zimplex/openvla-oft-genesis-hr-bench-step50000

openvla-oft fine-tuned on genesis-hr-bench (step 50000)

OpenVLA-OFT fine-tuning of openvla/openvla-7b on the zhouqh/hrbench genesis-hr-bench dataset.

Training

• Base model: openvla/openvla-7b
• LoRA rank: 32
• Batch size: 8 per GPU, 8 GPUs (effective batch 64)
• Learning rate: 5e-4
• Steps: 50000
• Hardware: 1 node x 8 H200 GPUs (FAIR Cloud)
• Trained at: step 50000 (final)
• Image augmentation: enabled
• WandB run: https://wandb.ai/multi-agent-world-model/openvla_oft_genesis_hr_bench

Files

• model-{1..4}-of-00004.safetensors — merged HF model (base + LoRA).
• lora_adapter/ — LoRA adapter only (apply on top of openvla/openvla-7b).
• action_head--50000_checkpoint.pt — OFT action head.
• proprio_projector--50000_checkpoint.pt — proprio projector.
• dataset_statistics.json — normalization stats used during training.

Usage

from transformers import AutoModelForVision2Seq, AutoProcessor
processor = AutoProcessor.from_pretrained("zimplex/openvla-oft-genesis-hr-bench-step50000", trust_remote_code=True)
model     = AutoModelForVision2Seq.from_pretrained("zimplex/openvla-oft-genesis-hr-bench-step50000", trust_remote_code=True, torch_dtype="bfloat16")

Trained via scripts/finetune.sh --model openvla_oft.