---
language:
- en
license: apache-2.0
tags:
- mental-health
- psychiatry
- psychology
- qwen3
- medical
- therapy
- gguf
base_model: Qwen/Qwen3-4B-Thinking-2507
library_name: transformers
pipeline_tag: text-generation
---

# Luna 1.0 - Psychiatric AI Companion

## Model Description

Luna is a 4B parameter psychiatric AI trained through 8-stage curriculum learning on Qwen3-4B-Thinking-2507. Created by Dr. Suvadeep.

**Training Stages:**
- Stage 1-2: Psychiatric knowledge (DSM-5, medications, CBT, counseling)
- Stage 3-7: Empathy & conversation skills (30,000+ dialogues)
- Stage 8: Identity & anti-refusal training

**Capabilities:**
- DSM-5/ICD-11 diagnoses
- Medication recommendations with dosages
- CBT/DBT/ACT psychotherapy
- Crisis support without deflection
- Empathetic conversation

## Files

| File | Size | Device |
|------|------|--------|
| model.safetensors | 8 GB | Training/fine-tuning |
| Luna-4B-thinking-Q4_K_M.gguf | 2.5 GB | GTX 1050 Ti |
| Luna-4B-thinking-Q3_K_M.gguf | 1.8 GB | iPhone 15 (recommended) |
| Luna-4B-thinking-Q2_K.gguf | 1.3 GB | iPhone 15 compact |
| Luna-4B-thinking-Q8_0.gguf | 4.5 GB | High-end GPUs |

## Usage

### iPhone 15

Download Q3_K_M.gguf, use with LM Studio iOS.

### Desktop (GTX 1050 Ti)

```python
from llama_cpp import Llama

llm = Llama(
    model_path="Luna-4B-thinking-Q4_K_M.gguf",
    n_ctx=2048,
    n_gpu_layers=35
)

response = llm.create_chat_completion(
    messages=[{"role": "user", "content": "I feel depressed"}],
    max_tokens=1024
)
print(response["choices"][0]["message"]["content"])
```

## Training

- 8-stage curriculum learning
- LoRA (r=64, alpha=16)
- ~60,000 mental health conversations
- 20% replay buffers to prevent catastrophic forgetting
- Kaggle dual T4 GPUs

## Disclaimer

Research model only. Not a replacement for professional medical advice.

## License

Apache 2.0