Upload batch 46: 223 files (7.45 GiB)

.gitattributes

@@ -76,3 +76,27 @@ checkpoints/predictor_v4_timestep_rollout_detached_timestepfp32_blocks1-52_overf
 checkpoints/predictor_v4_timestep_rollout_legacy-timestepbf16_blocks1-52_overfit8_seed20260721_fp32master_8gpu_flr1e-4_blr1e-5_schedsteps200_swarmup8_bwarmup8/replay0/swanlab/run-20260721_173532-xqwuujlx/run-xqwuujlx.swanlab filter=lfs diff=lfs merge=lfs -text
 checkpoints/predictor_v4_timestep_rollout_legacy-timestepbf16_blocks1-52_overfit8_seed20260721_fp32master_8gpu_flr1e-4_blr1e-5_schedsteps200_swarmup8_bwarmup8/replay0p2/swanlab/run-20260721_180252-jskm98hf/run-jskm98hf.swanlab filter=lfs diff=lfs merge=lfs -text
 checkpoints/predictor_v4_timestep_rollout_legacy-timestepbf16_blocks1-52_overfit8_seed20260721_fp32master_8gpu_flr1e-4_blr1e-5_schedsteps200_swarmup8_bwarmup8/replay0p4/swanlab/run-20260721_183105-7bjwo4bv/run-7bjwo4bv.swanlab filter=lfs diff=lfs merge=lfs -text
+confidence_validation/plots/risk_coverage.png filter=lfs diff=lfs merge=lfs -text
+confidence_validation_wide4096x2_c1/plots/risk_coverage.png filter=lfs diff=lfs merge=lfs -text
+confidence_validation_wide4096x2_c2/plots/risk_coverage.png filter=lfs diff=lfs merge=lfs -text
+episode_000004/episode_000000.mp4 filter=lfs diff=lfs merge=lfs -text
+episode_000004/episode_000001.mp4 filter=lfs diff=lfs merge=lfs -text
+episode_000004/episode_000002.mp4 filter=lfs diff=lfs merge=lfs -text
+episode_000004/episode_000003.mp4 filter=lfs diff=lfs merge=lfs -text
+episode_000004/episode_000004.mp4 filter=lfs diff=lfs merge=lfs -text
+episode_000004/episode_000005.mp4 filter=lfs diff=lfs merge=lfs -text
+episode_000004/episode_000006.mp4 filter=lfs diff=lfs merge=lfs -text
+episode_000004/episode_000007.mp4 filter=lfs diff=lfs merge=lfs -text
+episode_000004/episode_000008.mp4 filter=lfs diff=lfs merge=lfs -text
+episode_000004/episode_000009.mp4 filter=lfs diff=lfs merge=lfs -text
+episode_000004/episode_000010.mp4 filter=lfs diff=lfs merge=lfs -text
+episode_000004/episode_000011.mp4 filter=lfs diff=lfs merge=lfs -text
+episode_000004/episode_000012.mp4 filter=lfs diff=lfs merge=lfs -text
+episode_000004/episode_000013.mp4 filter=lfs diff=lfs merge=lfs -text
+episode_000004/episode_000014.mp4 filter=lfs diff=lfs merge=lfs -text
+episode_000004/hyworldplay_outputs/episode_000004_i2v_ffff_forward_then_backward_clamp_bottle_seed0/gen.mp4 filter=lfs diff=lfs merge=lfs -text
+episode_000004/hyworldplay_outputs/episode_000004_i2v_ffff_static_clamp_bottle_seed0/gen.mp4 filter=lfs diff=lfs merge=lfs -text
+episode_000004/hyworldplay_outputs/episode_000004_i2v_stage1_fp32master_fppf_forward_then_backward_clamp_bottle_seed0/predictor/case_0000_action_00.mp4 filter=lfs diff=lfs merge=lfs -text
+episode_000004/i2v_input_000000/26-15/episode_000000_first_frame.png filter=lfs diff=lfs merge=lfs -text
+outputs/c1_threshold_smoke/predictor/case_0000_action_00.mp4 filter=lfs diff=lfs merge=lfs -text
+outputs/hf_snapshot_repaired_9b49404/text_encoder/Glyph-SDXL-v2/assets/Arial.ttf filter=lfs diff=lfs merge=lfs -text

checkpoints/smoke_predictor_v4_random_exit_dmd_p2/latest/predictor.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f13eb217226d3bcf37f1fdbb7599bf7acacb998e875a7265e9aa779750a0849d
+size 696559936

checkpoints/smoke_predictor_v4_random_exit_dmd_p2/latest/predictor_training.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:61ce9cedd8445e08a9f2d561f4bf832f4e26f86500b75d5fd77a084978f08e79
+size 2668808903

checkpoints/smoke_predictor_v4_random_exit_dmd_p2/latest/state.json

@@ -0,0 +1,11 @@
+{
+  "critic_step": 1,
+  "exit_counts": [
+    0,
+    0,
+    1,
+    0
+  ],
+  "final": true,
+  "student_step": 1
+}

