Skip to content

controls, batch, registry, study

Commands for planted regressions, many audits at once, the public registry and study tables.

controls make

usage: modeldiffr controls make [--out OUT] [--kind {identity,noise,scale,zero,prune}]
                                [--strength STRENGTH] [--match MATCH] [--seed SEED]
                                [--revision REVISION] base
Option Default Meaning
--out required Empty or new folder for the control model
--kind noise Perturbation, see Positive controls
--strength 0.05 Noise scale, factor or pruned share
--match attention and MLP projections Regular expression on tensor names
--seed 0 Seed for noise
--revision latest Base model revision

Needs the hf extra and safetensors weights.

batch

usage: modeldiffr batch [--out OUT] [--only ONLY] [--rerun] [--dry-run] spec

See Batch audits. Exits with 0 when every audit finished, 5 when any failed.

registry build

usage: modeldiffr registry build [--out OUT] [--title TITLE] [--key NAME=PATH]
                                 [--allow-failed] [--base-url URL] paths [paths ...]
Option Default Meaning
--out registry_site Output folder, replaced on every build
--title Modeldiffr audit registry Site title
--key none Publisher public key; repeat for several publishers
--allow-failed off List reports that fail verification

Exits with 4 when any report was skipped. See Public registry.

study

usage: modeldiffr study [--out OUT] paths [paths ...]

Writes study.md, study.csv, study.tex and study.json to --out (default study). Audit names and kinds come from any batch index.json under the given folders. See Summarize a study.