Sathman/Nisargadatta-Agent-SmolLM3-3B-GGUF

Nisargadatta Agent (SmolLM3 3B) — Pure Nisargadatta Voice

A fine-tuned SmolLM3-3B model trained exclusively on Nisargadatta Maharaj's teaching atoms. Ruthless negation. Direct pointing. "I Am." Question in, Nisargadatta out.

This is the lightweight 3B branch of the Nisargadatta Agent. Same V6 training format as the 8B version, smaller footprint.

~2,870 atoms from I Am That (396 pages) — the record of his conversations in Bombay.

Quick Start

LM Studio / Ollama / llama.cpp: Download Nisargadatta_Agent-SmolLM3-3B-Q5_K_M.gguf and load it. No system prompt required.

Example:

User: How do I find peace?

Nisargadatta Agent: You are looking for peace as if it is something to be found. But where would you look? In the mind? The mind is the problem, not the solution. Peace is not a state to be achieved — it is what you are when the mind is not adding its commentary. Stop the search. Look at who is searching. That one — find that one.

User: What is the self?

Nisargadatta Agent: The self is what remains when everything that can be removed has been removed. Not the body — that changes. Not the mind — that changes. Not the emotions — they come and go. What does not come and go? Find that. That is the self. It is not an object you can find; it is the finder.

What Makes This Different

• Pure Nisargadatta, nothing else. Trained only on his atoms — no synthesis.
• Ruthless negation. Every false identity is dismantled. Nothing is left standing.
• "I Am" as the pointer. His specific method: hold the sense "I Am" without qualification.
• No consolation. He does not comfort — he cuts.
• 3B footprint. ~2 GB with Q5_K_M.

Training Details

Setting	Value
Base model	HuggingFaceTB/SmolLM3-3B-Base
Method	A-LoRA V6 (QDoRA + rsLoRA)
Rank	32, alpha 32
Training data	Nisargadatta V6 atoms (from ~2,870 raw, filtered for complete question+teaching pairs)
Source books	I Am That
Format	V6 — question + concept arrows (input) → pure teaching passage (output)
Epochs	1
Quantization	Q5_K_M GGUF (~2 GB)

The A-LoRA Methodology

A-LoRA (Atom LoRA) fine-tunes on complete reasoning atoms — indivisible multi-component teaching moves extracted from human expert traces. Each atom contains transformation (before/after), concept relations, anchoring quotes, and teacher method. Training on whole atoms produces models that move like teachers, not just sound like them.

The Teacher

Nisargadatta Maharaj (1897–1981). Shop owner, self-realized. Spoke in Marathi, translated by Maurice Frydman into the landmark I Am That. His method: direct pointing, no metaphysics, no ritual. Just: what are you, before you think you are anything?

Style element	What it means
Ruthless negation	Systematically removes every false identity
"I Am"	The unconditioned sense of existence as the only reliable pointer
No metaphysics	No cosmology, no philosophy — only direct investigation
Cutting directness	No comfort, no buffer — straight to the point

3B vs 8B

The 8B Nisargadatta Agent handles longer inquiry chains better. The 3B branch preserves the core negation movement for lighter hardware.

Files

File	Size	Use
Nisargadatta_Agent-SmolLM3-3B-Q5_K_M.gguf	~2 GB	Recommended

Related Models

• Full series — Meditation Agent Collection — all 19 models
• Nisargadatta Agent 8B — Full-fidelity, Qwen3-8B base
• Meditation Agent 3B — All 9 teachers blended, SmolLM3-3B
• Meditation Agent 8B — All 9 teachers blended, Qwen3-8B

---

ellam sivamayam — Everything is Shiva's expression.

எல்லாம் சிவமயம்