06 jun Micro essay: What is Large Language Model (LLM)?
Time to read: 1 mins 45 secs
Before 2017, if you wanted to use a machine to handle text that humans wrote, such as news, social media posts, books, reports, etc., in a similar way to how humans do (reading, interpreting, describing, summarizing, making comparisons, etc.), you would generally train a specific language model to perform just one task. Imagine you would like a model to analyze the reviews on a restaurant page to determine whether people like or dislike the restaurant. That is a sentiment analysis task, and you would need a database of many restaurant reviews to train the model. An important detail is that each review in the training database should be labeled by a human, indicating whether it is a good or bad review. Constructing a training database, in general, is laborious and boring work. However, you are a hardworking person who labeled 2000 reviews as either ‘love it’ or ‘hate it.’ You trained the model, and… zap! You have your sentiment analysis model that can determine if a new review is good or bad. Everything is working just fine. But now, if you plan to summarize the restaurant reviews, can you use the sentiment analysis model? NO! You have to build a new model. The sentiment analysis model doesn’t work for summarization because it wasn’t trained for that. Once again, you have to label all 2000 reviews, but this time, provide a short phrase that describes the review. Then, you can train a model specifically for generating summaries. (If you haven’t quit your job at this point!) Wouldn’t it be nice if we had a general-purpose model that could understand the basics of any language, and we only needed to fine-tune it for our specific task? Well, in 2017, Large Language Models (LLMs) were created for exactly that purpose. LLMs are large, general-purpose language models that can be pre-trained and then fine-tuned to serve specific purposes. A large language model (LLM) is a neural network consisting of many parameters, typically billions of weights or more. It is trained on vast amounts of text, learning from billions or trillions of words, enabling it to generate text on any topic or domain. Unlike models trained for specific tasks like sentiment analysis or named entity recognition, LLMs are designed to be versatile and excel at a wide range of tasks. With LLMs, we can harness the power of machines to understand, summarize, generate, and predict new content. They provide a valuable tool for various language-related applications #llm #neuralnetworks #microessay #datascience #learning