SabaPivot's picture
|
download
raw
390 Bytes

Attention-head stability reproduction

python reproduce.py is a toy metric-level check of Hungarian head matching, middle-layer broadening, query-norm correlation, and residual-stream CKA. It is not a checkpoint-level reproduction. Official code and weights: https://github.com/karanbali/attention_head_seed_stability and https://huggingface.co/karanbali/attention_head_seed_stability.

Xet Storage Details

Size:
390 Bytes
·
Xet hash:
f815152ac322879e272c043fd9a32b4e84298a716c216f27d0dc4a663591ef19

Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.