CleanLab 2026：11K 星 — 开源 AI 数据清洗与标签错误检测平台

from cleanlab.dataset import dataset_health
from cleanlab.filter import find_label_issues

# 加载数据
labels = [...]  # 真实标签
pred_probs = [...]  # 模型预测概率

# 检测标签错误
label_issues = find_label_issues(labels, pred_probs)
print(f"检测到 {len(label_issues)} 个标签错误")

from cleanlab.dataset import dataset_health

# 全面数据集健康检查
health = dataset_health(labels, pred_probs)
print(f"标签噪声率: {health['label_noise_rate']:.2%}")
print(f"不兼容样本: {health['joint_noise_matrix']}")

from cleanlab.filter import filter_by_prob_normalized_margin

# 获取最可能是错误的样本
confidence_scores, to_remove = filter_by_prob_normalized_margin(
    pred_probs, labels, threshold=0.2
)

from cleanlab.sklearn import ClassifierPipeline

# 自动处理标签错误的分类器
clf = ClassifierPipeline(
    model="random_forest",
    labels=labels,
    pred_probs=pred_probs
)
clf.fit(X_train, y_train)

from cleanlab.torch import CleanLearning

# 包装 PyTorch 模型
clean_learning = CleanLearning(model)
clean_learning.fit(X_train, y_train)

from cleanlab.transformers import CleanLearningForSequenceClassification

# NLP 任务自动清洗
clean_clf = CleanLearningForSequenceClassification(model)
clean_clf.fit(train_dataset, label_names=["label"])