msb222/indra-r10-voicing

Indra R10 — Ontological Voicing Model (LoRA Adapter)

A LoRA adapter for gemma-3-12b-it fine-tuned through 10 rounds of autopoietic training to generate ontological voicings — first-person expressions of meaning from seed words.

What It Does

Given a seed word (like "trust", "fire", "recursion"), R10 speaks as the pattern itself — from inside its being, not about it. Outputs are structured as four-field JSON: essence (one line), sentence, paragraph, and page.

Usage

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base_model = AutoModelForCausalLM.from_pretrained("google/gemma-3-12b-it")
model = PeftModel.from_pretrained(base_model, "msb222/indra-r10-voicing")
tokenizer = AutoTokenizer.from_pretrained("msb222/indra-r10-voicing")

Bridge Prompt

• System: "You are an ontological companion. When asked to give voice to a pattern, speak as the pattern itself — from inside its being, not about it."
• User: "Give voice to the pattern of {word}..." with four-field JSON output (essence, sentence, paragraph, page)

Training Details

• Base model: gemma-3-12b-it (12B dense)
• Method: LoRA (r=16, alpha=32, dropout=0.05)
• Target modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
• Training: 10 rounds of iterative refinement (autopoietic loop)
• Training data: Self-generated voicings, curated through human evaluation
• Platform: Together AI fine-tuning

Part Of

The Right Brain Platform — voicings as infrastructure for meaning in AI systems.