Exploring the World of Language Models: LLMs vs Chat Models

Introduction

Large Language Models have made significant advancements in the field of Natural Language Processing (NLP), enabling AI systems to understand and generate human-like text. ChatGPT is a popular language model based on Transformers architecture, enabling it to understand long texts and figure out how words or ideas are connected. It's great at making predictions about language and relationships between words.

LLMs and Chat Models are two types of models in LangChain, serving different purposes in natural language processing tasks. This lesson will examine the differences between LLMs and Chat Models, their unique use cases, and how they are implemented within LangChain.

Understanding LLMs and Chat Models

LLMs

LLMs, such as GPT-3, Bloom, PaLM, and Aurora genAI, take a text string as input and return a text string as output. They are trained on language modeling tasks and can generate human-like text, perform complex reasoning, and even write code. LLMs are powerful and flexible, capable of generating text for a wide range of tasks. However, they can sometimes produce incorrect or nonsensical answers, and their API is less structured compared to Chat Models.

Pre-training these models involves presenting large-scale corpora to them and allowing the network to predict the next word, which results in learning the relationships between words. This learning process enables LLMs to generate high-quality text, which can be applied to an array of applications, such as automatic form-filling and predictive text on smartphones.

Most of these models are trained on general purpose training dataset, while others are trained on a mix of general and domain-specific data, such as Intel Aurora genAI, which is trained on general text, scientific texts, scientific data, and codes related to the domain. The goal of domain specific LLMs is to increase the performance on a particularly domain, while still being able to solve the majority of tasks that general LLM can manage.

LLMs have the potential to infiltrate various aspects of human life, including the arts, sciences, and law. With continued development, LLMs will become increasingly integrated into our educational, personal, and professional lives, making them an essential technology to master.

You can follow these steps to use a large language model (LLM) like GPT-3 in LangChain. Import the OpenAI wrapper from the langchain.llms module and Initialize it with the desired model name and any additional arguments. For example, set a high temperature for more random outputs. Then, create a PromptTemplate to format the input for the model. Lastly, define an LLMChain to combine the model and prompt. Run the chain with the desired input using .run(). As mentioned before, remember to set your OpenAI key saved in the “OPENAI_API_KEY” environment variable before running the following codes. Remember to install the required packages with the following command: pip install langchain==0.1.4 deeplake openai==1.10.0 tiktoken.

from langchain.llms import OpenAI
from langchain.chains import LLMChain
from langchain.prompts import PromptTemplate

# Before executing the following code, make sure to have
# your OpenAI key saved in the “OPENAI_API_KEY” environment variable.
llm = OpenAI(temperature=0)

prompt = PromptTemplate(
  input_variables=["product"],
  template="What is a good name for a company that makes {product}?",
)

chain = LLMChain(llm=llm, prompt=prompt)

print( chain.run("wireless headphones") )

Wireless Audio Solutions

Here, the input for the chain is the string "wireless headphones". The chain processes the input and generates a result based on the product name.

Chat Models

Chat Models are the most popular models in LangChain, such as ChatGPT that can incorporate GPT-3 or GPT-4 at its core. They have gained significant attention due to their ability to learn from human feedback and their user-friendly chat interface.

Chat Models, such as ChatGPT, take a list of messages as input and return an AIMessage. They typically use LLMs as their underlying technology, but their APIs are more structured. Chat Models are designed to remember previous exchanges with the user in a session and use that context to generate more relevant responses. They also benefit from reinforcement learning from human feedback, which helps improve their responses. However, they may still have limitations in reasoning and require careful handling to avoid generating inappropriate content.

API Differences in LangChain

Chat Message Types

In LangChain, three main types of messages are used when interacting with chat models: SystemMessage, HumanMessage, and AIMessage.

SystemMessage: These messages provide initial instructions, context, or data for the AI model. They set the objectives the AI should follow and can help in controlling the AI's behavior. System messages are not user inputs but rather guidelines for the AI to operate within.

HumanMessage: These messages come from the user and represent their input to the AI model. The AI model is expected to respond to these messages. In LangChain, you can customize the human prefix (e.g., "User") in the conversation summary to change how the human input is represented.

