Inspector

---

ICLAD 2026 Paper:
Inspector: Conversational and Lightweight Analyzer of Analog Circuit Layouts Using LLM and CNNs

This repository contains training scripts for:

• LLM fine-tuning/classification (Inspector_llm.py)
• YOLO training/evaluation (Inspector_yolo.py)

The Inspector dataset is available on 🤗 Hugging Face.

1) Setup

Use Python 3.10+.

python -m venv .venv
# Linux/macOS
source .venv/bin/activate
# Windows (PowerShell)
# .venv\Scripts\Activate.ps1

pip install --upgrade pip
pip install -r requirements.txt

If you need a CUDA-enabled PyTorch build, install PyTorch from the official selector first: https://pytorch.org/get-started/locally/

2) Fine-tune LLM

Set your Hugging Face token (for gated model access):

# Linux/macOS
export HF_AUTH_TOKEN=<your_token>

# Windows PowerShell
# $env:HF_AUTH_TOKEN="<your_token>"

Run fine-tuning:

python llms/llm_ft.py \
  --data llms/task_identification_dataset.json \
  --length 2048 \
  --batch 16 \
  --epochs 10 \
  --max-new-tokens 128 \
  --seed 42

Outputs are written to:

• results_llm_ft/ (training/eval artifacts)
• ft_model/ (merged standalone model)

3) Train YOLO

Train with your YOLO data config:

python train/yolo_train.py \
  --model yolov8m \
  --data /path/to/data.yaml \
  --epochs 100 \
  --batch 16 \
  --imgsz 608 1600 \
  --seed 42

Useful optional flags:

• --device 0 or --device 0,1
• --project ./runs
• --name experiment_name
• --no-val and/or --no-test

Outputs include run logs, model weights, and summary.json inside the training run directory.

4) Notes

• Both scripts support reproducible seeds (--seed).
• Inspector_llm.py expects JSON rows with a messages field in chat format.
• Inspector_yolo.py expects a valid Ultralytics-style data.yaml.