FisherSketch is the reference implementation of The Geometry of Updates (ICML 2026). It turns a fine-tuning task into a ~192 KB signature of its update geometry — an aggregate sketch with no per-example records. Exchange a task's fingerprint through a registry where you could never share the dataset itself: compare, deduplicate, diagnose, and route to the best source — with no model and no raw data at compare time.
One streaming pass turns a task into a 192 KB .fsig — a pooled sketch with no per-example records. Share the fingerprint, not the corpus.
Deduplicate near-identical tasks, find related ones, and rank candidate sources from a registry of signatures — no model and no raw data at compare time.
Works at ≈128k-token output scale, where label-aware scores (LEEP, LogME, H-Score) can't run at all.
QUICKSTART
pip install fishersketch # synthetic surrogate, no model — top-1 source selection: python -m fishersketch.evaluation # error geometry 0.83 vs activation-only 0.17 (≈ chance)
python -m fishersketch inspect sigs/ # what's in a registry of .fsig fingerprints python -m fishersketch dedup sigs/ # near-duplicate task pairs python -m fishersketch nearest sigs/ my_task # the tasks most related to my_task
torch-backed; runs on CPU by default. Paper-scale numbers are measured on Llama-3.1-8B — the repo ships no weights, so bring your own model and data. Early release (v0.2, pre-1.0). Bug reports and package questions: GitHub issues.