Llama 2 is an updated version of Llama 1 that increased its pretraining corpus by $$40\%$$. This expansion in training data allowed it to achieve performance levels that are competitive with many closed-source product models.

Llama 2

Llama 1 is an open-sourced large language model designed with a focus on computational efficiency during inference. By training on a larger number of tokens than was typical for its size, it managed to outperform much larger language models.

Claude

Large Language Models (LLMs) represent one of the most significant recent advances in NLP, enabling the creation of systems with human-like capabilities for understanding and generating natural language. A key strength of LLMs is their ability to overcome the limitations of traditional models that require task-specific training. Instead, LLMs learn from vast amounts of text through the simple objective of next-token prediction. This process allows them to acquire extensive general knowledge, which can be prompted for a wide array of tasks. Notably, these models have also demonstrated the ability to reason, a capability that addresses a traditionally challenging problem in AI.

Large Language Models (LLMs)

Dive into Deep Learning

The exceptional proficiency in token prediction developed by pre-trained large language models makes it feasible to reframe a wide range of NLP problems as text generation tasks. This is achieved by using a prompt to instruct the model, which then leverages its predictive power to generate the desired output. This approach effectively converts diverse problems into a unified text generation format, allowing a single LLM to perform many different tasks.

Transforming NLP Tasks into Text Generation with LLMs

In discussions about modern AI, the term 'Large Language Models' (LLMs) is frequently used as a shorthand for generative models like GPT. While the definition of LLM is broad enough to include other model types, such as BERT, the common focus is often on the generative category.

Generative LLMs as a Focus of Study

The field of Large Language Models encompasses several key areas of focus. These include the foundational steps involved in constructing such models and the significant challenges related to scaling them. Two prominent scaling issues are the methodologies for training models on vast amounts of data and the strategies for improving their capacity to process and manage very long texts.

Core Topics in LLM Development and Scaling

In the context of language modeling, the terms 'word' and 'token' are frequently used interchangeably to denote the fundamental units of text being processed. This convention is adopted for simplicity, despite the fact that the terms have distinct original meanings.

Interchangeable Use of 'Word' and 'Token' in Language Modeling

A significant paradigm shift has occurred in the application of language models. Traditionally, they functioned as integrated components within larger systems, for instance, by scoring candidate translations in statistical machine translation. In contrast, modern Large Language Models (LLMs) in generative AI are utilized as complete, standalone systems. They tackle NLP problems directly by leveraging their generative nature, typically by processing a textual description of a task and generating the desired output.

Comparison of Traditional vs. Modern Language Model Applications

A defining characteristic of Large Language Models is the combination of their immense power and the significant expense required to build them. While they possess advanced capabilities for solving complex tasks, their development is a costly process.

Power and Cost of Large Language Models

Contrary to the traditional view of diminishing returns, the modern perspective in NLP is that continued scaling of computational resources and training data volume consistently leads to better-performing language models. This sustained improvement has driven the community to develop increasingly larger models. Evidence supports this view, showing that even models trained on trillions of tokens can still achieve performance gains from additional data.

Modern View on Continued Performance Gains from Scaling

The impressive power of Large Language Models has ignited significant interest in both their foundational techniques and their real-world applications. This has fueled a rapid expansion of research, leading to a vast number of new models and methods. The field is evolving at such a high velocity that creating comprehensive literature reviews has become impractical. This rapid pace makes it challenging to stay current, but researchers can leverage general and topic-specific review papers to keep up with key developments.

Rapid Evolution and Research Landscape of LLMs

Large language models are fundamentally trained on the task of next-token prediction. This objective involves training the model to predict the most likely subsequent token in a sequence, given all the tokens that came before it.

Next-Token Prediction as the Training Objective for LLMs

The perception of language modeling's role in artificial intelligence has undergone a significant transformation. Initially regarded as a fundamental NLP technique with no clear path to achieving broader AI goals, its potential is now viewed differently. The practice of training models on large-scale data using simple word prediction tasks has unexpectedly led to the emergence of intelligent systems. These systems demonstrate the ability to acquire a degree of general knowledge, suggesting that this approach is a promising step towards more advanced artificial intelligence and inspiring further research into powerful foundation models.

Shift in Perspective on Language Modeling's Role in AI

A key capability of a single, well-trained Large Language Model is its ability to handle a wide variety of tasks and generalize to new ones with only minimal adaptation. This versatility stems from the general knowledge acquired during large-scale pre-training, where the model learns from vast amounts of text through the fundamental objective of word prediction.

Versatility and Generalization of LLMs

Soft prompting is a technique used to adapt a pre-trained Large Language Model to specific tasks. It involves prepending a sequence of trainable vectors, called 'prompt embeddings' (e.g., p₀, p₁), to the user's input embeddings. Unlike discrete text prompts, these soft prompts are continuous vectors that are optimized directly via backpropagation to steer the model's output, often without altering the LLM's original weights.

