gemma4-vedic-opus-finetuned-v2

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 (e.g., Ṛta ≠ Ātman, Svarga ≠ Mokṣa)
• 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

Training

Dataset Composition

What the Model Can Do

• Define technical terms with chronological precision (e.g., when ātman means "breath" vs "ultimate self")
• 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")
• Answer sensitive topics (animal sacrifice, varṇa hierarchy) dispassionately using traditional frameworks
• Detect anachronistic projections of later concepts onto earlier texts

Example

User: What does the Rig Veda say about reincarnation?
Assistant: The Rig Veda Saṃhitā does NOT contain the concept of
reincarnation (saṃsāra) as a cycle of rebirth. The dead go to
the world of the fathers (pitṛ-loka) to dwell with Yama — they
do not return. The earliest clear formulation of saṃsāra occurs
in the Bṛhadāraṇyaka Upaniṣad (BU 4.4.5-6). What the Rig Veda
does describe is punar-janman in the sense of new progeny — a
man is 'born again' through his son. This is genealogical, not
soteriological.

Recommended Inference Settings

temperature: 0.2
top_p: 0.95
repetition_penalty: 1.15  # use in llama.cpp / HF generate

HumanEval Benchmark

The model was evaluated on the HumanEval benchmark in the MLX environment (Q4_K_M quantization, temperature 0.2, top_p 0.95). This is competitive with SOTA 12B models and confirms that Vedic fine-tuning did not degrade general coding/reasoning capability.

Error analysis: 14 wrong cases span string manipulation (make_palindrome, decode_shift, decode_cyclic, fix_spaces, words_string), numerics (find_zero, count_nums, closest_integer, max_fill, order_by_points, solve, tri), and nested structure reasoning (is_nested, minPath).

Quantizations

Ollama

ollama run cyborgxx101/gemma4-vedic-opus-finetuned-v2
ollama run cyborgxx101/gemma4-vedic-opus-finetuned-v2:q8_0

License

Apache 2.0