transformed_embedding_scanner

Modules:

Name	Description
`scanner`
`stainedglass_core`
`universal`	Model-agnostic `Mapper` classes (designed to be compatible with datasets.Dataset.map) useful for building LLM prompts for Stained
`utils`
`version`	Constants storing the version numbers for major changes to the codebase.

Classes:

Name	Description
`TransformedEmbeddingScanner`	Scans LLM input embeddings for unauthorized prompts.

TransformedEmbeddingScanner ¶

Scans LLM input embeddings for unauthorized prompts.

Methods:

Name	Description
`__init__`	Construct a `TransformedEmbeddingScanner`.
`scan`	Scan and detect unauthorized embeddings in the input.

__init__(tokenizer: PreTrainedTokenizer | PreTrainedTokenizerFast, model: CausalModelT) -> None

Construct a TransformedEmbeddingScanner.

Parameters:

Name	Type	Description	Default
`tokenizer` ¶	`PreTrainedTokenizer \| PreTrainedTokenizerFast`	Tokenizer for the LLM model.	required
`model` ¶	`CausalModelT`	The LLM model whose input embeddings to check against.	required

scan(input_embeddings: Tensor) -> None

Scan and detect unauthorized embeddings in the input.

Parameters:

Name	Type	Description	Default
`input_embeddings` ¶	`Tensor`	Input embeddings for LLM generation.	required

Raises:

Type	Description
`ValueError`	If the dtype, embedding dimension, or batch size of the input embeddings are invalid.
`RuntimeError`	If any of the input embeddings are invalid.