How ahiruneko/leflow works
LeFlow trains a latent-path flow planner and inverse-dynamics model over a frozen LeWorldModel, then evaluates it by turning generated latent paths into action chunks and reranking them with world-model rollouts. The graph also includes the repository’s LeWorldModel training path and benchmark configuration. Training and evaluation internals are described from the supplied excerpts; subsystem links without sampled call sites are limited to README-documented behavior.
Main entry points: eval.py Evaluation Runner · train_latent_planner.py Planner Training · latent_planner.py Latent Path Solver · train.py LeWM Training
Components 19
Latent Planner
- Flow and Inverse Dynamics
latent_planner.py - Planner Checkpoint
latent_planner.py - Latent Path Solver
latent_planner.py
Model Training
- Planner Training
train_latent_planner.py - Dataset Loading
train_latent_planner.py - Planner Losses
latent_planner.py
Evaluation
- Evaluation Runner
eval.py - Benchmark Config
config/eval - Latent Flow Config
config/eval/solver/latent_flow.yaml - World Model Policy
- Benchmark World
- Evaluation Results
eval.py
World Model
- Frozen LeWM
latent_planner.py - LeWM Training
train.py - JEPA Model
jepa.py - Model Layers
module.py
Other components
- Engineer
- LeWM Datasets
- Frozen LeWM Checkpoints
Connections 26
- Engineer → Planner Training: starts training
- Planner Training → Dataset Loading: builds loaders
- Dataset Loading → LeWM Datasets: reads episodes
- Planner Training → Frozen LeWM: loads backbone
- Frozen LeWM → Frozen LeWM Checkpoints: loads reference
- Planner Training → Flow and Inverse Dynamics: constructs models
- Planner Training → Planner Losses: computes losses
- Planner Losses → Flow and Inverse Dynamics: trains parameters
- Planner Training → Planner Checkpoint: saves state
- Engineer → LeWM Training: starts training
- LeWM Training → LeWM Datasets: loads dataset
- LeWM Training → JEPA Model: constructs model
- LeWM Training → Model Layers: builds components
- Engineer → Evaluation Runner: starts evaluation
- Evaluation Runner → Benchmark Config: loads settings
- Evaluation Runner → LeWM Datasets: reads episodes
- Evaluation Runner → Benchmark World: creates world
- Evaluation Runner → Latent Flow Config: instantiates solver
- Evaluation Runner → World Model Policy: creates policy
- World Model Policy → Latent Path Solver: dispatches plans
- Latent Path Solver → Planner Checkpoint: loads planner state
- Latent Path Solver → Frozen LeWM: loads backbone
- Latent Path Solver → Flow and Inverse Dynamics: generates paths and actions
- Latent Path Solver → Benchmark World: scores rollouts
- Evaluation Runner → World Model Policy: runs episodes
- Evaluation Runner → Evaluation Results: writes metrics