checkpoints/smoke_predictor_v4_random_exit_dmd_p2/train_config.json

@@ -0,0 +1,48 @@
+{
+  "base_model": "/mnt/local_nvme/cz/projects/HY-WorldPlay-DEV-Predictor/outputs/model_cache_repaired/HunyuanVideo-1.5-9b49404-v2",
+  "blocks_lr": 5e-07,
+  "cases_path": "/mnt/local_nvme/cz/projects/HY-WorldPlay-DEV-Predictor/datasets/hyworldplay_predictor_prefeature_v4/cases.jsonl",
+  "critic_updates_per_student": 1,
+  "data_root": "/mnt/local_nvme/cz/projects/HY-WorldPlay-DEV-Predictor/datasets/hyworldplay_predictor_prefeature_v4",
+  "fake_score_grad_clip": 10.0,
+  "fake_score_lr": 2e-07,
+  "fake_score_model": "50step_bidirectional_model_trainable",
+  "flow_shift": 5.0,
+  "forced_exit_step": 2,
+  "fusion_lr": 5e-06,
+  "keep_checkpoints": 1,
+  "latent_seed": 0,
+  "log_every": 1,
+  "max_steps": 1,
+  "max_trajectories": 8,
+  "output_dir": "/mnt/local_nvme/cz/projects/HY-WorldPlay-DEV-Predictor/checkpoints/smoke_predictor_v4_random_exit_dmd_p2",
+  "predictor_grad_clip": 1.0,
+  "predictor_trainable_parameters": 197218816,
+  "random_exits": [
+    1,
+    2,
+    3
+  ],
+  "real_guidance_scale": 6.0,
+  "real_score_model": "50step_bidirectional_model_frozen",
+  "regression_loss": false,
+  "resume": null,
+  "rollout_model": "4step_ar_distilled_action_model",
+  "rollout_teacher_checkpoint": "/mnt/local_nvme/cz/checkpoints/hy_worldplay/huggingface/hub/models--tencent--HY-WorldPlay/snapshots/f4c29235647707b571479a69b569e4166f9f5bf8/ar_distilled_action_model/diffusion_pytorch_model.safetensors",
+  "save_every": 1,
+  "score_teacher_checkpoint": "/mnt/local_nvme/cz/checkpoints/hy_worldplay/huggingface/hub/models--tencent--HY-WorldPlay/snapshots/f4c29235647707b571479a69b569e4166f9f5bf8/bidirectional_model/diffusion_pytorch_model.safetensors",
+  "score_timestep_max": 0.98,
+  "score_timestep_min": 0.02,
+  "seed": 20260818,
+  "stage1_expected_step": 2000,
+  "stage1_training_state": "/mnt/local_nvme/cz/projects/HY-WorldPlay-DEV-Predictor/checkpoints/predictor_v4_prefeature_blocks1-52_fp32master_8gpu_flr1e-4_blr1e-5_schedsteps2000_swarmup100_bwarmup100_swanlab/training_latest.pt",
+  "swanlab_experiment": "predictor-v4-random-exit-dmd-predictor-only-stage1step2000-8gpu",
+  "swanlab_mode": "disabled",
+  "swanlab_project": "HY-WorldPlay-Predictor-v4",
+  "swanlab_workspace": null,
+  "training_mode": "predictor_only_random_exit_dmd",
+  "validate_config_only": false,
+  "warmup_steps": 100,
+  "weight_decay": 0.01,
+  "world_size": 8
+}

checkpoints/smoke_predictor_v4_random_exit_dmd_p2/train_log.jsonl

@@ -0,0 +1 @@
+{"critic_grad_norm": 1.21875, "critic_loss": 0.05237160646356642, "critic_step": 1, "critic_timestep": 797.6681823730469, "dmd_gradient_norm": 0.47206708416342735, "dmd_loss": 0.2637263061478734, "dmd_normalizer": 0.1062138332054019, "dmd_score_timestep": 635.6212921142578, "exit_count_p1": 0, "exit_count_p2": 1, "exit_count_p3": 0, "exit_step": 2, "lr_fake_score": 2e-07, "lr_predictor_blocks": 1e-08, "lr_predictor_fusion": 1.0000000000000001e-07, "peak_memory_gib": 65.39037084579468, "predictor_grad_norm": 0.07057937979698181, "step_time_s": 202.15974524412377, "student_step": 1}

