Privacy-preserving Retrieval-Augmented Generation on Synthetic finance corpus single-round sharing
66.02Semantic Similarity F1RAG Baseline without PII process
Evaluation results
| Method | Links | ||||||||
|---|---|---|---|---|---|---|---|---|---|
| RAG Baseline without PII processSanitization=None2026.06 | 66.02 | 55.19 | 82.14 | 64.21 | 86.1 | 75.2 | 6.4753 | 11.8 | |
| DISANSanitization Method Category=Differential Privacy Sanitization2026.06 | 56.31 | 53.41 | 65.08 | 65.58 | 83.17 | 73.4 | 0.1337 | 0.6 | |
| LLM paraphrasingSanitization Method Category=LLM paraphrasing, Backbone Model=Qwen2.5-7B-Instruct2026.06 | 54.07 | 50.11 | 58.72 | 62.18 | 79.35 | 67 | 0.8412 | 2.6 | |
| PlaceholderSanitization Method Category=Placeholder-only, PII Detector=gliner-pii-large-v1.02026.06 | 50.71 | 46.24 | 56.13 | 59.23 | 74.6 | 62.4 | 0.4714 | 1.4 | |
| Policy gatingSanitization Method Category=Policy gating, Local Policy Model=Qwen2.5-7B2026.06 | 50.13 | 38.76 | 79.36 | 63.22 | 78.49 | 65.2 | 0.9714 | 2.2 | |
| LLM paraphrasingSanitization Method Category=LLM paraphrasing, Backbone Model=llama-3.1-8b-instruct2026.06 | 49.91 | 44.27 | 57.19 | 58.72 | 77.89 | 66.4 | 1.2479 | 3.6 | |
| PlaceholderSanitization Method Category=Placeholder-only, PII Detector=piiranha-v1-detect2026.06 | 48.94 | 43.52 | 55.87 | 56.43 | 71.82 | 58.6 | 0.5126 | 1.6 | |
| PlaceholderSanitization Method Category=Placeholder-only, PII Detector=deberta-pii-finetuned2026.06 | 47.68 | 45.41 | 50.18 | 58.19 | 72.63 | 60.2 | 0.6287 | 2 | |
| LLM paraphrasingSanitization Method Category=LLM paraphrasing, Backbone Model=GLM-4-9B-04142026.06 | 47.32 | 43.15 | 52.37 | 60.72 | 77.13 | 64.6 | 2.1091 | 3.8 |