saurabh-singh-rajput/green-tea-qwen2.5-coder-14b-energy-sft

TEXT GENERATIONConcurrent Unit Cost:1Model Size:14.8BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jul 5, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

saurabh-singh-rajput/green-tea-qwen2.5-coder-14b-energy-sft is a 14.8 billion parameter Qwen2.5-Coder model developed by Saurabhsingh Rajput and Tushar Sharma. It is specifically fine-tuned for energy-aware code generation, achieving 4.45% CARET (Correctness-Adjusted Reduction in Energy Total) on a 143-problem benchmark. This model focuses on optimizing code for energy efficiency rather than just speed or correctness, making it suitable for resource-constrained environments.

Loading preview...

Overview

saurabh-singh-rajput/green-tea-qwen2.5-coder-14b-energy-sft is a 14.8 billion parameter model based on the unsloth/Qwen2.5-Coder-14B architecture. Developed by Saurabhsingh Rajput and Tushar Sharma, this model is part of the Green Tea replication package, focusing on energy-aware code generation.

Key Capabilities

  • Energy-Contrastive Supervised Fine-Tuning: The model has undergone specialized fine-tuning to prioritize energy efficiency in generated code.
  • CARET Performance: Achieves a 4.45% CARET (Correctness-Adjusted Reduction in Energy Total) on a 143-problem held-out benchmark, indicating its effectiveness in reducing energy consumption while maintaining code correctness.
  • Code Generation: Built upon a coder-specific base model, it retains strong code generation capabilities.

Good For

  • Energy-Efficient Code Generation: Ideal for developers and researchers focused on minimizing the energy footprint of software.
  • Resource-Constrained Environments: Particularly useful for applications where energy consumption is a critical factor, such as edge computing, IoT devices, or sustainable software development.
  • Research in Sustainable AI: Provides a valuable tool and benchmark for exploring energy-aware AI and code optimization.

Citation

This model is associated with the preprint "Beyond the Need for Speed: Energy-Aware Code Generation via Simulation-Guided Reinforcement Learning" by Rajput and Sharma. The full replication package and code are available on GitHub, and the dataset is hosted on Zenodo.

Finetunes
1 models