sy000/dreamdojo-repa

DreamDojo GR1 distillation checkpoints

• Warmup: 20,000 updates
• Self-forcing: 1,000, 2,000, and 3,000 updates
• Self-forcing + REPA: 1,000, 2,000, and 3,000 updates
• Self-forcing + REPA, student LR 2e-6: 1,000 / 2,000 / 3,000 updates
• Ordinary self-forcing, student LR 2e-6: 1,000 / 2,000 / 3,000 updates

Warmup: 20,000 updates

This repository contains the final checkpoint from gr1_warmup_newdata_20261025. Training completed on 2026-10-26 after 20,000 cumulative optimizer updates across GCP interactive and normal allocations. The run used eight GPUs, four samples per GPU, no gradient accumulation (global batch 32), learning rate 3e-5, and weight decay 0.1. This is the warmup stage of DreamDojo distillation.

The checkpoint is at warmup/gr1_warmup_newdata_20261025/checkpoints/iter_000020000/. It is a PyTorch Distributed Checkpoint (DCP), rather than a Transformers from_pretrained or safetensors export. Its contents are:

• model/: model weights in their original DreamDojo checkpoint layout.
• optim/, scheduler/, and trainer/: optimizer, learning-rate schedule, and completed update count as saved by DreamDojo.
• rng_rank_00000.pt through rng_rank_00007.pt: per-rank Python, NumPy, PyTorch CPU, and CUDA random states for resuming training.
• COMMITTED.json: record that all training state was durably saved.

warmup/gr1_warmup_newdata_20261025/config.yaml records the resolved training configuration; its dataset, tokenizer, asset, output, and code paths refer to the original GCP environment and must be adjusted for another machine. The source snapshot was c4807df3487d in the training workspace. Use the DreamDojo DCP checkpointer with the checkpoint directory above; model-only consumers use its model/ subdirectory. Resuming the included sampler/RNG state requires the same eight-rank, batch-size-four layout and the warmup autoresume implementation. Training already reached the configured max_iter=20000; increase that limit only if intentionally continuing training.

Download the checkpoint and configuration with:

hf download sy000/dreamdojo-repa --include "warmup/gr1_warmup_newdata_20261025/**" --local-dir ./dreamdojo-repa

Training data and objective

The 10,000 samples came from csusupergear/dreamdojo_warmup_data, revision 7bfb50dbd780953e558b43a17d9ac2caaf07625e. Input is the first RGB frame and the following robot actions. The recorded 13-frame clip is ground truth. Warmup supervises the student against the teacher's final generated video latent, initialized from nvidia/DreamDojo, revision 89d029e10816d2995d700cb8ba06f171e0504203, 2B_GR1_post-train/iter_000050000/model/.

The training loss is mean squared error between student and teacher clean latents, excluding the first latent frame and averaging the batch, channels, remaining temporal frames, and spatial positions. Its units are squared normalized-latent values; lower means closer teacher imitation. It is not a perceptual quality score or robot-control success rate. The default warmup run does not provide a held-out quality evaluation or generated-video panels.

Provenance and integrity

The warmup portion includes only its final 20,000-update checkpoint. Data, source assets, earlier checkpoints, and environment credentials are not part of this upload. warmup/gr1_warmup_newdata_20261025/manifest.sha256.json records file sizes, SHA-256 digests, and Git blob SHA-1 digests for the uploaded checkpoint and supporting files (excluding the manifest itself). Upload verification compares these to Hub LFS SHA-256 or Git blob SHA-1 metadata, as appropriate.

The original run is recorded in W&B. Refer to upstream DreamDojo and Cosmos materials for their applicable terms and asset dependencies.

Self-forcing: 1,000, 2,000, and 3,000 updates

Full self-forcing checkpoints are under self_forcing/gr1_self_forcing_20261028. They include student/EMA and fake-score weights, both optimizer/scheduler states, trainer/scaler state, and four ranks' RNG states. The global batch is 16 (four GPUs, microbatch 1, accumulation 4). The final 3,000-update checkpoint represents 600 student updates and 2,400 fake-score updates. See the linked restore notes and provenance.json for the original config, initialization sources, and data revision. The existing warmup checkpoint remains available above.

Self-forcing with REPA: 1,000, 2,000, and 3,000 updates

The corrected REPA production checkpoints are under self_forcing_repa/gr1_self_forcing_repa_20261030. They use REPA weight 0.5, layer-18 features, a 1024-wide projector, frozen OmniStream targets, and global batch 16 (four GPUs, microbatch 2, accumulation 2). All three checkpoints include trained projector/EMA weights and full optimizer, scheduler, trainer, and RNG resume state. See the linked restore notes, configuration, and provenance. The existing warmup and non-REPA self-forcing archives remain available above.

Self-forcing + REPA, student LR 2e-6

Checkpoints at 1,000, 2,000, and 3,000 updates are under self_forcing_repa/gr1_self_forcing_repa_lr2e6_20261101. The student rate is 2e-6, fake-score rate 1e-5, and global batch 16. REPA weight is 0.5; full state includes the trained student/EMA projectors. The frozen OmniStream encoder is reloaded separately. All three snapshots retain full optimizer/scheduler, trainer/scaler, and four-rank RNG state. Existing earlier archives remain available.

Ordinary self-forcing, student LR 2e-6

Checkpoints at 1,000, 2,000, and 3,000 updates are under self_forcing/gr1_self_forcing_lr2e6_20261102. The student rate is 2e-6, fake-score rate 1e-5, and global batch 16. This is the uppercase official-based non-REPA implementation. It has no REPA projector or OmniStream encoder. All three snapshots retain full optimizer/scheduler, trainer/scaler, and four-rank RNG state. Existing earlier archives remain available.