Iterate over (key, item) tuples

Get an item for the given layer and module

Add a batched tensor for the given layer and module

Get all layer module keys

Get cached completion for a conversation if it exists.

Repeat each element in the activation lists n times before moving to the next element. This is equivalent to the pattern: [p

Get the batch size (validates consistency across all tensors)

Example usage: ```python neurons = db.get( SQLANeuronDescription, filter=and_(SQLANeuron.layer == 5,

Pad a list of sequences to the same length and return input_ids and attention_mask tensors. If `max_length` is provided, AND if the `max_len

Cache a completion for a conversation.

Construct a ``WHERE`` clause (plus its bound parameters) from the given ``conditions`` mapping. - scalar → equality check (``col = ?

Concatenate multiple ActivationsList into a single ActivationsList

Get a colored logger for the specified namespace. Args: namespace: The namespace for the logger Returns: A configured l

Single-call convenience wrapper around ``get_llm_completions_async``. Returns a single ``LLMOutput`` instead of a list.

Loads NeuronExplanations for given neuron if it exists.

Insert a batch of model instances. All rows share the same ``batch_time``. ``batch_time`` defaults to current UTC ISO timestamp.

Asynchronously iterate over rows that satisfy ``conditions``, yielding Pydantic objects one-by-one.

Used for experiments for blog post.

Loads exemplar data for a layer.

Returns NeuronExemplars for a neuron, given a split.

Support for a single write layer to read from multiple read layers ("flattening").

Checks that: - Role/content formatting is correct - Chat starts with system or user - Chat alternates between system/

All costs are in dollars per million tokens.

Construct multiple choice options for evaluation data. Uses uppercase letters (A, B, C, ...) as labels with period delimiter (e.g., "A. Opti

Assumes activations are normalized and discretized to the range [0, 10].

Generate text using the model with optional neuron and hidden state interventions.

This method may return None in cases where the score is undefined, for example if the normalized activations were all zero, yielding

Load reading eval results from a jsonl file. If there are any duplicate results for the same data_id, keep the latest result.

s is the output of calling str(UserAttribute)

Run belief evaluation on a subject model and save results to a database. Instantiates the appropriate evaluator based on the eval type (gene

Set and apply a new filter. You may presume that the neuron mask is always fresh, since it's updated when the filter is set.

Create a deterministic hash key from messages and model.

Returns a function that computes activations for a given input: input_ids: torch.Tensor attn_mask: torch.Tensor

Returns a list of PyTorch Datasets one for each HuggingFace dataset.