Superseded simulation-only staging PR

simulation/rmbench/full_attention/README.md

@@ -14,8 +14,8 @@ tags:
 Canonical Pi0.5 IMPRINT checkpoint for RMBench using the full-attention
 memory encoder.
 
-- Stage 1: 160,000 updates
-- Stage 2: 100,000 updates
+- Stage 1: 160,000 updates, W&B `ylinwu/Robocasa/j8t4ebpo`
+- Stage 2: 100,000 updates, W&B `liqhuiha/Robocasa/grruxd5p`
 - memory tokens: 256 x 1024
 - policy horizon: 50
 - visual memory: three-view pooled DINOv3 features
@@ -27,11 +27,10 @@ portable history configuration, and provenance. The checkpoint is a Gemma
 model derivative and is distributed subject to the included Gemma Terms of
 Use.
 
-Use `Amazon-FAR/imprint_datasets`, path `simulation/rmbench/450`, and the
-evaluation contract in the IMPRINT GitHub repository. Evaluation requires
-strict causal actions, online recent DINOv3 features,
-`EVAL_CONSOLIDATE_ALL=1`, BGR policy inputs, and 50-step
-prediction/execution chunks.
+Use the matching `yilin-wu/imprint-rmbench-450` dataset and the evaluation
+contract in the IMPRINT GitHub repository. Evaluation requires strict causal
+actions, online recent DINOv3 features, `EVAL_CONSOLIDATE_ALL=1`, BGR policy
+inputs, and 50-step prediction/execution chunks.
 
 The matching training store and reproducibility code are publicly available
 under the IMPRINT release.

simulation/rmbench/full_attention/SHA256SUMS

@@ -1,8 +1,8 @@
 3e2c24001f9ef57bf7ec959a3658fbb49cdad113cdf394c264da9d16f9bdd132  ./LICENSE_GEMMA.txt
 5b21d7116ac42a0ab6141488cb654c2411557ac977e2276accb88ef774f28c19  ./NOTICE
-b0609e31f727298fa56264762353a0a3e8eeeef24db72241fbc94160d3d240f7  ./README.md
+32a0a8a3422765c68c10bbd944cbb0d96d4781fab507b061f02f2a4099085a05  ./README.md
 a4c8e95585b546442350a3e3e7b88939569a69a092409fc32ef1203e9589d73b  ./config/history_config.yaml
-1979b352da1f54cad27419908f768c41a6b4d5feef2418978eeba90b4df84f1b  ./config/input_provenance.json
+6569bbc55da67ba7f89a5250ae7261f192754bb27a7fc5db8b579b013452d61a  ./config/input_provenance.json
 372a92664fadef29d988c33d083c0c2594227ae0078272d80d1d31560086f994  ./config/norm_stats.json
 5ea9c49f1b975f0aa1c4d7275381be406954d7d5e9bf6c359c9ee6dbe32359a9  ./stage1/stage1_step160000.npz
 3280d8311fd54fd78143091908db4051671ef7dd6de4665e0298438e428b5cac  ./stage2/step-100000/_CHECKPOINT_METADATA

simulation/rmbench/full_attention/config/input_provenance.json

@@ -5,7 +5,7 @@
   "checkpoint_metadata_sha256": "3280d8311fd54fd78143091908db4051671ef7dd6de4665e0298438e428b5cac",
   "checkpoint_path": "stage2/step-100000",
   "checkpoint_step": 100000,
-  "dataset": "Amazon-FAR/imprint_datasets:simulation/rmbench/450",
+  "dataset": "yilin-wu/imprint-rmbench-450",
   "eval_consolidate_all": 1,
   "execution_horizon": 50,
   "freeze_variant": "image_and_language_embedding",
@@ -17,7 +17,9 @@
   "prediction_horizon": 50,
   "stage1_checkpoint": "stage1/stage1_step160000.npz",
   "stage1_sha256": "5ea9c49f1b975f0aa1c4d7275381be406954d7d5e9bf6c359c9ee6dbe32359a9",
+  "stage1_wandb_run": "ylinwu/Robocasa/j8t4ebpo",
   "stage2_auxiliary_loss_weight": 0.0,
   "stage2_memory_encoder": "finetuned",
+  "stage2_wandb_run": "liqhuiha/Robocasa/grruxd5p",
   "visual_feature_stream": "dinov3"
 }

simulation/rmbench/molmoact2/README.md

@@ -10,8 +10,8 @@ tags:
 # IMPRINT RMBench MolmoAct2
 
 This directory contains the model-only 100k Stage-2 checkpoint used for the
-IMPRINT MolmoAct2 RMBench experiment. Optimizer and trainer state are
-excluded.
+IMPRINT MolmoAct2 RMBench experiment (`mit0jkx7`). Optimizer and trainer state
+are excluded.
 
