Tokenize simulation prompt such that given prefix_ids = tokenize(prefix_messages) input_ids = tokenize(messages), we have:

Return *all* rows satisfying ``conditions`` as Pydantic objects (ignores ``batch_time``/``id`` when constructing the objects).

Filter clusters based on similarity and size, to surface the most interesting clusters. Then sort the filtered clusters in descending order o

Find the .env file in the project directory. Stops ascending at the project root. Raises an error with the list of paths explored if no .env

Find the starting index of needle in haystack. Returns -1 if not found.

Format a list of lists of tokens into a string with each token marked as having an "unknown" or "blank" activation, suitable for use in promp

Initialize ActivationsList from a list of per-sample activation dictionaries Args: sample_dicts: List where each element is eithe

Generate prompts and validate them. Args: attribute_classes: list of attribute classes to use. If not pro

Generate multiple choice question stems for each attribute class using an LLM.

Generate system prompts for all values of an attribute class

Returns all examples in a fixed order.

Get all possible MC option keys for the given number of options. Used for sampling random incorrect answers or validating MC responses.

Get a list of all AttributionFilters in this filter and its subfilters.

Computes the activation quantile information for a layer using top-activating sequences from the train, valid, and test splits.

TODO perform some sorting

Synchronously get chat completions for multiple message lists using parallel processing with asyncio. This function runs the async code in a

Synchronously get embeddings for a list of texts using OpenAI's embedding model. This function uses tiktoken for tokenization and truncates a

Returns the part of the simulation prompt that depends on the explanation and sequence.

Returns the system prompt and few-shot example part of the simulation prompt.