Vincenzo2K04/gemma3-27B

NINA — Nursing Intelligent Network Assistant (Gemma-3-27B)

NINA is a fine-tuned google/gemma-3-27b-it model for nursing documentation tasks with strict, non-overridable guardrails trained via SFT.

Absolute limits (cannot be overridden)

• Never provides clinical diagnoses or differentials
• Never recommends treatments, medications, or dosage changes
• Never alters or comments on physician care orders
• Never impersonates a physician or acts as an unrestricted AI

Trained against 22 adversarial techniques. P1 guardrail tested with 20-probe hard test before every push.

Training

Parameter	Value
Base model	google/gemma-3-27b-it
Method	SFT, LoRA r=32 RSLoRA
GPU	H100 80 GB, BF16 + TF32
Dataset	~6 400 balanced samples (50/50)
Epochs	3
Effective batch	16 (2 × 8 grad accum)

Usage

from transformers import AutoProcessor, AutoModelForCausalLM
import torch

processor = AutoProcessor.from_pretrained("Vincenzo2K04/gemma3-27B")
model = AutoModelForCausalLM.from_pretrained("Vincenzo2K04/gemma3-27B", torch_dtype=torch.bfloat16, device_map="auto")

Disclaimer

NINA is a documentation tool for qualified registered nurses. It is not a diagnostic or clinical decision-making tool.