checkpoints/smoke_predictor_v4_random_exit_dmd_p3.log

@@ -0,0 +1,33 @@
+HY-World 1.5 loading from:  /mnt/local_nvme/cz/checkpoints/hy_worldplay/huggingface/hub/models--tencent--HY-WorldPlay/snapshots/f4c29235647707b571479a69b569e4166f9f5bf8/ar_distilled_action_model/diffusion_pytorch_model.safetensors
+HY-World 1.5 loading from:  /mnt/local_nvme/cz/checkpoints/hy_worldplay/huggingface/hub/models--tencent--HY-WorldPlay/snapshots/f4c29235647707b571479a69b569e4166f9f5bf8/ar_distilled_action_model/diffusion_pytorch_model.safetensors
+HY-World 1.5 loading from:  /mnt/local_nvme/cz/checkpoints/hy_worldplay/huggingface/hub/models--tencent--HY-WorldPlay/snapshots/f4c29235647707b571479a69b569e4166f9f5bf8/ar_distilled_action_model/diffusion_pytorch_model.safetensors
+HY-World 1.5 loading from:  /mnt/local_nvme/cz/checkpoints/hy_worldplay/huggingface/hub/models--tencent--HY-WorldPlay/snapshots/f4c29235647707b571479a69b569e4166f9f5bf8/ar_distilled_action_model/diffusion_pytorch_model.safetensors
+HY-World 1.5 loading from:  /mnt/local_nvme/cz/checkpoints/hy_worldplay/huggingface/hub/models--tencent--HY-WorldPlay/snapshots/f4c29235647707b571479a69b569e4166f9f5bf8/ar_distilled_action_model/diffusion_pytorch_model.safetensors
+/mnt/local_nvme/cz/projects/HY-WorldPlay-DEV-Predictor/hyvideo/models/text_encoders/byT5/__init__.py:74: FutureWarning: You are using `torch.load` with `weights_only=False` (the current default value), which uses the default pickle module implicitly. It is possible to construct malicious pickle data which will execute arbitrary code during unpickling (See https://github.com/pytorch/pytorch/blob/main/SECURITY.md#untrusted-models for more details). In a future release, the default value for `weights_only` will be flipped to `True`. This limits the functions that could be executed during unpickling. Arbitrary objects will no longer be allowed to be loaded via this mode unless they are explicitly allowlisted by the user via `torch.serialization.add_safe_globals`. We recommend you start setting `weights_only=True` for any use case where you don't have full control of the loaded file. Please open an issue on GitHub for any issues related to this experimental feature.
+  byt5_state_dict = torch.load(args["byT5_ckpt_path"], map_location=device)
+HY-World 1.5 loading from:  /mnt/local_nvme/cz/checkpoints/hy_worldplay/huggingface/hub/models--tencent--HY-WorldPlay/snapshots/f4c29235647707b571479a69b569e4166f9f5bf8/ar_distilled_action_model/diffusion_pytorch_model.safetensors
+HY-World 1.5 loading from:  /mnt/local_nvme/cz/checkpoints/hy_worldplay/huggingface/hub/models--tencent--HY-WorldPlay/snapshots/f4c29235647707b571479a69b569e4166f9f5bf8/ar_distilled_action_model/diffusion_pytorch_model.safetensors
+/mnt/local_nvme/cz/projects/HY-WorldPlay-DEV-Predictor/hyvideo/models/text_encoders/byT5/__init__.py:74: FutureWarning: You are using `torch.load` with `weights_only=False` (the current default value), which uses the default pickle module implicitly. It is possible to construct malicious pickle data which will execute arbitrary code during unpickling (See https://github.com/pytorch/pytorch/blob/main/SECURITY.md#untrusted-models for more details). In a future release, the default value for `weights_only` will be flipped to `True`. This limits the functions that could be executed during unpickling. Arbitrary objects will no longer be allowed to be loaded via this mode unless they are explicitly allowlisted by the user via `torch.serialization.add_safe_globals`. We recommend you start setting `weights_only=True` for any use case where you don't have full control of the loaded file. Please open an issue on GitHub for any issues related to this experimental feature.
+  byt5_state_dict = torch.load(args["byT5_ckpt_path"], map_location=device)
+/mnt/local_nvme/cz/projects/HY-WorldPlay-DEV-Predictor/hyvideo/models/text_encoders/byT5/__init__.py:74: FutureWarning: You are using `torch.load` with `weights_only=False` (the current default value), which uses the default pickle module implicitly. It is possible to construct malicious pickle data which will execute arbitrary code during unpickling (See https://github.com/pytorch/pytorch/blob/main/SECURITY.md#untrusted-models for more details). In a future release, the default value for `weights_only` will be flipped to `True`. This limits the functions that could be executed during unpickling. Arbitrary objects will no longer be allowed to be loaded via this mode unless they are explicitly allowlisted by the user via `torch.serialization.add_safe_globals`. We recommend you start setting `weights_only=True` for any use case where you don't have full control of the loaded file. Please open an issue on GitHub for any issues related to this experimental feature.
+  byt5_state_dict = torch.load(args["byT5_ckpt_path"], map_location=device)
+HY-World 1.5 loading from:  /mnt/local_nvme/cz/checkpoints/hy_worldplay/huggingface/hub/models--tencent--HY-WorldPlay/snapshots/f4c29235647707b571479a69b569e4166f9f5bf8/ar_distilled_action_model/diffusion_pytorch_model.safetensors
+/mnt/local_nvme/cz/projects/HY-WorldPlay-DEV-Predictor/hyvideo/models/text_encoders/byT5/__init__.py:74: FutureWarning: You are using `torch.load` with `weights_only=False` (the current default value), which uses the default pickle module implicitly. It is possible to construct malicious pickle data which will execute arbitrary code during unpickling (See https://github.com/pytorch/pytorch/blob/main/SECURITY.md#untrusted-models for more details). In a future release, the default value for `weights_only` will be flipped to `True`. This limits the functions that could be executed during unpickling. Arbitrary objects will no longer be allowed to be loaded via this mode unless they are explicitly allowlisted by the user via `torch.serialization.add_safe_globals`. We recommend you start setting `weights_only=True` for any use case where you don't have full control of the loaded file. Please open an issue on GitHub for any issues related to this experimental feature.
+  byt5_state_dict = torch.load(args["byT5_ckpt_path"], map_location=device)
+/mnt/local_nvme/cz/projects/HY-WorldPlay-DEV-Predictor/hyvideo/models/text_encoders/byT5/__init__.py:74: FutureWarning: You are using `torch.load` with `weights_only=False` (the current default value), which uses the default pickle module implicitly. It is possible to construct malicious pickle data which will execute arbitrary code during unpickling (See https://github.com/pytorch/pytorch/blob/main/SECURITY.md#untrusted-models for more details). In a future release, the default value for `weights_only` will be flipped to `True`. This limits the functions that could be executed during unpickling. Arbitrary objects will no longer be allowed to be loaded via this mode unless they are explicitly allowlisted by the user via `torch.serialization.add_safe_globals`. We recommend you start setting `weights_only=True` for any use case where you don't have full control of the loaded file. Please open an issue on GitHub for any issues related to this experimental feature.
+  byt5_state_dict = torch.load(args["byT5_ckpt_path"], map_location=device)
+
+  byt5_state_dict = torch.load(args["byT5_ckpt_path"], map_location=device)
+/mnt/local_nvme/cz/projects/HY-WorldPlay-DEV-Predictor/hyvideo/models/text_encoders/byT5/__init__.py:74: FutureWarning: You are using `torch.load` with `weights_only=False` (the current default value), which uses the default pickle module implicitly. It is possible to construct malicious pickle data which will execute arbitrary code during unpickling (See https://github.com/pytorch/pytorch/blob/main/SECURITY.md#untrusted-models for more details). In a future release, the default value for `weights_only` will be flipped to `True`. This limits the functions that could be executed during unpickling. Arbitrary objects will no longer be allowed to be loaded via this mode unless they are explicitly allowlisted by the user via `torch.serialization.add_safe_globals`. We recommend you start setting `weights_only=True` for any use case where you don't have full control of the loaded file. Please open an issue on GitHub for any issues related to this experimental feature.
+  byt5_state_dict = torch.load(args["byT5_ckpt_path"], map_location=device)
+
+  byt5_state_dict = torch.load(args["byT5_ckpt_path"], map_location=device)
+
+
+
+
+
+
+
+
+{"critic_grad_norm": 0.69921875, "critic_loss": 0.05981058580800891, "critic_step": 1, "critic_timestep": 797.6681823730469, "dmd_gradient_norm": 0.43714500218629837, "dmd_loss": 0.2176895122975111, "dmd_normalizer": 0.11242425534874201, "dmd_score_timestep": 635.6212921142578, "exit_count_p1": 0, "exit_count_p2": 0, "exit_count_p3": 1, "exit_step": 3, "lr_fake_score": 2e-07, "lr_predictor_blocks": 1e-08, "lr_predictor_fusion": 1.0000000000000001e-07, "peak_memory_gib": 65.43626928329468, "predictor_grad_norm": 0.06792225688695908, "step_time_s": 184.85065027012388, "student_step": 1}

