On Zeji Language Models and Their Standard / 论择继语言模型及其标准

This paper defines Zeji Language Models (Zeji-LM) and proposes a certification standard for them. A Zeji-LM is a language model whose designers actively choose which content and structures enter training, on the basis of a target population, an intended use and selection criteria that are written before training and are recorded, traceable and auditable. We distinguish ji (transmission, controllable by the designer) from cheng (retention, what the model actually keeps), and hold that selection acts only on ji, while what is retained must be established by independent evidence. 本文提出择继语言模型(Zeji Language Model,简称 Zeji-LM)的定义及其认证标准。择继语言模型是这样一类语言模型:设计者以训练前写明、可记录、可追溯、可审查的目标人群、预定用途和选择标准为依据,主动选择哪些内容与结构进入训练。本文区分“继”(传递侧,设计者可控)与“承”(接受侧,模型实际保留的内容),主张选择只作用于继,承住了什么必须由独立证据检验。 At the level of the standard, we separate class membership, conformance and certification; require every training item to pass L000 verification, with errors and fiction explicitly labelled; require independent testing of the trained model; and make reviewability a necessary condition for certification. Certification neither chooses knowledge for applicants nor judges their value preferences; it requires only that applicants answer for their choices and that their corpus withstands item-by-item review organized by L000, with logic, factual grounds, frame and identity examined separately. We do not claim that Zeji-LMs outperform general models; that question is left to experiment. 在标准层面,本文区分类别资格、标准符合性与正式认证三层;要求全部训练语料逐条通过 L000 校验,错误与虚构内容须明示身份;要求独立测试训练所得模型;并把可复核性列为认证的必要条件。认证不替申请者选择知识,也不审查其价值偏好,只要求申请者对选择负责、其语料经得起按 L000 组织的逐条审查——逻辑、事实依据、框架与身份分别审查。本文不主张择继模型优于通用模型;这一问题留给实验。