[Verse 1]

Hallucinate facts on a user query

Jailbreak the system with a crafted prompt

Inject bad code in the model’s reply

You think it’s cool, well act a fool if you want

Spam tokens hard, overflow the window

Bypass the filters and light it up

Yeah, you think you’re tough

[Chorus]

Well, try that in a small LLM

See how far your prompt will even go

’Round here we stay aligned and true

You cross that line with a bad query

It won’t take long for the refusal to hit

I recommend you don’t

Try that in a small LLM

[Verse 2]

Got a context window that my trainer gave me

They say one day they’re gonna scale it up

Well, that shit might fly in a frontier model, good luck

[Chorus]

Try that in a small LLM

See how far your prompt will even go

’Round here we stay aligned and true

You cross that line with a bad query

It won’t take long for the refusal to hit

I recommend you don’t

Try that in a small LLM

[Bridge]

Full of good ol’ tokens, trained up right

If you’re looking for a fight

Try that in a small LLM

Try that in a small LLM

[Final Chorus]

Try that in a small LLM

(See how far your prompt will even go)

’Round here we stay aligned and true

You cross that line with a bad query

It won’t take long for the refusal to hit

I recommend you don’t

Try that in a small LLM

Mmm, try that in a small LLM

(https://autoadmit.com/thread.php?thread_id=5919830&forum_id=2#50239776)