dejan.ai
Unlocking hidden skills inside Google's model
Training as few as 30 to 1,000 randomly chosen weights in a language model can unlock hidden skills that no prompt can elicit, recovering up to 95% of the gap to full fine-tuning.
One line. Many voicesSeek and you shall find
dejan.ai
Training as few as 30 to 1,000 randomly chosen weights in a language model can unlock hidden skills that no prompt can elicit, recovering up to 95% of the gap to full fine-tuning.