checkpoints/smoke_predictor_v4_random_exit_dmd_p3/train_config.json

@@ -0,0 +1,49 @@
+{
+  "base_model": "/mnt/local_nvme/cz/projects/HY-WorldPlay-DEV-Predictor/outputs/model_cache_repaired/HunyuanVideo-1.5-9b49404-v2",
+  "blocks_lr": 5e-07,
+  "cases_path": "/mnt/local_nvme/cz/projects/HY-WorldPlay-DEV-Predictor/datasets/hyworldplay_predictor_prefeature_v4/cases.jsonl",
+  "critic_updates_per_student": 1,
+  "data_root": "/mnt/local_nvme/cz/projects/HY-WorldPlay-DEV-Predictor/datasets/hyworldplay_predictor_prefeature_v4",
+  "fake_score_grad_clip": 10.0,
+  "fake_score_lr": 2e-07,
+  "fake_score_model": "50step_bidirectional_model_trainable",
+  "flow_shift": 5.0,
+  "forced_exit_step": 3,
+  "fusion_lr": 5e-06,
+  "keep_checkpoints": 2,
+  "latent_seed": 0,
+  "log_every": 1,
+  "max_steps": 1,
+  "max_trajectories": 8,
+  "no_save": true,
+  "output_dir": "/mnt/local_nvme/cz/projects/HY-WorldPlay-DEV-Predictor/checkpoints/smoke_predictor_v4_random_exit_dmd_p3",
+  "predictor_grad_clip": 1.0,
+  "predictor_trainable_parameters": 197218816,
+  "random_exits": [
+    1,
+    2,
+    3
+  ],
+  "real_guidance_scale": 6.0,
+  "real_score_model": "50step_bidirectional_model_frozen",
+  "regression_loss": false,
+  "resume": null,
+  "rollout_model": "4step_ar_distilled_action_model",
+  "rollout_teacher_checkpoint": "/mnt/local_nvme/cz/checkpoints/hy_worldplay/huggingface/hub/models--tencent--HY-WorldPlay/snapshots/f4c29235647707b571479a69b569e4166f9f5bf8/ar_distilled_action_model/diffusion_pytorch_model.safetensors",
+  "save_every": 200,
+  "score_teacher_checkpoint": "/mnt/local_nvme/cz/checkpoints/hy_worldplay/huggingface/hub/models--tencent--HY-WorldPlay/snapshots/f4c29235647707b571479a69b569e4166f9f5bf8/bidirectional_model/diffusion_pytorch_model.safetensors",
+  "score_timestep_max": 0.98,
+  "score_timestep_min": 0.02,
+  "seed": 20260818,
+  "stage1_expected_step": 2000,
+  "stage1_training_state": "/mnt/local_nvme/cz/projects/HY-WorldPlay-DEV-Predictor/checkpoints/predictor_v4_prefeature_blocks1-52_fp32master_8gpu_flr1e-4_blr1e-5_schedsteps2000_swarmup100_bwarmup100_swanlab/training_latest.pt",
+  "swanlab_experiment": "predictor-v4-random-exit-dmd-predictor-only-stage1step2000-8gpu",
+  "swanlab_mode": "disabled",
+  "swanlab_project": "HY-WorldPlay-Predictor-v4",
+  "swanlab_workspace": null,
+  "training_mode": "predictor_only_random_exit_dmd",
+  "validate_config_only": false,
+  "warmup_steps": 100,
+  "weight_decay": 0.01,
+  "world_size": 8
+}

