---
license: apache-2.0
base_model: HuggingFaceTB/SmolLM3-3B-Base
tags:
  - contemplative-ai
  - fine-tuned
  - gguf
  - lora
  - qlora
  - smollm3
  - nondual
  - teaching
  - spirituality
  - awareness
  - advaita
  - meditation
language:
  - en
pipeline_tag: text-generation
---

# Meditation Agent (SmolLM3 3B) — Contemplative Teaching AI

This is the 3B branch of the Meditation Agent series, built on
[HuggingFaceTB/SmolLM3-3B-Base](https://huggingface.co/HuggingFaceTB/SmolLM3-3B-Base)
and fine-tuned with the A-LoRA V6 recipe for contemplative teaching.

All 9 teachers blended — Osho, Thich Nhat Hanh, Nisargadatta, Krishnamurti, Eckhart Tolle, Alan Watts, Atmananda, Rupert Spira, Pema Chodron. No system prompt required. Question in, teaching out.

## 50-question eval summary

This 3B branch was run through a raw 50-question eval after GGUF conversion.

- `Q8_0`: completed `50/50` with `0` request failures and is the
  highest-fidelity public quant
- `Q5_K_M`: completed `50/50` with `0` request failures and is the recommended
  default public quant
- `Q3_K_M`: completed `50/50` with `0` request failures, but is weaker and more
  generic than `Q5_K_M`
- overall read: strong stability for a 3B model, but still below the larger
  Meditation Agent branches in teacher-specific nuance and factual reliability

## Final training setup

| Setting | Value |
|---------|-------|
| Base model | [HuggingFaceTB/SmolLM3-3B-Base](https://huggingface.co/HuggingFaceTB/SmolLM3-3B-Base) |
| Method | A-LoRA V6 |
| Format | Question + concept arrows in, pure teaching passage out |
| Data exported | 24,031 atoms |
| V6 formatted set | 17,088 examples after opener cap |
| Train / eval split | 16,233 / 855 |
| Adapter recipe | QDoRA + rsLoRA, rank 32, alpha 32 |
| Epochs | 1 |
| Max sequence length | 1536 |
| Completion-only loss | Yes |
| NEFTune | alpha 5 |

## Training result

Merged checkpoint: `checkpoint-2000`

| Checkpoint | Eval loss | Eval token accuracy |
|------------|-----------|---------------------|
| 500 | 1.7580 | 0.5554 |
| 1000 | 1.6840 | 0.5686 |
| 1500 | 1.6396 | 0.5771 |
| 2000 | 1.6338 | 0.5781 |

## Files

| File | Size | Use |
|------|------|-----|
| `Meditation_Agent-SmolLM3-3B-Q8_0.gguf` | 3.05 GB | Highest fidelity |
| `Meditation_Agent-SmolLM3-3B-Q5_K_M.gguf` | 2.06 GB | Recommended default |
| `Meditation_Agent-SmolLM3-3B-Q3_K_M.gguf` | 1.46 GB | Smallest, most brittle |
| `Meditation_Agent-SmolLM3-3B-BF16.gguf` | 5.74 GB | Archive / conversion source |

## Individual Teacher 3B Specialists

Each teacher also has their own dedicated 3B model — same SmolLM3-3B base, trained on single-teacher data only. Use these when you want one specific voice rather than the blended multi-teacher model.

| Teacher | Repo |
|---------|------|
| Osho | [Osho-Agent-SmolLM3-3B-GGUF](https://huggingface.co/Sathman/Osho-Agent-SmolLM3-3B-GGUF) |
| Thich Nhat Hanh | [TNH-Agent-SmolLM3-3B-GGUF](https://huggingface.co/Sathman/TNH-Agent-SmolLM3-3B-GGUF) |
| Nisargadatta | [Nisargadatta-Agent-SmolLM3-3B-GGUF](https://huggingface.co/Sathman/Nisargadatta-Agent-SmolLM3-3B-GGUF) |
| Atmananda | [Atmananda-Agent-SmolLM3-3B-GGUF](https://huggingface.co/Sathman/Atmananda-Agent-SmolLM3-3B-GGUF) |
| Krishnamurti | [Krishnamurti-Agent-SmolLM3-3B-GGUF](https://huggingface.co/Sathman/Krishnamurti-Agent-SmolLM3-3B-GGUF) |
| Eckhart Tolle | [Tolle-Agent-SmolLM3-3B-GGUF](https://huggingface.co/Sathman/Tolle-Agent-SmolLM3-3B-GGUF) |
| Alan Watts | [Watts-Agent-SmolLM3-3B-GGUF](https://huggingface.co/Sathman/Watts-Agent-SmolLM3-3B-GGUF) |
| Rupert Spira | [Spira-Agent-SmolLM3-3B-GGUF](https://huggingface.co/Sathman/Spira-Agent-SmolLM3-3B-GGUF) |

## Release recommendation

Recommended default: `Meditation_Agent-SmolLM3-3B-Q5_K_M.gguf`

Use `Q8_0` when you want the strongest public 3B quant, `Q5_K_M` as the
balanced default, and `Q3_K_M` when size matters most. `BF16` is the archive
and further-conversion source.

## Positioning

This is the lightweight 3B Meditation Agent:

- much smaller than the 8B/Phi4 branches
- capable of direct contemplative answers without prompt scaffolding
- best suited for local inference where memory footprint matters

## Related Models

- [Full series — Meditation Agent Collection](https://huggingface.co/collections/Sathman/meditation-agent-contemplative-teacher-series-69c0ceca6e74d6f18c1445a8) — all 19 models
- [GitHub Source / Training Repo](https://github.com/Sathman-1/Alora---Expert-Voice) — training pipeline, configs, and release scripts
- [Meditation Agent 8B](https://huggingface.co/Sathman/Meditation-Agent-8B-GGUF) — larger Qwen3 branch with stronger teacher fidelity
- [Meditation Agent Phi4 14B](https://huggingface.co/Sathman/Meditation-Agent-Phi4-GGUF) — strongest larger branch with richer cross-tradition depth

---

*ellam sivamayam* — Everything is Shiva's expression.

*எல்லாம் சிவமயம்*