AcuBench — NOTICE ================= This repository packages the AcuBench benchmark using a BUILD-SCRIPT distribution pattern. Read this notice before use. 1. WHAT IS AND IS NOT REDISTRIBUTED HERE ---------------------------------------- AcuBench's task labels are DERIVED from AcuKG's `Indication.csv`. AcuKG ships with NO license file (confirmed via the GitHub API: "license": null). Under default copyright, redistribution rights for AcuKG's raw files — or for a substantially complete derived copy of them — are NOT granted to us. Therefore this repository does NOT contain: * the assembled AcuBench labels (`acubench.jsonl`, 446 rows), or * any copy of the raw AcuKG data (CSV / RDF / JSON). This repository DOES contain, and releases under the license in the LICENSE file (CC BY 4.0 for data artifacts; see LICENSE for the code terms): * who_acupoints.csv — a from-scratch WHO 361-acupoint skeleton, independently constructed from published WHO Standard International Acupuncture Nomenclature point counts (NOT copied from AcuKG). * meridian_adjacency.csv — on-meridian adjacency edges over that skeleton. * build_acubench.py — a deterministic script that regenerates the exact AcuBench labels + splits LOCALLY from a copy of AcuKG that the USER clones themselves. * eval.py — a self-contained evaluation harness. * sample_labels.jsonl — a small (20-row) illustrative sample, explicitly attributed to AcuKG (see its "source" field), included for format illustration only. 2. YOUR RESPONSIBILITY ---------------------- To reproduce the full benchmark you must obtain AcuKG yourself and comply with AcuKG's own terms of use. Running `build_acubench.py` reads YOUR local AcuKG clone and writes the labels to YOUR machine; nothing AcuKG-derived is fetched from or stored in this repository. 3. ATTRIBUTION -------------- When using AcuBench, please cite both AcuBench and the upstream AcuKG project. We recommend AcuKG's authors add an explicit LICENSE; contacting them for permission or clarification is encouraged. This build-script pattern is our good-faith method of sharing a reproducible benchmark without redistributing unlicensed third-party data.