Improve LLM Issue-Detection Hallucinations

[Softdev1]

Description

Improve reliability of the outputs to avoid excessive false positives.

[brunneis]

All the LLM I've tried seem to produce hallucinations for reasons that remain unclear. I've tried various strategies to mitigate this, such as modifying prompts, simplifying contexts, and employing multi-step processes like providing step-by-step code explanations rather than presenting the code outright. Despite these efforts, there appears to be an inherent bias for false positives when a specific issue is proposed for evaluation, affecting the ability to accurately interpret requests that deviate from conventional implementations.

WIP...

[kesar]

are we using gpt4 ? or its gpt3 ? if its gpt4 we could give it a try to other LLMs like the new claude 3 or even llamacode. If its gpt3 we can give it a try to 4 (even if its way lower)

[brunneis]

are we using gpt4 ? or its gpt3 ? if its gpt4 we could give it a try to other LLMs like the new claude 3 or even llamacode. If its gpt3 we can give it a try to 4 (even if its way lower)

GPT 3.5 by default, GPT 4 is very slow. But the issue happens also with it. Trying to get access to Claude 3.

[Royal-lobster]

Found we can use Claude 3 opus via https://sdk.vercel.ai tho not an API. Probably it's possible to inspect network tab and use it's api for testing

[brunneis]

I start to see the light... GPT 3.5 is a lost battle for some cases. Got GPT 4 to work by massively minimizing the context, avoiding to add excessive details about the issue to analize. It's a very bizarre behavior.

Sample case. Asking for unchecked external calls. The IF will be completely ignored.

function withdraw() public returns (bool) {
    uint amount = pendingReturns[msg.sender];
    if (amount > 0) {
        pendingReturns[msg.sender] = 0;

>>>>>>> if (!payable(msg.sender).send(amount)) {
            pendingReturns[msg.sender] = amount;
            return false;
        }
    }
    return true;
}

[brunneis]

❌ claude-3-opus (WTF) ✅ claude-3-sonnet ❌ claude-3-haiku ✅ gpt-4 ❌ gpt-3.5

Will try with some coding-specific LLMs...

[kesar]

aweeesome

[Royal-lobster]

❌ claude-3-opus (WTF) ✅ claude-3-sonnet ❌ claude-3-haiku ✅ gpt-4 ❌ gpt-3.5

Will try with some coding-specific LLMs...

Wow claude-3-opus failing but sonnet succeeding is wild

[brunneis]

❌ claude-3-opus (WTF) ✅ claude-3-sonnet ❌ claude-3-haiku ✅ gpt-4 ❌ gpt-3.5 Will try with some coding-specific LLMs...

Wow claude-3-opus failing but sonnet succeeding is wild

When increasing temperature it works well, though.

[brunneis]

• Works better now with Sonnet, Opus and GPT-4.
• Promising results with Mistral 7B Instruct v0.2 and Haiku by removing hallucination triggers (asking for severity seems to bias some models to detect the issue).

[kesar]

is the demo url working with that to test it out a bit?

[brunneis]

is the demo url working with that to test it out a bit?

Yes, deployed with Claude 3 Sonnet right now.