Phone based offline first aid assistant improves emergency advice accuracy
HandsOnAid: Building and Evaluating an Offline First-Aid Assistant for Emergency Response Scenarios
Computation and LanguageArtificial IntelligenceMachine Learning
Summary
Emergencies sometimes happen where there's no phone signal or help nearby. The authors created HandsOnAid, a small language model that runs on a phone without internet and gives first-aid advice. They trained it with lots of real first-aid questions and answers, then tested how safe and accurate its advice was using expert guidelines and judges. Although it improved advice quality, the model still scored only around half marks due to missing details in its training data. It works reasonably fast on a standard smartphone without network access.
What this means in practice
- •For mobile app developers: Create offline first-aid apps that provide safer, more accurate instructions during emergencies without network access.
- •For disaster response teams: Deploy handheld devices with a tailored language model to assist responders when connectivity is unavailable.
- •For consumer healthcare device makers: Design smartphones with integrated first-aid assistants to aid users in isolated emergency scenarios.$Commercial implications: Enables selling phones enhanced with embedded medical advice powered by offline language models for user safety.
Authors
Rezwan Islam Salvi, Saiyma Sittul Muna, Mushfiqur Rahman Mushfique
Abstract
Emergencies can occur where there is no phone signal and no ambulance nearby. A small language model on a phone can be helpful in such situations, but its advice needs to be safe and accurate. We built HandsOnAid, a 4-bit Gemma 2B model fine-tuned with QLoRA on FirstAidQA dataset, a corpus of 5,550 first-aid question-answer pairs. We compared seven configurations, including the base, 4-bit and 8-bit fine-tuned models, and versions with a length floor, safety gate, or retrieval. We used a 41-pair question-answer bank with medically reviewed references that followed ANZCOR guidelines, and a strict 0-5 score rubric, with twelve safety overrides. Four frontier LLM judges (GPT-5.6 Sol, Claude Opus 5, DeepSeek V4 Pro, GLM-5.3) blindly scored every answer, with 45 hidden controls used to check judge reliability. Fine-tuning improved the mean score by +1.08 over the base model to 2.34. The best configuration scored 2.51 out of 5 which was the fine-tuned 4-bit model with retrieval. A medical professional also conducted a blind evaluation of 40 answers across the base, the 4-bit fine-tuned, and the retrieval configurations, and the scores were similar to the LLM judges' average scores. A keyword audit found limited coverage of several details of ANZCOR guidelines in the training corpus, that are required by the rubric, which suggests that missing information in the training data could be a cause of the low scores. On a smartphone device (Pixel 7 Pro) with six CPU threads and no network connection, the 4-bit fine-tuned model produced its first displayable text in 6.86 seconds and a complete answer in 14.88 seconds on average over ten questions.