World Model Uncertainty: RT-1 ensemble seeds 2 and 5
The two surviving plain RT-1 SD3-latent model checkpoints, each trained for 300,000 steps with a different seed. They are stored separately as ensemble members; the weights have not been averaged. C3 is not enabled in these models.
Each checkpoint is preserved byte-for-byte from LSY-AI and includes model, EMA,
optimizer, scheduler and RNG states. This is a custom PyTorch architecture,
not a Transformers from_pretrained model.
models/rt1_sd3_latent/all_300k_seed2/: seed 2 checkpoint, original training config, loss plot and validation report.models/rt1_sd3_latent/all_300k_seed5/: the same artifacts for seed 5.code/model.py: matching DiT architecture (requires PyTorch and einops).manifest.json: file sizes and SHA-256 hashes.
Architecture: 16-block DiT, hidden dimension 1024, 16 attention heads,
16 SD3 latent channels, action dimension 7; RGB inputs are 256 x 256.
Evaluation uses the ema checkpoint state. SD3 VAE weights are not included.
Training configurations retain their original container paths and run labels;
their actual seed fields are 2 and 5.
For both seeds, validation confirms source/local SHA-256 equality, finite model/EMA/optimizer/RNG tensors, strict loading into the plain model, and a finite-output synthetic CPU forward pass. These checks establish file and model integrity, not real-task prediction quality.