Skip to content

README.md

Latest commit

 

History

History
34 lines (26 loc) · 1.6 KB

README.md

File metadata and controls

34 lines (26 loc) · 1.6 KB

Tapestry Consortium Training PoC

This package is a small executable model of the Tapestry consortium-training vision. It is not a generic FedAvg demo: the PoC keeps the N+1 model outcome front and center.

  • 1 shared base model is governed by the consortium coordinator.
  • N sovereign model artifacts are produced and owned by participant nodes.
  • Nodes train locally on sovereign corpora and share only weight deltas.
  • A governed contribution policy applies a quality floor and anti-capture cap before integrating accepted deltas into the shared base.

Modules

Module Purpose
model.py TinyCausalModel, a small next-token model for fast tests and demos.
node.py SovereignTrainingNode, which runs local continued pretraining and keeps a sovereign model artifact.
coordinator.py ConsortiumCoordinator, which evolves the shared base from governed contributions.
policy.py ContributionPolicy, a minimal quality-floor and anti-capture weighting policy.
messages.py Data classes for sovereign artifacts, contributions, and round results.

What This Demonstrates

The PoC demonstrates the architecture-level invariants from the ADRs:

  1. Raw sovereign data stays local.
  2. Each participant gets a persistent sovereign model artifact.
  3. The coordinator integrates only governed weight deltas.
  4. Low-quality contributions can be rejected.
  5. A single participant's influence can be capped.

It intentionally does not claim frontier-scale training, formal privacy, production governance, or a complete post-training alignment pipeline.