{
  "id": 583434,
  "title": "BirdCLEF 2025 in a Nutshell: Winning Recipes on One Table",
  "url": "/competitions/birdclef-2025/discussion/583434",
  "author_name": "",
  "post_date": "2025-06-06T20:12:33.499521900Z",
  "votes": 4,
  "comment_count": 1,
  "views": 0,
  "content": "<table>\n<thead>\n<tr>\n<th>Rank</th>\n<th>Public / Private LB</th>\n<th>Core Models (backbone × count)</th>\n<th><strong>Spectrogram Params</strong><br>(mel / freq / FFT / hop)</th>\n<th>Data &amp; Label Strategy</th>\n<th>Secret Sauce</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td><strong>1</strong></td>\n<td>0.933 / 0.930</td>\n<td>SED + CNN – EffNetB0/B3/B4, RegNetY, ECA-NFNet</td>\n<td>224 / 0–16kHz / 4096 / 1252</td>\n<td>2025 + Xeno-Canto (target &amp; extra species); multi-iterative pseudo-labeling (up to 4 rounds); Amphibia/Insecta-specific model</td>\n<td>Multi-iterative Noisy Student; Mixup+Stochastic Depth; custom ensembling of 7 models; OpenVINO</td>\n</tr>\n<tr>\n<td><strong>2</strong></td>\n<td>0.922 / 0.918</td>\n<td>SED + CNN – EfficientNetV2-s, ECA-NFNet</td>\n<td>224 / 0–16kHz / 4096 / 1024</td>\n<td>2025 + Xeno-Canto (pretraining excl. current species); 3-stage pseudo-labeling with filtered soft labels; no iNat used</td>\n<td>Pretrained backbones; soft pseudo-labels (&gt;0.5), multi-iteration pseudo mix; postprocessing via chunk top-prob scaling; no TTA in final</td>\n</tr>\n<tr>\n<td><strong>3</strong></td>\n<td>0.927 / 0.927</td>\n<td>CNN + SED – EffNet-B0-NS, V2-B3, V2-S, MNASNet, SPNASNet</td>\n<td>96–128 / 0–16kHz / 2048 / 512</td>\n<td>Used 2025 + 80% of BirdCLEF 2023 + Xeno-Canto + iNaturalist + pseudo-labeled soundscapes</td>\n<td>Rank-aware postproc + model soup + ONNX; ensemble of 20 models (10 CNN + 10 SED)</td>\n</tr>\n<tr>\n<td><strong>5</strong></td>\n<td><strong>0.928 / 0.924</strong></td>\n<td>SED – EffNet-V2-S × 4, V2-B3 × 3, B3-NS × 4, B0-NS × 2</td>\n<td>192 / 20–15kHz / 2048 / 768</td>\n<td>Self-distillation 5× on train_audio, then 2× with train_soundscapes (1:1 mix)</td>\n<td>Iterative KD on unlabeled bird calls + 2.5s overlap + OpenVINO</td>\n</tr>\n<tr>\n<td><strong>6</strong></td>\n<td>0.928 / 0.923</td>\n<td>SED – EffNet-B3-JFT, B5-JFT, V2-B3</td>\n<td><em>Not specified</em></td>\n<td>Multi-round PL cycles with segment logits</td>\n<td>Dual-head BCE loss (clip + segment) despite sigmoid mismatch</td>\n</tr>\n<tr>\n<td><strong>8</strong></td>\n<td>0.924 / 0.924</td>\n<td>2× SED (V2-B3 &amp; ECA-NFNet-L0) + CNN (SeResNeXt-26T)</td>\n<td>192 / 40–140kHz / 2048 / 512</td>\n<td>Pretrained on 2021–2024 data, PL (thr=0.4), 2-way KD</td>\n<td>FilterAug + hard mixup + weighted BCE; dropped postproc trick at end</td>\n</tr>\n<tr>\n<td><strong>9</strong></td>\n<td>0.913 / 0.925</td>\n<td>SED + CNN – V2-B3, V2-S, SeResNeXt-26T, B0</td>\n<td>256 / 20–16kHz / 2048 / ~400</td>\n<td>Two-stage: SED/CNN → PL → retrain</td>\n<td>ONNX + single-thread workaround avoided timeout bug</td>\n</tr>\n<tr>\n<td><strong>10</strong></td>\n<td>0.915 / 0.921</td>\n<td>3× ConvNeXt-T + SED (V2-S/B3)</td>\n<td>320–384 / 0–16kHz / 1536–2048 / varied</td>\n<td>Confident pos/neg filtering → pp_data_clean</td>\n<td>Hybrid CE + negative-logit penalty; α-blend recall trick</td>\n</tr>\n<tr>\n<td><strong>11</strong></td>\n<td>0.920 / 0.919</td>\n<td>5× V2-B3 + 1× V2-S</td>\n<td>256 / 50–14kHz / 1024–1536 / 535</td>\n<td>Entropy-top 20% PL + CE loss</td>\n<td>Postproc kernel [0.1, 0.4, 0.1]; CE boosted perf by 0.05</td>\n</tr>\n<tr>\n<td><strong>13</strong></td>\n<td>0.901 / 0.90*</td>\n<td>SED – tf_efficientnetv2_s, tf_efficientnetv2_b3, SEResNeXt26t</td>\n<td>N/A / 0–16kHz / N/A / 512</td>\n<td>2025 + train_soundscapes pseudo-labeled; rare species model (+0.01); cleaned audio, voice removal, duration recalculated</td>\n<td>Stepwise 4-stage strategy; Sumix, no TTA; raw audio mixup; rare-species submodel ensembling; minimal boost from backbone diversity</td>\n</tr>\n<tr>\n<td><strong>14</strong></td>\n<td>0.922 / n/a</td>\n<td>3× V2-M (in21k)</td>\n<td>128 / 40–15kHz / 1024 / 512</td>\n<td>Iterative KD (1s avg soft 0.3 + 10s chunk soft 0.7)</td>\n<td>Pure SED + OpenVINO pipeline for speed</td>\n</tr>\n<tr>\n<td><strong>26</strong></td>\n<td>0.909 / 0.908</td>\n<td>4× CNN + 1× SED (EffNet family)</td>\n<td>384–448 / 0–16kHz / 3072–4096 / ~370</td>\n<td>Stage-2 PL from confident 5s chunks</td>\n<td>“15s RMS → 3×5s max-pool” trick + model soup</td>\n</tr>\n<tr>\n<td><strong>28</strong></td>\n<td>0.893 / 0.909</td>\n<td>5-backbone SED (SeResNeXt-26T heavy)</td>\n<td><em>Not specified</em></td>\n<td>Segment-based voice removal + progressive PL (0.2→0.5)</td>\n<td>Tuned smoothing [0.2, 0.6, 0.2]; large batch size (128) helped</td>\n</tr>\n<tr>\n<td><strong>29</strong></td>\n<td>0.902 / 0.906</td>\n<td>EfficientNet-B3 • ResNeSt-50 • NFNet-L0</td>\n<td>256 / not stated / ? / ?</td>\n<td>Trained on all data; mixup &amp; segment variety (10–30s); infer on 5s</td>\n<td>Multi-stage SpecAugment + 3-model blend + ONNX for speed</td>\n</tr>\n</tbody>\n</table>\n<h2>What's next - What to Take into BirdCLEF 2026 : <a href=\"https://www.kaggle.com/competitions/birdclef-2025/discussion/583437\" target=\"_blank\">https://www.kaggle.com/competitions/birdclef-2025/discussion/583437</a></h2>",
  "messages": [
    {
      "id": "3218859",
      "postDate": "06/06/2025 20:12:33",
      "content": "<table>\n<thead>\n<tr>\n<th>Rank</th>\n<th>Public / Private LB</th>\n<th>Core Models (backbone × count)</th>\n<th><strong>Spectrogram Params</strong><br>(mel / freq / FFT / hop)</th>\n<th>Data &amp; Label Strategy</th>\n<th>Secret Sauce</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td><strong>1</strong></td>\n<td>0.933 / 0.930</td>\n<td>SED + CNN – EffNetB0/B3/B4, RegNetY, ECA-NFNet</td>\n<td>224 / 0–16kHz / 4096 / 1252</td>\n<td>2025 + Xeno-Canto (target &amp; extra species); multi-iterative pseudo-labeling (up to 4 rounds); Amphibia/Insecta-specific model</td>\n<td>Multi-iterative Noisy Student; Mixup+Stochastic Depth; custom ensembling of 7 models; OpenVINO</td>\n</tr>\n<tr>\n<td><strong>2</strong></td>\n<td>0.922 / 0.918</td>\n<td>SED + CNN – EfficientNetV2-s, ECA-NFNet</td>\n<td>224 / 0–16kHz / 4096 / 1024</td>\n<td>2025 + Xeno-Canto (pretraining excl. current species); 3-stage pseudo-labeling with filtered soft labels; no iNat used</td>\n<td>Pretrained backbones; soft pseudo-labels (&gt;0.5), multi-iteration pseudo mix; postprocessing via chunk top-prob scaling; no TTA in final</td>\n</tr>\n<tr>\n<td><strong>3</strong></td>\n<td>0.927 / 0.927</td>\n<td>CNN + SED – EffNet-B0-NS, V2-B3, V2-S, MNASNet, SPNASNet</td>\n<td>96–128 / 0–16kHz / 2048 / 512</td>\n<td>Used 2025 + 80% of BirdCLEF 2023 + Xeno-Canto + iNaturalist + pseudo-labeled soundscapes</td>\n<td>Rank-aware postproc + model soup + ONNX; ensemble of 20 models (10 CNN + 10 SED)</td>\n</tr>\n<tr>\n<td><strong>5</strong></td>\n<td><strong>0.928 / 0.924</strong></td>\n<td>SED – EffNet-V2-S × 4, V2-B3 × 3, B3-NS × 4, B0-NS × 2</td>\n<td>192 / 20–15kHz / 2048 / 768</td>\n<td>Self-distillation 5× on train_audio, then 2× with train_soundscapes (1:1 mix)</td>\n<td>Iterative KD on unlabeled bird calls + 2.5s overlap + OpenVINO</td>\n</tr>\n<tr>\n<td><strong>6</strong></td>\n<td>0.928 / 0.923</td>\n<td>SED – EffNet-B3-JFT, B5-JFT, V2-B3</td>\n<td><em>Not specified</em></td>\n<td>Multi-round PL cycles with segment logits</td>\n<td>Dual-head BCE loss (clip + segment) despite sigmoid mismatch</td>\n</tr>\n<tr>\n<td><strong>8</strong></td>\n<td>0.924 / 0.924</td>\n<td>2× SED (V2-B3 &amp; ECA-NFNet-L0) + CNN (SeResNeXt-26T)</td>\n<td>192 / 40–140kHz / 2048 / 512</td>\n<td>Pretrained on 2021–2024 data, PL (thr=0.4), 2-way KD</td>\n<td>FilterAug + hard mixup + weighted BCE; dropped postproc trick at end</td>\n</tr>\n<tr>\n<td><strong>9</strong></td>\n<td>0.913 / 0.925</td>\n<td>SED + CNN – V2-B3, V2-S, SeResNeXt-26T, B0</td>\n<td>256 / 20–16kHz / 2048 / ~400</td>\n<td>Two-stage: SED/CNN → PL → retrain</td>\n<td>ONNX + single-thread workaround avoided timeout bug</td>\n</tr>\n<tr>\n<td><strong>10</strong></td>\n<td>0.915 / 0.921</td>\n<td>3× ConvNeXt-T + SED (V2-S/B3)</td>\n<td>320–384 / 0–16kHz / 1536–2048 / varied</td>\n<td>Confident pos/neg filtering → pp_data_clean</td>\n<td>Hybrid CE + negative-logit penalty; α-blend recall trick</td>\n</tr>\n<tr>\n<td><strong>11</strong></td>\n<td>0.920 / 0.919</td>\n<td>5× V2-B3 + 1× V2-S</td>\n<td>256 / 50–14kHz / 1024–1536 / 535</td>\n<td>Entropy-top 20% PL + CE loss</td>\n<td>Postproc kernel [0.1, 0.4, 0.1]; CE boosted perf by 0.05</td>\n</tr>\n<tr>\n<td><strong>13</strong></td>\n<td>0.901 / 0.90*</td>\n<td>SED – tf_efficientnetv2_s, tf_efficientnetv2_b3, SEResNeXt26t</td>\n<td>N/A / 0–16kHz / N/A / 512</td>\n<td>2025 + train_soundscapes pseudo-labeled; rare species model (+0.01); cleaned audio, voice removal, duration recalculated</td>\n<td>Stepwise 4-stage strategy; Sumix, no TTA; raw audio mixup; rare-species submodel ensembling; minimal boost from backbone diversity</td>\n</tr>\n<tr>\n<td><strong>14</strong></td>\n<td>0.922 / n/a</td>\n<td>3× V2-M (in21k)</td>\n<td>128 / 40–15kHz / 1024 / 512</td>\n<td>Iterative KD (1s avg soft 0.3 + 10s chunk soft 0.7)</td>\n<td>Pure SED + OpenVINO pipeline for speed</td>\n</tr>\n<tr>\n<td><strong>26</strong></td>\n<td>0.909 / 0.908</td>\n<td>4× CNN + 1× SED (EffNet family)</td>\n<td>384–448 / 0–16kHz / 3072–4096 / ~370</td>\n<td>Stage-2 PL from confident 5s chunks</td>\n<td>“15s RMS → 3×5s max-pool” trick + model soup</td>\n</tr>\n<tr>\n<td><strong>28</strong></td>\n<td>0.893 / 0.909</td>\n<td>5-backbone SED (SeResNeXt-26T heavy)</td>\n<td><em>Not specified</em></td>\n<td>Segment-based voice removal + progressive PL (0.2→0.5)</td>\n<td>Tuned smoothing [0.2, 0.6, 0.2]; large batch size (128) helped</td>\n</tr>\n<tr>\n<td><strong>29</strong></td>\n<td>0.902 / 0.906</td>\n<td>EfficientNet-B3 • ResNeSt-50 • NFNet-L0</td>\n<td>256 / not stated / ? / ?