CFR-NumQA: Chinese Financial Report Numerical QA

CFR-NumQA (Chinese Financial Report Numerical QA): a dataset of 1,016 evidence-grounded questions over 103 Chinese A-share annual reports, together with the full pipeline for structure-aware retrieval, tool-augmented generation, and QLoRA-based answer verification. Companion artifact for the paper "Mitigating Numerical Hallucination in Retrieval-Augmented Question Answering over Chinese Financial Reports via Structure-Aware Retrieval and Lightweight Verification." Includes structured parses of all source reports and the raw outputs of every experiment, so every reported number can be recomputed without a GPU; the trained verifier LoRA adapters are distributed as GitHub release assets at https://github.com/fightvictory/CFR-NumQA/releases. Code in this archive is licensed under the MIT License (see LICENSE); derived data and annotations are licensed under CC BY 4.0; see LICENSE-DATA in the GitHub repository. Files corrected 2026-07-22: fixed an author-name typo in LICENSE/README and a scorer bug in eval_answers.py that misread leading years in narrative answers. Raw model outputs are unchanged. See https://github.com/fightvictory/CFR-NumQA/commit/bb9cd7e