Gradient-free adaptation of frozen 1-bit language models via discrete search over binary weight groups.
Check-out the Colab notebook in the README that demonstrates proof-of-concept for post-deployment adaptation on Bonsai 1.7B. A 13 KB patch (3,019 sign-group flips, 0.023% of weights) corrects two verbatim text extraction failures. You can run it for free on T4 after waiting for Bonsai to download, applying the patch itself happens in microseconds.
The patch was found by a targeted search that concentrates on probes nearest the decision boundary (focal-loss-style weighting). The search crossed both targeted probe boundaries by iteration 6,059.
Patch searches can be performed relatively cheaply or on a home setup in 20 hours or less.
sbenjam1n•1h ago
Check-out the Colab notebook in the README that demonstrates proof-of-concept for post-deployment adaptation on Bonsai 1.7B. A 13 KB patch (3,019 sign-group flips, 0.023% of weights) corrects two verbatim text extraction failures. You can run it for free on T4 after waiting for Bonsai to download, applying the patch itself happens in microseconds.
The patch was found by a targeted search that concentrates on probes nearest the decision boundary (focal-loss-style weighting). The search crossed both targeted probe boundaries by iteration 6,059.
Patch searches can be performed relatively cheaply or on a home setup in 20 hours or less.