AIMessage: These messages are sent from the AI's perspective as it interacts with the human user. They represent the AI's responses to human input. Like HumanMessage, you can also customize the AI prefix (e.g., "AI Assistant" or "AI") in the conversation summary to change how the AI's responses are represented.

An example of using ChatOpenAI with a HumanMessage: In this section, we are trying to use the LangChain library to create a chatbot that can translate an English sentence into French. This particular use case goes beyond what we covered in the previous lesson. We'll be employing multiple message types to differentiate between users' queries and system instructions instead of relying on a single prompt. This approach will enhance the model's comprehension of the given requirements.

First, we create a list of messages, starting with a SystemMessage that sets the context for the chatbot, informing it that its role is to be a helpful assistant translating English to French. We then follow it with a HumanMessage containing the user’s query, like an English sentence to be translated.

from langchain.chat_models import ChatOpenAI
from langchain.schema import (
  HumanMessage,
  SystemMessage
)

chat = ChatOpenAI(model_name="gpt-4", temperature=0)

messages = [
	SystemMessage(content="You are a helpful assistant that translates English to French."),
	HumanMessage(content="Translate the following sentence: I love programming.")
]

chat(messages)

AIMessage(content="J'aime la programmation.", additional_kwargs={}, example=False)

As you can see, we pass the list of messages to the chatbot using the chat() function. The chatbot processes the input messages, considers the context provided by the system message, and then translates the given English sentence into French.

Using the generate method, you can also generate completions for multiple sets of messages. Each batch of messages can have its own SystemMessage and will perform independently. The following code shows the first set of messages translate the sentences from English to French, while the second ones do the opposite.

batch_messages = [
  [
    SystemMessage(content="You are a helpful assistant that translates English to French."),
    HumanMessage(content="Translate the following sentence: I love programming.")
  ],
  [
    SystemMessage(content="You are a helpful assistant that translates French to English."),
    HumanMessage(content="Translate the following sentence: J'aime la programmation.")
  ],
]
print( chat.generate(batch_messages) )

LLMResult(generations=[[ChatGeneration(text="J'aime la programmation.", generation_info=None, message=AIMessage(content="J'aime la programmation.", additional_kwargs={}, example=False))], [ChatGeneration(text='I love programming.', generation_info=None, message=AIMessage(content='I love programming.', additional_kwargs={}, example=False))]], llm_output={'token_usage': {'prompt_tokens': 65, 'completion_tokens': 11, 'total_tokens': 76}, 'model_name': 'gpt-4'})

As a comparison, here's what LLM and Chat Model APIs look like in LangChain.

llm_output:  {'product': 'Translate the following text from English to French: Hello, how are you?', 'text': '\n\nBonjour, comment allez-vous?'}

chat_output:  content='Bonjour, comment ça va ?' additional_kwargs={} example=False

Conclusion

LLMs and Chat Models each have their advantages and disadvantages. LLMs are powerful and flexible, capable of generating text for a wide range of tasks. However, their API is less structured compared to Chat Models.

On the other hand, Chat Models offer a more structured API and are better suited for conversational tasks. Also, they can remember previous exchanges with the user, making them more suitable for engaging in meaningful conversations. Additionally, they benefit from reinforcement learning from human feedback, which helps improve their responses. They still have some limitations in reasoning and may require careful handling to avoid hallucinations and generating inappropriate content.

In the next lesson we’ll see how GPT-4 and ChatGPT can be used for context-aware chat applications via APIs.

Intel and the Intel logo are trademarks of Intel Corporation or its subsidiaries.

RESOURCES:

ChatGPT and the Large Language Models (LLMs).

ENGLISH — LEARN ABOUT ChatTGPT TECHNOLOGY AND APPLICATIONS

medium.com

Large language models (LLMs) vs. ChatGPT

ChatGPT is a large language model but not every LLM is ChatGPT. Discover interesting applications, how models are trained, and what this tech means for society.

www.thoughtspot.com

A Complete Guide to LangChain: Building Powerful Applications with Large Language Models

LangChain is a powerful framework that simplifies the process of building advanced language model applications.

notes.replicatecodex.com

You can find the code of this lesson in this online Notebook.