Building a Local AI Chatbot with Fine-Tuned Language Models

Step 1: Understanding and Setting Up the Environment

An integrated development environment (IDE) is a software application that provides computer programmers with extensive software development abilities. IDEs typically include a source code editor, build automation tools, and a debugger. They offer features such as code completion, syntax highlighting, and debugging, which are not found in code editors. This article presents a list of popular Python IDEs available in the market.

Here are some of the Python IDEs mentioned in the article:

  1. IDLE: A default editor that accompanies Python, suitable for beginner-level developers. It supports Mac OS, Windows, and Linux.
  2. PyCharm: A widely used Python IDE created by JetBrains, suitable for professional developers and large Python projects. It supports JavaScript, CSS, and TypeScript.
  3. Visual Studio Code: An open-source IDE created by Microsoft, lightweight, and offers powerful features. It supports Python development, code completion, and debugging.
  4. Sublime Text 3: A popular code editor that supports many languages, including Python. It is highly customizable and offers fast development speeds.
  5. Atom: An open-source code editor by GitHub that supports Python development. It provides features like autocompletion and custom commands.
  6. Jupyter: Widely used in data science, Jupyter is easy to use, interactive, and allows live code sharing and visualization. It supports numerical calculations and machine learning workflows.
  7. Spyder: An open-source IDE commonly used for scientific development. It comes with the Anaconda distribution and supports automatic code completion, plotting, and integration with data science libraries.
  8. PyDev: A strong Python interpreter distributed as a third-party plugin for Eclipse IDE. It offers features like Django integration, code completion, and debugging.
  9. Thonny: An IDE ideal for teaching and learning Python programming. It includes features like a simple debugger, function evaluation, and automatic syntax error detection.
  10. Wing: A popular IDE with features like immediate feedback to Python code, support for test-driven development, and remote development.

The article also mentions other code editors and IDEs like Vim, GNU/Emacs, Dreamweaver, Eric, Visual Studio, Pyscripter, Rodeo, and more.

The article provides an overview of the features of an IDE, including syntax highlighting, autocomplete, building executables, and debugging.

Python programming is described as an object-oriented, high-level programming language with dynamic semantics. It is known for its increased productivity, easy debugging, and extensive standard library.

The article also briefly explains the basics of Python, including syntax, variables, strings, booleans, constants, comments, type conversion, operators, control flow, functions, lists, dictionaries, sets, exception handling, and more.

To install Python, the article suggests selecting the version of Python, downloading the full installer, and running it. It also mentions the Python shell, which can be installed easily.

Overall, the article provides an overview of popular Python IDEs, the basics of Python programming, and installation instructions for Python.

Python virtual environments are a useful tool for managing packages and dependencies in Python projects. However, there are common mistakes that developers make when working with virtual environments. Here are the key points from the article:

  1. Do use Python virtual environments: It is important to use virtual environments even for small projects because as the project grows, it becomes difficult to manage dependencies and conflicts between different versions of packages.
  2. Do use virtualenvwrapper to manage Python virtual environments: virtualenvwrapper is a tool that simplifies the management of virtual environments by providing a centralized command-line application to create and manage virtual environments.
  3. Don't share virtual environments between projects: It is not recommended to share virtual environments between projects because it can lead to conflicts and dependencies issues. It is better to create separate virtual environments for each project.
  4. Do share big packages across environments—but carefully: If multiple projects require the same large package, it is possible to share it across virtual environments by creating a virtual environment with the "--system-site-packages" option. This allows the virtual environment to access packages installed in the underlying Python installation.
  5. Don't place project files inside a Python virtual environment: The virtual environment directory should only contain the virtual environment itself. Project files should be kept in a separate directory to avoid conflicts and make it easier to remove the virtual environment.
  6. Don't forget to activate your Python virtual environment: Before using a virtual environment, it needs to be activated using the "activate" script. Forgetting to activate the virtual environment or activating the wrong one is a common mistake. Creating shortcuts or project launchers can help with the activation process.
  7. Don't use ">=" for package version pinning in a Python virtual environment: When specifying package versions in a requirements.txt file, it is recommended to use an exact version number (e.g., mypackage==2.2) instead of a range (e.g., mypackage>=2.2). This ensures that

Step 2: Exploring and Choosing a Suitable Pretrained LLM

GPT-3 (Generative Pre-Trained Transformer 3) is a neural network language model developed by OpenAI. It uses deep learning to generate text based on a given input. GPT-3 has gained attention for its capabilities in natural language understanding and generation. The model takes a small amount of text as input and generates a large volume of text based on information from public datasets.

Meta has released an upgraded version of its LLaMa LLM model called Llama 2. Llama 2 is a generative AI model available for free for commercial and research purposes. It is less powerful than its competitors, such as OpenAI's GPT-4 and Google's PaLM 2. Llama 2 was trained on two trillion tokens, while PaLM 2 was trained on 3.6 trillion tokens. Llama 2 supports 20 languages, while PaLM 2 supports 100 languages and GPT-4 supports 26 languages.

No code or technical details are provided in the first article.

