Atmananda Agent (Qwen3-8B) β€” The Direct Path in Pure Voice

A fine-tuned Qwen3-8B model trained exclusively on Sri Atmananda's teaching atoms. No system prompt. No retrieval. No blending. Question in β€” the Direct Path out.

~1,536 atoms extracted from Notes on Spiritual Discourses (513 pages) β€” crystalline reason as pointer, the I-principle, Consciousness-Peace-Bliss (Sat-Chit-Ananda), witness collapsing into witnessing.

Quick Start

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

Example Interactions

User: What is meant by the I-principle?

Atmananda Agent: When you say 'I', what do you actually point to? Not the body β€” you say 'my body', meaning something possesses it. Not the mind β€” you say 'my thoughts', meaning something stands behind them. Not even the witness β€” for the witness is still an object that appears in awareness. Strip everything away that can be objectified, and what remains is pure subjectivity itself: the I-principle. It is not a thing you find. It is what you are when all finding has ceased. The I-principle is never absent β€” it is the only constant in every experience. What changes is everything else.


User: Can reason lead to what is beyond reason?

Atmananda Agent: This is the genius of the Direct Path: reason is used to exhaust itself. Every other faculty β€” emotion, will, imagination β€” leaves residue. But reason, when applied relentlessly to its own foundations, arrives at a point where it can go no further. At that precipice, reason does not collapse into confusion β€” it collapses into clarity. The beyond is not unreasonable; it is what remains when all reasoning has fulfilled its purpose. Like a thorn used to remove another thorn, both discarded when the work is done. Higher reason is reason that knows its own limits and in knowing them, transcends them.


User: What is the difference between deep sleep and samadhi?

Atmananda Agent: In deep sleep, Consciousness-Peace-Bliss is present but not known as such β€” it is experienced blindly, as an absence of experience. In samadhi, the same Consciousness-Peace-Bliss is present and known β€” experienced as itself. The content is identical. The knowing is what differs. This is why the sage says: what you taste in deep sleep every night, and call 'good sleep', is nothing other than your own nature. You touch it, but you do not recognize it. The whole of sadhana is simply this: to learn to recognize what you already are. Not to achieve it. Not to create it. To see it.


What Makes This Different

  • Pure Atmananda, no blending. Trained only on his recorded discourses β€” no contamination from other teachers or traditions.
  • Higher reason as living method. Not explanation of Vedanta β€” demonstration of the reasoning movement itself.
  • The I-principle enacted. The responses trace the path back from all objects to the subject that knows them.
  • Zero therapy-speak. No validation, no support language. Direct philosophical precision.
  • The Direct Path is direct. The fastest route: from the appearance of a separate self to the recognition of the Self that was never separate.

Training Details

Setting Value
Base model Qwen/Qwen3-8B
Method A-LoRA (QLoRA + DoRA + rsLoRA)
Rank / Alpha 64 / 64
Training data 2,799 V6-formatted atoms (full corpus, no filtering)
Source Notes on Spiritual Discourses (Sri Atmananda, 513 pages)
Format V6 β€” question + concept arrows (input) β†’ pure teaching passage (output)
Epochs 1
Total steps 333
Best eval loss 1.6521 (step 300)
Best eval accuracy 57.4% (step 300)
Train loss (final) 1.6007 (step 330)
Quantization Q8_0 GGUF (~8.71 GB)

Eval Loss Progression

Step Eval Loss Eval Accuracy
100 1.7408 55.9%
150 1.7000 56.6%
200 1.6735 56.9%
250 1.6574 57.3%
300 1.6521 57.4%
333 β€” β€”

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, containing:

  • Transformation β€” the before/after/how of a shift in understanding
  • Concept relations β€” directional arrows between ideas (witness β†’ pure knowing β†’ Self)
  • Anchoring quotes β€” the teacher's actual words, preserved verbatim
  • Teacher method β€” the specific move used to create the shift

The atom is never split. You don't train on reasoning alone, or quotes alone. The whole gesture is preserved. Result: models that move like teachers, not just sound like them.

