shawber2/rdt-agilex-best

RDT AgileX Best Fine-tuned Checkpoint

This repository contains a deployable package for a fine-tuned Robotics Diffusion Transformer model.

Selected checkpoint

• Run: rdt-1b-finetune-agilex-3000
• Best step: 3700
• Selection criterion: lowest observed overall_avg_sample_mse
• Package checkpoint path: checkpoint/

What is included

• checkpoint/: selected fine-tuned checkpoint
• configs/: relevant RDT config files
• lang_embed/: precomputed language embeddings and instruction JSON files
• patches/: modified local loader/dataset files used during training
• scripts/setup_deploy.sh: helper to link this package into an RDT project
• scripts/verify_package.py: quick validation script

What is not included

• Full training HDF5 dataset
• T5-v1_1-xxl model
• SigLIP model
• Original RDT-1B base model
• Conda environment

Deployment note

For single-GPU deployment, use precomputed language embeddings in lang_embed/. Do not load T5-v1_1-xxl at inference time unless you have enough CPU/GPU memory.