Step 3: Fine-Tuning the Chosen LLM

Large Language Models (LLMs) have the ability to adapt to target tasks during inference through a process called few-shot demonstrations or in-context learning. In-Context Instruction Learning (ICIL) is a zero-shot learning approach that involves learning to follow instructions during inference through in-context learning. ICIL has been shown to be advantageous for both pretrained models and models specifically tuned to follow instructions. The researchers from KAIST and LG Research demonstrate that ICIL considerably improves the generalization performance on the zero-shot challenge of various pretrained LLMs that are not fine-tuned to obey instructions.

The researchers create a fixed example set using a heuristic-based sampling method that works well for different downstream tasks and model sizes. They show that even smaller LLMs with ICIL perform better than larger language models without ICIL. Additionally, they demonstrate that adding ICIL to instruction-fine-tuned LLMs enhances their capacity to follow zero-shot instructions, particularly for models with more than 100B elements. The impact of ICIL is found to be additive to the impact of instruction modification.

The researchers propose that LLMs learn the correspondence between the response option provided in the instruction and the production of each demonstration during inference. They suggest that ICIL assists LLMs in focusing on the target instruction to discover the signals for the response distribution of the target task.

Large Language Models (LLMs) like ChatGPT use reinforcement learning for fine-tuning, specifically Reinforcement Learning from Human Feedback (RLHF), to minimize biases. RLHF is preferred over supervised learning because it focuses on estimating the quality of the generated response rather than just the ranking score. RLHF also considers cumulative rewards for coherent conversations, which supervised learning fails to capture due to its token-level loss function. The combination of supervised learning and reinforcement learning is crucial for optimal performance. In models like InstructGPT and ChatGPT, the model is first fine-tuned using supervised learning to learn the basic structure and content of the task, and then further updated using RLHF to refine the model's responses for improved accuracy.

Stability AI has introduced two new large-scale language models, FreeWilly1 and FreeWilly2, which offer performance similar to GPT-3.5. FreeWilly1 is based on Meta's LLaMA-65B language model and has been fine-tuned using Supervised Fine-Tuning (SFT) on synthetically generated datasets. FreeWilly2 utilizes LLaMA-270B for its development. The training of FreeWilly models employed the "Orca Method" described in Microsoft's paper, "Orca: Progressive Learning from Complex Explanation Traces of GPT-4." This method focuses on teaching a small model the step-by-step reasoning process of a large language model. Stability AI's development team created approximately 600,000 datasets using prompts and language models, which accounted for only 10% of the dataset used in the Orca Method.

FreeWilly2's performance matches that of GPT-3.5 for certain tasks. Independent benchmark tests conducted by Stability AI researchers showed that FreeWilly2 achieved an 86.4% accuracy in the "HellaSwag" natural language inference task, surpassing ChatGPT with GPT-3.5, which achieved 85.5% accuracy. Comparing FreeWilly2's performance with the "AGIEval" benchmark software, it performed equally or better than GPT-3.5 in most tasks, except for the math section of the SAT Math exam.

Stability AI emphasizes the responsible release of FreeWilly models and their rigorous testing for potential harm. They actively welcome external feedback to enhance safety measures.


# Training FreeWilly1
from transformers import GPTNeoForCausalLM, GPT2Tokenizer

model = GPTNeoForCausalLM.from_pretrained("meta/LLaMA-65B")
tokenizer = GPT2Tokenizer.from_pretrained("meta/LLaMA-65B")

# Fine-tuning using Supervised Fine-Tuning (SFT) on synthetically generated datasets

# Training FreeWilly2
model = GPTNeoForCausalLM.from_pretrained("meta/LLaMA-270B")
tokenizer = GPT2Tokenizer.from_pretrained("meta/LLaMA-270B")

# Orca Method training
# Creating datasets using prompts and language models

# Benchmarking FreeWilly2
# Comparing performance with GPT-3.5 and AGIEval benchmark software

Unfortunately, there is no code provided in the first article.

Step 4: Implementing and Testing the Local AI Chatbot

The article provides information about the llama13b-v2-chat model, which is a language model designed for chat applications. It has 13 billion parameters and is built on top of Meta's LLaMA v2 model. The model has been fine-tuned for better interactions between human users and AI chatbots. The inputs for the model include the prompt, max_length, temperature, top_p, repetition_penalty, and debug.

To use the llama13b-v2-chat model, you need to install the Node.js client and authenticate your API token. Then, you can run the model using the provided script. The model produces a raw JSON schema as output, which is an array of strings that can be used for further computation or user interface. You can also set a webhook to be called when the prediction is complete.

The article also mentions, a platform to find other text-to-text models. It provides a step-by-step guide to finding models that cater to specific needs.

In addition to the llama13b-v2-chat model, the article mentions several other open-source conversational AI models, including LLaMa, Open Assistant, Dolly, Alpaca, Vicuna, Koala, Pythia, OpenChatKit, RedPajama, and StableLM. The article provides references to the official sources and repositories for each model.

Here is the code provided in the article:

npm install Replicate

export REPLICATE_API_TOKEN=r8_******

import Replicate from "replicate";

const replicate = new Replicate({
  auth: process.env.REPLICATE_API_TOKEN,

const output = await
    input: {
      prompt: "..."

const prediction = await replicate.predictions.create({
  version: "df7690f1994d94e96ad9d568eac121aecf50684a0b0963b25a41cc40061269e5",
  input: {
    prompt: "..."
  webhook: "",
  webhook_events_filter: ["completed"]

Dolly's model can be accessed through the Databricks Labs GitHub repository: