Log trajectories and/or metrics. Can be used in two ways: 1. Log trajectories: `await model.log(trajectory_groups, split="tr

Return the name that should be sent to the inference endpoint. Args: step: If provided, returns name for specific checkpoint.

Split the collection into train and validation sets.

Get the model's current training step. For non-trainable models, returns 0.

Decode tokens to string, buffering incomplete UTF-8 sequences. Args: tokens: New tokens to decode. Returns:

Recursively remove 'logprobs' keys from nested data structures to reduce data storage size. Args: obj: Any nested data structure

Validate dedicated mode GPU configuration. Raises ValueError if the configuration is invalid. Does nothing if neither trainer_gpu_ids nor inf

Read trajectory groups from a Parquet file. Args: path: Path to the Parquet file. Returns: List of TrajectoryGroup obje

Train the model on the given trajectory groups. This is the recommended way to train models. Unlike model.train(), this method does N

Migrate a single JSONL trajectory file to Parquet format. Args: jsonl_path: Path to the JSONL file to migrate. delete_origin

Convert the tool to a dictionary.

Write trajectory groups to a Parquet file with ZSTD compression. Args: trajectory_groups: List of TrajectoryGroup objects to seriali

Upload the model directory from local storage to S3.

Create learning rate schedule for training with optional warmup. Args: total_steps: Total number of training steps peak_lr:

Prepare trajectories in chunks for multiple model.train_sft() calls. Yields SFTChunk objects so that each call to model.train_sft() produces

Reinforce fine-tune the model with a batch of trajectory groups. .. deprecated:: Use ``backend.train(model, trajectory_g

Download the model directory from S3 into local Backend storage. Right now this can be used to pull trajectory logs for processing or model checkpoint

Get or create the wandb run for this model.

Convert token count to cost in dollars.

After moving an object to a child process, you can use this function to close the proxy object and terminate the child process. Args:

Assert that content is text-only and return it as a string. Raises ValueError if content contains images or multiple parts. Use this to

Move an object to a child process and return a proxy to it. This function creates a proxy object that runs in a separate process. Method cal

Return a new collection with shuffled scenarios.

Return the fully-qualified S3 URI for this model directory.

Format JSON data for display

Format JSON data for display

If running vLLM in a separate process, this will kill that process and close the communication threads.

Fetch data from Alpha Vantage API

Generate scenarios for MCP agent evaluation using OpenAI's o3 model. Args: server_params_path: Relative path to the server_params file (e

Return True if the config specifies dedicated mode (separate training and inference GPUs).

Stream trajectories from a JSONL file for one or more epochs. Uses buffer-based shuffling to randomize order without loading all data in

Run a function on all workers in a distributed setup. Args: llm: The AsyncLLM instance with workers. func: The function to r

Supervised fine-tune the model with an iterable of trajectories. Args: trajectories: An iterable of Trajectory objects.

Merge configuration into the W&B run config for this model. This can be called before the W&B run exists, in which case the config is

Get the output directory for this model.

Create minimal TrainInputs-like dict for loss_fn.

Create n dummy trajectories.

Return the inference name for a model checkpoint. Args: model: The model. step: If provided, returns name for specifi

Deletes a model from the Backend. Args: model: An art.Model instance to delete.