 The checkpoint uses the full-attention IMPRINT Stage-1 memory encoder with 256
 memory tokens. Evaluation uses online WAN and DINOv3 history features and the

simulation/rmbench/molmoact2/checkpoint/config.yaml

@@ -3127,16 +3127,18 @@ batch_divisor: global_batch
 max_grad_norm_ratio: null
 precision: amp_bf16
 wandb:
-  project: null
-  entity: null
+  project: Robocasa
+  entity: ylinwu
   group: null
-  name: null
-  tags: []
+  name: molmoact2-stage2-dinov3-memory-reference-fixed-film-100k-v3
+  tags:
+  - watching
   log_artifacts: false
   rank_zero_only: true
   log_interval: 20
   allow_resume: true
   finish_on_sigterm: true
+beaker_log_interval: 50
 speed_monitor:
   window_size: 20
   gpu_flops_available: null

simulation/rmbench/molmoact2/manifest.json

@@ -9,7 +9,7 @@
   },
   "checkpoint_config": {
     "path": "checkpoint/config.yaml",
-    "sha256": "4fd5a9fbb27d500ff8047b96f0739c6ba7cde23a4df0ed687f146b3340340ce0"
+    "sha256": "ed4573be786a951ccb3319a090b8443db03acb1a204c827327eb6164ecfb905d"
   },
   "checkpoint_step": 100000,
   "eval_assets_path_prefix": "simulation/rmbench/molmoact2_eval_assets",
@@ -19,5 +19,6 @@
   "offline_visual_features": false,
   "path_prefix": "simulation/rmbench/molmoact2",
   "schema_version": 1,
+  "source_run": "mit0jkx7",
   "visual_history": "online_wan_and_dinov3"
 }

simulation/rmbench/pi0/README.md

@@ -8,8 +8,8 @@ RMBench backbone comparison.
 - Policy backbone: native Pi0
 - Memory encoder: full-attention IMPRINT encoder
 - Visual history: pooled DINOv3 features
-- Stage-1 checkpoint: full-attention memory encoder, step 160,000
-- Stage-2 checkpoint: step 100,000
+- Stage-1 source: `j8t4ebpo`, step 160,000
+- Stage-2 source: `s746sl7k`, step 100,000
 - Memory bottleneck: 256 tokens of width 1024
 - Action horizon: 50
 - Image encoder: frozen in Stage 2

simulation/rmbench/pi0/config/input_provenance.json

@@ -16,9 +16,11 @@
   "model_type": "pi0",
   "num_latents": 256,
   "pi05": false,
+  "stage1_checkpoint": "j8t4ebpo_160k",
   "stage1_checkpoint_hf_path": "simulation/rmbench/full_attention/stage1/stage1_step160000.npz",
   "stage1_checkpoint_hf_repo": "Amazon-FAR/imprint_models",
   "stage2_memory_encoder": "finetuned",
+  "training_run": "s746sl7k",
   "training_store_hf_path": "simulation/rmbench/450",
   "training_store_hf_repo": "Amazon-FAR/imprint_datasets",
   "visual_feature_stream": "dinov3",

simulation/rmbench/pi0/manifest.json

@@ -13,7 +13,10 @@
   "payload_bytes": 11534439221,
   "payload_file_count": 22,
   "schema_version": 1,
+  "source_eval_bundle": "s746sl7k_100k_pi0_dinov3_memft_causal_c1",
+  "source_run": "s746sl7k",
   "stage1_checkpoint_step": 160000,
+  "stage1_source_run": "j8t4ebpo",
   "train_state_included": false,
   "training_store": {
     "path_prefix": "simulation/rmbench/450",

simulation/rmbench/recurrent/README.md

@@ -15,8 +15,8 @@ tags:
 Canonical Pi0.5 IMPRINT checkpoint for RMBench using the recurrent gated-delta
 memory encoder.
 
