Four final LIBERO DEFT checkpoints
Inspection bundle for the four training runs completed on 2026-11-08. All checkpoints are the original, unmodified native PyTorch payloads at optimizer step 21,700. Each variant has its own matching source snapshot, training configuration, inference configuration, dataset normalization statistics, and provenance.
| Directory | Variant | Training job | Training checkpoint layers |
|---|---|---|---|
teacher_picture/ |
Backbone teacher, action + picture prediction, current view enabled | 37504034 | None |
no_current/ |
Teacher-picture ablation with current-view projections disabled | 37496658 | None |
joint/ |
FastWAM-Joint + DEFT, PCA-initialized trainable MLP | 37504036 | None |
idm/ |
FastWAM-IDM + DEFT, PCA-initialized trainable MLP | 37510021 | 1, 3, 7, 11, 15, 17, 19, 23, 27, 29 (zero based) |
All four trainings used 8 GPUs, batch size 16 per GPU, gradient accumulation 1, and BF16. The Joint and IDM projection is a learned 3072 -> 384 -> 192 GELU MLP. PCA initialized its parameters; optimization continued to update those parameters during training. The checkpoint contains those learned parameters. no_current retains its actual ablation implementation in its matching source.zip.
Files
Each variant directory contains:
step_021700.pt: final model weights and native checkpoint metadata.train_config.yaml: resolved configuration saved by the completed training.eval_config.yaml: inference configuration prepared for the matching model implementation.dataset_stats.json: that training run's normalization statistics.source.zip: matching frozenrift/,configs/,experiments/and supporting scripts.source_files_sha256.json: hashes of the files inside the source archive.training_success.json: completion receipt for the original training job.provenance.json: checkpoint size, SHA256, source revision and training identity.
The root manifest.json records all four checkpoint hashes. environment_versions.json records the currently validated evaluation library versions. Configurations preserve original cluster paths for provenance; adapt filesystem paths when running on another host. External Wan VAE, text encoder and tokenizer assets must be supplied separately for inference.
Download and inspect
Download a selected variant:
from huggingface_hub import snapshot_download
directory = snapshot_download(
repo_id="research-vla/deft-wam-transfer",
allow_patterns=["manifest.json", "environment_versions.json", "no_current/*"],
)
Inspect the native payload without initializing a GPU model:
from pathlib import Path
import torch
checkpoint = Path(directory) / "no_current/step_021700.pt"
payload = torch.load(checkpoint, map_location="cpu", mmap=True, weights_only=True)
print(payload["step"])
print(payload.get("prediction_training_policy"))
print(list(payload["mot"])[:10])
For Joint and IDM, inspect payload["mot"]["representation.space.mlp.0.weight"], representation.space.mlp.2.weight, their biases and representation.space.ready.
For model loading, extract the selected variant's source.zip to a private source directory and place that directory first on PYTHONPATH. Use its experiments.checkpoint_utils.load_rift_checkpoint_exact with the corresponding configuration. The inference configuration enables the text encoder and disables training compilation/fused paths. Each checkpoint is tied to its own source snapshot and statistics.
Evaluation protocol
The ongoing native LIBERO-Plus evaluation uses 10,030 tasks, seed 42, one trial per task, initial state index 0, 30 warmup steps, replanning every 10 steps, 10 denoising steps, an action horizon of 32 and BF16. Partial evaluation results are not included as final benchmark scores in this bundle.