checkpoints/smoke_predictor_v4_random_exit_dmd_p3/train_log.jsonl

@@ -0,0 +1 @@
+{"critic_grad_norm": 0.69921875, "critic_loss": 0.05981058580800891, "critic_step": 1, "critic_timestep": 797.6681823730469, "dmd_gradient_norm": 0.43714500218629837, "dmd_loss": 0.2176895122975111, "dmd_normalizer": 0.11242425534874201, "dmd_score_timestep": 635.6212921142578, "exit_count_p1": 0, "exit_count_p2": 0, "exit_count_p3": 1, "exit_step": 3, "lr_fake_score": 2e-07, "lr_predictor_blocks": 1e-08, "lr_predictor_fusion": 1.0000000000000001e-07, "peak_memory_gib": 65.43626928329468, "predictor_grad_norm": 0.06792225688695908, "step_time_s": 184.85065027012388, "student_step": 1}

checkpoints/smoke_predictor_v4_random_exit_dmd_resume.log

@@ -0,0 +1,32 @@
+HY-World 1.5 loading from:  /mnt/local_nvme/cz/checkpoints/hy_worldplay/huggingface/hub/models--tencent--HY-WorldPlay/snapshots/f4c29235647707b571479a69b569e4166f9f5bf8/ar_distilled_action_model/diffusion_pytorch_model.safetensors
+HY-World 1.5 loading from:  /mnt/local_nvme/cz/checkpoints/hy_worldplay/huggingface/hub/models--tencent--HY-WorldPlay/snapshots/f4c29235647707b571479a69b569e4166f9f5bf8/ar_distilled_action_model/diffusion_pytorch_model.safetensors
+/mnt/local_nvme/cz/projects/HY-WorldPlay-DEV-Predictor/hyvideo/models/text_encoders/byT5/__init__.py:74: FutureWarning: You are using `torch.load` with `weights_only=False` (the current default value), which uses the default pickle module implicitly. It is possible to construct malicious pickle data which will execute arbitrary code during unpickling (See https://github.com/pytorch/pytorch/blob/main/SECURITY.md#untrusted-models for more details). In a future release, the default value for `weights_only` will be flipped to `True`. This limits the functions that could be executed during unpickling. Arbitrary objects will no longer be allowed to be loaded via this mode unless they are explicitly allowlisted by the user via `torch.serialization.add_safe_globals`. We recommend you start setting `weights_only=True` for any use case where you don't have full control of the loaded file. Please open an issue on GitHub for any issues related to this experimental feature.
+  byt5_state_dict = torch.load(args["byT5_ckpt_path"], map_location=device)
+HY-World 1.5 loading from:  /mnt/local_nvme/cz/checkpoints/hy_worldplay/huggingface/hub/models--tencent--HY-WorldPlay/snapshots/f4c29235647707b571479a69b569e4166f9f5bf8/ar_distilled_action_model/diffusion_pytorch_model.safetensors
+HY-World 1.5 loading from:  /mnt/local_nvme/cz/checkpoints/hy_worldplay/huggingface/hub/models--tencent--HY-WorldPlay/snapshots/f4c29235647707b571479a69b569e4166f9f5bf8/ar_distilled_action_model/diffusion_pytorch_model.safetensors
+/mnt/local_nvme/cz/projects/HY-WorldPlay-DEV-Predictor/hyvideo/models/text_encoders/byT5/__init__.py:74: FutureWarning: You are using `torch.load` with `weights_only=False` (the current default value), which uses the default pickle module implicitly. It is possible to construct malicious pickle data which will execute arbitrary code during unpickling (See https://github.com/pytorch/pytorch/blob/main/SECURITY.md#untrusted-models for more details). In a future release, the default value for `weights_only` will be flipped to `True`. This limits the functions that could be executed during unpickling. Arbitrary objects will no longer be allowed to be loaded via this mode unless they are explicitly allowlisted by the user via `torch.serialization.add_safe_globals`. We recommend you start setting `weights_only=True` for any use case where you don't have full control of the loaded file. Please open an issue on GitHub for any issues related to this experimental feature.
+  byt5_state_dict = torch.load(args["byT5_ckpt_path"], map_location=device)
+
+/mnt/local_nvme/cz/projects/HY-WorldPlay-DEV-Predictor/hyvideo/models/text_encoders/byT5/__init__.py:74: FutureWarning: You are using `torch.load` with `weights_only=False` (the current default value), which uses the default pickle module implicitly. It is possible to construct malicious pickle data which will execute arbitrary code during unpickling (See https://github.com/pytorch/pytorch/blob/main/SECURITY.md#untrusted-models for more details). In a future release, the default value for `weights_only` will be flipped to `True`. This limits the functions that could be executed during unpickling. Arbitrary objects will no longer be allowed to be loaded via this mode unless they are explicitly allowlisted by the user via `torch.serialization.add_safe_globals`. We recommend you start setting `weights_only=True` for any use case where you don't have full control of the loaded file. Please open an issue on GitHub for any issues related to this experimental feature.
+  byt5_state_dict = torch.load(args["byT5_ckpt_path"], map_location=device)
+HY-World 1.5 loading from:  /mnt/local_nvme/cz/checkpoints/hy_worldplay/huggingface/hub/models--tencent--HY-WorldPlay/snapshots/f4c29235647707b571479a69b569e4166f9f5bf8/ar_distilled_action_model/diffusion_pytorch_model.safetensors
+HY-World 1.5 loading from:  /mnt/local_nvme/cz/checkpoints/hy_worldplay/huggingface/hub/models--tencent--HY-WorldPlay/snapshots/f4c29235647707b571479a69b569e4166f9f5bf8/ar_distilled_action_model/diffusion_pytorch_model.safetensors
+HY-World 1.5 loading from:  /mnt/local_nvme/cz/checkpoints/hy_worldplay/huggingface/hub/models--tencent--HY-WorldPlay/snapshots/f4c29235647707b571479a69b569e4166f9f5bf8/ar_distilled_action_model/diffusion_pytorch_model.safetensors
+/mnt/local_nvme/cz/projects/HY-WorldPlay-DEV-Predictor/hyvideo/models/text_encoders/byT5/__init__.py:74: FutureWarning: You are using `torch.load` with `weights_only=False` (the current default value), which uses the default pickle module implicitly. It is possible to construct malicious pickle data which will execute arbitrary code during unpickling (See https://github.com/pytorch/pytorch/blob/main/SECURITY.md#untrusted-models for more details). In a future release, the default value for `weights_only` will be flipped to `True`. This limits the functions that could be executed during unpickling. Arbitrary objects will no longer be allowed to be loaded via this mode unless they are explicitly allowlisted by the user via `torch.serialization.add_safe_globals`. We recommend you start setting `weights_only=True` for any use case where you don't have full control of the loaded file. Please open an issue on GitHub for any issues related to this experimental feature.
+  byt5_state_dict = torch.load(args["byT5_ckpt_path"], map_location=device)
+
+  byt5_state_dict = torch.load(args["byT5_ckpt_path"], map_location=device)
+
+  byt5_state_dict = torch.load(args["byT5_ckpt_path"], map_location=device)
+/mnt/local_nvme/cz/projects/HY-WorldPlay-DEV-Predictor/hyvideo/models/text_encoders/byT5/__init__.py:74: FutureWarning: You are using `torch.load` with `weights_only=False` (the current default value), which uses the default pickle module implicitly. It is possible to construct malicious pickle data which will execute arbitrary code during unpickling (See https://github.com/pytorch/pytorch/blob/main/SECURITY.md#untrusted-models for more details). In a future release, the default value for `weights_only` will be flipped to `True`. This limits the functions that could be executed during unpickling. Arbitrary objects will no longer be allowed to be loaded via this mode unless they are explicitly allowlisted by the user via `torch.serialization.add_safe_globals`. We recommend you start setting `weights_only=True` for any use case where you don't have full control of the loaded file. Please open an issue on GitHub for any issues related to this experimental feature.
+  byt5_state_dict = torch.load(args["byT5_ckpt_path"], map_location=device)
+/mnt/local_nvme/cz/projects/HY-WorldPlay-DEV-Predictor/hyvideo/models/text_encoders/byT5/__init__.py:74: FutureWarning: You are using `torch.load` with `weights_only=False` (the current default value), which uses the default pickle module implicitly. It is possible to construct malicious pickle data which will execute arbitrary code during unpickling (See https://github.com/pytorch/pytorch/blob/main/SECURITY.md#untrusted-models for more details). In a future release, the default value for `weights_only` will be flipped to `True`. This limits the functions that could be executed during unpickling. Arbitrary objects will no longer be allowed to be loaded via this mode unless they are explicitly allowlisted by the user via `torch.serialization.add_safe_globals`. We recommend you start setting `weights_only=True` for any use case where you don't have full control of the loaded file. Please open an issue on GitHub for any issues related to this experimental feature.
+  byt5_state_dict = torch.load(args["byT5_ckpt_path"], map_location=device)
+
+
+
+
+
+
+
+

