Sidecar files

trimbed.sidecar

Carrying the source repository's vocabulary-independent files into the output.

sentence-transformers models are what make this necessary. Their pooling and dense modules live in files the trim itself never writes, and without them the output reopens with default mean pooling and no error to say so.

DEFAULT_SIDECAR_PATTERNS module-attribute

DEFAULT_SIDECAR_PATTERNS: tuple[str, ...] = (
    "modules.json",
    "config_sentence_transformers.json",
    "sentence_bert_config.json",
    "[0-9]_*/*",
)

Files copied from the source repository when sidecar_patterns says nothing else.

copy_sidecar_files

copy_sidecar_files(
    source: str,
    output_dir: str | Path,
    patterns: Sequence[str] = DEFAULT_SIDECAR_PATTERNS,
    revision: str | None = None,
) -> list[str]

Copy the source repository's non-vocabulary files next to the trimmed artefacts.

Parameters:

Name	Type	Description	Default
source	str	Hub model id or local path the trim started from.	required
output_dir	str | Path	Directory the trimmed artefacts were written to.	required
patterns	Sequence[str]	Glob patterns matched against repository-relative paths, e.g. "modules.json" and "[0-9]_*/*".	DEFAULT_SIDECAR_PATTERNS
revision	str | None	Revision to copy from, for a Hub id.	None

Returns:

Type	Description
list[str]	The relative paths that were copied, in order, e.g. ["1_Pooling/config.json", "config_sentence_transformers.json", "modules.json", "sentence_bert_config.json"]. Empty for a plain checkpoint, which keeps everything in the files the trim writes.

View source on GitHub: src/trimbed/sidecar.py lines 76–112