transformer_lens.tools.analysis.jacobian_lens_causal_swap_benchmark_cli module¶
Artifact-generation entry point for the causal coordinate-swap benchmark.
Kept separate from
transformer_lens.tools.analysis.jacobian_lens_causal_swap_benchmark because the analysis
package’s __init__ imports that module: running it as a script would execute it twice and
runpy would warn. This module is not imported by the package, so python -m runs it
once.
Reproduce the frozen artifact with (no HF_TOKEN needed – GPT-2 is not gated):
uv run python -m transformer_lens.tools.analysis.jacobian_lens_causal_swap_benchmark_cli
- transformer_lens.tools.analysis.jacobian_lens_causal_swap_benchmark_cli.main(argv: Sequence[str] | None = None) None¶
Generates the frozen GPT-2 causal-swap benchmark artifact.
Loads the published GPT-2-small lens (the same artifact
tests/integration/test_jacobian_lens.pyuses), runs the reused country corpus over every one of the lens’s fitted source layers, and writes the versioned, fingerprinted JSON artifact consumed bydemos/Jacobian_Lens_Coordinate_Patch_Benchmark_Demo.ipynb.