🎒 Our inventory assistant will deliver a link to the best assistant for your use case
The goal is to accelerate capabilities of free programming assistants with GPT builder. Here you'll find instructions & knowledge files for creating next-gen programming assistants.
All programming-related assistants are welcome. If you would like a new GPT to be created, or would like to take custody of one, please open an issue with the title "New GPT request: " or "New GPT custody: <title>."
If you are experiencing an issue with one of our assistants: kindly open an issue and include the title of the assistant and links to relevant conversation history. If the conversation contains sensitive information, generalized plain text may be copied and pasted.
all assistants are hosted on ChatGPT which is 100% free to use for ChatGPT premium users. Assistants are held to the highest standards and are quality-tested to guaruntee a good user experience.
- Git assistant
- Regex assistant
- Flutter GPT
- Firebase GPT
- Python GPT
- Node.js GPT -- up for adoption
- C++ GPT -- up for adoption
- C# Coming soon
Industry-specific:
- Bioinformatics: Coming soon
Dorkotron for finding everything else.
-
If you would like to partner with Whitebox, fill out our form here
- Existing Models
- How does it work?
- Custodial Process
- Making and Maintaining Assistants
- Are Whitebox assistants safe for enterprise use?
- Getting Involved
- "I don't like reading is there a GPT that will spoonfeed this to me?"
-
What are Custom GPTs?
Custom GPT's allow experts to collaborate and condense their knowledge into a single assistant powered by GPT4. You can read OpenAI's announcement here
Because they're hosted on ChatGPT, all Code-GPT assistants can be used freely and require no installation.
If a user does not have ChatGPT premium, assistants may still be used by copying knowledge files to a different LLM. -
Background
AI assistants accelerate programming by suggesting improvments and providing context based on a wide training set of language and code. A key flaw is that they are not continuously up to date on best practices for every domain. Because of this, all models have blind spots that limit their full potential. To counteract this we must define the blindspots caused by training and create techniques to overcome them. -
Purpose and Function
expanded context: The latest generation of multimodal LLMs have the capacity to parse through massive files that would typically overwhelm its context window. If information is structured correctly, this can vastly increase the amount of knowledge availible to a model when working in a known field. For instance, by creating specific rule sets for each flavor of regex, we greatly improved our assistant's ability to create valid patterns that did not mix flavors.
Specialization: Each knowledge file is dedicated to a particular entity or topic, providing in-depth information about it. This could include historical data, technical specifications, or any relevant details that aids the assistant's understanding of a topic.
Integration with GPT: These files are designed to be integrated into the LLM's existing knowledge base, augmenting its ability to generate accurate and contextually relevant responses about the specific entities.
Content Organization: Information within these files is usually organized in a hierarchical or relational manner, allowing the model to understand the connections between different pieces of data. -
Creation and Maintenance
Data Sourcing: The information in these files is compiled from reliable sources, ensuring accuracy and relevancy. Experts for given frameworks are welcome to contribute files or improvements.
Regular Updates: To maintain the relevance of the information, these knowledge files are regularly updated with the latest data.
Quality Assurance: Assistants are checked rigorously to ensure accuracy of the information. A secondary goal of this project is to develop automated testing to ensure widespread functionality can be guarunteed for all models. -
Impact on GPT Performance
Enhanced Accuracy: By having direct access to detailed information, the GPT model can provide better and more accurate responses.
Efficiency: Since the data is structured and tailored for quick retrieval, the response time can be faster for queries related to these entities.
Customization: This approach allows for customization of the GPT model’s responses based on the specific requirements of the application or domain.
Each assistant is assigned a custodian to manage its state and field questions. They are a subject matter expert for their given technology and are the sole decider of what content is included in the official model.
custodian: If you are interested in becoming a custodian, open an issue for the language or framework you wish to claim, and begin preparing your assistant. In the issue, comment links to chat instances that demonstrate effectiveness of the assistant. Once you are granted access, set up the folder for your assistant and create a pull request.
admin: The admin will assess possible candidates and grant custodianship to the most qualified candidate. The admin is the sole decider of who is the official custodian of a assistant but should seek out the opinions of the community before adding or revoking custodianship.
admin: Once the assistant is complete and a link is provided, the admin will confirm the directory in this file is updated and then merge the pull request.
revoking custodianship: If a custodian wishes to forfeit custodianship of an assistant, we ask that they participate in finding a suitable replacement. Once found, we will grant them access and update the directory to reflect the change of ownership.
revoking adminship: we'll cross that bridge when we come to it 😧
Activity: Once custodianship is granted, you're free to update your assistant however you see fit. We just ask that you make a reasonable effort to seek and aggregate user requests and improve your assistant, especially during periods of high activity such as when OpenAI updates their models, or a new major revision of a language is released.
Standards: The custodian has the final say in the name and description of a assistant but we ask that they are both descriptive and that the description features a link to this repo. For instance: "Python development made easy. Maintained by Whitebox at https://github.com/Decron/Whitebox"
Experimentation: It may be beneficial to create a backup assistant to experiment with to avoid disrupting users of the primary assistant.
Conversation training: For now we ask that you disable conversation training for the models under your purvue. There are pros and cons of leaving it disabled, and the topic can be addressed later if the community believes conversation training is important.
Less is more: If your assistant is struggling with too many files or over-generlization, you can always split it into multiple assistants.
For the most part yes, here are the facts:
- This project is enirely open-source so you may repurpose this repo however you see fit. In return giving credit for our files is appreciated but the decision is ultimately yours.
- We've asked all custodians to disable conversation training. This setting cannot be truly verified so it is not reccomended to include information you would not want OpenAI to see. Whitebox does not have access to your conversation history.
- Training based on knowledge files and uploaded documents cannot be disabled with GPT builder. Because of this you should not include sensitive material in knowledge files for our assistants, and you should not upload sensitive files when using them.
- Unconsented storage of user data by model creators is absolutely prohibbited and will lead to irrevocable dismissal from the project.
- If you have a custom OpenAI endpoint or you are using our knowledge files on a different LLM, rules about conversation and document training may not apply. Talk to your system administrator.
- If you would like our assistance creating personalized assistants for your enterprise, please message us at hadriogroup@gmail.com.
Contributing
- The most important thing is to understand GPT4's weaknesses and blind spots. If you find it struggling with certain topics or see complaints online, open an issue or a discussion to help us understand the problem. The custodians of the relevant bot will fact find and test possible solutions, then update the knowledge files to counteract that specific issue.
- Secondly, we need to get the word out about this new technology. Share this repo with people you think would be interested, and invite domain experts to contribute by claiming assistants.
- If you're reading this we want to hear your use case. What annoys you most about programming assistants? Go open a discussion and we'll do our best to improve your experience.
- If you don't have access to ChatGPT premium, we'd love to collaborate on other applications for our knowledge files.
- If you'd like to hear announcements about new assistant releases and partnered agents, follow us for free on Substack
Support
- Whitebox is maintained entirely by volunteers. If you would like to donate to the project, see our Donation Link
- If you're interested in Whitebox swag, we have a merch page here
Yes: https://chat.openai.com/g/g-cwigWCh11-code-gpt-gpt
We are actively seeking organizations to sponsor this project in order to deliver the best possible programming assistants. If you're interested in sponsoring us please send all inquiries to hadriogroup@gmail.com
This project was brought to you by The Hadrio Group. We are a San Francisco based community of MIT and UC Berkeley alumni that focuses on quality and data stewardship in AI.
This project is geared to optimize assistants for the custom GPT marketplace provided by OpenAI. If you find that our knowledge files transfer effectively to other models, we would be very interested in hearing more about it.
Wander with confidence.