Orchestrated Research, Benchmarking, and Iterative Training
ORBIT is a personal research workspace for turning experiments into repeatable remote runs. It keeps local planning, configuration, and audit records separate from the machines that execute the work, so training, evaluation, and data collection jobs can be launched, inspected, and reproduced without relying on one-off shell sessions.
The project is built around explicit execution templates, bundle artifacts, and clear control-plane / execution-plane boundaries. It is intended for practical model and environment iteration rather than as a hosted platform or an organization-branded product.
The main workflow is straightforward: operate jobs locally, execute them on Targon rental machines, and collect logs and artifacts through explicit templates instead of ad-hoc remote orchestration.
ORBIT is organized around four concerns:
control plane: experiment records, task orchestration, template selection, and run inspectionexecution plane: generic bundles, placement backends, launch modes, and artifact collectiontask plugins: training, evaluation, and collection request shapingsidecars: operational helpers such as remote ops and monitoring
The default documented workflow is:
- local
control - remote
targon_rental - launch mode
host_process - template
targon-rental-host
- Targon-first remote execution from a local control plane
- explicit execution templates instead of hidden runtime branching
- bundle-based execution with runtime audit logs
- separate control-plane and execution-plane responsibilities
- official config-driven remote training example
- native
ms-swiftSFT and RLHF workflows throughorbit control launch train uv-based setup as the default environment workflow
Start here:
- Getting Started: first remote run on Targon
- User Guide: how to think about workflows, targets, and command families
Reference:
- Documentation Hub
- Architecture
- CLI Guide
- Operations Guide
- Official Remote Examples
- Testing Guide
- Test Runbook
Supported execution matrix:
local + host_processlocal + docker_imagetargon_rental + host_processtargon_rental + docker_image
Primary documented and validated path:
- local
control->targon_rental + host_process - this path has been real-validated for config-driven remote training,
including native
ms-swiftSFT and GKD configs submitted throughlaunch train
Other paths remain available but are documented as secondary.
- Direct dependencies are declared in pyproject.toml and resolved in uv.lock.
- Training uses upstream
ms-swiftdirectly. ORBIT's role is to validate config, build bundles, provision execution targets, and submit runs.