Skip to content

gpt2 family fisher information

Latest
Compare
Choose a tag to compare
@smpanaro smpanaro released this 06 Mar 04:48
· 3 commits to main since this release

Pre-computed Fisher information sensitivities ("weighting" modification from the blog post).

gpt2-large and gpt2-xl are too large for GitHub releases. To re-assemble them download all parts and run:
cat gpt2-large-grads.safetensors.tar.gz.* | tar -xzvf - or cat gpt2-xl-grads.safetensors.tar.gz.* | tar -xzvf -