ThoughtfulThings/phi-2-light-gguf

Thoughtful Lamp

These are the weights in gguf format for the "thoughtful lamp" from our paper Teaching Things To Think: Bootstrapping Local Reasoning for Smart(er) Devices.

The base model has been fine-tuned to take "action" and "explanation" oriented commands that a user might give to a smart lamp and generate either JSON actions or natural language explanations in response.

Please refer to the GitHub repo for more information.

Prompt Structure

The models are fine-tuned to accept prompts in this format:

Action

User: {command}
[SENSORS] {sensor json} [/SENSORS]
Output: [SETTINGS]

Introspection

User: {command}
[SENSORS] {sensor json} [/SENSORS]
[SETTINGS] {settings json} [/SETTINGS]
Output: [EXPLANATION]

Citation

If our work is helpful, please cite us:

@inproceedings{king2025teaching,
  title={Teaching Things To Think: Bootstrapping Local Reasoning for Smart (er) Devices},
  author={King, Evan and Yu, Haoxiang and Vartak, Sahil and Jacob, Jenna and Lee, Sangsu and Julien, Christine},
  booktitle={2025 IEEE International Conference on Pervasive Computing and Communications (PerCom)},
  pages={78--88},
  year={2025},
  organization={IEEE}
}

Contact

Contact Evan King with questions, or open an issue on the repository.