Skip to content

Lensemble

Research stack for reproducible federated training and evaluation of action-conditioned JEPA world models.

Lensemble implements the model, participant, coordinator, aggregation, privacy, artifact, and evaluation machinery needed to study federated JEPA world models while participant trajectories remain local.

Research status

The code can execute joint encoder-and-predictor training through the federation stack. That software capability is not an achieved useful and gauge-stable scientific result. The checked-in full-model evidence predates correction of an outer-step sign defect and must be rerun before it supports new claims. The corrected research loop remains open in #335.

The optional browser experience is narrower: it performs federated continuation of a bounded residual adapter on a frozen LeWorldModel checkpoint. It is not full-model browser training.

Install

From a clean checkout:

uv venv .venv --python 3.11
uv pip install "torch>=2.4,<3" --index-url https://download.pytorch.org/whl/cpu
uv pip install -e ".[dev,docs]"

Validate the core loop

uv run pytest \
  tests/e2e/test_toy_pipeline.py::test_federated_round_commits_and_advances_the_global_hash \
  -q

This small CPU test runs two real participants through the in-process coordinator, closes a deterministic aggregation round, and verifies that the committed global hash advances. It is a software smoke test, not benchmark evidence.

Read by task

Optional browser demo

uv run lensemble demo federated --port 8765

Follow the printed local URL. Read the demo card before presenting its result.

Claim boundary

Lensemble is a research codebase. It is not a production federation service, a cryptographic proof system, a full-model browser trainer, or evidence of closed-loop physical robot success.