Abstract
Large language models are trained on corpora rich in mathematical proof, code, scientific argument and dialogue, and their rule application is measured by benchmarks for statutory, legal, moral-exception and defeasible reasoning. This paper describes a training-data and evaluation design for one part of that capability, applying a rule to a new case while recognising when the rule does not apply, built from the Vinaya Piṭaka, the monastic legal code of the Theravāda Buddhist canon. The Vinaya's rule-by-rule analysis, the Suttavibhaṅga, treats each rule it analyses in a recurring order: the case that prompted the rule; the rule text and any later amendment; definitions of the operative terms; variants by object, means, perception and intention, each with a graded verdict; an exemption clause listing the conditions under which the rule does not apply; and further adjudicated cases. By the authors' count over one printed edition of the Pāli text, its two books carry 341 exemption clauses listing about 1,900 conditions, and some 3,400 written verdicts. Four uses of this structure are described: including the text in training data; supervising intermediate reasoning steps in the template's order; a benchmark of rule variants scored against the canon's own verdicts; and a test set of exemption cases for non-application. The paper reports no training run and no measurement. Whether the design improves a model's rule application is an open empirical question, and use of any edition or translation is subject to that edition's licence.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Ly, Thon, "The Vinaya Piṭaka as Training Corpus for Rule-with-Exception Reasoning in AI Systems: Training Data and Evaluation Benchmarks for Rule Application and Exception Handling in Large Language Models", Technical Disclosure Commons, ()
https://www.tdcommons.org/dpubs_series/12032