ZILtoid1991@lemmy.world to 196@lemmy.blahaj.zoneEnglish · 2 months agoRule2346757123926472142107lemmy.worldimagemessage-square12linkfedilinkarrow-up1235arrow-down11
arrow-up1234arrow-down1imageRule2346757123926472142107lemmy.worldZILtoid1991@lemmy.world to 196@lemmy.blahaj.zoneEnglish · 2 months agomessage-square12linkfedilink
minus-squarePhobosAnomaly@feddit.uklinkfedilinkEnglisharrow-up22·2 months agoI’m trying to figure out why it has output incremental numbers. It seems like an oddly specific pattern to push out.
minus-squareitslilith@lemmy.blahaj.zonelinkfedilinkEnglisharrow-up30·2 months agoThat’s probably the most common way numbers are arranged in the training data
minus-squaregetFrog@piefed.sociallinkfedilinkEnglisharrow-up9·2 months agoWhy tf is it training on the switch/case statement of my calculator program? Friggin plagiarism man
minus-squarePhobosAnomaly@feddit.uklinkfedilinkEnglisharrow-up8·2 months agoAh fair enough. Makes sense that it’s something straightforward. Cheers.
minus-squareWirlocke@lemmy.blahaj.zonelinkfedilinkEnglisharrow-up2·2 months agoLLMs don’t see numbers as numbers, they see them as tokens which is like a word or piece of a word. So “123456789” is like a single word to the LLM because it’s a common enough string of characters. This is also why they struggle with math.
minus-squareRugnjr@lemmy.blahaj.zonelinkfedilinkEnglisharrow-up1·edit-22 months agoPretty sure it’s been edited.
I’m trying to figure out why it has output incremental numbers.
It seems like an oddly specific pattern to push out.
That’s probably the most common way numbers are arranged in the training data
Why tf is it training on the switch/case statement of my calculator program? Friggin plagiarism man
Ah fair enough. Makes sense that it’s something straightforward. Cheers.
LLMs don’t see numbers as numbers, they see them as tokens which is like a word or piece of a word.
So “123456789” is like a single word to the LLM because it’s a common enough string of characters. This is also why they struggle with math.