This resource contains sense glosses and generated commentary derived from Mark Dyczkowski's copyrighted Kashmir Śaiva translations and lectures. Access is granted for non-commercial academic research only.

Exegetical Generation — Evaluation Experiments (ExeScore-Lex + Tradition-Drift)

Code and artifacts for two evaluation studies supporting the Exegetical Generation paper (Anamavajra Labs). Gated because the Śaiva sense glosses and generated explanations derive from Mark Dyczkowski's corpus (research-only).

A. ExeScore-Lex — linguistically-grounded generation metric

code/exescore_lex/ — a lemma/synset-grounded replacement for a script-dependent surface info-gain counter, built on the Digital Corpus of Sanskrit (lemma inventory), Monier-Williams + BHS (attestation), a WordNet synset projection, and ByT5-Sanskrit neural lemmatization (run on SageMaker). Metrics: G_lex (grounded info gain), G_syn (sense-level gain), F_lex (lemma-coverage faithfulness), and a lexicon-validated hallucination rate.

Key result: the naïve counter is biased both ways — it over-counts LLM output (surface-form inflation) and under-counts Devanāgarī-emitting fine-tuned output; neural lemmatization raises recall +22–38 % and improves hallucination precision.

B. Tradition-drift — measuring tradition-coherence (T)

code/tradition_drift/ — does a model keep the Kashmir-Śaiva technical sense of a polysemous term, or drift to the mainstream classical / Buddhist sense?

Pipeline: (1) surface candidate polysemous terms (corpus ∩ MW/BHS); (2) build a three-pole sense inventory (classical MW / Buddhist BHS / Śaiva, the Śaiva pole extracted from the corpus by a hybrid Haiku→Sonnet step); (3) score each system's term-explanation against the poles — by embedding (Titan v2) and by an independent LLM-judge (Qwen2.5-32B on H100, length/style-robust).

Final LLM-judge sense-bleed (n=20 terms)

System	Śaiva-rate	bleed-rate
B2 zero-shot	0.40	0.15
B3.1 RAG (classical dict.)	0.30	0.45
B4-ft QLoRA (Śaiva-tuned)	0.50	0.20

Findings: RAG with a classical dictionary triples sense-bleed (0.15→0.45) — grounding on the wrong tradition harms fidelity; fine-tuning on Śaiva data best preserves the tradition sense (0.50). The embedding metric gave a length artifact (ranked B4-ft worst) that the LLM-judge reverses — verbose systems need a judge, not cosine.

Layout

code/exescore_lex/ — A metric + ByT5 (SageMaker) + synset builder
code/tradition_drift/ — B pipeline (candidates → inventory → embedding + judge)
results/ — candidates, sense inventory, per-system explanations, bleed results

Caveats

n=20; local Cologne MW dump is abbreviated (stub entries for some headwords); single-author (Dyczkowski) corpus; awaiting domain-expert gold validation.

📄 Paper
🤗 Model · Corpus

Downloads last month: 12

Total file size:

403 kB

You need to agree to share your contact information to access this dataset