How meftah416/llm-energyplus works
Project type and purpose
• This repository is an ML pipeline / open-source tool for fine-tuning a large language model (LLM) to assist with interactive energy model development (EnergyPlus, IDF file generation, parametrics, etc.).
• The README confirms that the focus is on preparing datasets, prompt-engineering scripts, notebooks, and ultimately fine-tuning an LLM, with datasets hosted on Hugging Face.
File structure
Top-level files & directories:
• BuildingSurface.ipynb, BuildingSurface.py, fraction_geom_data.py, etc. – core scripts and notebooks processing building geometry and surfaces
• material_prompt.py, construction_prompt.py, full_idf_prompt.py, eplusPrompt.txt, etc. – prompt-engineering modules
• datasets/ – raw text docs of EnergyPlus I/O groups, plus helper notebooks (rag_example.ipynb, parametric_prompt.py)
• eplus_idf/ – sample .idf files used for testing or as examples
• web/ – web-scraping scripts to gather external data
• results/ – output IDFs and weather files
• train_test.py & trial models/ – code for training, evaluation, and dependency specification
• .gitignore, README.md – configuration and documentation
Patterns & components:
• Modular data ingestion → data preprocessing → prompt assembly → LLM fine-tuning → evaluation
• Notebooks (Jupyter) for interactive exploration
• Python scripts for batch processing and automation
• External integration: Hugging Face (datasets & model hosting), EnergyPlus (simulation engine)
README for additional insights
• States “all necessary python codes, datasets…to finetune an LLM model for human interactive energy model development”
• Points to Hugging Face for dataset storage
a. Main components:
1. Data Storage
– datasets/InputOutput text docs
– building-type prototypes
– eplus_idf examples
2. Data Ingestion & Preprocessing
– pdf_to_dataset.ipynb, importjson.py, web_scrap.py
– fraction_geom_data.py, parametric_prompt.py
3. Prompt Engineering
– material_prompt.py, construction_prompt.py, full_idf_prompt.py, eplusPrompt.txt
– IDF_prompt.txt, EnergyPlusEssentials.txt
4. LLM Fine-tuning & Training
– train_test.py
– trial models/context_increase01.py & requirements.txt
5. Evaluation & Results
– BuildingSurface.ipynb, datasets/ollamaRAG.ipynb, rag_example.ipynb
– results/*.idf, results/*.epw
6. External Services
– Hugging Face (dataset & model storage)
– EnergyPlus (simulation engine called via Python)
b. Relationships & interactions:
• Ingestion feeds Preprocessing modules → outputs cleaned training samples
• Samples + Prompt-Engineering feed into Fine-tuning
• Fine-tuned model pulled/pushed from Hugging Face
• Evaluation modules query the fine-tuned model → generate IDFs → EnergyPlus runs simulations → results stored
• Web-scraping may feed additional raw data into Data Storage
c. Architectural patterns & principles:
– Modular pipeline pattern (clear staging: ingest → transform → train → evaluate)
– Separation of concerns: notebooks for exploration vs. scripts for automation
– Extensibility: new prompt scripts or data sources can attach to the pipeline
d. Key technologies & libraries:
– Python 3.x, Jupyter Notebooks
– Hugging Face Datasets & Transformers
– EnergyPlus (via API or command-line calls)
– Standard data-processing libs (pandas, NumPy)