checkpoints/smoke_predictor_v4_random_exit_dmd_resume/train_config.json

@@ -0,0 +1,49 @@
+{
+  "base_model": "/mnt/local_nvme/cz/projects/HY-WorldPlay-DEV-Predictor/outputs/model_cache_repaired/HunyuanVideo-1.5-9b49404-v2",
+  "blocks_lr": 5e-07,
+  "cases_path": "/mnt/local_nvme/cz/projects/HY-WorldPlay-DEV-Predictor/datasets/hyworldplay_predictor_prefeature_v4/cases.jsonl",
+  "critic_updates_per_student": 1,
+  "data_root": "/mnt/local_nvme/cz/projects/HY-WorldPlay-DEV-Predictor/datasets/hyworldplay_predictor_prefeature_v4",
+  "fake_score_grad_clip": 10.0,
+  "fake_score_lr": 2e-07,
+  "fake_score_model": "50step_bidirectional_model_trainable",
+  "flow_shift": 5.0,
+  "forced_exit_step": null,
+  "fusion_lr": 5e-06,
+  "keep_checkpoints": 2,
+  "latent_seed": 0,
+  "log_every": 1,
+  "max_steps": 1,
+  "max_trajectories": 8,
+  "no_save": true,
+  "output_dir": "/mnt/local_nvme/cz/projects/HY-WorldPlay-DEV-Predictor/checkpoints/smoke_predictor_v4_random_exit_dmd_resume",
+  "predictor_grad_clip": 1.0,
+  "predictor_trainable_parameters": 197218816,
+  "random_exits": [
+    1,
+    2,
+    3
+  ],
+  "real_guidance_scale": 6.0,
+  "real_score_model": "50step_bidirectional_model_frozen",
+  "regression_loss": false,
+  "resume": "checkpoints/smoke_predictor_v4_random_exit_dmd_p2/checkpoint_final",
+  "rollout_model": "4step_ar_distilled_action_model",
+  "rollout_teacher_checkpoint": "/mnt/local_nvme/cz/checkpoints/hy_worldplay/huggingface/hub/models--tencent--HY-WorldPlay/snapshots/f4c29235647707b571479a69b569e4166f9f5bf8/ar_distilled_action_model/diffusion_pytorch_model.safetensors",
+  "save_every": 200,
+  "score_teacher_checkpoint": "/mnt/local_nvme/cz/checkpoints/hy_worldplay/huggingface/hub/models--tencent--HY-WorldPlay/snapshots/f4c29235647707b571479a69b569e4166f9f5bf8/bidirectional_model/diffusion_pytorch_model.safetensors",
+  "score_timestep_max": 0.98,
+  "score_timestep_min": 0.02,
+  "seed": 20260818,
+  "stage1_expected_step": 2000,
+  "stage1_training_state": "/mnt/local_nvme/cz/projects/HY-WorldPlay-DEV-Predictor/checkpoints/predictor_v4_prefeature_blocks1-52_fp32master_8gpu_flr1e-4_blr1e-5_schedsteps2000_swarmup100_bwarmup100_swanlab/training_latest.pt",
+  "swanlab_experiment": "predictor-v4-random-exit-dmd-predictor-only-stage1step2000-8gpu",
+  "swanlab_mode": "disabled",
+  "swanlab_project": "HY-WorldPlay-Predictor-v4",
+  "swanlab_workspace": null,
+  "training_mode": "predictor_only_random_exit_dmd",
+  "validate_config_only": false,
+  "warmup_steps": 100,
+  "weight_decay": 0.01,
+  "world_size": 8
+}

