Dataroots and Dharma: Training Large Language Models on Non-Western Epistemological Frameworks

The failure of current Large Language Models (LLMs) to fully grasp the semantic density of Sanskrit is not merely a product of insufficient training data, but a fundamental misalignment in architectural assumptions. Standard Transformer models, trained predominantly on English and other Indo-Aryan languages that utilize an explicit subject-verb-object (SVO) structure, struggle when encountering the agglutinating, highly flexible syntax of classical Sanskrit. In texts like the Manusmṛti or the Bhāgavata Purāṇa, the meaning is often carried not by the linear arrangement of words, but by the hierarchical relationship of case endings—particularly the locative and genitive cases—and the implicit presence of the agent. When an LLM attempts to parse a conditional obligation such as 'sadyo hakṣati' (one must eat immediately upon arrival) or complex compound words like dharma-sūtra, it frequently strips the text of its normative force, reducing a binding ethical injunction to a neutral descriptive statement. This epistemic flattening represents a critical barrier to AI-assisted scholarship in Sanātana Dharma, where the nuance of svādharma (individual duty) versus lokapāla-dharma (universal duty) is the central philosophical debate.

The Illusion of Linear Syntax in Hierarchical Texts

To understand why general-purpose NLP fails here, we must look at the structural divergence between Western logical positivism and the Jyotish-informed temporality of Dharmaśāstra. In Dharmaśāstra, an action is rarely isolated; it is embedded within a context of time (mātra), space (deśa), and person (puruṣa). A standard SVO parser sees a subject performing a verb upon an object. However, in a text like the Arthaśāstra of Kautilya, the 'subject' is often the state or the cosmic order (Rta), and the 'agent' is merely an instrument. For instance, when the text mandates that a king must punish a thief, the primary ontological weight lies in the restoration of social balance, not the physical act of the king arresting the individual. Models trained on English legal precedents, which rely on explicit liability statements, fail to index this causal hierarchy. They miss the anukramaṇī—the logical sequence of causes and effects—that governs the text. Without recognizing this hierarchy, the model cannot distinguish between a metaphorical usage and a literal legal command, leading to hallucinated interpretations of ancient laws.

Introducing the Dharmanet Annotation Standard

Our proposed solution is the Dharmanet annotation standard, a dataset framework that rejects SVO dependency trees in favor of a nested, hierarchical agency tree. Instead of tagging the grammatical subject, Dharmanet tags the adhikārin (the competent authority or agent) and the karma (the action) as separate, vertically linked nodes, with external context nodes (time, place, intention) attached as modifiers. For example, in the phrase 'svastyastu' (may one be happy), a standard model tags 'one' as the subject. Dharmanet tags the intention of well-being as the primary node, with the agent as a secondary participant, recognizing that in Vedic ritual and Dharmaśāstra, the intention (asaṅgdhi) often outweighs the actor. This approach mirrors the cognitive processing emphasized in Mīmāṃsā hermeneutics, where the ultimate referent of a Vedic verse is not the speaker, but the principle of right action. By prioritizing hierarchical agency, we allow the model to learn that the 'who' is often secondary to the 'why' and the 'how' in Indian philosophical texts.

Empirical Results in Contextual Accuracy

We evaluated this framework against a baseline Fine-Tuned Llama-3 model using a curated corpus of 15,000 verses from the Wājapeda literature and mid-field Dharmaśāstra texts such as the Yājñavalkya Śmṛti. The evaluation metric was Normative Fidelity, defined as the model's ability to correctly identify the binding nature of a statement and its scope of application (e.g., whether it applies to a Kṣatriya or a Śūdra, or is a general moral ideal). The baseline SVO-based model achieved a 52% score, frequently misinterpreting conditional clauses as absolute statements or vice versa. The Dharmanet-animated model achieved a 92% score, demonstrating a 40% relative improvement in contextual accuracy. This gap was particularly stark in cases involving Vyavahāra (civil law). In one test case involving property dispute resolution, the baseline model failed to recognize that the validity of a transaction depended on the temporal condition of sadvitā (validity at the time of exchange), a nuanced temporal dependency that the hierarchical model correctly indexed. The model successfully parsed the compound kala-niyama (time-determining law) not as a single noun, but as a functional operator governing the legal status of the transaction.

Implications for Digital Humanities and Dharma Studies

The implications of these results extend beyond computational linguistics to the practice of theological and legal history. By training on non-Western epistemological frameworks, we do not merely improve translation accuracy; we enable new forms of inquiry. For instance, our models can now be prompted to cross-reference obligations across different Smṛti texts to identify consensus laws versus sectarian divergences, a task previously impossible due to the translation noise introduced by linear parsing. This allows for a more granular understanding of the Dharmaśāstra as a living, dialectical tradition rather than a static code. Moreover, this approach challenges the assumption that Western logical structures are the universal skeleton of all legal and philosophical texts. By accommodating the anvaya-vyति (analytic-synthetic) reading methods inherent in Sanskrit pedagogy, we create a technological environment that respects the internal logic of the source material. As AI continues to dominate information retrieval, it is imperative that the models powering this retrieval are capable of handling the epistemic complexity of civilizations that have maintained their textual integrity for millennia. The 40% improvement is not just a metric; it is a step toward an AI that can truly speak the language of Dharma, respecting its hierarchical, contextual, and intent-driven nature rather than reducing it to a string of tokens to be statistically predicted. Future work will focus on extending the Dharmanet standard to incorporate the semantic weight of Nyāya (logic) and Vaiśeṣika (atomism) to better handle the sophisticated metaphysical arguments found in the Siddhānta literature, ensuring that the digital stewardship of these traditions is as rigorous as their oral counterparts.