AnukeerthiReddy/llama-3.1-8b-clinical-fhir-lora

LLaMA 3.1 8B — Clinical Notes to FHIR

Fine-tuned adapter converting unstructured clinical notes into structured FHIR R4 JSON.

Model Details

• Base model: unsloth/meta-llama-3.1-8b-bnb-4bit
• Method: QLoRA (4-bit quantization + LoRA adapters)
• LoRA config: r=16, alpha=32, dropout=0.05
• Trainable params: ~20M / 8B (0.25%)
• Hardware: NVIDIA L40 (47.7 GB VRAM)
• Training time: ~4 minutes

Training Metrics

Metric	Value
Train loss (epoch 1)	2.398
Train loss (epoch 3)	1.394
Eval loss	1.389
Peak VRAM	8.3 GB

Usage

from unsloth import FastLanguageModel
from peft import PeftModel

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name="unsloth/meta-llama-3.1-8b-bnb-4bit",
    max_seq_length=2048,
    load_in_4bit=True,
)
model = PeftModel.from_pretrained(model, "AnukeerthiReddy/llama-3.1-8b-clinical-fhir-lora")

note = "Patient: 58M with chest pain radiating to left arm x 2h. HTN, T2DM. BP 158/92."
inputs = tokenizer(note, return_tensors="pt").to("cuda")
outputs = model.generate(**inputs, max_new_tokens=512)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Dataset

ai-galileo/clinical-notes-to-fhir

Training Infrastructure

• GPU: NVIDIA L40 (47.7 GB) — Georgia State University
• Framework: Unsloth + TRL SFTTrainer
• Logging: Weights & Biases