HandsOnAid：4-bit Gemma 2B 离线急救助手实测
HandsOnAid: Building and Evaluating an Offline First-Aid Assistant for Emergency Response Scenarios
完全离线的手机急救助手，Gemma 2B 量化微调后 15 秒内出答案，还请了真医生盲评。论文把七种配置的得分差距摊开讲，训练数据缺了 ANZCOR 细节导致分数上不去。做端侧垂直应用的人能从这套评测方法直接抄作业。
一个研究团队构建了 HandsOnAid，将 Gemma 2B 以 QLoRA 4-bit 微调，用于无信号、无救护车场景下的离线急救建议。训练数据是自建的 FirstAidQA 数据集，包含 5,550 条急救问答对；团队对比了 7 种配置，包括基础模型、4-bit 与 8-bit 微调版本，以及加入长度下限、安全门控或检索增强的变体。评估采用 41 对经医学审核、遵循 ANZCOR 指南的问答库，配合 0-5 分评分细则与 12 条安全否决项，由 GPT-5.6 Sol、Claude Opus 5、DeepSeek V4 Pro、GLM-5.3 四个前沿模型盲评，并用 45 个隐藏对照检验评委可靠性。
结果显示微调使平均分从基础模型提升 +1.08 至 2.34，最佳配置为“4-bit 微调 + 检索”，得分 2.51/5。一名医学专业人员对 40 条答案做了盲评，分数与 LLM 评委均值接近。关键词审计发现训练语料对 ANZCOR 若干细节覆盖不足，可能是低分的成因。在 Pixel 7 Pro 上用 6 个 CPU 线程完全离线运行，4-bit 微调模型平均 6.86 秒输出首段文本，14.88 秒生成完整答案。
这项工作的价值在于给出了小模型在移动端离线跑垂直任务的完整评测框架：微调收益明确但绝对分数仍低，训练数据质量与覆盖度成为瓶颈，对做端侧垂直助手的开发者有直接参考意义。
HandsOnAid: Building and Evaluating an Offline First-Aid Assistant for Emergency Response Scenarios
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.
This report is categorized under AITOP's "Papers" desk. Within our cross-regional analytical framework, it directly maps to AI developer toolchains, coding agents, and engineering workflows. Key entities and technical topics under continuous tracking include Gemma, QLoRA, 急救, 移动端部署.
Source attribution: curated from arXiv cs.LG (@Rezwan Islam Salvi, Saiyma Sittul Muna, Mushfiqur Rahman Mushfique).
