| license: mit | |
| base_model: openvla/openvla-7b | |
| tags: | |
| - robotics | |
| - vla | |
| - openvla | |
| - openvla-oft | |
| - xarm | |
| library_name: transformers | |
| # oft_setting2_chunksize25_batch32_10k | |
| OpenVLA-OFT checkpoint fine-tuned on real-world XArm data, **setting 2: cup stacking**. | |
| **Intermediate checkpoint at 10000 / 30000 training steps** (see [oft_setting2_chunksize25_batch32](https://huggingface.co/AAyano/oft_setting2_chunksize25_batch32) for the final 30k-step model). | |
| LoRA weights (rank 32) are already merged into the [openvla/openvla-7b](https://huggingface.co/openvla/openvla-7b) base — this repo is a standalone model. | |
| ## Training | |
| - Recipe: OpenVLA-OFT (L1 regression, FiLM, parallel decoding) | |
| - Steps: **10000** of 30000 (LR 5e-4, decay after 20000), effective batch size 32 | |
| - Action chunk: 25, action dim: 7, proprio dim: 6 (BOUNDS_Q99 normalization) | |
| - Inputs: 2 images (3rd-person + wrist) + proprio | |
| ## Contents | |
| - Merged model shards (`model-*.safetensors`) | |
| - `action_head--10000_checkpoint.pt`, `proprio_projector--10000_checkpoint.pt`, `vision_backbone--10000_checkpoint.pt` — OFT components needed for deployment | |
| - `dataset_statistics.json` — action/proprio normalization statistics | |
| - `oft_training_config.json` — training configuration snapshot | |
| Use with the [openvla-oft](https://github.com/moojink/openvla-oft) codebase; set XArm constants (chunk 25, action dim 7, proprio dim 6) in `prismatic/vla/constants.py` before deployment. | |