Soft Prompting

The training or fine-tuning of a Large Language Model involves adjusting its trainable parameters to improve performance on a task. This is achieved by calculating a 'Loss' value, which quantifies the difference between the model's predictions and the correct target outputs. This loss is then used in an optimization algorithm, like backpropagation, to update the parameters, such as the model's internal weights or the embeddings of a soft prompt.

LLM Training and Fine-Tuning

A technology firm needs to build systems for three different language-based tasks: summarizing long articles, translating user interface text, and answering frequently asked questions. They are evaluating two approaches. Approach 1 involves building a single, very large system trained on a vast and diverse collection of text from the internet, with the simple objective of learning to predict the next piece of text in a sequence. This one system would then be guided to perform all three tasks. Approach 2 involves developing three separate, specialized systems, each trained exclusively on a dataset tailored to one specific task (e.g., a dataset of article-summary pairs for the summarization system). Which statement best analyzes the core principle that distinguishes these two approaches?

Large Language Models (LLMs) are recognized for their powerful capabilities, but their creation is a financially demanding endeavor. Scaling up the training process requires immense computing resources—often hundreds or thousands of GPUs to train a model with tens of billions of parameters from scratch. This vast requirement drastically increases the cost, which is further compounded by the necessity of performing numerous training runs during the model development process.

High Cost of Building LLMs

A medical diagnostics company needs to develop a system that automatically classifies short patient reports into one of five predefined, highly specific disease categories. The system's accuracy must be extremely high and its behavior must be very predictable, as it will be used in a critical clinical workflow. The company has a large, high-quality dataset of reports, each expertly labeled with the correct category.

Two proposals are being considered:

1.  **Proposal A:** Use a state-of-the-art, massive language model that has been pre-trained on a vast corpus of general internet text. This model, which learned by predicting the next word in a sentence, has demonstrated impressive general knowledge and reasoning abilities. It would be adapted for the classification task.

2.  **Proposal B:** Design and train a new, smaller neural network from scratch, using only the company's own labeled dataset of patient reports. This model's architecture would be specifically tailored for this single classification task.

Evaluate the two proposals. Which proposal is more suitable for this specific application, and why? Justify your decision by contrasting the core strengths and potential weaknesses of each approach in the context of this high-stakes, narrow-domain problem.

Choosing the Right NLP Approach for a Specialized Task

Before the widespread adoption of very large, general-purpose language models, a common approach to solving a specific natural language processing task (like sentiment analysis or machine translation) was to train a model exclusively on a dataset curated for that single task. Contrast this traditional, task-specific approach with the modern approach. In your answer, analyze the key differences in terms of (1) the training objective and data, and (2) how the model is applied to solve a variety of tasks.

Paradigm Shift in Natural Language Processing

The development of Large Language Models (LLMs) has ushered Natural Language Processing (NLP) into a new era of research, enabling the field to make significant strides in solving historically difficult AI problems. By leveraging the advanced capabilities of LLMs to understand, generate, and reason with natural language, researchers are now able to build highly sophisticated conversational systems that can communicate with humans smoothly.

Solving Difficult NLP Problems with LLMs

With the significant achievements brought by Large Language Models (LLMs), Natural Language Processing (NLP) has entered a new era where previously difficult problems are actively being solved. A prominent example of this progress is the successful development of conversational systems capable of smoothly and naturally communicating with humans.

LLM-Powered Conversational Systems

Despite varying implementation details, many Large Language Models (LLMs) share a common foundational Transformer architecture designed for language modeling. These models earn the designation 'large' because they feature significant scale in both their depth (the number of stacked layers or blocks) and their width (the dimensionality of their internal representations).

Dimensions of Large Language Models: Depth and Width

Galactica is a large language model designed specifically for scientific tasks. Unlike many other models, it is not trained on a general corpus, yet it shows promising quantitative and scientific reasoning capabilities.

Galactica

OPT (Open Pretrained Transformers) is an open-sourced large language model that contributed to democratizing the research and practical application of advanced natural language processing models.

OPT (Open Pretrained Transformers)

BLOOM is an open-sourced large language model that helped democratize the use and research of large language models, making advanced text generation capabilities accessible to a broader research community.

BLOOM

Llama 1

Emergent abilities in large language models are complex capabilities that spontaneously appear in larger models but are absent in smaller ones. However, simply scaling up the model size does not inherently guarantee that the model will become better at following human instructions.

Emergent Abilities of Large Language Models

The BIG-Bench benchmark is a standard evaluation dataset used to assess and quantify the capabilities of large language models across diverse tasks. It serves as a rigorous testing ground to compare model performance against human baselines. For example, the 540-billion-parameter PaLM (Pathway Language Model) demonstrated its advanced capabilities by outperforming average human performance on the BIG-Bench benchmark.

BIG-Bench Benchmark

A visual language model is a multimodal architecture that extends traditional language modeling capabilities to process visual inputs alongside text. These models are designed to reason over multiple modalities, enabling them to perform tasks such as few-shot learning on visual data. They are often created by augmenting existing large language models with visual understanding components.

Learn Before

Related

Learn After