25 lines
1.4 KiB
Markdown
25 lines
1.4 KiB
Markdown
# MLTrail: Verifiable Provenance Ledger for Federated ML Experiments (MVP)
|
|
|
|
This repository contains a minimal, working MVP of MLTrail, a light-weight,
|
|
open-source ledger platform for recording machine-learning experiments across
|
|
organizations. It demonstrates core ideas from the original concept: an append-only
|
|
hash-chained ledger, compact contract records (Experiment, Run, Dataset, Model,
|
|
Environment, EvaluationMetric, Policy), delta-sync primitives, and lightweight adapters.
|
|
|
|
What you get in this MVP:
|
|
- Core ledger with cryptographic hash chaining (no external dependencies required)
|
|
- Data contracts (Experiment, Run, Dataset, Model, Environment, EvaluationMetric, Policy)
|
|
- Reproducibility helpers (environment fingerprint) and a small governance hook
|
|
- Two starter adapters (MLFlow-like and WandB-like) to publish records
|
|
- A minimal contract registry for schemas and conformance tests scaffold
|
|
- Basic delta-sync primitive to simulate cross-partition reconciliation
|
|
- CLI/test scaffold for local verification
|
|
|
|
How to run the MVP locally (quickstart):
|
|
- Install Python 3.9+ and run tests with pytest
|
|
- See test files under tests/ for guidance
|
|
|
|
This is a foundational MVP intended for stepping stones into a broader ecosystem and governance model. Extend it to implement more sophisticated delta-sync, secure anchoring, and adapter ecosystems as needed.
|
|
|
|
Hook the package into a Python packaging workflow via pyproject.toml (provided).
|