oceanflowlab/taro

How oceanflowlab/taro works

TaRO is a framework and set of model assets for training video-temporal-grounding models and evaluating or demoing them. Its documented training approach moves from supervised imitation of constructed reasoning toward self-exploration, with temporal-sensitivity rewards; the user-facing result is a grounded temporal answer. The graph also captures the verl training runtime and checkpoint/model handling. Some TaRO source files and most verl modules were not excerpted, so their internal call relationships are omitted where the evidence does not show them.

Main entry points: ray_trainer.py PPO training loop · workers Training workers · evaluate_vtg.py VTG evaluation · decorator.py Worker dispatch

Components 24

TaRO application

Interactive demo demo.py
VTG data src/vtg_dataset.py
Constructed reasoning src/utils.py
Temporal reward src/rewards.py
VTG evaluation standalone_eval/evaluate_vtg.py
Dataset configuration standalone_eval/dataset_config.py

Training runtime

PPO training entry verl/trainer/main_ppo.py
PPO training loop verl/trainer/ppo/ray_trainer.py
Agent loops verl/experimental/agent_loop
Reward loops verl/experimental/reward_loop
Training workers verl/workers
Inference rollouts verl/workers/rollout
Tools and interactions verl/tools

Distributed execution

Batch protocol verl/protocol.py
Worker dispatch verl/single_controller/base/decorator.py
Ray controller verl/single_controller/ray/base.py

Model runtime

Engine configuration verl/workers/config/engine.py
Model adapters verl/models/transformers/monkey_patch.py
Checkpoint merger verl/model_merger/base_model_merger.py

Shared infrastructure

Checkpoint management verl/utils/checkpoint/checkpoint_manager.py
FSDP checkpoints verl/utils/checkpoint/fsdp_checkpoint_manager.py
Megatron checkpoints verl/utils/checkpoint/megatron_checkpoint_manager.py
Shared utilities verl/utils

Other components

Researcher

Connections 25

Researcher → Interactive demo: runs
Researcher → VTG evaluation: evaluates
Researcher → PPO training entry: trains
VTG data → Constructed reasoning: grounds paths
Constructed reasoning → Temporal reward: informs reward
PPO training entry → PPO training loop: starts verl/trainer/ppo/ray_trainer.py
PPO training loop → Training workers: orchestrates verl/trainer/ppo/ray_trainer.py
PPO training loop → Batch protocol: uses batches verl/trainer/ppo/ray_trainer.py
PPO training loop → Ray controller: uses verl/trainer/ppo/ray_trainer.py
Training workers → Worker dispatch: registers calls verl/workers/engine_workers.py
Training workers → Engine configuration: configures engine verl/workers/engine_workers.py
Training workers → Inference rollouts: invokes verl/workers/engine_workers.py
VTG evaluation → Dataset configuration: loads settings standalone_eval/evaluate_vtg.py
VTG evaluation → Model adapters: uses standalone_eval/evaluate_vtg.py
VTG evaluation → Inference rollouts: generates predictions
Agent loops → Tools and interactions: dispatches tools
Reward loops → Temporal reward: scores outputs
Checkpoint management → FSDP checkpoints: specializes verl/utils/checkpoint/fsdp_checkpoint_manager.py
Checkpoint management → Megatron checkpoints: specializes verl/utils/checkpoint/megatron_checkpoint_manager.py
PPO training loop → Checkpoint management: manages checkpoints verl/trainer/ppo/ray_trainer.py
Checkpoint merger → Megatron checkpoints: converts weights
Worker dispatch → Batch protocol: splits batches verl/single_controller/base/decorator.py
Worker dispatch → Ray controller: dispatches workers verl/single_controller/base/decorator.py
Training workers → Shared utilities: uses runtime helpers verl/workers/engine_workers.py
Engine configuration → Model adapters: configures execution

More Python diagrams: noaa-gfdl/pace · notbvdr/etsy-keywords-research · nyelidl/pacs-q · openseespro/e2o-seaoc2020 · ozturkoktay/insurance-llm-framework

Drawn by GitDiagram on September 30, 2026 from the repository's file tree, README and source files. Read as Markdown ·