Feed a single token and yield any resulting deltas. Args: token: A single token ID from the model output. Yields:

Filter scenarios by difficulty range.

Filters dataset based on tokenized prompt length.

Filters dataset entries based on tokenized prompt length.

Force decode any remaining tokens. Call this at end of stream. May produce replacement characters for incomplete sequences.

Format a message into a readable string.

Return a clone of *tool* whose parameters schema permits *only* `fixed_args`. Each field is cast to typing.Literal[value] so Pydantic emits a

Geocode an address to coordinates

Extract current version from pyproject.toml.

Count the number of non-empty rows in a JSONL file. Args: file_path: Path to JSONL file Returns: Number of non-empty li

Given a completion from a teacher model, returns chat completion params that can be used to guide a student model's response. Useful for RL-b

Get instruction and response parts for a model. Checks for explicit model configuration first, then falls back to auto-detection from the tok

Return the factor that you should scale the full fine-tuning learning rate by to get the equivalent LoRA learning rate. Previously we had a m

Factory function to create renderers by name. Args: name: Renderer name. Supported values: - "role_colon": Simple role:conten

Return the stop sequences used when sampling from this renderer.

Get a summary of the scenario collection.

Returns a Uvicorn logging config that writes to the given path.

Vectorised group-wise reduction over sequence dimension. Parameters ---------- values: Tensor of shape ``[B, S, *F]`` where ``*F`` can be

Initialize server configuration. Returns: StdioServerParameters for the specific server implementation

Load trajectories and aggregate metrics at the step level. This function builds on top of load_trajectories to provide step-level aggreg

Load all logs from the pickle file.

Run all logging demonstrations.

Run all tests.