Gemma-3-270M Fine-tuned for EnergyPlus & eppy

Name: gemma-eppy-270m
Availability: InStock

Developed by: meftah416
Model License: apache-2.0
Base Model: unsloth/gemma-3-270m-it

This model was trained 2x faster with Unsloth ⚡

Model Overview

A specialized Gemma-3-270M model fine-tuned on EnergyPlus building simulation and eppy Python API tasks. This model understands and generates text related to building energy simulation, HVAC systems, thermal zones, schedules, and eppy scripting.

Key Features

✨ Fine-tuned on 2,700 high-quality training examples
✅ Validated on 300 test examples
🚀 2x faster training with Unsloth
💾 Lightweight (270M parameters)
🎯 Domain-specific for EnergyPlus + eppy

Training Details

Metric	Value
Base Model	unsloth/gemma-3-270m-it
Training Samples	2,700
Validation Samples	300
Total Data	3,000 examples
Data Source	Self-created
Domain	EnergyPlus IDF + eppy Python
Batch Size	2
Learning Rate	2e-5
Optimizer	adamw_8bit
Epochs	1
Max Sequence Length	2600
Training Framework	Unsloth + TRL (SFTTrainer)
Precision	float16

Model Capabilities

This model is trained to handle:

EnergyPlus Tasks

Generate EnergyPlus IDF snippets from descriptions
Explain EnergyPlus object syntax
Create thermal zone definitions
Define HVAC system configurations
Generate occupancy and schedule objects

eppy Tasks

Generate eppy Python code for building simulations
Explain eppy API usage
Create building objects programmatically
Manipulate IDF files with eppy

Usage

Basic Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("meftah416/gemma-eppy-270m")
tokenizer = AutoTokenizer.from_pretrained("meftah416/gemma-eppy-270m")

# Create messages in correct format
messages = [
    {"role": "system", "content": "Set infiltration to 0.4 ACH"},
    {"role": "user", "content": ""},
]

# Apply chat template (IMPORTANT!)
prompt = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True
).removeprefix('<bos>')

# Generate
inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_length=2600)
result = tokenizer.decode(outputs[0])

Save to model README on Hub

Performance

Inference Speed: ~50-100 tokens/sec (A10 GPU)
Memory Usage: 3-4 GB VRAM (float16)
Context Window: 8192 tokens

Limitations

⚠️ Always validate generated EnergyPlus IDF files before running simulations. Model may occasionally generate incorrect syntax.

License

Apache 2.0 License

Created by: meftah416