Orion

Research / training / benchmarking for Constellation’s research stack.

Orion is the package researchers actively wield. Built on top of Ursa and Virgo, it wraps torch_brain modules in orion.models.* so iteration speed isn’t gated on upstream PRs.

• Pydantic + Tyro configs — typed, validated, CLI-overridable, serialized into the run record

• Lightning trainer with pre-wired callbacks (OrionTrainer); DDP default, FSDP exposed, PyTorch Monarch opt-in

• Lance-streaming dataloader that resolves an Ursa QuerySpec to a streaming scan

• SkyPilot for cloud orchestration; Slurm on Polaris

• ClearML for run tracking + model registry (self-hosted, R2-backed)

• Rich checkpoints with bit-exact resume state plus a data_hashes/manifest.json answering “what data was this trained on?”

• Benchmarks as first-class artifacts — content-addressed in Ursa, configurable compute boundaries, partial subsets for fast in-training metrics

• Multi-stage training pipelines with full lineage propagation

• Aggressive pre-flight validation so doomed runs never reach a GPU

Contents

• Architecture
  • Goals
  • Configuration
  • Rich checkpoints
  • Benchmarks
  • Multi-stage training
  • Phasing
• Concepts
• Quickstart
• Tutorials
• API Reference

Where this fits

Orion is one of three packages in Constellation’s research stack:

• Ursa — database / data layer

• Virgo — DAG-based preprocessing

• Orion (this site) — research / training / benchmarking

Full architecture: Research Stack Architecture (Notion).

Status

🌱 Early bootstrap. Implementation tracked in the Linear Orion project.

Phasing (mirrors the Linear project milestones):

• M1 — Foundations (in progress)

• M2 — MVP (Phase 3)

• M3 — Benchmarking Framework

• M4 — Multi-stage Training & Lineage

• M5 — Production-scale (Phase 4)

• M6 — Polish & Onboarding