The Nalanda Method: Dharmakīrti-Grounded Three-Tier Corpus Architecture for Domain-Specialized Small Language Model Training

Read full text Summarise with VIRAT
Authors: Anil Kumar Sharma
Cited by: 0

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

We present the Nalanda Method, a principled architecture for constructing training corpora and designing confidence scoring systems for domain-specialized Small Language Models (SLMs). The method grounds corpus stratification in the epistemological framework of Dharmakīrti (c. 600–660 CE), the Nalanda University logician whose *Pramāṇavārttika* (Commentary on Valid Cognition) classified sources of valid knowledge into three types: *Agama* (authoritative testimony), *Anumana* (inference), and *Pratyaksha* (direct perception). We argue that these three pramana map precisely and usefully onto three tiers of domain training data — statutory text, expert reasoning corpora, and production case outcomes — and that explicit stratification by pramana type is both theoretically justified and practically superior to undifferentiated fine-tuning. We further describe the Viveka Promotion Pipeline, which automatically elevates validated production cases to inference rules upon meeting evidence thresholds, and the DHARMA layer, an architecturally encoded ceiling at 85% autonomy that reserves moral judgment, creative originality, and contextual wisdom for humans. Implementations in maritime operations and Indian legal compliance domains are described, covering 5,336 rules across 35 categories. We also present a dual-track open source strategy that separates an Apache 2.0 Tier 1-only SLM release from a private production model incorporating all three tiers.

Fields of Study & Subjects

Artificial Intelligence in Law Topic Modeling Natural Language Processing Techniques

Publication details

Journal / source
Zenodo (CERN European Organization for Nuclear Research)
Publication date
25 March 2026
Language
EN
Document type
Preprint