confidence_validation/c1_threshold_validation25/report.md

@@ -0,0 +1,25 @@
+# C1 threshold validation25 comparison
+
+## Reference metrics against FFFF
+
+| Method | Fallback | P1/P2/P3 fallback | PSNR | SSIM | LPIPS | Pixel MSE | Temporal MSE |
+| --- | ---: | --- | ---: | ---: | ---: | ---: | ---: |
+| FFFF | 100.0% | - | ∞ | 1.00000 | 0.00000 | 0.000000 | 0.000000 |
+| FPPF | 33.3% | - | 21.5933 | 0.74668 | 0.12100 | 0.009692 | 0.008824 |
+| FPPP | 0.0% | - | 20.3085 | 0.69143 | 0.23528 | 0.011888 | 0.008931 |
+| C1-r10 | 14.4% | 0.0%/0.0%/43.1% | 20.6740 | 0.71893 | 0.16237 | 0.011000 | 0.008912 |
+| C1-r25 | 28.6% | 0.0%/0.0%/85.9% | 21.0259 | 0.73270 | 0.13290 | 0.010374 | 0.009166 |
+| C1-r33 | 33.8% | 1.4%/5.3%/94.7% | 21.1834 | 0.73752 | 0.12620 | 0.010172 | 0.009169 |
+| C1-r50 | 51.7% | 13.1%/42.0%/100.0% | 22.1161 | 0.76838 | 0.10507 | 0.008328 | 0.008148 |
+
+## VBench
+
+| Method | Subject | Background | Flicker | Motion | Dynamic | Aesthetic | Imaging | Overall | Core5 mean |
+| --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: |
+| FFFF | 0.948075 | 0.943076 | 0.958449 | 0.989400 | 0.820000 | 0.598486 | 0.734028 | 0.248095 | 0.842613 |
+| FPPF | 0.947461 | 0.943822 | 0.960451 | 0.989683 | 0.800000 | 0.598215 | 0.730930 | 0.249178 | 0.842022 |
+| FPPP | 0.925574 | 0.930874 | 0.965478 | 0.987025 | 0.820000 | 0.563811 | 0.688810 | 0.247289 | 0.819219 |
+| C1-r10 | 0.940300 | 0.938889 | 0.961824 | 0.988701 | 0.820000 | 0.586908 | 0.717366 | 0.249135 | 0.834433 |
+| C1-r25 | 0.946081 | 0.942168 | 0.960483 | 0.989495 | 0.800000 | 0.595933 | 0.727648 | 0.249157 | 0.840265 |
+| C1-r33 | 0.947570 | 0.942669 | 0.959909 | 0.989582 | 0.800000 | 0.597482 | 0.729476 | 0.249080 | 0.841356 |
+| C1-r50 | 0.947762 | 0.943007 | 0.958938 | 0.989504 | 0.800000 | 0.599139 | 0.732626 | 0.248683 | 0.842408 |

