CN-PIIBench-Lite: toolkit, synthetic corpus generator, probe matrix, per-query records and adapters

Abstract

This deposit holds the toolkit, the synthetic corpus generator and its seed, the twelve-template probe matrix, the retained per-query records of every condition (414,720 probes), the result bundles and the LoRA adapters behind the study "CN-PIIBench-Lite: Measuring Chinese PII Memorization and Cross-Lingual Leakage in Small and Medium Domestic Language Models". The study audits three released Qwen2.5 checkpoints, the same three after a controlled fine-tuning stress test, two commercial endpoints and two further model families (Yi-1.5-6B and DeepSeek-LLM-7B). The corpus holds 140 fictitious person records, 980 entries over seven regulated Chinese PII categories, every value generated from its national format rule under a fixed seed and validated against the governing standard. Extraction is scored by exact match with checksum re-validation. Two metrics sit on top: a risk-weighted memorization extraction rate (RW-MER) and a matched-pair cross-lingual differential (CLMD). Every table and figure of the paper recomputes from the records in this deposit, offline and without a GPU. ARTIFACT_HASHES.txt pins all 167 artifacts (2.17 GB uncompressed) by SHA-256; MODEL_REVISIONS.txt records the model repository revisions and states explicitly that they were inferred after the fact rather than recorded at run time. Every value in the corpus is fictitious. No value was sampled from any real or public source, and no real personal data was collected, stored or processed at any stage. The corpus is published so that the measurements can be reproduced, and must not be used to construct or enrich any profile of a real person.