-- Stage 1: 160,000 updates
-- Stage 2: 100,000 updates
+- Stage 1: 160,000 updates, W&B `liqhuiha/Robocasa/9bm5uvc7`
+- Stage 2: 100,000 updates, W&B `liqhuiha/Robocasa/0f4k18rw`
 - memory tokens: 256 x 1024
 - policy horizon: 50
 - training scan: chunkwise WY
@@ -29,11 +29,10 @@ portable history configuration, and provenance. The checkpoint is a Gemma
 model derivative and is distributed subject to the included Gemma Terms of
 Use.
 
-Use `Amazon-FAR/imprint_datasets`, path `simulation/rmbench/450`, and the
-evaluation contract in the IMPRINT GitHub repository. Evaluation requires
-strict causal actions, online recent DINOv3 features,
-`EVAL_CONSOLIDATE_ALL=1`, BGR policy inputs, and 50-step
-prediction/execution chunks.
+Use the matching `yilin-wu/imprint-rmbench-450` dataset and the evaluation
+contract in the IMPRINT GitHub repository. Evaluation requires strict causal
+actions, online recent DINOv3 features, `EVAL_CONSOLIDATE_ALL=1`, BGR policy
+inputs, and 50-step prediction/execution chunks.
 
 The matching training store and reproducibility code are publicly available
 under the IMPRINT release.

simulation/rmbench/recurrent/SHA256SUMS

@@ -1,8 +1,8 @@
 3e2c24001f9ef57bf7ec959a3658fbb49cdad113cdf394c264da9d16f9bdd132  ./LICENSE_GEMMA.txt
 5b21d7116ac42a0ab6141488cb654c2411557ac977e2276accb88ef774f28c19  ./NOTICE
-6a1513e4dd336322a88f2909e0e0f8fdf0905ccecfa9d5be49e84da41cfb722a  ./README.md
+ad0a0597a9e8433ca73ddafe3e9ae5869845525f1e98a6b048f6956252c8ea51  ./README.md
 aa75ca05051d41efc9d146c1b022e15759f153a078f60bb6ae73c3eb59f80fdb  ./config/history_config.yaml
-ba328bb8990998c9355dff0bbfe3f93e636c863324d56e3490e6cb78b1edb426  ./config/input_provenance.json
+1b63e66aa24c04f109fe0764b25d0dbf1d9e4425330eca4957fd8218d7a06b9a  ./config/input_provenance.json
 372a92664fadef29d988c33d083c0c2594227ae0078272d80d1d31560086f994  ./config/norm_stats.json
 76f7ecea50d65c374c79bba071074fc35aded1dea22846d86cfe90b772f4bea3  ./stage1/stage1_step160000.npz
 dd12980babae314a5097df31c47da055e3ced591c4fc44e1a394442365cb4323  ./stage2/step-100000/_CHECKPOINT_METADATA

simulation/rmbench/recurrent/config/input_provenance.json

@@ -5,7 +5,7 @@
   "checkpoint_metadata_sha256": "dd12980babae314a5097df31c47da055e3ced591c4fc44e1a394442365cb4323",
   "checkpoint_path": "stage2/step-100000",
   "checkpoint_step": 100000,
-  "dataset": "Amazon-FAR/imprint_datasets:simulation/rmbench/450",
+  "dataset": "yilin-wu/imprint-rmbench-450",
   "eval_consolidate_all": 1,
   "execution_horizon": 50,
   "freeze_variant": "image_only",
@@ -19,7 +19,9 @@
   "prediction_horizon": 50,
   "stage1_checkpoint": "stage1/stage1_step160000.npz",
   "stage1_sha256": "76f7ecea50d65c374c79bba071074fc35aded1dea22846d86cfe90b772f4bea3",
+  "stage1_wandb_run": "liqhuiha/Robocasa/9bm5uvc7",
   "stage2_auxiliary_loss_weight": 0.0,
   "stage2_memory_encoder": "finetuned",
+  "stage2_wandb_run": "liqhuiha/Robocasa/0f4k18rw",
   "visual_feature_stream": "dinov3"
 }

simulation/robocasa/full_attention/README.md

@@ -14,18 +14,18 @@ tags:
 Canonical Pi0.5 IMPRINT checkpoint for the customized 12-task RoboCasa
 benchmark using the full-attention memory encoder.
 
