MainframeGPT

[venkatzhub]

Project description

The project entails the training and fine tuning of an open source Gen AI Large Language Model (LLM) designed to function as a primary mentor and educator within the mainframe domain. Its primary objectives include:

• Facilitating the onboarding process for new personnel to the mainframe platform by providing comprehensive guidance on understanding mainframe fundamentals and associated concepts.
• Serving as a quick reference resource for experienced professionals within the mainframe ecosystem, offering assistance in accessing how-to guides and best practice recommendations.

Value Proposition: The implementation of this initiative offers significant value through the creation of a centralized repository of mainframe knowledge. This resource is instrumental in expediting the onboarding process for new recruits and streamlining the search for pertinent information. Consequently, it serves to enhance overall efficiency and productivity within the mainframe environment.

Statement on alignment with Open Mainframe Project Mission and Vision statements

Aligns with OMP vision of ensuring the mainframe remains as an integral and indispensable part of Enterprise IT by democratizing the access to mainframe concepts in an easy-to-consume fashion

Are there similar/related projects out there?

Not that we are aware of. Many organizations are considering building something similar on their own to help the onboarding process; hence, having something driven by the mainframe ecosystem in the open-source world would save a lot of time and resources.

Sponsor from TAC

To be appointed

Proposed Project Stage

Incubation

License and contribution guidelines

License: Apache2

Current or desired source control repository

GitHub

External dependencies (including licenses)

• Open source LLM Mistral2 licensed under Apache 2.0 as the foundational model: https://huggingface.co/mistralai/Mistral-7B-Instruct-v0.2
• Buy-in from vendors in the mainframe ecosystem
• A curated set of documentation that would form the basis of the MainframeGPT

Initial committers

TBD

Infrastructure requests

TBD

Communication channels

Slack and Mailing lists

Communication channels

Github Issues

Website

TBD

Release methodology and mechanics

TBD

Social media accounts

TBD

Community size and any existing sponsorship

TBD

[v1gnesh]

Hi @venkatzhub, can I reach you on email? Would like to get involved.

[markbsigler]

I've found Mistral and Mixtral to be much more expensive for inference than Llama3. While I'd agree on using an open source LLM as a foundation, I would suggest evaluating before final selection.

Fine tuning could be very effective but also can also be costly. Retrieval-Augmented Generation (RAG) could be more cost effective. This approach keeps the LLM separate from a large knowledge base. When responding to a prompt, the LLM first searches the knowledge base for relevant information and then uses that information to generate a response. This allows the LLM to access and process much more information than it could on its own.

It would be great if IBM could contribute RedBooks to the training data.
Interskill would certainly not be thrilled about this project, since mainframe education is their core offering.

[venkatzhub]

Hi @venkatzhub, can I reach you on email? Would like to get involved.

Sure.

[venkatzhub]

I've found Mistral and Mixtral to be much more expensive for inference than Llama3. While I'd agree on using an open source LLM as a foundation, I would suggest evaluating before final selection.

Fine tuning could be very effective but also can also be costly. Retrieval-Augmented Generation (RAG) could be more cost effective. This approach keeps the LLM separate from a large knowledge base. When responding to a prompt, the LLM first searches the knowledge base for relevant information and then uses that information to generate a response. This allows the LLM to access and process much more information than it could on its own.

It would be great if IBM could contribute RedBooks to the training data. Interskill would certainly not be thrilled about this project, since mainframe education is their core offering.

We can definitely look at the different models. We experimented heavily with mistral and mixtral, and can speak to the quality of output from those models. I bet Llama3 would perform the same. The goal is to pick an open source model that is well governed.

re: How we implement, RAG has to be in the picture and that is how we have to structure ourselves.

re: Interskill, I do not think this would replace the instructor led learning, like any other AI tool, this would be an assistant. I can see ways Interskill can benefit from the use of this model to enhance their offerings.

[venkatzhub]

@jmertic : Some names to consider

Mainframe Mentor
Mainframe Mind
Mainframe Insight
Mainframe Assistant

[v1gnesh]

Easy one - wiZard :)
EDIT: Actually, IBM used the name DocBuddy too early; that would have been perfect.

[jmertic]

Thanks - I've requested our trademark council to review and provide feedback. Will keep you posted.

[anushkasingh98]

Hey @venkatzhub , I would also love to get involved in this project.