Single inference call returns N completions for the group.

Run the ART CLI.

Dump the model, but remove the config field to prevent serialization errors in the backend.

Dump the model, but remove the config field to prevent serialization errors in the backend.

Create sample trajectories for testing.

Create a sample trajectory for testing.

Create a sample trajectory group for testing.

Create sample trajectory groups for testing.

Create sample trajectory groups for testing.

Serialize trajectory groups to a JSONL string.

Return bytes saved by compression.

Split a sequence into a list of lists, where the sizes are as equal as possible, and the long and short lists are as uniformly distributed as

Start the MCP server and establish connection.

Start the API session.

Stop the MCP server and clean up connections.

Stop the API session.

Verify base_path can be customized.

Verify base_path defaults to '.art'.

Test trajectories with Choice format (finish_reason) are flattened to messages.

Test copying an empty trajectory group.

Test that create_sft_dataset_iterator is deterministic with the same seed.

Test that empty trajectories yields no chunks.

Test that initial_step skips completed chunks.

Test that concatenated chunk LRs match the full schedule from create_lr_schedule.

Test that chunk_size >= dataset produces one chunk with full schedule.

Test that step, epoch, and epoch_step are correct on each chunk.

fast_inference=False is fine in dedicated mode (it's the intended state).

Test that deep copy works correctly.

Test that deep copy handles circular references correctly.

Test that deep copy preserves trajectory metadata.

Test that deep copy works with exceptions.

Test that the function creates a deep copy.

Test that deepcopy failure returns original object and logs warning.

Test handling of RecursionError during deepcopy.

Test handling of empty file.

Test empty trajectory groups.

Verify exception_rate is calculated correctly for successful trajectories.

Verify files are written to expected paths.

Test all fixture files round-trip correctly.

Train model A, fork its checkpoint to model B via W&B, then train model B.

Explicitly passing step=None should behave same as no step.

With step, should append @step suffix.

With step, should use inference_model_name as base if set.

Without step, should prefer inference_model_name if set.

Without step, should return the model name.