-- Stage 1: trained 180,000 updates, selected at 160,000
-- Stage 2: 150,000 updates
+- Stage 1: trained 180,000 updates, selected at 160,000, W&B `far-wandb/Robocasa/yw0l1a20`
+- Stage 2: 150,000 updates, W&B `far-wandb/Robocasa/c0t2f1cg`
 - memory tokens: 256 x 1024
 - policy horizon: 20
 - Stage 2: memory encoder finetuned; image tower frozen
 
-Use `Amazon-FAR/imprint_datasets`, paths `simulation/robocasa/600` and
-`simulation/robocasa/environment`, for the training store and patched
-RoboCasa 1.0.1 simulator source. This release contains the complete Stage-2
-Orbax parameter checkpoint, including the PaliGemma/Gemma backbone. It is a
-Gemma model derivative and is distributed subject to the included Gemma Terms
-of Use. Large simulator assets are not duplicated.
+Use `yilin-wu/imprint-robocasa-600` for training and
+`yilin-wu/imprint-robocasa-environment` for the exact patched RoboCasa 1.0.1
+simulator source. This release contains the complete Stage-2 Orbax parameter
+checkpoint, including the PaliGemma/Gemma backbone. It is a Gemma model
+derivative and is distributed subject to the included Gemma Terms of Use.
+Large simulator assets are not duplicated.
 
 The matching training store, patched simulator source, and reproducibility
 code are publicly available under the IMPRINT release.

simulation/robocasa/full_attention/SHA256SUMS

@@ -1,8 +1,8 @@
 3e2c24001f9ef57bf7ec959a3658fbb49cdad113cdf394c264da9d16f9bdd132  ./LICENSE_GEMMA.txt
 5b21d7116ac42a0ab6141488cb654c2411557ac977e2276accb88ef774f28c19  ./NOTICE
-b8a3b8c7bbdce1dd57bbf0b084888bc72cddb13878576feb0a70ab8bb9f94009  ./README.md
+688f34b0256eb63dff227c29811b7e1b916e503a2eb56888e5b52430707ee2bc  ./README.md
 6ed9405ca99c33d57f8101c1c98130e74a33602dadbc2aeb1e0cac72e03a9788  ./config/history_config.yaml
-3527e705b84214f306083935265e5f37c227d0f750b2fa7fec72af5fe542ae73  ./config/input_provenance.json
+a7dda1ef87035e50a39390c549d7e4de8fb6359d3f972f28f140f02dc2029138  ./config/input_provenance.json
 39a8f9a475d18c9167a3872c7679a348b81a165776fe3e16c391f354c7ad8514  ./config/norm_stats.json
 34387cbbd3891cc7bb082516dfb4993ee9d2319e32344a364ea32bd611cd924e  ./stage1/stage1_step160000.npz
 d962b86d3894c8afdbadc56c986b8c6f9d96a453989e2fe77bc9794ef4c125e6  ./stage2/step-150000/_CHECKPOINT_METADATA

simulation/robocasa/full_attention/config/input_provenance.json

@@ -5,7 +5,7 @@
   "checkpoint_metadata_sha256": "d962b86d3894c8afdbadc56c986b8c6f9d96a453989e2fe77bc9794ef4c125e6",
   "checkpoint_path": "stage2/step-150000",
   "checkpoint_step": 150000,
-  "dataset": "Amazon-FAR/imprint_datasets:simulation/robocasa/600",
+  "dataset": "yilin-wu/imprint-robocasa-600",
   "eval_consolidate_all": 1,
   "execution_horizon": 20,
   "freeze_variant": "image_only",
@@ -17,7 +17,9 @@
   "prediction_horizon": 20,
   "stage1_checkpoint": "stage1/stage1_step160000.npz",
   "stage1_sha256": "34387cbbd3891cc7bb082516dfb4993ee9d2319e32344a364ea32bd611cd924e",
+  "stage1_wandb_run": "far-wandb/Robocasa/yw0l1a20",
   "stage2_auxiliary_loss_weight": 0.0,
   "stage2_memory_encoder": "finetuned",
+  "stage2_wandb_run": "far-wandb/Robocasa/c0t2f1cg",
   "visual_feature_stream": "dinov3"
 }

simulation/robocasa/recurrent/README.md

@@ -15,8 +15,8 @@ tags:
 Canonical Pi0.5 IMPRINT checkpoint for the customized 12-task RoboCasa
 benchmark using the recurrent gated-delta memory encoder.
 
