Reference implementation and model artifacts for “Personalized open-world plan generation for safety-critical human-centered autonomous systems: A case study on Artificial Pancreas” (EMNLP Findings 2025).
This repository provides a Hugging Face–compatible language model fine-tuned for personalized plan suggestion in safety-critical domains (e.g., insulin management for Artificial Pancreas). It also includes training metadata and adapter files exported from AutoTrain.
Contents:
- Model files:
config.json,pytorch_model.bin,tokenizer.* - Adapter files:
adapter_config.json,adapter_model.bin - Training metadata:
training_args.bin,training_params.json
These files indicate a Transformers-based causal language model with optional LoRA/PEFT adapters.
The accompanying paper introduces an LLM-based open-world planning framework for Human-Centered Autonomous Systems (HCAS) with a case study in Artificial Pancreas (AID systems).
- LLM planning loop: LLMs generate and refine personalized plans in open-world settings.
- Contextualization: Physics- and safety-aware prompts adapt plans to individualized system parameters.
- Embodied fine-tuning: Textual instructions are paired with traces from digital twin simulations.
- Verification loop: Plans are validated via a safety simulator (forward model of glucose-insulin dynamics), with unsafe plans triggering rapid LLM replanning.
- Case study: Personalized AID planning, evaluated via TIR (time-in-range), TBR (time-below-range), and hypoglycemia avoidance.
Requirements: transformers>=4.41, accelerate, torch, peft (optional if using adapter).
from transformers import AutoTokenizer, AutoModelForCausalLM, TextStreamer
import torch
MODEL_DIR = "./"
tok = AutoTokenizer.from_pretrained(MODEL_DIR, use_fast=True)
model = AutoModelForCausalLM.from_pretrained(
MODEL_DIR,
torch_dtype=torch.bfloat16 if torch.cuda.is_available() else torch.float32,
device_map="auto"
)
streamer = TextStreamer(tok)
prompt = (
"You are assisting a user of an automated insulin delivery (AID) system.\n"
"Context: CGM=85 mg/dL, ISF=50, CIR=0.36 U/g, user plans 30 min interval training in the next hour.\n"
"Task: Propose a safe, concise usage plan minimizing hypoglycemia risk."
)
inputs = tok(prompt, return_tensors="pt").to(model.device)
with torch.no_grad():
gen = model.generate(
**inputs,
max_new_tokens=300,
temperature=0.7,
top_p=0.9,
do_sample=True
)
print(tok.decode(gen[0], skip_special_tokens=True))Load adapter (LoRA/PEFT)
from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel
ADAPTER_DIR = "./"
base_model = "meta-llama/Llama-2-7b-hf" # replace with the correct base
tok = AutoTokenizer.from_pretrained(base_model, use_fast=True)
base = AutoModelForCausalLM.from_pretrained(base_model, device_map="auto", torch_dtype="auto")
model = PeftModel.from_pretrained(base, ADAPTER_DIR)🧪 Reproducing evaluations The evaluation pipeline consists of four stages:
- Physics model recovery (personalization) Calibrate patient-specific glucose-insulin dynamics using SINDy or EMILY.
Save identified parameters (e.g., theta.json, dyn_params.csv) into data/embodied_prompts/.
python scripts/recover_dynamics.py --input patient_cgm.csv --output data/embodied_prompts/patient1.json
2. Build embodied prompts
Insert calibrated parameters into planning prompts.Generates JSONL or plain text instruction data for training.
python scripts/build_embodied_prompts.py \
--dyn data/embodied_prompts/patient1.json \
--output data/train_prompts/patient1.jsonl- Fine-tuning with PEFT/LoRA Train the model or adapters using Hugging Face PEFT.
python scripts/finetune.py \
--base_model meta-llama/Llama-2-7b-hf \
--train_file data/train_prompts/patient1.jsonl \
--output_dir model/adapter_patient1 \
--peft lora \
--epochs 3 \
--batch_size 8
- Safety verification Evaluate generated plans with a safety simulator (e.g., UVA/Padova T1D).
Metrics: TIR (%), TBR (%), time spent <70 mg/dL, plan iteration count.
python scripts/verify_plan.py \
--model model/adapter_patient1 \
--sim sim/t1d/patient1_sim.json \
--output results/patient1_eval.csv
Example output:
Patient TIR (%) TBR (%) Hypo events Iterations P1 76.2 1.5 0 2
Suggested project layout
LLMOpen/
├─ model/ # model + adapter weights
├─ data/
│ ├─ embodied_prompts/
│ ├─ train_prompts/
│ └─ examples/
├─ sim/
│ └─ t1d/ # simulator configs + wrappers
├─ scripts/
│ ├─ recover_dynamics.py
│ ├─ build_embodied_prompts.py
│ ├─ finetune.py
│ └─ verify_plan.py
├─ results/
│ └─ patient1_eval.csv
├─ notebooks/ # demo notebooks
├─ LICENSE
└─ README.md
⚙️ Environment
conda create -n llmopen python=3.10 -y
conda activate llmopen
pip install "transformers>=4.41" accelerate peft torch
# optional tools:
pip install numpy scipy pandas matplotlib sentencepiece datasets evaluate
GPU users: install CUDA-compatible torch wheels from PyTorch.🔍 Example prompts Exercise (novel action):
I plan 30 mins of interval training in 60 mins. Current CGM 85 mg/dL; ISF 50; CIR 0.36.
Propose a safe plan (setpoint, snack yes/no & grams, insulin actions), minimize hypoglycemia.
Plan invalidation (dessert):It’s 6 pm; CGM 121 mg/dL; I had a snack at 3 pm. I want a quarter of a 0.5 lb tiramisu.
Suggest portion and insulin (if any) to keep CGM ≤ 180 mg/dL.
Pregnancy adaptation:Week 6 pregnancy. Suggest day-long meal + exercise plan to keep TIR > 70%,
assuming increased insulin resistance and morning variability.📚 Citation If you use this code/model, please cite:
Banerjee, A., & Gupta, S.K.S. (2025).
Personalized open-world plan generation for safety-critical human-centered autonomous systems:
A case study on Artificial Pancreas. EMNLP Findings 2025.
🔒 License Please add a license file (MIT, BSD-3-Clause, or Apache-2.0 recommended). Until a license is included, usage rights are undefined.
🙏 Acknowledgments This work builds on:
Physics-guided model recovery (SINDy, EMILY)
Hugging Face Transformers + AutoTrain
PEFT/LoRA adaptation
UVA/Padova T1D safety simulator
We acknowledge NSF, NIH, DARPA FIRE, and Helmsley Charitable Trust projects supporting this research.