Few-shot prompting
هەوێنی چەندنموونەیی
Proposed Kurdish term — under reviewزاراوەی پێشنیارکراو — لە ژێر پێداچوونەوەدایە Definitionپێناسە
Putting a few worked examples of the task (an input and the wanted output) into the , so that the model follows their pattern. The model's do not change: it uses the examples only inside this .
دانانی چەند نموونەیەکی ئامادەکراوی ئەرکەکە ( و ئەو دەتەوێت) لە ناو ، بۆ ئەوەی مۆدێلەکە پەیڕەوی شێوازیان بکات. ناگۆڕێن: تەنیا لە ناو ئەم نموونەکان بەکار دەهێنێت.
Picturesوێنەکان
A few-shot prompt: inside examples tags, three customer messages each with its label (question, order, complaint), then a new message that gets the label complaint.هەوێنی چەندنموونەیی: لە ناو نیشانەکانی examples، سێ پەیامی کڕیار هەر یەکەیان بە ناوی خۆیەوە (question، order، complaint)، پاشان پەیامێکی نوێ کە ناوی complaint وەردەگرێت.
Exampleنموونە
Three labelled customer messages ("order", "question", "complaint") before a new message make a few-shot prompt.
سێ نامەی کڕیار بە نیشانەکانیانەوە («داواکردن»، «پرسیار»، «سکاڵا») پێش نامەیەکی نوێ پێکدەهێنن.
Note on the Kurdish termتێبینی لەبارەی زاراوە کوردییەکە
Proposed: هەوێن + چەندنموونەیی (with a few examples). Brown et al. 2020 ("Language Models are Few-Shot Learners") made the idea known; Anthropic's docs also say "multishot prompting".