# DailyLLM: Project Summary

## Problem and Motivation

Human activity recognition systems commonly map sensor measurements to isolated labels such as *walking*, *running*, or *sitting*. These labels are useful, but they omit much of the context needed to describe everyday life: where an activity occurred, what the surrounding environment was like, how the person moved, and what physiological state accompanied the activity. DailyLLM addresses the broader task of automatically generating readable, context-aware activity logs and higher-level summaries from mobile and wearable sensing.

## Main Idea

DailyLLM combines four context dimensions: **location, motion, environment, and physiology**. The system uses heterogeneous sensors found on smartphones and smartwatches. The paper groups these into spatial sensing such as GPS, barometer, Wi-Fi, and Bluetooth; motion sensing such as accelerometers, gyroscopes, and magnetometers; environmental sensing such as microphones, temperature, and light; and physiological sensing such as GSR, PPG, thermistors, and infrared sensors.

The method first aligns sensor streams within configurable time windows. It performs reverse geocoding and localization, annotates environmental measurements, and refreshes motion and physiological readings. Efficient feature extraction converts raw time-series signals into compact representations. Structured prompts then combine task instructions, semantic feature descriptions, time, location, and feature values. A locally fine-tuned 1.5B-parameter DeepSeek-R1-Distill-Qwen model uses these prompts to predict activity context, generate activity logs, and summarize logs over longer windows. LoRA is used for efficient fine-tuning.

## Released Dataset

The **DailyLLM Dataset** is released as JSONL dialogue records for four task groups: activity prediction, acoustic scene understanding, location description, and activity-log generation and summarization. It derives sensor and contextual examples from public resources including UCI, HHAR, MotionSense, Shoaib, DCASE, and StudentLife. The release supports classification, natural-language generation, and multimodal-fusion research and is distributed for research use under CC BY-NC-SA 4.0, subject to the original sources' terms.

## Contributions and Results

The paper presents an integrated activity-log generation and summarization system spanning all four context dimensions, releases a fine-grained activity-context dataset, and evaluates feature extraction and prompt-engineering strategies for multimodal sensor interpretation. For activity-log generation, DailyLLM reports BERTScore precision of 81.65%, recall of 85.38%, and F1 of 83.47%. The reported SOTA baseline achieves 65.0%, 78.2%, and 70.4%, respectively. The paper describes the precision difference as a 17% improvement over the SOTA approach using LLaMA3-70B.

On an RTX 4090 PC, average DailyLLM log-generation latency is 2.22 seconds, compared with 22.09 seconds for the reported SOTA approach, nearly a 10× speedup under the reported setup. The edge evaluation uses a Raspberry Pi 5 with 8 GB memory. It reports 180 seconds for log generation and 240 seconds for summarizing a two-hour activity-log window. These measurements demonstrate feasibility on constrained hardware, not real-time performance for every task.

## Publication and Resources

DailyLLM was published in the **2025 IEEE 22nd International Conference on Mobile Ad-Hoc and Smart Systems (MASS)**, pages 372–380. DOI: [10.1109/MASS66014.2025.00060](https://doi.org/10.1109/MASS66014.2025.00060).

- [Project page](https://gdfwj.github.io/DailyLLM/)
- [Paper and arXiv record](https://arxiv.org/abs/2507.13737)
- [Code repository](https://github.com/gdfwj/DailyLLM)
- [DailyLLM dataset](https://huggingface.co/datasets/YeTianCS/DailyLLMDataset)