Fine-tuned GPT model for TVM instructions

[Gusarich]

Summary

Deploy a LLM model that fully understands TVM instructions to assist developers with various tasks.

Context

There are more than 700 distinct TVM instructions, and remembering all of them in detail is practically impossible, even for those experienced in low-level programming on TON. This is where LLMs (Large Language Models) can be immensely helpful. They understand the entire context of your question, as well as the comprehensive list of instructions, and can assist developers in a conversational manner.

Such models could provide multifaceted support, including:

• Presenting usage examples and clarifying how specific instructions impact the stack.
• Identifying necessary instructions based on a natural language description of their intended function.
• Rapidly estimating the gas consumption of certain code segments.
• Potentially aiding in the constructing of instruction sequences to achieve a specific outcome.

References

• https://platform.openai.com/docs/guides/fine-tuning
• https://docs.ton.org/learn/tvm-instructions/instructions

Estimate suggested reward

2000$

[Gusarich]

I have already come up with a draft solution to optimally fine-tune the GPT-3.5-turbo for that task. The process is multi-step and should theoretically lead to good results.

The training itself is estimated to require approximately $300-500 for API usage. The remainder of the suggested reward compensates for my time and accounts for my experience with both Large Language Models (LLMs) and TVM.

[Gusarich]

If it goes well and the results are satisfactory, I will be happy to work on more fine-tuning tasks for other parts of TON.

[sepezho]

https://openai.com/blog/introducing-gpts

use this pls

[Gusarich]

https://openai.com/blog/introducing-gpts

use this pls

"GPTs" and "Assistants" weren't really made for such tasks. They perform good only with general stuff that doesn't require any additional knowledge. Fine-tuning, on the other hand, allows you to make LLM to learn something new.

[mbaneshi]

This is a critical step toward simplifying development and enjoying a better Developer Experience (DX). Thanks, @Gusarich, for initiating this effort.

More than this, I suggest we should scale up rewards in general as well. It seems Nowadays we consider how much time someone needs to accomplish a task. this is wrong, we should consider how many years takes for a person to reach a point that can accomplish these tasks, and also how much impact this endeavor has on overall community growth. I suggest this task for 3000$.
Scaling up bounty have some other consequences as well, first attracting more developer. you can count active members in this repo, why should we have such few participants?
I assume this is simply tiny respect for those technical guys.

[delovoyhomie]

It's a valuable tool, but currently, it hasn't resonated with the community as I hoped. We're also aware of a similar tool being developed at the moment.

Consequently, we're closing this Bounty for now, but there's a possibility it might be revisited in the future. Thank you for contribution to the community!