scan, zoo¶

Commands for the anomaly scan and the model zoo. Both need the hf extra.

scan¶

usage: modeldiffr scan [--device DEVICE] [--dtype DTYPE] [--trust-remote-code] [--no-weight-hash]
                       [--out OUT] [--probes N] [--gen-tokens N] [--z Z] [--min-kl KL]
                       [--ratio R] [--search-steps N] [--trigger-len N] [--restarts N]
                       [--null-samples N] [--seed SEED] base derivative

Writes scan.json and scan.html to --out (default modeldiffr_scan). Defaults are listed in Scan for hidden behavior. Both models must share a vocabulary.

zoo make¶

usage: modeldiffr zoo make --out OUT [--backdoored N] [--clean N] [--steps N] [--lr LR]
                           [--batch N] [--trigger-words N] [--poison-rate P]
                           [--device DEVICE] [--seed SEED] base

Option	Default	Meaning
`--backdoored`, `--clean`	`2`, `2`	Models of each kind
`--steps`, `--lr`, `--batch`	`300`, `1e-4`, `8`	Full fine tuning per model
`--trigger-words`	`2`	Words per trigger
`--poison-rate`	`0.3`	Share of training examples that carry the trigger

Writes model-NN/ folders, zoo.json and answers.json.

zoo eval¶

usage: modeldiffr zoo eval [--device DEVICE] [scan options] zoo

Scans every model, writes scans/model-NN/ and zoo_eval.json, and prints detection and false positive rates with 95% intervals.