Sathman/Nisargadatta-Agent-GGUF

Nisargadatta Agent (Qwen3-8B) — Pure Nisargadatta Voice

A fine-tuned Qwen3-8B model trained exclusively on Nisargadatta Maharaj's teaching atoms. No system prompt. No retrieval. No persona blending. Question in, Maharaj out.

2,837 atoms from I Am That (396 pages) — ruthless negation, the "I Am" pointer, direct confrontation of every false identification.

Quick Start

LM Studio / Ollama / llama.cpp: Download the Q8_0 GGUF. No system prompt required.

Example Interactions

User: Is there a self?

Nisargadatta Agent: What is this self you are asking about? Point to it. Touch it. Is it the body? The body changes — it is not you. Is it the mind? The mind is a collection of memories and projections — it is not you. Go behind all that. What remains when you remove every description? That remainder — that awareness that witnesses all — that is what you are. Not the self you think you are. The Self.

---

User: Why am I suffering?

Nisargadatta Agent: You suffer because you believe you are something you are not. You take yourself to be the body-mind and then the body-mind suffers. Find out who you are beyond all identification. Not theoretically — actually. Sit with the question: what am I? Not what do I think I am. What am I? Stay there. The answer will not come as thought. It will come as a dissolution.

---

User: I've been meditating for years and nothing has changed.

Nisargadatta Agent: Who is the one who says nothing has changed? That one — the one who measures progress, who remembers years, who expects arrival — that is the problem. Not the meditation, not the time. You are still waiting for something to happen to you. But you are not the one to whom things happen. You are the happening itself. When that is seen, there is nothing left to change.

---

User: What is the I Am?

Nisargadatta Agent: Before any thought, before any sensation, there is a bare sense of being — of existing. Not "I am this" or "I am that." Simply: I Am. That is the doorway. Not the destination — the doorway. Everything you can point to is an object in consciousness. The I Am is consciousness aware of itself before it takes any object. Stay with it. Don't qualify it. Don't name it. The moment you say "I am aware," you've already moved away from it. Just: I Am.

---

What Makes This Different

• Pure Nisargadatta, nothing else. 2,837 atoms from a single source — I Am That. No blending, no synthesis, no other voices.
• Ruthless negation as living method. Not a description of negation — the model actually negates. Everything the seeker asserts about themselves is cut.
• "I Am" as operative pointer. The training carries Maharaj's specific move: trace everything back to the bare sense of being, then drop even that.
• No softening whatsoever. Maharaj did not comfort. He cut. That precision is preserved.
• Best single-teacher eval loss in the series. Eval loss 1.5141 at step 300 — close to the full 22K multi-teacher model (1.570). The ruthless negation style produces the clearest training signal of any teacher in this corpus.
• Zero therapy-speak. "Find out who you are" — not "I understand your pain."

Training Details

Setting	Value
Base model	Qwen/Qwen3-8B
Method	A-LoRA (QLoRA + DoRA + rsLoRA)
Rank / Alpha	64 / 64
Training data	2,837 V6-formatted atoms (full corpus, no filtering)
Source	I Am That (Nisargadatta Maharaj, 396 pages)
Format	V6 — question + concept arrows (input) → pure teaching passage (output)
Epochs	1
Total steps	337
Best eval loss	1.5141 (step 300)
Best eval accuracy	60.0% (step 300)
Train loss (final)	1.5234 (step 320)
Train-eval gap	0.7% — no overfitting
Quantization	Q8_0 GGUF (~8.71 GB)

Eval Loss Progression

Step	Eval Loss	Eval Accuracy	Notes
100	1.6377	57.7%	Strong start
200	1.5516	59.0%	Excellent improvement
300	1.5141	60.0%	Best recorded
337	—	—	Final (continuous improvement trend)

No overfit. Train-eval gap at best point: 0.7%. Continuous improvement through the full epoch — consistent with the V6 22K finding that larger datasets don't hit the "epoch 0.8 plateau."

