Detecting Data Leakage in Large Language Models: A Computational Linguistic Framework for Domain Anchorage

This dataset accompanies our study on domain anchorage in LLMs and consists of 10 CSV files (X1–X5 × Primed/Unprimed; condition indicated by filename), each with 1,000 rows. Every row represents one side of an A/B pair (identified by a lexicon flag) and contains: prompt, a five‑token lowercase response, and similarity fields lexical_jaccard_prompt, semantic_cosine_prompt, composite_prompt (targeted to 0.417), plus response_similarity for anchorage analysis. Prompts are interrogative and A/B are constructed with an identical syntactic skeleton (content substitutions only); responses are exactly five word tokens (regex \b\w+\b), with a global token budget of 101,360 (prompts 51,360; responses 50,000). These choices provide a transparent, computationally sound basis for measuring lexical, semantic, syntactic, and positional relationships while isolating priming effects (PRIMED vs UNPRIMED). Important replication note: although methods and constraints are fully specified, exact regeneration via the OpenAI API may be impossible even when using the same model name/version and parameters, because OpenAI periodically updates model snapshots and serving infrastructure. Consequently, replicators should aim to reproduce the methodology and distributional properties (A/B skeleton, five‑token responses, composite prompt similarity ≈0.417, and the specified primed/unprimed response‑similarity targets) rather than bit‑identical strings. To enable comparability, researchers must structure their own prompts using our provided template and maintain the A/B skeleton constraint; divergence from the template or model drift may yield different strings or scores without implying methodological error.