Birdclef2026 Rajuly Xc Pseudo Top3514

--- library_name: none license: cc-by-4.0 tags: - birdclef - bioacoustics - pseudo-labels - multilabel - weak-supervision --- # BirdCLEF 2026 Rajuly XC Pseudo Labels (Top 3514) This repository provides soft pseudo labels for a curated subset of external XC clips, packaged as a model-style snapshot for easy download. ## Files - `manifest.csv` Row-level metadata for each sample. Includes `filepath` (path to audio file **inside the zip**). - `soft_probs.npy` Soft labels as `float32` probabilities with shape `(3514, 234)`. - `primary_labels.npy` Ordered class labels (length `234`) matching columns in `soft_probs.npy`. - `dataset_info.json` Dataset summary statistics. - `top10k_noleak_audio.zip` Audio files package for all rows in `manifest.csv`. ## Audio zip usage Download: - `top10k_noleak_audio.zip` Then read audio by joining extracted folder with `manifest.csv` column `filepath`. Example: ```python import numpy as np import pandas as pd from pathlib import Path manifest = pd.read_csv("manifest.csv") soft_probs = np.load("soft_probs.npy") labels = np.load("primary_labels.npy", allow_pickle=True) assert len(manifest) == soft_probs.shape[0] assert soft_probs.shape[1] == len(labels) audio_root = Path("./top10k_noleak_audio") audio_path = audio_root / manifest.loc[0, "filepath"] print(audio_path) ``` ## Data summary - Samples: **3,514** - Classes: **234** - Max-prob mean: **0.9224** - Max-prob median: **0.9207**