46 1 month ago

A fine-tuned version of Gemma 4 12B IT on Vedic wisdom literature blended with general reasoning data.

tools thinking
ollama run cyborgxx101/gemma4-vedic-opus-finetuned-v2

Applications

Claude Code
Claude Code ollama launch claude --model cyborgxx101/gemma4-vedic-opus-finetuned-v2
Codex App
Codex App ollama launch codex-app --model cyborgxx101/gemma4-vedic-opus-finetuned-v2
OpenClaw
OpenClaw ollama launch openclaw --model cyborgxx101/gemma4-vedic-opus-finetuned-v2
Hermes Agent
Hermes Agent ollama launch hermes --model cyborgxx101/gemma4-vedic-opus-finetuned-v2
Codex
Codex ollama launch codex --model cyborgxx101/gemma4-vedic-opus-finetuned-v2
OpenCode
OpenCode ollama launch opencode --model cyborgxx101/gemma4-vedic-opus-finetuned-v2

Models

View all →

Readme

Gemma 4 12B — Vedic Opus Finetuned v2

A scholarly-grade fine-tune of Gemma 4 12B (text-only) for rigorous Vedic studies. Built to answer questions about the Ṛgveda, Brāhmaṇas, Upaniṣads, Vedāṅgas, Pūrva Mīmāṃsā, and Vedānta with philological precision.

Key Features

  • Chronological awareness — training examples tagged with [Epoch: Saṃhitā], [Epoch: Upaniṣad], etc. so the model distinguishes Vedic layers
  • Contrastive pairs — explicit cross-epoch comparisons
  • Anachronism detection — trained to say “this concept does not appear in that text” when appropriate
  • Mīmāṃsā hermeneutics — proper handling of śāstra-vihita-hiṃsā, apūrva, vidhi/niṣedha
  • Scholarly system prompt — mandates citation of specific texts, epochs, and technical terms
  • Zero New Age content — purely philological, no devotional filler

What the Model Can Do

  • Define technical terms with chronological precision
  • Distinguish Ṛta (cosmic order in the Saṃhitā) from Brahman (Vedāntic absolute)
  • Explain Mīmāṃsā hermeneutics (vidhi, apūrva, arthāpatti, śāstra-vihita-hiṃsā)
  • Trace semantic evolution (e.g., brahman from “sacred utterance” to “ultimate reality”)
  • Detect anachronistic projections of later concepts onto earlier texts

HumanEval Benchmark

Accuracy: 91.5% (150164) — competitive with SOTA 12B models; Vedic fine-tuning did not degrade general coding/reasoning.

Recommended Settings

temperature: 0.2, top_p: 0.95, repetition_penalty: 1.15

Quantizations

  • Q4_K_M (~7 GB) — :latest on Ollama
  • Q8_0 (~12 GB) — :q8_0 on Ollama
  • F16 (~23 GB) — GGUF repo