LLM Self-Taught Reasoning - Explained!

[@EricKaysan]

Great job!

[@EobardUchihaThawne]

why do you think task spesific models (basic, non llm models) cant do arithmetic without cot dataset?

[@willw4957]

How does the back-propagation, fine-tuning and inference work though? The rationale is a more detailed answer, this is bootstrapping the dataset with model outputs hoping theres enough context in the question answer to generate a rationale? which is probably why the rationales are still wrong.

[@CodeEmporium]

For backprop, the output answer (without the rationale) generated during the rationale generation phase is compared to the label output. From this, we can get a loss and hence backprop comes in for the network to learn during the fine tuning phase. The issue here, and with STaR is that even though the answer may be right, the rationale could be wrong

[@CyberwizardProductions]

it was a good video - until you got to quiz time and decided to try to click your tongue and make incredibly annoying jeopardy sounds. that cost you a like and a subscribe