</td>\n<td>Trained on all data; mixup &amp; segment variety (10–30s); infer on 5s</td>\n<td>Multi-stage SpecAugment + 3-model blend + ONNX for speed</td>\n</tr>\n</tbody>\n</table>\n<h2>What's next - What to Take into BirdCLEF 2026 : <a href=\"https://www.kaggle.com/competitions/birdclef-2025/discussion/583437\" target=\"_blank\">https://www.kaggle.com/competitions/birdclef-2025/discussion/583437</a></h2>",
      "rawMarkdown": "| Rank | Public / Private LB | Core Models (backbone × count) | **Spectrogram Params**<br>(mel / freq / FFT / hop) | Data & Label Strategy | Secret Sauce |\n|------|----------------------|-------------------------------|----------------------------------------------------|------------------------|----------------|\n| **1** | 0.933 / 0.930 | SED + CNN – EffNetB0/B3/B4, RegNetY, ECA-NFNet | 224 / 0–16kHz / 4096 / 1252 | 2025 + Xeno-Canto (target & extra species); multi-iterative pseudo-labeling (up to 4 rounds); Amphibia/Insecta-specific model | Multi-iterative Noisy Student; Mixup+Stochastic Depth; custom ensembling of 7 models; OpenVINO |\n| **2** | 0.922 / 0.918 | SED + CNN – EfficientNetV2-s, ECA-NFNet | 224 / 0–16kHz / 4096 / 1024 | 2025 + Xeno-Canto (pretraining excl. current species); 3-stage pseudo-labeling with filtered soft labels; no iNat used | Pretrained backbones; soft pseudo-labels (>0.5), multi-iteration pseudo mix; postprocessing via chunk top-prob scaling; no TTA in final |\n| **3** | 0.927 / 0.927 | CNN + SED – EffNet-B0-NS, V2-B3, V2-S, MNASNet, SPNASNet | 96–128 / 0–16kHz / 2048 / 512 | Used 2025 + 80% of BirdCLEF 2023 + Xeno-Canto + iNaturalist + pseudo-labeled soundscapes | Rank-aware postproc + model soup + ONNX; ensemble of 20 models (10 CNN + 10 SED) |\n| **5** | **0.928 / 0.924** | SED – EffNet-V2-S × 4, V2-B3 × 3, B3-NS × 4, B0-NS × 2 | 192 / 20–15kHz / 2048 / 768 | Self-distillation 5× on train_audio, then 2× with train_soundscapes (1:1 mix) | Iterative KD on unlabeled bird calls + 2.5s overlap + OpenVINO |\n| **6** | 0.928 / 0.923 | SED – EffNet-B3-JFT, B5-JFT, V2-B3 | _Not specified_ | Multi-round PL cycles with segment logits | Dual-head BCE loss (clip + segment) despite sigmoid mismatch |\n| **8** | 0.924 / 0.924 | 2× SED (V2-B3 & ECA-NFNet-L0) + CNN (SeResNeXt-26T) | 192 / 40–140kHz / 2048 / 512 | Pretrained on 2021–2024 data, PL (thr=0.4), 2-way KD | FilterAug + hard mixup + weighted BCE; dropped postproc trick at end |\n| **9** | 0.913 / 0.925 | SED + CNN – V2-B3, V2-S, SeResNeXt-26T, B0 | 256 / 20–16kHz / 2048 / ~400 | Two-stage: SED/CNN → PL → retrain | ONNX + single-thread workaround avoided timeout bug |\n| **10** | 0.915 / 0.921 | 3× ConvNeXt-T + SED (V2-S/B3) | 320–384 / 0–16kHz / 1536–2048 / varied | Confident pos/neg filtering → pp_data_clean | Hybrid CE + negative-logit penalty; α-blend recall trick |\n| **11** | 0.920 / 0.919 | 5× V2-B3 + 1× V2-S | 256 / 50–14kHz / 1024–1536 / 535 | Entropy-top 20% PL + CE loss | Postproc kernel [0.1, 0.4, 0.1]; CE boosted perf by 0.05 |\n| **13** | 0.901 / 0.90* | SED – tf_efficientnetv2_s, tf_efficientnetv2_b3, SEResNeXt26t | N/A / 0–16kHz / N/A / 512 | 2025 + train_soundscapes pseudo-labeled; rare species model (+0.01); cleaned audio, voice removal, duration recalculated | Stepwise 4-stage strategy; Sumix, no TTA; raw audio mixup; rare-species submodel ensembling; minimal boost from backbone diversity |\n| **14** | 0.922 / n/a | 3× V2-M (in21k) | 128 / 40–15kHz / 1024 / 512 | Iterative KD (1s avg soft 0.3 + 10s chunk soft 0.7) | Pure SED + OpenVINO pipeline for speed |\n| **26** | 0.909 / 0.908 | 4× CNN + 1× SED (EffNet family) | 384–448 / 0–16kHz / 3072–4096 / ~370 | Stage-2 PL from confident 5s chunks | “15s RMS → 3×5s max-pool” trick + model soup |\n| **28** | 0.893 / 0.909 | 5-backbone SED (SeResNeXt-26T heavy) | _Not specified_ | Segment-based voice removal + progressive PL (0.2→0.5) | Tuned smoothing [0.2, 0.6, 0.2]; large batch size (128) helped |\n| **29** | 0.902 / 0.906 | EfficientNet-B3 • ResNeSt-50 • NFNet-L0 | 256 / not stated / ? / ? | Trained on all data; mixup & segment variety (10–30s); infer on 5s | Multi-stage SpecAugment + 3-model blend + ONNX for speed |\n\nWhat's next - What to Take into BirdCLEF 2026 : https://www.kaggle.com/competitions/birdclef-2025/discussion/583437\n----",
      "votes": null
    },
    {
      "id": "3218864",
      "postDate": "06/06/2025 20:17:03",
      "content": "<p>Congratulations to all the winners.</p>\n<p><strong>References</strong></p>\n<ol>\n<li>26th - <a href=\"https://www.kaggle.com/competitions/birdclef-2025/discussion/583306\" target=\"_blank\">https://www.kaggle.com/competitions/birdclef-2025/discussion/583306</a> </li>\n<li>9th - <a href=\"https://www.kaggle.com/competitions/birdclef-2025/discussion/583365\" target=\"_blank\">https://www.kaggle.com/competitions/birdclef-2025/discussion/583365</a></li>\n<li>8th - <a href=\"https://www.kaggle.com/competitions/birdclef-2025/discussion/583324\" target=\"_blank\">https://www.kaggle.com/competitions/birdclef-2025/discussion/583324</a></li>\n<li>14th - <a href=\"https://www.kaggle.com/competitions/birdclef-2025/discussion/583344\" target=\"_blank\">https://www.kaggle.com/competitions/birdclef-2025/discussion/583344</a></li>\n<li>6th - <a href=\"https://www.kaggle.com/competitions/birdclef-2025/discussion/583381\" target=\"_blank\">https://www.kaggle.com/competitions/birdclef-2025/discussion/583381</a></li>\n<li>11th - <a href=\"https://www.kaggle.com/competitions/birdclef-2025/discussion/583384\" target=\"_blank\">https://www.kaggle.com/competitions/birdclef-2025/discussion/583384</a></li>\n<li>28th - <a href=\"https://www.kaggle.com/competitions/birdclef-2025/discussion/583377\" target=\"_blank\">https://www.kaggle.com/competitions/birdclef-2025/discussion/583377</a></li>\n<li>10th - <a href=\"https://www.kaggle.com/competitions/birdclef-2025/discussion/583310\" target=\"_blank\">https://www.kaggle.com/competitions/birdclef-2025/discussion/583310</a></li>\n<li>5th - <a href=\"https://www.kaggle.com/competitions/birdclef-2025/discussion/583312\" target=\"_blank\">https://www.kaggle.com/competitions/birdclef-2025/discussion/583312</a></li>\n<li>29th - <a href=\"https://www.kaggle.com/competitions/birdclef-2025/discussion/583387\" target=\"_blank\">https://www.kaggle.com/competitions/birdclef-2025/discussion/583387</a></li>\n<li>3rd - <a href=\"https://www.kaggle.com/competitions/birdclef-2025/discussion/583477\" target=\"_blank\">https://www.kaggle.com/competitions/birdclef-2025/discussion/583477</a></li>\n<li>1st - <a href=\"https://www.kaggle.com/competitions/birdclef-2025/discussion/583577\" target=\"_blank\">https://www.kaggle.com/competitions/birdclef-2025/discussion/583577</a></li>\n<li>2nd - <a href=\"https://www.kaggle.com/competitions/birdclef-2025/discussion/583699\" target=\"_blank\">https://www.kaggle.com/competitions/birdclef-2025/discussion/583699</a></li>\n<li>13th - <a href=\"https://www.kaggle.com/competitions/birdclef-2025/discussion/583457\" target=\"_blank\">https://www.kaggle.com/competitions/birdclef-2025/discussion/583457</a></li>\n</ol>",
      "rawMarkdown": "Congratulations to all the winners.\n\n**References**\n1. 26th - https://www.kaggle.com/competitions/birdclef-2025/discussion/583306 \n2. 9th - https://www.kaggle.com/competitions/birdclef-2025/discussion/583365\n3. 8th - https://www.kaggle.com/competitions/birdclef-2025/discussion/583324\n4. 14th - https://www.kaggle.com/competitions/birdclef-2025/discussion/583344\n5. 6th - https://www.kaggle.com/competitions/birdclef-2025/discussion/583381\n6. 11th - https://www.kaggle.com/competitions/birdclef-2025/discussion/583384\n7. 28th - https://www.kaggle.com/competitions/birdclef-2025/discussion/583377\n8. 10th - https://www.kaggle.com/competitions/birdclef-2025/discussion/583310\n9. 5th - https://www.kaggle.com/competitions/birdclef-2025/discussion/583312\n10. 29th - https://www.kaggle.com/competitions/birdclef-2025/discussion/583387\n11. 3rd - https://www.kaggle.com/competitions/birdclef-2025/discussion/583477\n12. 1st - https://www.kaggle.com/competitions/birdclef-2025/discussion/583577\n13. 2nd - https://www.kaggle.com/competitions/birdclef-2025/discussion/583699\n14. 13th - https://www.kaggle.com/competitions/birdclef-2025/discussion/583457",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 3218864,
      "author_name": "aayush26",
      "author_url": "",
      "post_date": "06/06/2025 20:17:03",
      "content": "<p>Congratulations to all the winners.</p>\n<p><strong>References</strong></p>\n<ol>\n<li>26th - <a href=\"https://www.kaggle.com/competitions/birdclef-2025/discussion/583306\" target=\"_blank\">https://www.kaggle.com/competitions/birdclef-2025/discussion/583306</a> </li>\n<li>9th - <a href=\"https://www.kaggle.com/competitions/birdclef-2025/discussion/583365\" target=\"_blank\">https://www.kaggle.com/competitions/birdclef-2025/discussion/583365</a></li>\n<li>8th - <a href=\"https://www.kaggle.com/competitions/birdclef-2025/discussion/583324\" target=\"_blank\">https://www.kaggle.com/competitions/birdclef-2025/discussion/583324</a></li>\n<li>14th - <a href=\"https://www.kaggle.com/competitions/birdclef-2025/discussion/583344\" target=\"_blank\">https://www.kaggle.com/competitions/birdclef-2025/discussion/583344</a></li>\n<li>6th - <a href=\"https://www.kaggle.com/competitions/birdclef-2025/discussion/583381\" target=\"_blank\">https://www.kaggle.com/competitions/birdclef-2025/discussion/583381</a></li>\n<li>11th - <a href=\"https://www.kaggle.com/competitions/birdclef-2025/discussion/583384\" target=\"_blank\">https://www.kaggle.com/competitions/birdclef-2025/discussion/583384</a></li>\n<li>28th - <a