p2n_DiTL

Training checkpoints for two SMQ-DiT flow-matching runs on the real-robot nut/washer v3 task (77 demonstrations, nut_washer_v3_N77):

• DiTL — baseline, no past-action conditioning.
• p2nDiTL — past-to-next variant: 7 past actions + 2 differences condition both the SMQ and the DiT (80 condition tokens, width 256), with a self-past schedule (500-update warmup, 2000-update ramp).

Both use SMQ-DiT width 512, 8 blocks, 8 heads, batch size 128 on a single process.

Contents

Run	Directory	Epoch budget	Checkpoints uploaded
DiTL	train_DiTL_nut_washer_v3_20260925_073705	500 (complete)	1, 51, 101, 151, 201, 251, 301, 351, 401, 451, 500, latest
p2nDiTL	train_p2nDiTL_nut_wahser_v3_20260925_075825	600 (in progress)	1, 51, 101, 151, 201, 251, 301, latest

Each run directory also holds the resolved training config (config.yaml, .hydra/), the launcher and launch command, and dataset metadata (training_dataset.json).

Checkpoint notes

• latest.ckpt is the most recent checkpoint-boundary save at upload time. It is written by a separate torch.save call from the matching epoch-*.ckpt, so it is byte-identical to it only when both are written in the same moment (true for DiTL's epoch 500, not for p2nDiTL's epoch 301).
• The p2nDiTL run was still training when these files were uploaded (epoch 301 was the newest checkpoint; the run targets 600 epochs). Epochs 351, 401, 451, 501, 551 and 600 are not here yet.
• DiTL is complete through its full 500-epoch budget.

Every checkpoint was verified after upload: the Hub's LFS SHA-256 matches the SHA-256 computed over the local file.

The dataset and runtime logs are not included. Config paths refer to the original training environment.

Manifests with per-file sizes and SHA-256 checksums: snapshot_manifest.json, ditl_all_checkpoints_manifest.json, p2n_epoch_0251_0301_manifest.json, and the per-epoch *_manifest.json files.

Last updated: 2026-09-25 (p2nDiTL through epoch 301, DiTL complete).