2 ms·
Summary: Fine tune a foundation model for a task. How did all weights change? That's called the "parameter delta." These changes are highly redundant. You can c
by a_wild_dandan 3y ago
Summary: Fine tune a foundation model for a task. How did all weights change? That's called the "parameter delta." These changes are highly redundant. You can carefully (use DARE to) revert like 80% of them, yet maintain fine-tuned task accuracy! But only if the tuned weights didn't shift much. Otherwise DARE fails. Maybe you can make an LM polymath by melting together many fine-tunes of some base model. No GPU needed.