-- Stage 1: trained 180,000 updates, selected at 160,000
-- Stage 2: 150,000 updates
+- Stage 1: trained 180,000 updates, selected at 160,000, W&B `far-wandb/Robocasa/ag6ytzrm`
+- Stage 2: 150,000 updates, W&B `far-wandb/Robocasa/1ozs9qy8`
 - memory tokens: 256 x 1024
 - policy horizon: 20
 - training scan: chunkwise WY
@@ -27,12 +27,12 @@ The Stage-2 model-level auxiliary weight was zero. The archived history YAML
 retains a legacy `aux_weight: 0.25` preload flag, which caused auxiliary
 targets to be read but did not contribute an auxiliary loss.
 
-Use `Amazon-FAR/imprint_datasets`, paths `simulation/robocasa/600` and
-`simulation/robocasa/environment`, for the training store and patched
-RoboCasa 1.0.1 simulator source. This release contains the complete Stage-2
-Orbax parameter checkpoint, including the PaliGemma/Gemma backbone. It is a
-Gemma model derivative and is distributed subject to the included Gemma Terms
-of Use. Large simulator assets are not duplicated.
+Use `yilin-wu/imprint-robocasa-600` for training and
+`yilin-wu/imprint-robocasa-environment` for the exact patched RoboCasa 1.0.1
+simulator source. This release contains the complete Stage-2 Orbax parameter
+checkpoint, including the PaliGemma/Gemma backbone. It is a Gemma model
+derivative and is distributed subject to the included Gemma Terms of Use.
+Large simulator assets are not duplicated.
 
 The matching training store, patched simulator source, and reproducibility
 code are publicly available under the IMPRINT release.

simulation/robocasa/recurrent/SHA256SUMS

@@ -1,8 +1,8 @@
 3e2c24001f9ef57bf7ec959a3658fbb49cdad113cdf394c264da9d16f9bdd132  ./LICENSE_GEMMA.txt
 5b21d7116ac42a0ab6141488cb654c2411557ac977e2276accb88ef774f28c19  ./NOTICE
-89aeee3c2bbdc50d23e1eec66805c20d3db55f74c62fa5836b30d2cc63d1a87a  ./README.md
+1c4d7719baef48359e81dc02e729ec9361c51acf5ba3bd788f21dcd3268b5277  ./README.md
 3f5ee9416db5224d6effaea03e799540023817301a0ba2677922a31709216b1d  ./config/history_config.yaml
-5916c46494b1d56c6b05d3620a66634f8972eabe469cdd24336f0094824d33de  ./config/input_provenance.json
+4143882de7d4e41230b03bd95a7001745139f6a757e64b4db3afa2857ec5b9e7  ./config/input_provenance.json
 39a8f9a475d18c9167a3872c7679a348b81a165776fe3e16c391f354c7ad8514  ./config/norm_stats.json
 7567aca14012be88045a10203432c7117c128b4b1455718143c429cea8155fc1  ./stage1/stage1_step160000.npz
 fe7ad82eeb9e1096c90f433a50fb4b2fabe000779a6d14bfe8ee317f6aea9c72  ./stage2/step-150000/_CHECKPOINT_METADATA

simulation/robocasa/recurrent/config/input_provenance.json

@@ -5,7 +5,7 @@
   "checkpoint_metadata_sha256": "fe7ad82eeb9e1096c90f433a50fb4b2fabe000779a6d14bfe8ee317f6aea9c72",
   "checkpoint_path": "stage2/step-150000",
   "checkpoint_step": 150000,
-  "dataset": "Amazon-FAR/imprint_datasets:simulation/robocasa/600",
+  "dataset": "yilin-wu/imprint-robocasa-600",
   "eval_consolidate_all": 1,
   "execution_horizon": 20,
   "freeze_variant": "image_only",
@@ -20,7 +20,9 @@
   "prediction_horizon": 20,
   "stage1_checkpoint": "stage1/stage1_step160000.npz",
   "stage1_sha256": "7567aca14012be88045a10203432c7117c128b4b1455718143c429cea8155fc1",
+  "stage1_wandb_run": "far-wandb/Robocasa/ag6ytzrm",
   "stage2_auxiliary_loss_weight": 0.0,
   "stage2_memory_encoder": "finetuned",
+  "stage2_wandb_run": "far-wandb/Robocasa/1ozs9qy8",
   "visual_feature_stream": "dinov3"
 }