# HCUP LLM Humanizer Adapter

## Model Overview
- **Base Model:** mistralai/Mistral-7B-v0.3
- **Adapter Type:** LoRA (QLoRA 4-bit)
- **Purpose:** Humanized, non-AI-sounding clinical manuscript generation from HCUP administrative data

## Training Data
- **400 instruction pairs** from HCUP corpus (trinetx, nis, neds, hcup_general)
- **50 DPO preference pairs** for humanization style transfer

## Humanization Rules
- AVOID: Furthermore, It is important to note, In conclusion, Delve into, Notably
- USE: short declarative sentences, clinical connectors (Then, But, So)
- TONE: confident, direct, patient-centered clinical framing

## LoRA Config
- rank=16, alpha=32
- target_modules: q_proj, v_proj
- dropout: 0.05

## Usage
```python
from peft import PeftModel, AutoModelForCausalLM, AutoTokenizer
import torch

base = AutoModelForCausalLM.from_pretrained("mistralai/Mistral-7B-v0.3", torch_dtype=torch.float16, device_map="auto")
model = PeftModel.from_pretrained(base, "Sharpener9290/hcup-llm-humanizer")
tokenizer = AutoTokenizer.from_pretrained("mistralai/Mistral-7B-v0.3")
```