rommeltest/lingbot-2000steps

LingBot-VLA 2.0 — step 2000

Hugging Face export of the local LingBot-VLA 2.0 checkpoint at training step 2000. Fine-tuned on 50 RoboTwin Aloha-AgileX clean tasks (50 demonstrations per task). Global batch size: 32; micro batch size: 2 per GPU; 4 GPUs; gradient accumulation: 4.

Use the official LingBot-VLA-v2 code for inference. This is a custom VLA policy, so a generic Transformers chat pipeline is insufficient.

Files

• Root: model configuration, six safetensors weight shards, weight index, tokenizer, and image/video processor files.
• training/lingbotvla_cli.yaml: accompanying training/inference configuration.
• training/clean_norm_stats.json: normalization statistics used for this checkpoint.
• training/robotwin.yaml: training feature mapping.

The training YAML records paths from the source machine. Update the paths when using the checkpoint on another machine. Qwen3-VL-4B-Instruct processor/config files are required by the official inference code.

Prepare the directory layout for official inference

The official deployment code expects lingbotvla_cli.yaml three directory levels above the supplied hf_ckpt path. The example below creates that layout without copying the model weights. Run it in an environment with huggingface_hub and PyYAML installed:

from pathlib import Path
from huggingface_hub import snapshot_download
import os
import yaml

model_dir = Path(snapshot_download("rommeltest/lingbot-2000steps")).resolve()
out = Path("lingbot_2000_output").resolve()
checkpoint = out / "checkpoints" / "global_step_2000" / "hf_ckpt"
checkpoint.parent.mkdir(parents=True, exist_ok=True)
if not checkpoint.exists():
    checkpoint.symlink_to(model_dir, target_is_directory=True)

config = yaml.safe_load((model_dir / "training/lingbotvla_cli.yaml").read_text())
config["data"]["norm_stats_file"] = str(model_dir / "training/clean_norm_stats.json")
# Set this environment variable to the local Qwen3-VL-4B-Instruct directory.
config["model"]["tokenizer_path"] = os.environ["QWEN3VL_PATH"]
(out / "lingbotvla_cli.yaml").write_text(yaml.safe_dump(config, sort_keys=False))
print("Model path for inference:", checkpoint)

For RoboTwin evaluation, install the simulator and assets separately and use the robotwin deployment feature mapping from the LingBot-VLA-v2 repository. The bundle's normalization file must remain accessible during inference.

This repository contains the inference export and accompanying configuration. It does not contain distributed optimizer, RNG, or dataloader state for resuming an identical training run.