layers, directions, features, causal¶

Commands for the representational and causal layers. All four need the hf extra and load both models with transformers.

Shared options¶

Option	Default	Meaning
`--device`	`auto`	`cpu`, `cuda`, `cuda:0` or `mps`
`--dtype`	`auto`	`float32`, `float16` or `bfloat16`
`--base-revision`, `--derivative-revision`	latest	Pin revisions; Hub URLs with `/tree/REF` also work
`--trust-remote-code`	off	Allow custom model code

Corpus options for layers and features:

Option	Default	Meaning
`--corpus`	`builtin`	`builtin`, a `.jsonl` or `.txt` file, or `hf:DATASET[:CONFIG]:SPLIT`
`--field`	`text`	Field with the text
`--texts`	500 for layers, 5000 for features	Texts to read
`--max-tokens`	20,000 for layers, 50,000 for features	Tokens to capture
`--max-length`	`256`	Tokens per text

layers¶

usage: modeldiffr layers [--layers LAYERS] [--json] base derivative

--layers takes all, middle, an index (negative counts from the end), ranges such as 2-5, or a comma separated mix.

directions¶

usage: modeldiffr directions --positive FILE --negative FILE [--layers LAYERS] [--json] base derivative

features¶

usage: modeldiffr features [--layer LAYER] [--dict-size N] [--k K] [--variant {standard,delta}]
                           [--steps N] [--batch-size N] [--lr LR] [--seed SEED]
                           [--per-class N] [--examples N] [--report REPORT]
                           [--label-api URL] [--label-model NAME] [--no-weight-hash]
                           [--out OUT] base derivative

Option	Default	Meaning
`--layer`	`middle`	One layer
`--dict-size`	`4096`	Features in the crosscoder
`--k`	`32`	Average active features per token
`--variant`	`standard`	`delta` ties the decoders and penalizes the difference
`--steps`, `--batch-size`, `--lr`, `--seed`	`2000`, `2048`, `5e-4`, `0`	Training
`--per-class`	`20`	Features per class shown with examples
`--examples`	`8`	Top contexts per feature
`--report`	none	Link to an audit and check the weights match
`--label-api`, `--label-model`	none	Automatic, unverified labels
`--out`	`modeldiffr_features`	Folder for `features.json`, `features.html` and `crosscoder.safetensors`

causal¶

usage: modeldiffr causal [--layer LAYER] [--suite SUITE] [--tasks TASKS] [--limit N]
                         [--data-dir DIR] [--features DIR] [--top 10,50,200]
                         [--alphas 0.5,1,2] [--resamples N] [--seed SEED]
                         [--report REPORT] [--no-weight-hash] [--out OUT] base derivative

Option	Default	Meaning
`--layer`	`middle`	One layer
`--suite`, `--tasks`, `--limit`, `--data-dir`	`quick`, all, `200`	Items to use; only multiple choice tasks are kept
`--features`	none	Folder with `features.json` and `crosscoder.safetensors` from the same layer
`--top`	`10,50,200`	Feature set sizes to patch
`--alphas`	`0.5,1,2`	Steering strengths; empty to skip
`--resamples`, `--seed`	`2000`, `0`	Bootstrap
`--report`	none	Link to an audit and check the weights match
`--out`	`modeldiffr_causal`	Folder for `causal.json` and `causal.html`

modeldiffr verify features.json and modeldiffr verify causal.json check these files. Errors such as a wrong layer, a missing corpus or models with different vocabularies exit with 2.