World Model Uncertainty: RT-1 C3, seed 4
Archive of the surviving RT-1 C3 checkpoint trained for 300,000 steps.
The checkpoint is preserved byte-for-byte from LSY-AI and includes model, EMA,
optimizer, scheduler and RNG states. It is a custom PyTorch model, not a
Transformers from_pretrained model.
- Dataset: RT-1, SD3 latents from 256 x 256 RGB inputs.
- Seed: 4, as recorded in the original training configuration.
- Architecture: 16-block DiT, hidden dimension 1024, 16 heads, 16 latent channels, action dimension 7, C3 confidence head enabled.
models/c3/rt1_sd3_latent/all_300k_seed4/: checkpoint, original config, and training loss plot.c3_ref_latents_sd3/sd3_256/: original red/green/blue C3 reference latents.code/model.py: matching DiT architecture (requires PyTorch and einops).manifest.json: exact file sizes and SHA-256 hashes.validation.json: integrity and model-load checks performed before upload.
The original config retains its original container paths and a stale seed-0
W&B run label; its actual seed field is 4. SD3 VAE weights are not included.
The evaluation code uses the ema state for inference.
Validation: source/local SHA-256 match; all stored model, EMA, optimizer and RNG tensors are finite; strict C3 EMA loading succeeds; a synthetic CPU forward pass produces finite prediction and confidence tensors. This does not measure real-task prediction quality or confidence calibration.