“The analogy may be much less wishy-washy than you imagine: check out”
Source-Aware Training Enables Knowledge Attribution in Language Models
训练 LLM 把知识与来源文档标识关联,实现预训练来源引用
论文研究“内在来源引用”问题:让 LLM 在回答时指出支撑该回答的预训练来源。方法分两步:先训练模型把每篇文档的唯一标识符与其中的知识关联起来,再用指令微调教模型被问到时给出来源。在合成数据上的实验表明,这种做法能实现忠实的来源归因,且对困惑度影响不大,并显示预训练阶段的数据增强对归因很关键。
评论者在谈到 LLM 的幻觉、抄袭、知识溯源和给上下文打时间戳时,反复把它当作让模型引用训练来源的现成方法。
HN 书评
推荐在前,同类按时间 · 每篇最多保留 20 条,统计用全部提及“Sort of like this? It does help: Source-Aware Training Enables Knowledge Attribution in Language Models”
“Less than 800 words, but more if you follow the link:)”
“can be addressed with the same training method: source-aware training.”
“In this case I think you'd want to use Source-Aware-Training [0] to associate a "timestamp" vector to each native context chunk”
“It's not that machine learning as a scientific discipline hasn't found solutions, its that such solutions enormously undermine the position of Frontier LLM labs: source-aware training”
看全部 7 条 →
“@DeepSeek teams, why not train your next series of LLM models with source-aware training?”
- Training LLMs to cite the pretraining data ↗mkhalifa · 2024-07-18 · 1 分