GPTparser for OPENAI Fine-Tune API

• GPTparser is an AI parsing tool to create datasets for OpenAI Fine-Tuning.
• GPTparser allows you to scrape and parse webpage text into individual JSON files in Chat Completions format.
• OpenAI Fine-Tune API requires data to be in valid JSON, Chat Completions API format, with a .jsonl dataset where each line represents a unique JSON object containing a topic within a message array.

GPTparser Overview

• GPTparser works great for scraping / parsing text content directly from any website into individual JSON files.

• GPTparser is very simple to use:

  1. Recommended: Activate Miniconda Environment
  2. Recommended: export OpenAI API within Miniconda Session
  3. To use i.e.: GPTparser https://url.com output_file.json

• GPTparser script utilizes 'Few Shot Prompting', to adjust data output, adjust the examples & prompt (follow 'modifying for local use' directions).

  • Current prompt is carefully curated to accurately parse text into Chat Completions, JSON API.

• I suggest utilizing Miniconda Envrionments for security & version / dependency control.

  • FYI, for beginners this isn't required, and if you choose not to use these virutal environments then be sure to:
    1. Install Dependency(1) & Dependency(2) below before using GPTparser 2. Securely activate your OpenAI API Key, and store it.

• This was built using Linux Ubuntu, please adjust the below directions per your OS.

• GPTparser shares no professional affiliation with OpenAI.

Why

• I created GPTparser because I couldn't find a tool that enabled me to efficiently scrape and parse content directly from URL's into structured JSON files.
• Working directly from your Linux CLI with individually parsed files allows for more granular control over the data.
• GPTparser was created for ease of use, to be cost effective, and to enable effective quality control to enhance access to OpenAI's Fine-tuning service and dataset curation techniques.
  1. GPTparser enables you to create a large 50k+ token dataset and finetune via OpenAI API for <$5.
• See the /scripts directory for further tools to help you format, validate, and combine these files into one dataset in proper Chat Completions JSON API format.
• I'd like to gather a small community of people interested in making complex AI workflows more accessible, & in the near future I'll be designing a website / UI to host this project.
  • If you're interested connecting, reach out! -- websitegithub.happily959@passinbox.com

GPTparser Installation

Pip Linux Directions:

• Create a new Miniconda environment (Conda Env not required, but highly recommended):
  1. $ conda create --name GPTparser python=3.8
• Activate Environment:
  1. $ conda activate GPTparser
• Install Package:
  1. $ pip install GPTparser
• Export your OpenAI API Key within your Miniconda environment - will expire when session ends:
  • Must be done everytime you start Miniconda session / start using GPTparser
  1. $ export OPENAI_API_KEY=<enter_api_key>
• To use GPTparser - first create and cd into directory that will host your parsed files - then:
  1. $ GPTparser https://url.com output_file.json

Github Linux Directions:

• Clone Repo and Install Package:

  1. $ git clone git@github.com:dtflare/GPTparser.git

• Navigate into Root Directory of Cloned Repo:

  1. $ cd GPTparser

• Choose one of the two below options to create a new Miniconda environment (Conda Env not required, but highly recommended):

  1. $ conda env create -f environment.yml
  2. $ conda create --name GPTparser python=3.8

• Activate Environment:

  1. $ conda activate GPTparser

• Once in /GPTparser install the package/dependencies (Miniconda env recommended but not required):

  1. $ pip install .

• Export your OpenAI API Key within your Miniconda environment - will expire when session ends:

  • Must be done everytime you start Miniconda session / start using GPTparser
  1. $ export OPENAI_API_KEY=<enter_api_key>

• To use GPTparser - first create and cd into directory that will host your parsed files - then:

  1. $ GPTparser https://url.com output_file.json

• Anytime in the future where you use GPTparser, all you have to do is activate the correct Miniconda env, and export your API key.

For those modifying the GPTparser for local use, follow directions below

• Adjust prompts as needed, currently it will output OPENAI's Chat Completions JSON format for Fine-Tuning.
  1. For those planning on editing the examples/prompt for different output, and/or create a new Miniconda Env.
• Once changes are applied add GPTparser to $PATH.
• Active Miniconda Environment
  1. With YAML file in GPTparser repo:
     • $ conda env create -f environment.yml
  2. Create your own:
     • $ conda create --name GPTparser python=3.8
• Once GPTparser Miniconda session is activated, launch the below 1 & 2 dependencies.
  1. If you use other Python tools within your Cona Env, launch the dependencies everytime at start of every session.
     • For best results only don't use other tools within your GPTparser Conda Env.
  2. Dependency(1)
     • $ pip install langchain==0.1.4 deeplake openai==1.10.0 tiktoken
  3. Dependency(2), Langchain's newspaper module:
     • $ !pip install -q newspaper3k python-dotenv
• Add API KEY:
  1. $ export OpenAI_API_KEY
• mkdir & cd into parsed files host directory, then with GPTparser in $PATH:
  1. $ GPTparser https://url.com output_file.json
• IF GPTparser is not in $PATH, simply use:
  1. $ ./GPTparser https://url.com output_file.json

Reminder

• To run GPTparser after installation, or once added to your global path:
  1. $ GPTparser website_url.com file_name.json
• OR:
  1. $ GPTparser website_url.com

After you have used GPTparser to parse your webpages into documents, utilize the scripts in /scripts to turn them into a usable dataset.

View comments inside individual scripts in /scripts for use directions

• j_val = JSON validator for individual .json files within Directory
• combinR = combines all individual files into 1 .jsonl dataset
• wordcount = use to get accurate word count of your combined dataset

Contributions

Please feel free to contribute! Whether it's code, directions, etc., credit will always be given. Please reach out if you'd like to collab on any AI related projects: websitegithub.happily959@passinbox.com

To contribute:

Open a new issue to start a discussion around a feature idea or a bug.
Fork the repo and start making your changes to a new branch.
Include a test showing the fix & features working properly.
Send a pull request!

Submit feature requests via the issues tab. Submit Security Issues directly to my email: websitegithub.happily959@passinbox.com

If you use GPTparser for your dataset please share it - or your experience - with the community!

Citation

If you use GPTparser or associated tools to create a dataset or wish to refer to the baseline results published here, please use the following citation:

@dtflare{GPTparser, author = {Daniel Flaherty}, title = {GPTparser}, year = {2024}, publisher = {GitHub}, journal = {GitHub repository}, howpublished = {\url{https://github.com/dtflare/GPTparser}} }