|
所在平台: Udemy |
课程主页: https://www.udemy.com/course/nlp-engineer-interview-questions-small-language-models/
课程评论：没有评论
课程名称：生成式人工智能面试问题 - 小语言模型 - 第2部分 课程概述： 本课程精心设计的实践测试紧跟AI行业的最新进展，涵盖了深度与广度，重点关注小语言模型（SLM）的模型架构，如BERT、T5、BART、RoBERTa、ALBERT、ELECTRA、Flan-T5、Reformer、DistilBERT、MobileBERT、多语言BERT、SentenceBERT和SpanBERT等。课程内容还包括模型预训练、微调、知识蒸馏、Hugging Face库、Transformer架构、注意力机制、模型压缩技术、标记化与GLUE等重要主题。此外，课程中提供了来自顶尖科技公司的真实面试问题，以帮助学员更好地准备面试。 示例面试问题包括： 1. 多语言BERT（mBERT）如何处理跨语言任务？ 2. RoBERTa在训练期间与BERT在标记掩蔽方面有什么不同？ 3. BERT中Q、K和V矩阵的维度是如何确定的？ 4. WordPiece标记化是如何处理未知单词的？ 5. 对于一个嵌入大小为1024且有24层的BERT模型，如果词汇表大小为50000，嵌入层有多少个参数？ 6. 字节级BPE如何改进传统的基于单词的标记化方法？ 7. ELECTRA模型生成组件的主要训练目标是什么？ 8. FLAN-T5的哪些方面提高了其性能？ 9. 注意力掩码在Transformer模型中的作用是什么？ 10. 在知识蒸馏中通常使用什么作为“软目标”？ 课程涵盖的话题： - 小语言模型（SLMs）的模型架构，如BERT、T5等 - Hugging Face Transformers库 - 模型压缩技术 - 量化、剪枝、参数共享与知识蒸馏 - 模型的预训练、微调与特征提取技术 - 嵌入模型与编码器-解码器模型 - 深度学习在自然语言处理（NLP）中的应用 小语言模型（SLMs）因其更快的推理时间和较少的计算资源需求而越来越受到重视，尤其是在生成式AI和NLP工程师岗位中。这使得SLMs在资源受限的环境（如移动设备）中理想部署，训练和微调的成本效益提升，促使其更广泛的可定制化和可访问性。此外，SLMs经过适当训练后，可以在专业任务上实现高准确率，从而在性能与效率之间提供实用平衡。通过我们的在线课程“生成式AI面试问题 - 小语言模型”，全面准备生成AI工程师面试。
Our meticulously designed practice tests keeps pace with the AI industry's latest advancements, covers both depth and breath, while concentrating on the important topics including Model Architectures of Small Language Models (SLM) like BERT, T5, BART, ROBERTa, ALBERT, ELECTRA, Flan-T5, Reformer, DistilBERT, MobileBERT, Multilingual BERT, SentenceBERT, SpanBERT, Model Pretraining, Fine-tuning, Knowledge Distillation, Hugging Face library, Transformers Architecture, Attention Mechanism, Model Compression techniques, Tokenization, GLUE. Additionally, the course features real questions that are asked by leading tech companies.Sample Questions:1. How does multilingual BERT (mBERT) handle cross-lingual tasks?2. How does RoBERTa handle token masking differently than BERT during training?3. How are the dimensions of the Q, K, and V matrices determined in BERT?4. How does WordPiece tokenization handle unknown words during tokenization?5. For a BERT model with an embedding size of 1024 and 24 layers, how many parameters are in the embedding layer if the vocabulary size is 50,000?6. How does byte-level BPE improve upon traditional word-based tokenization methods?7. What is the main training objective of the ELECTRA model's generator component?8. What aspect of FLAN-T5 improves its performance over T5?9. What is the role of an attention mask in Transformer models?10. What is typically used as the "soft targets" in knowledge distillation?Topics Covered in the Course:-Model Architectures of Small Language Models (SLM) like BERT, T5, BART, ROBERTa, ALBERT, ELECTRA, Flan-T5, Reformer, DistilBERT, MobileBERT, Multilingual BERT, SpanBERT, SentenceBERT, XLM-ROBERTa, LongformerHugging Face Transformers LibraryModel Compression Techniques - Quantization, Pruning, Parameter sharing & Knowledge DistillationModel - Pretraining, Fine-tuning & Feature Extraction TechniquesEmbedding Models Encoder-Decoder ModelsDeep Learning for NLPSmall language models (SLMs) are gaining traction and are equally important as large language models (LLMs) to master for Gen AI and NLP Engineer roles because they offer faster inference times and require fewer computational resources, making them ideal for deployment in resource-constrained environments like mobile devices. They are more cost-effective to train and fine-tune for specific tasks, enabling broader accessibility and customization. Additionally, SLMs can still achieve high accuracy on specialized tasks when properly trained, providing a practical balance between performance and efficiency.Prepare comprehensively for Generative AI Engineer interviews with our Udemy course, "Gen AI Interview Questions - Small Language Models"
