Part of the Green Tea replication package for Beyond the Need for Speed: Energy-Aware Code Generation via Simulation-Guided Reinforcement Learning (Rajput and Sharma). Energy-contrastive supervised fine-tuning. On the 143-problem held-out benchmark it achieves 1.29% CARET (Correctness-Adjusted Reduction in Energy Total).

Base model: deepseek-ai/deepseek-coder-6.7b-base
Training: Energy-contrastive supervised fine-tuning
Code and full replication: https://github.com/SMART-Dal/green-tea
Dataset (Zenodo): https://doi.org/10.5281/zenodo.21210099

Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer
m = AutoModelForCausalLM.from_pretrained('saurabh-singh-rajput/green-tea-deepseek-coder-6.7b-energy-sft', device_map='auto')
t = AutoTokenizer.from_pretrained('saurabh-singh-rajput/green-tea-deepseek-coder-6.7b-energy-sft')

Citation

bibtex

@misc{rajput2026greentea,
  title={Beyond the Need for Speed: Energy-Aware Code Generation via Simulation-Guided Reinforcement Learning},
  author={Rajput, Saurabhsingh and Sharma, Tushar},
  year={2026},
  note={Preprint}
}