_jamendo/manifests/catalog.bootstrap.json` 成功生成
+- `data/external_bootstrap/modelscope_music/manifests/catalog.bootstrap.json` 成功生成
+- `reports/smoke-v5/synthetic_v2/eval.json` 成功生成
+- 当前结果：top1=0.60, top5=0.90
+- hard-case 结果：
+  - humming_like top1=0.50（较 v4 有提升）
+  - confused top1=0.00（仍未解决）
+
+结论：
+- type-aware weighting 比 naive oversampling 更有效
+- 下一轮应专门针对 confused 类设计更强的 negative mining / confusion-aware 信号