# DailyLLM

> DailyLLM is a lightweight LLM-based system for generating context-rich activity logs and summaries from smartphone and smartwatch sensor data.

## Paper

Title: DailyLLM: Context-Aware Activity Log Generation Using Multi-Modal Sensors and LLMs
Authors: Ye Tian, Xiaoyuan Ren, Zihao Wang, Onat Gungor, Xiaofan Yu, Tajana Rosing
Venue: 2025 IEEE 22nd International Conference on Mobile Ad-Hoc and Smart Systems (MASS)
Pages: 372–380
Published: October 6, 2025
DOI: 10.1109/MASS66014.2025.00060

## Main Idea

DailyLLM integrates location, motion, environment, and physiological context from multimodal mobile and wearable sensors. It uses efficient feature extraction, structured prompting, and a lightweight 1.5B-parameter LLM to generate semantic activity logs and higher-level summaries.

## Key Contributions

- Context-aware multimodal activity understanding across four context dimensions.
- A public multimodal dataset for activity prediction, scene understanding, location description, activity-log generation, and summarization.
- Lightweight LLM-based generation using structured sensor features and prompt engineering.
- Measured deployment on personal computers and Raspberry Pi 5 edge hardware.

## Dataset

The DailyLLM Dataset is organized as JSONL dialogues across four task groups: Activity Prediction, Scene Understanding, Location Description, and Log Generation & Summarization. It derives data from public resources including UCI, HHAR, MotionSense, Shoaib, DCASE, and StudentLife. The release supports classification, natural-language generation, and multimodal-fusion research.

## Key Results

- Activity-log BERTScore: 81.65% precision, 85.38% recall, and 83.47% F1.
- Reported SOTA baseline BERTScore: 65.0% precision, 78.2% recall, and 70.4% F1.
- The paper reports an approximately 17-point BERTScore precision gain over the SOTA approach using LLaMA3-70B.
- Average PC log-generation latency: 2.22 seconds for DailyLLM versus 22.09 seconds for the reported SOTA approach, nearly a 10x speedup.
- On Raspberry Pi 5 with 8 GB memory, the paper reports 180 seconds for activity-log generation and 240 seconds for summarizing a two-hour log window.

## Resources

Project: https://gdfwj.github.io/DailyLLM/
Summary: https://gdfwj.github.io/DailyLLM/summary.md
Paper: https://arxiv.org/abs/2507.13737
PDF: https://arxiv.org/pdf/2507.13737
DOI: https://doi.org/10.1109/MASS66014.2025.00060
Code: https://github.com/gdfwj/DailyLLM
Dataset: https://huggingface.co/datasets/YeTianCS/DailyLLMDataset

This file is an additional machine-readable entry point for the project; it is not presented as a universal web standard.