hggforget/lingbotvla-robotwin-clean-sft

LingbotVLA — RoboTwin Clean SFT Checkpoints

Supervised fine-tuning checkpoints of LingbotVLA v2 (LingbotVlaV2Policy, model_type: lingbotvla) trained on the cleaned RoboTwin 2.0 LeRobot dataset (demo_clean configuration, clean-only subset, ~549k frames).

Repository layout

Each subdirectory is a complete, independently loadable HF-format checkpoint:

global_step_1000/   ...   global_step_12000/     (12 checkpoints, every 1000 steps)

Each contains config.json, sharded model-*.safetensors + model.safetensors.index.json, and the tokenizer / video preprocessor configs (Qwen3-VL-4B-Instruct tokenizer).

Model

• Architecture: LingbotVlaV2Policy — Qwen3-VL-4B VLM backbone with token-level MoE (32 experts, top-4) and a dedicated action head (dim 128, 32 attention heads), flex/block-causal attention.
• Action space: action_dim = 55, chunk_size = 50, n_obs_steps = 1, loss L1_fm.
• Normalization stats: robotwin_clean_norm.json (computed on the clean-only subset).

Evaluation (RoboTwin 2.0, global_step_10000)

• Protocol: 50 tasks × 10 episodes, demo_clean config, seen instructions, fp32 + compile, 7 GPU inference slots.
• Overall success rate: 370 / 500 = 74.0% (evaluated 2026-10-16).

Usage

Load with the lingbot-vla-v2 codebase (custom lingbotvla model type), e.g. point the policy/eval entry point at global_step_10000/ and pass the norm stats file. Inference requires transformers >= 4.57, flash-attention-2 for the vision tower, and flex attention support.

Training artifacts

Beyond the checkpoints, this repository includes the full training/eval artifacts at the repo root:

• lingbotvla_cli.yaml, launch/ — training/normalization launch configs and robot config
• robotwin_clean_norm.json (+ .sha256) — normalization statistics used by all checkpoints
• model_assets/ — tokenizer and preprocessor assets
• runs/ — TensorBoard training curves; images/ — training visualization snapshots
• eval/ — RoboTwin evaluation logs and stats (74.0% overall run)
• *.log — training / normalization run logs

Optimizer states and raw training-format weights (global_step_*/ subfolders only contain HF-format hf_ckpt content) are intentionally not included.