Update OpenWAM video pretraining weights (chunk-causal objective, step 366000)

SHA256SUMS

@@ -1,2 +1,2 @@
-46f1e19a37f8ff68f88b2895f026da10624ebe067f5ca85a13b3b35d16bfa9bf  model_state.pt
-9b55df75880e3bd926bfa92b91062cb2daa0e664f1b8f532d1f4424b904b64c7  resolved_config.yaml
+3116958037211db705c16dc6aac1b28af60a8efba315605b19758399dea40c17  model_state.pt
+9e226b63572d77e664e3494f6d2867cda7745e0abb12509f52347ea586eb4fb2  resolved_config.yaml

model_state.pt

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:46f1e19a37f8ff68f88b2895f026da10624ebe067f5ca85a13b3b35d16bfa9bf
+oid sha256:3116958037211db705c16dc6aac1b28af60a8efba315605b19758399dea40c17
 size 20433187287

release_metadata.json

@@ -2,7 +2,7 @@
   "project": "OpenWAM",
   "checkpoint_type": "video pretraining",
   "release_type": "weights_only",
-  "source_checkpoint_completed_utc": "2026-10-09T12:46:47+00:00",
+  "source_checkpoint_completed_utc": "2026-10-25T08:07:58+00:00",
   "repo_id": "OpenWAM-Stanford/OpenWAM-Pretraining",
   "model_payload_key": "model_state_dict",
   "optimizer_state_included": false,
@@ -11,11 +11,11 @@
   "checkpoint_files": {
     "model_state.pt": {
       "bytes": 20433187287,
-      "sha256": "46f1e19a37f8ff68f88b2895f026da10624ebe067f5ca85a13b3b35d16bfa9bf"
+      "sha256": "3116958037211db705c16dc6aac1b28af60a8efba315605b19758399dea40c17"
     },
     "resolved_config.yaml": {
-      "bytes": 13338,
-      "sha256": "9b55df75880e3bd926bfa92b91062cb2daa0e664f1b8f532d1f4424b904b64c7"
+      "bytes": 13655,
+      "sha256": "9e226b63572d77e664e3494f6d2867cda7745e0abb12509f52347ea586eb4fb2"
     }
   },
   "configuration": {

resolved_config.yaml

@@ -340,6 +340,18 @@ data:
     include_streams: []
     channel_mappings: []
     sampling_weight: 0.1
+  - source_id: VPT-02
+    manifest_csv: ./data/manifests/VPT-02.csv
+    repo_id: null
+    local_root: null
+    latent_root: null
+    source_format: latent
+    latent_key: video_latents
+    enabled: true
+    source_group: null
+    include_streams: []
+    channel_mappings: []
+    sampling_weight: 0.008459264733050064
   latent_encoding_mode: canonical
   latent_view_combinations: []
   decode_size_mode: aspect_ratio_bins
@@ -434,6 +446,7 @@ training:
   use_teacher_forcing: true
   chunk_size: 4
   window_size: 64
+  noisy_video_condition_prob: 0.5
   optimizer_name: adamw
   scheduler_name: constant
   learning_rate: 1.0e-05
@@ -443,7 +456,7 @@ training:
   warmup_steps: 100
   decay_steps: 0
   min_lr_ratio: 0.1
-  gradient_accumulation_steps: 2
+  gradient_accumulation_steps: 1
   max_grad_norm: null
   num_steps: null
   text_condition_dropout_prob: 0.1
@@ -458,7 +471,6 @@ training:
   trainable_components:
   - all
   frozen_components: []
-  noisy_video_condition_prob: 0.5
 inference:
   video_num_inference_steps: 25
   action_num_inference_steps: 50
@@ -496,7 +508,7 @@ trainer:
   validation_interval: null
   log_every_n_steps: 1
   accelerator: gpu
-  devices: 16
+  devices: 32
   precision: bf16-mixed
   enable_checkpointing: true
   enable_model_summary: false