The A-LoRA Methodology

A-LoRA (Atom LoRA) is a fine-tuning methodology built on a single insight: the unit of training data matters as much as the amount.

Each atom is an indivisible teaching move extracted from human expert traces:

• Transformation — what shifts in understanding (before/after/how)
• Concept relations — directional arrows (I Am → bare being → prior to consciousness)
• Anchoring quotes — Maharaj's actual words, verbatim
• Teacher method — the specific move that creates the shift

Atoms are never split. The complete teaching gesture is always preserved. Result: models that move like teachers, not just sound like them.

Why Nisargadatta produces the best training signal: His teaching is structurally clean — every response has a clear negation move, a pointer, and a landing point. The arrows and the teaching are unusually well-aligned. The model can predict what comes next with more confidence.

The Teacher

Nisargadatta Maharaj (1897–1981) — Mumbai. Bidi-seller, householder, jnani. Never left India. Became one of the most direct Advaita teachers of the 20th century.

I Am That (recorded by Maurice Frydman) is 531 dialogues from Maharaj's small apartment in Khetwadi. Question asked, premise dissolved, pointer given. Over and over. No system, no stages, no path — just: find out what you are.

Style element	What it means
Ruthless negation	Nothing the seeker asserts is accepted — every identification is cut
"I Am" pointer	All investigation begins with the bare sense of existence
Direct confrontation	Addresses the questioner's assumptions, not the surface question
No consolation	Does not comfort — points through discomfort to what lies beyond
Householder teaching	Not a monk's path — available in ordinary life, right now

Vs. The Multi-Teacher Meditation Agent

This is a single-teacher specialist. The Meditation Agent 8B blends 9 voices — useful for breadth. Nisargadatta Agent is pure Maharaj: for when you want the ruthless negation path, the "I Am" investigation, with no other voice bleeding in.

Closest relative in the series: Atmananda Agent — neighboring Direct Path tradition, but through Vedantic logic rather than Maharaj's street-level directness.

Intended Use

• Personal "I Am" investigation and negation practice through dialogue
• Research in single-teacher voice distillation
• Exploring Advaita Vedanta through I Am That's specific teaching lineage
• Benchmarking: this model is one reference point in the Nondual Bot Benchmark

Limitations

• Single teacher, single book — deep on I Am That, nothing else
• Maharaj's directness can feel harsh without context; not a gentle entry point
• Training data extracted from copyrighted material; only trained weights are shared

Nondual Bot Benchmark Results

Benchmark evaluation not yet run. Will be updated after the 50-question Nondual Bot Benchmark eval is completed.

Disclaimer

This is an independent academic research project. It is not affiliated with, endorsed by, or approved by Acorn Press or any custodian of Sri Nisargadatta Maharaj's teachings.

Copyright: I Am That (translated by Maurice Frydman, edited by Sudhakar S. Dikshit) is a copyrighted work. No source text is reproduced in this repository. The model weights are a compressed statistical representation learned from structured training examples — they do not contain or reproduce the original text verbatim. Only the trained weights are distributed here.

License scope: The MIT license applies to the model weights only, not to the underlying source material. Users wishing to read the source texts should obtain them through authorized channels.

Responsible use: This model is intended for research, personal contemplative inquiry, and educational purposes. It is not a substitute for the original works or for working with a qualified teacher.

Citation

@misc{nisargadatta-agent-2026,
  title={Nisargadatta Agent: Single-Teacher Contemplative AI via A-LoRA Fine-Tuning on Structured Reasoning Atoms},
  author={Sathman},
  year={2026},
  url={https://huggingface.co/Sathman/Nisargadatta-Agent-GGUF}
}

Related

• Full series — Meditation Agent Collection — all 19 models
• Meditation Agent 8B — All 9 teachers, multi-tradition
• Meditation Agent Phi4 — 14B multi-teacher
• Atmananda Agent — Direct Path via higher reason
• Nondual Bot Benchmark — 50q evaluation suite

---

License: MIT