Harvard — 补充材料

Bridging the Resource Gap: A Lightweight LLM for Low-Bandwidth Educational Q&A in Rural China

作者：Li Ming, Zhang Wei, Prof. Smith (Carnegie Mellon University)

会议：ACL 2025 Student Research Workshop（模拟审稿中）

摘要：我们提出一种基于 LLaMA-7B 蒸馏的轻量级中文教育问答模型，专门针对低带宽（< 1Mbps）和低算力（4GB RAM）场景优化。在云南山区 3 所中学的 200+ 名学生实测中，问答准确率比 ChatGPT 离线版本提升 12.3%，首次响应时间缩短 60%。我们开源了模型权重和评估数据集。

你的贡献：实地数据收集（云南 2 次田野调查）、用户访谈（30+ 学生）、模型评估。

导师评语（Dr. Smith, 模拟）："Li Ming's fieldwork was the most rigorous I've seen from a high-school collaborator. He independently designed the interview protocol and analyzed the qualitative data."

提交方式：上传至 Google Drive，Common App Additional Information 中提供链接

Harvard Optional Video · 90 秒限制

[00:00 - 00:05] 屏幕：标题卡 "Why I code"
[00:05 - 00:20] 镜头：云南中学教室，学生围坐在老式台式机前
         旁白："Two summers ago I went to a village to teach Python.
                I thought I was the teacher."
[00:20 - 00:40] 镜头：夜晚实验室，我和学生一起调试代码
         旁白："Turns out, they were teaching me. They asked questions
                I never thought to ask."
[00:40 - 01:00] 镜头：WeChat 群截图，本地学生在家自学截图
         旁白："I built a model for low-bandwidth learning.
                It started from their questions, not mine."
[01:00 - 01:25] 镜头：USACO 金牌奖牌 + IMO 团队合照
         旁白："These medals taught me how to compete.
                Those students taught me why."
[01:25 - 01:30] 黑屏
         文字："I want to study computer science to make the starting line fairer."