zihanqu/Haidass-bot-6b-v1

English｜中文

🤖 Haidass Bot 6B v1

Haidass Bot 6B v1 is a post-trained vision-language-action model based on the LingBot-VLA 2.0 architecture for robotic manipulation. This repository provides model weights, model and processor configuration, RoboTwin evaluation results, and example videos.

📊 Evaluation

Environment	Successes / Episodes	Success rate
clean	387 / 500	77.4%
randomized	398 / 500	79.6%
Overall	785 / 1000	78.5%

The evaluation covers 50 tasks and two environments, with 10 actual model episodes per combination. Inference used BF16, compilation disabled, and action execution length 25; weights are stored in FP32. Episodes were accumulated across batches, excluding initialization skips without model execution. These results are not the upstream official FP32 reproduction. See the evaluation report, including results for every task in both environments.

🚀 Files and loading

The six Safetensors shards and index contain 6,375,907,511 parameters stored in FP32. Tokenizer and processor files are included. Loading requires compatible LingBot-VLA 2.0 code and runtime.

Download this repository, install the compatible runtime, and point its loader at the model directory. config.json retains model architecture fields; external auxiliary asset paths are relative to assets/, and the tokenizer path is .. Auxiliary assets are not bundled: obtain the corresponding upstream resources and resolve paths against the model directory. inference_config.json records inference settings and is not a drop-in upstream CLI configuration.

🎬 Example videos

videos/ contains 12 clean success demonstrations from a separate recording run. The videos illustrate model behavior; they are not the complete evaluation or evidence of a 100% success rate. Available metadata is listed in the video manifest.

📜 Attribution and license

Architecture source: LingBot-VLA 2.0. This release uses Apache License 2.0, following the original model's license declaration. The license text is included in LICENSE, and attribution is retained in ATTRIBUTION.md. Third-party dependencies and auxiliary assets retain their applicable terms.