orbiter.train - Orbiter Docs

Training framework with trajectory capture, data synthesis, evolution pipelines, and VeRL integration.

Installation

bash

pip install "orbiter-train @ git+https://github.com/Midsphere-AI/orbiter-ai.git#subdirectory=packages/orbiter-train"

# For VeRL integration:
pip install "orbiter-train[verl] @ git+https://github.com/Midsphere-AI/orbiter-ai.git#subdirectory=packages/orbiter-train"

Module path

python

import orbiter.train

Package exports

Export	Module	Description
`TrajectoryItem`	`orbiter.train.trajectory`	A single step in an agent execution trajectory
`TrajectoryDataset`	`orbiter.train.trajectory`	Dataset of trajectory items with capture and export
`TrajectoryStrategy`	`orbiter.train.trajectory`	ABC for building trajectory items from messages
`DefaultStrategy`	`orbiter.train.trajectory`	Default strategy extracting input/output/tool_calls
`TrajectoryError`	`orbiter.train.trajectory`	Error during trajectory operations
`Trainer`	`orbiter.train.trainer`	Abstract base class for agent training
`TrainConfig`	`orbiter.train.trainer`	Minimal training configuration
`TrainerState`	`orbiter.train.trainer`	Trainer lifecycle state enum
`TrainMetrics`	`orbiter.train.trainer`	Metrics produced by training or evaluation
`TrainerError`	`orbiter.train.trainer`	Error during training operations
`DataSynthesiser`	`orbiter.train.synthesis`	ABC for custom data synthesisers
`TemplateSynthesiser`	`orbiter.train.synthesis`	Generate items via template transforms
`SynthesisPipeline`	`orbiter.train.synthesis`	Orchestrates data synthesis from trajectory items
`SynthesisConfig`	`orbiter.train.synthesis`	Configuration for a synthesis pipeline run
`SynthesisResult`	`orbiter.train.synthesis`	Output of a synthesis pipeline run
`SynthesisStrategy`	`orbiter.train.synthesis`	Strategy enum (llm, template, augment)
`SynthesisError`	`orbiter.train.synthesis`	Error during data synthesis
`augment_swap_io`	`orbiter.train.synthesis`	Augment by swapping input/output
`augment_add_noise`	`orbiter.train.synthesis`	Augment by adding noise to input
`filter_by_score`	`orbiter.train.synthesis`	Filter items by minimum score
`deduplicate`	`orbiter.train.synthesis`	Remove duplicate items
`split_dataset`	`orbiter.train.synthesis`	Split items into train/test sets
`EvolutionPipeline`	`orbiter.train.evolution`	Multi-epoch evolution pipeline
`EvolutionConfig`	`orbiter.train.evolution`	Configuration for an evolution run
`EvolutionStrategy`	`orbiter.train.evolution`	ABC for evolution phase strategies
`EvolutionState`	`orbiter.train.evolution`	Pipeline state enum
`EvolutionPhase`	`orbiter.train.evolution`	Evolution epoch phases enum
`EpochResult`	`orbiter.train.evolution`	Metrics for a single evolution epoch
`EvolutionResult`	`orbiter.train.evolution`	Aggregate result of a full evolution run
`EvolutionError`	`orbiter.train.evolution`	Error during evolution operations
`VeRLTrainer`	`orbiter.train.verl`	Concrete trainer integrating with VeRL
`VeRLConfig`	`orbiter.train.verl`	VeRL-specific training configuration
`VeRLAlgorithm`	`orbiter.train.verl`	Supported VeRL RL algorithms enum
`RewardSpec`	`orbiter.train.verl`	Descriptor for a reward function

Submodules

orbiter.train.trajectory — TrajectoryDataset, TrajectoryItem, strategies
orbiter.train.trainer — Trainer ABC, TrainConfig, TrainMetrics
orbiter.train.synthesis — DataSynthesiser, SynthesisPipeline, augment helpers
orbiter.train.evolution — EvolutionPipeline, EvolutionConfig, strategies
orbiter.train.verl — VeRLTrainer, VeRLConfig, RewardSpec