Definitionپێناسە
A training stage in which people compare or rate a model's answers, and the is then trained with to give the kind of answers people preferred. It helps turn a raw into a helpful assistant.
قۆناغێکی ڕاهێنانە کە تێیدا خەڵک وەڵامە جیاوازەکانی بەراورد دەکەن یان نمرەیان پێدەدەن، پاشان مۆدێلەکە بە ڕادەهێنرێت تاکو ئەو جۆرە وەڵامانە بداتەوە کە خەڵک پێیان باشتر بووە. یارمەتی دەدات ببێتە یاریدەدەرێکی بەسوود.
Exampleنموونە
Reviewers saw two answers to the same question and chose the clearer, more honest one; thousands of such choices trained the model.
هەڵسەنگێنەران دوو وەڵامی هەمان پرسیاریان بینی و ڕوونتر و ڕاستگۆترەکەیان هەڵبژارد؛ هەزاران هەڵبژاردنی لەو جۆرە مۆدێلەکەیان ڕاهێنا.
Note on the Kurdish termتێبینی لەبارەی زاراوە کوردییەکە
Proposed: the registry's reinforcement-learning «فێربوونی بەهێزکردنەوە» + لە هەڵسەنگاندنی مرۆڤەوە (from people's judgements: "feedback" said as evaluation by people). Method described by Ouyang et al. 2022 (InstructGPT). The abbreviation RLHF is what students will meet.