href=\"https://www.kaggle.com/competitions/birdclef-2025/discussion/583377\" target=\"_blank\">https://www.kaggle.com/competitions/birdclef-2025/discussion/583377</a></li>\n<li>10th - <a href=\"https://www.kaggle.com/competitions/birdclef-2025/discussion/583310\" target=\"_blank\">https://www.kaggle.com/competitions/birdclef-2025/discussion/583310</a></li>\n<li>5th - <a href=\"https://www.kaggle.com/competitions/birdclef-2025/discussion/583312\" target=\"_blank\">https://www.kaggle.com/competitions/birdclef-2025/discussion/583312</a></li>\n<li>29th - <a href=\"https://www.kaggle.com/competitions/birdclef-2025/discussion/583387\" target=\"_blank\">https://www.kaggle.com/competitions/birdclef-2025/discussion/583387</a></li>\n<li>3rd - <a href=\"https://www.kaggle.com/competitions/birdclef-2025/discussion/583477\" target=\"_blank\">https://www.kaggle.com/competitions/birdclef-2025/discussion/583477</a></li>\n<li>1st - <a href=\"https://www.kaggle.com/competitions/birdclef-2025/discussion/583577\" target=\"_blank\">https://www.kaggle.com/competitions/birdclef-2025/discussion/583577</a></li>\n<li>2nd - <a href=\"https://www.kaggle.com/competitions/birdclef-2025/discussion/583699\" target=\"_blank\">https://www.kaggle.com/competitions/birdclef-2025/discussion/583699</a></li>\n<li>13th - <a href=\"https://www.kaggle.com/competitions/birdclef-2025/discussion/583457\" target=\"_blank\">https://www.kaggle.com/competitions/birdclef-2025/discussion/583457</a></li>\n</ol>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "3218859": "| Rank | Public / Private LB | Core Models (backbone × count) | **Spectrogram Params**<br>(mel / freq / FFT / hop) | Data & Label Strategy | Secret Sauce |\n|------|----------------------|-------------------------------|----------------------------------------------------|------------------------|----------------|\n| **1** | 0.933 / 0.930 | SED + CNN – EffNetB0/B3/B4, RegNetY, ECA-NFNet | 224 / 0–16kHz / 4096 / 1252 | 2025 + Xeno-Canto (target & extra species); multi-iterative pseudo-labeling (up to 4 rounds); Amphibia/Insecta-specific model | Multi-iterative Noisy Student; Mixup+Stochastic Depth; custom ensembling of 7 models; OpenVINO |\n| **2** | 0.922 / 0.918 | SED + CNN – EfficientNetV2-s, ECA-NFNet | 224 / 0–16kHz / 4096 / 1024 | 2025 + Xeno-Canto (pretraining excl. current species); 3-stage pseudo-labeling with filtered soft labels; no iNat used | Pretrained backbones; soft pseudo-labels (>0.5), multi-iteration pseudo mix; postprocessing via chunk top-prob scaling; no TTA in final |\n| **3** | 0.927 / 0.927 | CNN + SED – EffNet-B0-NS, V2-B3, V2-S, MNASNet, SPNASNet | 96–128 / 0–16kHz / 2048 / 512 | Used 2025 + 80% of BirdCLEF 2023 + Xeno-Canto + iNaturalist + pseudo-labeled soundscapes | Rank-aware postproc + model soup + ONNX; ensemble of 20 models (10 CNN + 10 SED) |\n| **5** | **0.928 / 0.924** | SED – EffNet-V2-S × 4, V2-B3 × 3, B3-NS × 4, B0-NS × 2 | 192 / 20–15kHz / 2048 / 768 | Self-distillation 5× on train_audio, then 2× with train_soundscapes (1:1 mix) | Iterative KD on unlabeled bird calls + 2.5s overlap + OpenVINO |\n| **6** | 0.928 / 0.923 | SED – EffNet-B3-JFT, B5-JFT, V2-B3 | _Not specified_ | Multi-round PL cycles with segment logits | Dual-head BCE loss (clip + segment) despite sigmoid mismatch |\n| **8** | 0.924 / 0.924 | 2× SED (V2-B3 & ECA-NFNet-L0) + CNN (SeResNeXt-26T) | 192 / 40–140kHz / 2048 / 512 | Pretrained on 2021–2024 data, PL (thr=0.4), 2-way KD | FilterAug + hard mixup + weighted BCE; dropped postproc trick at end |\n| **9** | 0.913 / 0.925 | SED + CNN – V2-B3, V2-S, SeResNeXt-26T, B0 | 256 / 20–16kHz / 2048 / ~400 | Two-stage: SED/CNN → PL → retrain | ONNX + single-thread workaround avoided timeout bug |\n| **10** | 0.915 / 0.921 | 3× ConvNeXt-T + SED (V2-S/B3) | 320–384 / 0–16kHz / 1536–2048 / varied | Confident pos/neg filtering → pp_data_clean | Hybrid CE + negative-logit penalty; α-blend recall trick |\n| **11** | 0.920 / 0.919 | 5× V2-B3 + 1× V2-S | 256 / 50–14kHz / 1024–1536 / 535 | Entropy-top 20% PL + CE loss | Postproc kernel [0.1, 0.4, 0.1]; CE boosted perf by 0.05 |\n| **13** | 0.901 / 0.90* | SED – tf_efficientnetv2_s, tf_efficientnetv2_b3, SEResNeXt26t | N/A / 0–16kHz / N/A / 512 | 2025 + train_soundscapes pseudo-labeled; rare species model (+0.01); cleaned audio, voice removal, duration recalculated | Stepwise 4-stage strategy; Sumix, no TTA; raw audio mixup; rare-species submodel ensembling; minimal boost from backbone diversity |\n| **14** | 0.922 / n/a | 3× V2-M (in21k) | 128 / 40–15kHz / 1024 / 512 | Iterative KD (1s avg soft 0.3 + 10s chunk soft 0.7) | Pure SED + OpenVINO pipeline for speed |\n| **26** | 0.909 / 0.908 | 4× CNN + 1× SED (EffNet family) | 384–448 / 0–16kHz / 3072–4096 / ~370 | Stage-2 PL from confident 5s chunks | “15s RMS → 3×5s max-pool” trick + model soup |\n| **28** | 0.893 / 0.909 | 5-backbone SED (SeResNeXt-26T heavy) | _Not specified_ | Segment-based voice removal + progressive PL (0.2→0.5) | Tuned smoothing [0.2, 0.6, 0.2]; large batch size (128) helped |\n| **29** | 0.902 / 0.906 | EfficientNet-B3 • ResNeSt-50 • NFNet-L0 | 256 / not stated / ? / ? | Trained on all data; mixup & segment variety (10–30s); infer on 5s | Multi-stage SpecAugment + 3-model blend + ONNX for speed |\n\nWhat's next - What to Take into BirdCLEF 2026 : https://www.kaggle.com/competitions/birdclef-2025/discussion/583437\n----",
    "3218864": "Congratulations to all the winners.\n\n**References**\n1. 26th - https://www.kaggle.com/competitions/birdclef-2025/discussion/583306 \n2. 9th - https://www.kaggle.com/competitions/birdclef-2025/discussion/583365\n3. 8th - https://www.kaggle.com/competitions/birdclef-2025/discussion/583324\n4. 14th - https://www.kaggle.com/competitions/birdclef-2025/discussion/583344\n5. 6th - https://www.kaggle.com/competitions/birdclef-2025/discussion/583381\n6. 11th - https://www.kaggle.com/competitions/birdclef-2025/discussion/583384\n7. 28th - https://www.kaggle.com/competitions/birdclef-2025/discussion/583377\n8. 10th - https://www.kaggle.com/competitions/birdclef-2025/discussion/583310\n9. 5th - https://www.kaggle.com/competitions/birdclef-2025/discussion/583312\n10. 29th - https://www.kaggle.com/competitions/birdclef-2025/discussion/583387\n11. 3rd - https://www.kaggle.com/competitions/birdclef-2025/discussion/583477\n12. 1st - https://www.kaggle.com/competitions/birdclef-2025/discussion/583577\n13. 2nd - https://www.kaggle.com/competitions/birdclef-2025/discussion/583699\n14. 13th - https://www.kaggle.com/competitions/birdclef-2025/discussion/583457"
  },
  "source": "meta"
}