bad-case-analyzer

To analyze false positives and false negatives in NLP classification, you can process a predictions file to categorize bad cases, compute per-class metrics, and generate an interactive HTML report for error pattern detection.

You need a predictions file containing fields like text, true_label, and predicted_label, with an optional confidence score. The tool computes the confusion matrix and outputs a bad-case report HTML directly from these inputs.

Yes, the generated interactive HTML report supports filtering mispredictions by error type, true or predicted label, and confidence levels, allowing you to isolate specific bad cases for deeper model evaluation and confusion analysis.

You can identify top confusion pairs by running a bad-case analysis on your model predictions. The process evaluates the results, builds a confusion matrix, and highlights the most frequent misclassification patterns in the output report.

No dependencies are required to use this error analysis tool. It operates independently on a standard predictions file with fields like text, true_label, and predicted_label to compute accuracy and output the interactive HTML report.

The best way to visualize model mispredictions for data collection is to generate an interactive HTML report. It categorizes FP and FN cases, computes per-class F1 scores, and allows filtering to guide targeted model improvement.