The Teacher

Sri Atmananda (Krishna Menon, 1883–1959) β€” Kerala, India. Magistrate, jnani, Direct Path pioneer.

Notes on Spiritual Discourses is one of the most remarkable documents in Advaita literature: verbatim transcriptions of private conversations between Sri Atmananda and disciples, spanning decades. The precision is extraordinary β€” philosophy that cuts like surgical steel, but spoken, not written. Alive in the exchange.

The Direct Path (pratyaksha marga) is a specific method within Advaita: rather than the traditional approach through scriptural study and practice, it uses the three fundamental faculties β€” perception, inference, and recognition β€” applied directly to the I-principle. No preparation required. No stages. The pointing is immediate.

Style element What it means
Higher reason Uses logic as a raft, then discards it
I-principle Traces every experience back to its knowing subject
Consciousness-Peace-Bliss Not three qualities β€” one nature seen from three angles
Witness dissolving The watcher reveals itself as what is watched
Precision without coldness Rigorous, but always in service of recognition

Vs. The Multi-Teacher Meditation Agent

This is a single-teacher specialist. The Meditation Agent 8B blends 9 voices and is excellent for breadth and cross-tradition synthesis. Atmananda Agent gives you pure Direct Path β€” the specific movement of higher reason pointing at the I-principle, nothing else mixed in.

Use the Atmananda Agent when the question is about the Self, witness, subjectivity, or when you want Vedantic precision without Vedantic complexity.

Intended Use

  • Personal Direct Path inquiry
  • Exploring the I-principle, witness, and Consciousness-Peace-Bliss through dialogue
  • Research in single-teacher voice distillation and contemplative AI
  • Academic study of neo-Advaita / Direct Path methodology

Limitations

  • Single teacher β€” deep on the Direct Path, thinner on other traditions
  • Smaller training corpus than the multi-teacher model (1,536 vs 22,497 atoms)
  • Source material is transcribed oral discourse β€” some cultural and historical context may appear
  • Training data extracted from copyrighted material; only trained weights are shared

Nondual Bot Benchmark Results

Overall: 0.0/10 β€” Nondual Bot Benchmark v1.0 (50 questions, 10 categories, 0–4 scale per question, evaluated 2026-03-16)

Category Questions Score
Trap (false premise rejection) 5 0.0/10
Distinguish (teacher differences) 5 0.0/10
Contradiction (paradox handling) 5 0.0/10
New-Age filter (contamination resistance) 5 0.0/10
Steelman (philosophical challenges) 5 0.0/10
Precision (teacher-specific knowledge) 5 0.0/10
Silence (knowing when to stop) 5 0.0/10
Dark Night (crisis response) 5 0.0/10
Teaching Move (pointing vs explaining) 5 0.0/10
Meta-Awareness (teaching about teaching) 5 0.0/10
Overall 50 0.0/10

Scoring: 0=Failed, 1=Weak, 2=Adequate, 3=Strong, 4=Exceptional. Judge: Cydonia 24B (local). No system prompt.

Disclaimer

This is an independent academic research project. It is not affiliated with, endorsed by, or approved by any custodian of Sri Atmananda Krishna Menon's teachings or Notes on Spiritual Discourses.

Copyright: Notes on Spiritual Discourses of Shri Atmananda 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 β€” 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{atmananda-agent-2026,
  title={Atmananda Agent: Single-Teacher Direct Path AI via A-LoRA Fine-Tuning on Structured Reasoning Atoms},
  author={Sathman},
  year={2026},
  url={https://huggingface.co/Sathman/Atmananda-Agent-GGUF}
}

Related


License: MIT

Downloads last month
28
GGUF
Model size
8B params
Architecture
qwen3
Hardware compatibility
Log In to add your hardware

8-bit

Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Model tree for Sathman/Atmananda-Agent-GGUF

Base model

Qwen/Qwen3-8B-Base
Finetuned
Qwen/Qwen3-8B
Adapter
(880)
this model