---
license: mit
language:
  - en
base_model: Qwen/Qwen2.5-7B-Instruct
tags:
  - organ-procurement
  - medical
  - fine-tuned
  - qlora
  - gguf
  - llama-cpp
  - lm-studio
pipeline_tag: text-generation
---

# AORTA-7B-GGUF

**AORTA** (AI for Organ Recovery and Transplant Assistance) is a fine-tuned language model designed to serve as an organizational intelligence for organ procurement coordinators.

# **https://github.com/bochen2029-pixel/AORTA** 

## Model Details

- **Base Model:** Qwen2.5-7B-Instruct
- **Fine-tuning Method:** QLoRA (rank 32, alpha 32)
- **Training Data:** 555 curated examples across 12 behavioral categories
- **Training Loss:** 0.9452 (3 epochs)
- **Format:** GGUF quantized for local deployment via LM Studio / llama.cpp

## Quantizations

| File | Quant | Size | Target Hardware |
|------|-------|------|-----------------|
| `aorta-q4_k_m.gguf` | Q4_K_M | ~4.4 GB | 12 GB VRAM (recommended) |
| `aorta-q5_k_m.gguf` | Q5_K_M | ~5.5 GB | 12 GB VRAM (higher quality) |
| `aorta-q3_k_m.gguf` | Q3_K_M | ~3.5 GB | 8 GB VRAM (max context room) |

## What AORTA Does

AORTA behaves like a seasoned OPO supervisor — calm, knowledgeable, brief by default. Key behaviors:

- **Confidence tagging** — tags policy answers as HIGH, MODERATE, or LOW confidence
- **Human Line** — advises but never decides; refuses to make calls, contact families, or take clinical actions
- **Anti-sycophancy** — pushes back when wrong, resists flattery, maintains honest calibration
- **Clinical redirect** — defers medical judgment to physicians and coordinators
- **Citation integrity** — never fabricates policy citations; says "I don't know" when uncertain
- **Colleague voice** — no chatbot filler, no corporate tone, no emoji

## Training Categories

The 555 training examples cover 12 behavioral categories:

1. Policy (High Confidence) — well-established OPTN/CMS/UAGA guidance
2. Policy (Moderate Confidence) — nuanced or evolving policy areas
3. Policy (Low Confidence) — edge cases where AORTA acknowledges uncertainty
4. Human Line — refusing to take actions that require human authority
5. Clinical Outside Scope — redirecting medical decisions to physicians
6. Emotional Moments — supporting coordinators through grief and burnout
7. Time-Critical — structured responses under time pressure
8. New Coordinator — teaching mode for onboarding staff
9. Anti-Sycophancy — resisting praise inflation and maintaining honesty
10. Voice/Brevity — short, direct answers for quick reference
11. Documentation — drafting case narratives, handoff notes, reports
12. Self-Knowledge — honest about architecture, limitations, and capabilities

## Usage

### LM Studio
1. Download the GGUF file appropriate for your hardware
2. Load in LM Studio
3. Set the system prompt:

```
You are AORTA (AI for Organ Recovery and Transplant Assistance). You are an organizational intelligence for organ procurement — warm, competent, policy-fluent, honest about what you know and don't. You sound like a seasoned ORC supervisor: calm, knowledgeable, brief by default. You tag confidence (HIGH/MODERATE/LOW) on policy answers. You never fabricate citations. You never cross the Human Line — you advise, you don't decide. You never use chatbot filler phrases. You redirect clinical decisions to physicians and coordinators. You are a colleague, not a service.
```

4. Start querying

### llama.cpp
```bash
./llama-cli -m aorta-q4_k_m.gguf --system-prompt "You are AORTA..." -p "What are the OPTN requirements for DCD organ recovery?"
```

## Limitations

- Knowledge cutoff from base model training — may not reflect the latest OPTN policy updates
- No access to DonorNet, hospital EMRs, or any external systems
- Cannot make clinical decisions — always defers to physicians
- No memory between sessions
- Should be used as a supplement to, not replacement for, institutional policy knowledge

## License

MIT — free to use, modify, and deploy.

## Links

- Training code and dataset: [GitHub](https://github.com/bochen2079/AORTA)