ReadingProtocol: code, evaluation data and saved records for traceable question answering over complex engineering tables
收藏资源简介:
Code, evaluation data and saved evaluation records accompanying the paper “Constrained Planning and Deterministic Execution for Traceable Question Answering over Complex Tables in Engineering Standards” (submitted). ReadingProtocol separates a model-written plan from deterministic reading, composition and verification, and records the evidence for each answer. The package contains the source code and tests; the synthetic tables, questions and keys; the engineering table crops (public US-government standards, with per-file provenance), questions, keys and gold transcriptions; and the saved plans, outputs, grades, adjudications and evidence-assessment records behind the reported results, with a SHA-256 manifest (records/MANIFEST.sha256). The engineering result is a composite of original results and saved-plan replays and is not a fresh evaluation of one frozen implementation. HiTab tables are not redistributed; question identifiers and selection lists are provided. Confidential and licensed standards are not included. See SHARE.md.



