Training method and reproduction
E5 adds public state history as two tokens (stride 50) and an executed-action summary
of the last eight commands; visual history is disabled. Action/state history masks remain
independent. Public state history means prior drive-target observations at policy decisions,
not measured qpos. See scripts/e5_rollout_history.py and the history implementations under
third_party/lingbot-vla-v2/lingbotvla/.
C3 augmentation is enabled at probability 0.8; camera keys receive independent draws,
while temporal frames share parameters, maps and noise per camera. Photometric perturbations
and probability-0.5 Fourier amplitude jitter are enabled; background/style are disabled.
Current and future depth/video teacher pixels share the augmented images, with no cached teacher
features. The exact parameters are in configs/augmentation/c3_v1.json and actual configs.
Both runs use 2500 clean trajectories, 549787 frames, no holdout or validation selection, fresh official baseline initialization, seed 7, Muon, mixed BF16 computation and FP32 master and exported checkpoint parameters. Configs retain the actual LR schedule, FSDP2, action/state sizes, teacher loss weights, history settings, normalization and dependency layout.
| Run | Epochs | Steps | GPUs | Microbatch | Accumulation | Global batch | LR | Minimum LR |
|---|---|---|---|---|---|---|---|---|
| ep2 | 2 | 17180 | 4 | 8 | 2 | 64 | 0.0001 | 5e-05 |
| ep6 | 6 | 17178 | 8 | 12 | 2 | 192 | 0.00015 | 7.5e-05 |
Install the original runtime requirements from third_party/lingbot-vla-v2/ into a compatible
CUDA environment. Place the official base in checkpoints/base/lingbot-vla-v2-6b, Qwen3-VL
and MoGe dependencies in the paths specified by the configs, and the original RoboTwin LeRobot
trajectory directories under data/lerobot. Training manifests include both original train/val
directories intentionally: the actual all-clean runs have no held-out validation split.
These assets are dependencies and are not included in this Git repository.
Run from third_party/lingbot-vla-v2 so ../../ config references resolve to repository root.
Set MASTER_ADDR, MASTER_PORT, NODE_RANK and PHASE4_NODE_LOCAL_ROOT from the allocated
training environment. For epoch2 use NNODES=1, NPROC_PER_NODE=4 and NODE_RANK=0; epoch6 uses
NNODES=2 and NPROC_PER_NODE=4 on each node. TRAIN_CONFIG is ../../configs/ep2_training.yaml
or ../../configs/ep6_training.yaml; RANK_ENTRY is ../../deployment/finetune/e5_rank_entry.py
or ../../deployment/finetune/e5_8gpu_rank_entry.py respectively.
export PHASE4_HF_EXPORT_DTYPE=fp32
export PHASE4_STATE_HISTORY_REQUIRED=1
export PHASE4_ACTION_HISTORY_REQUIRED=1
export PHASE4_VALIDATION_CACHE=0
export PHASE4_HF_DEDUP=0
export PHASE4_PERF_MODE=0
export PHASE4_DEPTH_TEACHER_BATCH=0
python -m torch.distributed.run --nnodes="$NNODES" --nproc-per-node="$NPROC_PER_NODE" --node-rank="$NODE_RANK" --master-addr="$MASTER_ADDR" --master-port="$MASTER_PORT" --max-restarts=0 "$RANK_ENTRY" tasks/vla/train_lingbotvla.py "$TRAIN_CONFIG"
This portable invocation exposes the original training entry and recipe. It has not been rerun for this release and does not recreate the original scheduler, data staging or recovery orchestration. The actual run completion records report zero automatic restarts. Use the actual sanitized configs, not similarly named generic upstream examples.
Final training losses and elapsed training seconds are reported in RESULTS.json.
The plots show each actual optimizer-step loss and a labeled trailing 100-step mean (shorter
prefix window at the beginning). The rank0 metric records contain total loss all-reduced
over the FSDP group. Scalar-only CSVs are in loss/ep2.csv and loss/ep6.csv; coverage and
original record SHA256 values are in loss/summary.json. Both series cover every optimizer
step without gaps. Raw logs and all non-scalar metadata remain local: 原始记录保存在本地。