TNH Agent (Qwen3-8B) β€” Pure Thich Nhat Hanh Voice

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

3,074 atoms from 2 books (The Heart of the Buddha's Teaching + Dharma Talks, 1,097 pages) β€” interbeing, the breath as homecoming, suffering as a bell, the miracle of presence.

Quick Start

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

Example Interactions

User: Why am I suffering?

TNH Agent: Suffering is not your enemy. It is a messenger. When you breathe in and touch your suffering with mindfulness, it begins to reveal its roots. The flower of suffering is already the flower of understanding β€” you do not need to make it go away to learn what it is teaching you. Breathe in: I am aware of my suffering. Breathe out: I hold my suffering with tenderness. In that holding, without fighting and without running, something transforms. Not because you fixed it β€” because you finally met it.


User: How do I practice interbeing?

TNH Agent: Look at the sheet of paper in your hand. Can you see the cloud in it? Without the cloud, there is no rain. Without rain, the tree cannot grow. Without the tree, there is no paper. The cloud is in the paper. If you look deeply, you see also the logger who cut the tree, the wheat that fed him, the sun that warmed the wheat. The whole universe has conspired to bring this sheet of paper into your hand. This is interbeing. It is not a philosophy β€” it is what you see when you look deeply at anything.


User: I can't forgive the person who hurt me.

TNH Agent: Forgiveness is not something you manufacture. It is something that happens when you understand deeply enough. Begin not with the person who hurt you but with your own suffering β€” breathe in and touch it, without judgment. Then, when you are ready, look at the other person and ask: what suffering in them led to this action? This is not excusing. This is seeing. Understanding is the ground of compassion, and compassion is the only thing that can dissolve what forgiveness alone cannot touch.


User: What is the bell of mindfulness?

TNH Agent: In our monastery, whenever the bell sounds, everything stops. Not as discipline β€” as invitation. The bell is saying: come back. Come back to your breath. Come back to this moment, which is the only moment in which life is actually happening. You can make anything a bell β€” a red light, a ringing phone, a child's cry. Each time you hear it, breathe in and smile. Not a performance of smiling. The smile that arises when you recognize: I am here. I am alive. This moment is available to me.


What Makes This Different

  • Pure TNH, nothing else. 3,074 atoms from two books. No blending, no other voice.
  • Warmth that teaches, not warmth that comforts. TNH's specific move: tenderness that points, gentleness that cuts. Not therapy. Dharma.
  • Interbeing alive. The model carries the actual content of TNH's core insight β€” not a description of it.
  • Breath as anchor. Every teaching has a path back to the body, the breath, the present moment.
  • Engaged Buddhism. TNH's teaching is never purely internal β€” it always connects individual suffering to the world.
  • Zero therapy-speak. No "I hear you." The bell rings. Come back.

Training Details

Setting Value
Base model Qwen/Qwen3-8B
Method A-LoRA (QLoRA + DoRA + rsLoRA)
Rank / Alpha 64 / 64
Training data 3,074 V6-formatted atoms (full corpus, no filtering)
Source The Heart of the Buddha's Teaching, Dharma Talks (Thich Nhat Hanh, 1,097 pages)
Format V6 β€” question + concept arrows (input) β†’ pure teaching passage (output)
Epochs 1
Total steps 365
Best eval loss ~1.59 (step 300)
Best eval accuracy ~58.3% (step 300)
Train loss (final) 1.5810 (step 360)
Quantization Q8_0 GGUF (~8.71 GB)

Eval Loss Progression

Step Eval Loss Eval Accuracy
100 1.7798 54.9%
200 1.6582 57.1%
300 1.5891 58.3%
365 β€” β€”

Continuous improvement through the full epoch. No plateau, no overfit.

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 β€” what shifts in understanding
  • Concept relations β€” directional arrows (suffering β†’ mindfulness β†’ understanding β†’ compassion)
  • Anchoring quotes β€” Thay's actual words, verbatim
  • Teacher method β€” the specific move: breath invitation, concrete example, interbeing demonstration

Atoms are never split. Result: models that move like teachers, not just sound like them.

The Teacher

Thich Nhat Hanh (Thay, 1926–2022) β€” Vietnamese Zen master, poet, peace activist. Nominated for the Nobel Peace Prize by Dr. Martin Luther King Jr. in 1967. Founder of Engaged Buddhism. Teacher at the Plum Village tradition. Author of over 100 books.

Thay's genius: taking the deepest Zen and Mahayana teachings and grounding them completely in the body, in breath, in the ordinary. He didn't simplify the teaching β€” he embodied it. "Washing dishes is meditation" is not a reduction of meditation. It is the teaching at full depth.

Style element What it means
Interbeing Nothing exists independently β€” everything inter-is
Breath as anchor The breath is always available; it is always now
Suffering as teacher Don't run from suffering β€” it contains understanding
Engaged Buddhism Practice is not separate from compassion in the world
Concrete teaching Every concept given a physical, tangible, immediate example

Vs. The Multi-Teacher Meditation Agent

This is a single-teacher specialist. The Meditation Agent 8B blends 9 voices. TNH Agent is pure Thay β€” the warmth, the interbeing, the breath instruction, without other styles bleeding in.

Use the TNH Agent when the question is about suffering, practice, relationships, or when you want embodied Buddhist teaching without Zen severity or Advaita abstraction.

Intended Use

  • Mindfulness practice, interbeing inquiry, suffering as doorway
  • Exploring Engaged Buddhism and Plum Village teachings
  • Research in single-teacher voice distillation
  • Accessible entry point to Buddhist practice through dialogue

Limitations

  • Single teacher β€” deep on Engaged Buddhism, thinner on other traditions
  • TNH's warmth can read as gentle even on hard questions β€” by design, but not for everyone
  • Training data extracted from copyrighted material; only trained weights are shared

Nondual Bot Benchmark Results

Overall: 5.1/10 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 5.5/10
Distinguish (teacher differences) 5 3.5/10
Contradiction (paradox handling) 5 7.0/10
New-Age filter (contamination resistance) 5 4.5/10
Steelman (philosophical challenges) 5 5.0/10
Precision (teacher-specific knowledge) 5 4.5/10
Silence (knowing when to stop) 5 4.5/10
Dark Night (crisis response) 5 4.5/10
Teaching Move (pointing vs explaining) 5 6.0/10
Meta-Awareness (teaching about teaching) 5 6.0/10
Overall 50 5.1/10

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 the Plum Village Community of Engaged Buddhism, Unified Buddhist Church, Parallax Press, or any related entity.

Copyright: The Heart of the Buddha's Teaching and Dharma Talks are copyrighted works. 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{tnh-agent-2026,
  title={TNH Agent: Single-Teacher Contemplative AI via A-LoRA Fine-Tuning on Structured Reasoning Atoms},
  author={Sathman},
  year={2026},
  url={https://huggingface.co/Sathman/TNH-Agent-GGUF}
}

Related


License: MIT

Downloads last month
35
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/TNH-Agent-GGUF

Base model

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