CPC-PKL: Fine-Tuning, Retrieval, and Citation Hallucination in Legal Question Answering on Pakistan's Code of Civil Procedure

Large language models (LLMs) are increasingly used for legal research, yet they are known to fabricate case law, and most evaluation work targets a small number of well-resourced jurisdictions. We study legal question answering for Pakistan's Code of Civil Procedure 1908 (CPC). We introduce CPC-PKL, a dataset of 2,467 long-form question-answer pairs drafted with Claude Sonnet 5 from an OCR-processed CPC commentary, partly reviewed with the help of law students, and organised into five reasoning categories (definitional, procedural, scenario, comparative and consequence). We compare four systems built on Qwen2.5-7B-Instruct: the base model, a QLoRA fine-tuned model, and both models with hybrid retrieval-augmented generation (RAG). Beyond ROUGE and BERTScore, we measure whether answers name the correct statutory provisions and where each cited case comes from. Fine-tuning improves every quality metric significantly (ROUGE-L 0.234 to 0.266) but teaches the model to cite cases it cannot ground: only 1.4% of its 366 case citations match the reference answer and 39.1% appear nowhere in our data. Retrieval raises the grounded share of the fine-tuned model's citations from 1.4% to 64.9% and gives the best answer quality (ROUGE-L 0.278), while the base model with retrieval is the most faithful system (99.4% grounded citations). Finally, an audit of the dataset itself against the source text shows that 24.0% of its citation mentions do not occur in the source book, even though the generation prompt asked the LLM to cite only cases present in the text. LLM-generated legal training data can therefore carry fabricated authority that fine-tuning then learns. Our code, citation audit and fine-tuned adapter are publicly available.