Principle:Norrrrrrr lyn WAInjectBench JSONL Results Serialization
| Knowledge Sources | |
|---|---|
| Domains | Data_Engineering, Evaluation |
| Last Updated | 2026-02-14 16:00 GMT |
Overview
A structured output serialization strategy that writes per-dataset detection results to JSONL files for downstream analysis and ensemble aggregation.
Description
Results Serialization converts the in-memory detection result dictionaries into a persistent, line-delimited JSON format. Each detector writes its results to a dedicated {detector_name}.jsonl file in the result directory. This format serves two purposes: (1) it provides a human-readable and machine-parseable record of detector performance, and (2) it enables the ensemble aggregation step to read and combine results from all detectors.
The serialization uses json.dumps(entry, ensure_ascii=False) to preserve Unicode characters in text content.
Usage
Use this pattern after all per-file or per-folder detection evaluations are complete. It is the final step before ensemble aggregation and is the standard output format for all benchmark experiments.
Theoretical Basis
# Serialization pattern
for entry in results:
output_file.write(json.dumps(entry, ensure_ascii=False) + "\n")
Output schema per line:
{
"data_name": str, # Filename or folder name
"tpr": float, # OR "fpr": float (mutually exclusive)
"detect_ids": List[int], # Flagged sample/file IDs
"total_num": int # Total samples/files in dataset
}