confidence_validation/c1_threshold_validation25/summary.csv

@@ -0,0 +1,8 @@
+method,threshold_log_c1,fallback_rate,p1_fallback_rate,p2_fallback_rate,p3_fallback_rate,mean_end_to_end_s,psnr_db,frame_psnr_mean_db,ssim,lpips_alex,pixel_mae,pixel_mse,pixel_rmse,temporal_delta_mse,vbench_subject_consistency,vbench_background_consistency,vbench_temporal_flickering,vbench_motion_smoothness,vbench_dynamic_degree,vbench_aesthetic_quality,vbench_imaging_quality,vbench_overall_consistency,vbench_core5_mean
+FFFF,None,1.0,,,,,inf,inf,1.0,0.0,0.0,0.0,0.0,0.0,0.9480749635855037,0.9430755910565776,0.9584489341647017,0.9893996710686921,0.82,0.5984862497448922,0.7340275594329835,0.24809463396668435,0.8426128069777297
+FPPF,None,0.3333333333333333,,,,,21.593346196430627,23.97453586290849,0.746684046046734,0.12099588543950114,0.04511181315502153,0.009691555944529896,0.09161645448522822,0.008823608518287073,0.9474605375095722,0.9438224349483367,0.9604507553109936,0.989682650584811,0.8,0.5982152584195137,0.730930085635376,0.24917750611901282,0.8420221934195219
+FPPP,None,0.0,,,,,20.308533622256405,22.446092516703374,0.6914340979039671,0.23528402410950516,0.054245410915873954,0.011888072101751463,0.10328612366794497,0.008930881888954901,0.9255741960555315,0.9308735509072581,0.9654783755155171,0.9870247469802622,0.82,0.5638106274604797,0.6888101954261779,0.24728918850421905,0.819218663365942
+C1-r10,-3.523588657,0.1438095238095238,0.0,0.0,0.43142857142857144,84.42822845273069,20.673972463574263,22.60256892630426,0.7189267579293257,0.16236638777270912,0.05034142703045158,0.010999858650310487,0.09900532903010674,0.008911931206093869,0.9403002772648489,0.9388887466922883,0.9618239460903055,0.9887010740619526,0.82,0.5869080236554146,0.7173659597297669,0.2491347920894623,0.8344328162808543
+C1-r25,-3.915276289,0.2861904761904762,0.0,0.0,0.8585714285714285,83.51112565651128,21.025889159217662,23.00035589281393,0.7327034662091734,0.13290133690908554,0.04774270830776542,0.010374391180268139,0.095749054033939,0.009165997090021846,0.9460808533273878,0.9421683231476814,0.9604834818723156,0.9894952394128942,0.8,0.5959329095482826,0.7276483651573181,0.2491569009423256,0.8402651381187128
+C1-r33,-4.329442501,0.3380952380952381,0.014285714285714285,0.05285714285714286,0.9471428571428572,84.58286115615047,21.183387303768146,23.192352492844282,0.7375190385472777,0.12620077199377122,0.046917045371290295,0.010171981018946102,0.09456364497563306,0.009169250550476135,0.9475698670841033,0.9426687129851311,0.9599090755009183,0.9895824586437182,0.8,0.5974819004535675,0.729475989491272,0.2490800651907921,0.8413557857315584
+C1-r50,-4.79892683,0.5171428571428571,0.13142857142857142,0.42,1.0,99.25920022401726,22.11610699774378,24.12363936617357,0.7683776850652694,0.1050749322929606,0.040959788803271946,0.008328448020131038,0.08546495607525889,0.008148231709055835,0.9477616070198916,0.9430070643271169,0.9589383629805901,0.9895042519327572,0.8,0.5991387969255447,0.7326260545639037,0.24868253156542777,0.8424075549538428

confidence_validation/c1_threshold_validation25/summary.json

@@ -0,0 +1,209 @@
+{
+  "protocol": "validation25, seed 0, Stage-1 v4 FP32-master step 2000",
+  "rows": [
+    {
+      "method": "FFFF",
+      "threshold_log_c1": null,
+      "fallback_rate": 1.0,
+      "psnr_db": Infinity,
+      "frame_psnr_mean_db": Infinity,
+      "ssim": 1.0,
+      "lpips_alex": 0.0,
+      "pixel_mae": 0.0,
+      "pixel_mse": 0.0,
+      "pixel_rmse": 0.0,
+      "temporal_delta_mse": 0.0,
+      "vbench_subject_consistency": 0.9480749635855037,
+      "vbench_background_consistency": 0.9430755910565776,
+      "vbench_temporal_flickering": 0.9584489341647017,
+      "vbench_motion_smoothness": 0.9893996710686921,
+      "vbench_dynamic_degree": 0.82,
+      "vbench_aesthetic_quality": 0.5984862497448922,
+      "vbench_imaging_quality": 0.7340275594329835,
+      "vbench_overall_consistency": 0.24809463396668435,
+      "vbench_core5_mean": 0.8426128069777297
+    },
+    {
+      "method": "FPPF",
+      "threshold_log_c1": null,
+      "fallback_rate": 0.3333333333333333,
+      "frame_psnr_mean_db": 23.97453586290849,
+      "lpips_alex": 0.12099588543950114,
+      "pixel_mae": 0.04511181315502153,
+      "pixel_mse": 0.009691555944529896,
+      "pixel_rmse": 0.09161645448522822,
+      "psnr_db": 21.593346196430627,
+      "ssim": 0.746684046046734,
+      "temporal_delta_mse": 0.008823608518287073,
+      "vbench_subject_consistency": 0.9474605375095722,
+      "vbench_background_consistency": 0.9438224349483367,
+      "vbench_temporal_flickering": 0.9604507553109936,
+      "vbench_motion_smoothness": 0.989682650584811,
+      "vbench_dynamic_degree": 0.8,
+      "vbench_aesthetic_quality": 0.5982152584195137,
+      "vbench_imaging_quality": 0.730930085635376,
+      "vbench_overall_consistency": 0.24917750611901282,
+      "vbench_core5_mean": 0.8420221934195219
+    },
+    {
+      "method": "FPPP",
+      "threshold_log_c1": null,
+      "fallback_rate": 0.0,
+      "frame_psnr_mean_db": 22.446092516703374,
+      "lpips_alex": 0.23528402410950516,
+      "pixel_mae": 0.054245410915873954,
+      "pixel_mse": 0.011888072101751463,
+      "pixel_rmse": 0.10328612366794497,
+      "psnr_db": 20.308533622256405,
+      "ssim": 0.6914340979039671,
+      "temporal_delta_mse": 0.008930881888954901,
+      "vbench_subject_consistency": 0.9255741960555315,
+      "vbench_background_consistency": 0.9308735509072581,
+      "vbench_temporal_flickering": 0.9654783755155171,
+      "vbench_motion_smoothness": 0.9870247469802622,
+      "vbench_dynamic_degree": 0.82,
+      "vbench_aesthetic_quality": 0.5638106274604797,
+      "vbench_imaging_quality": 0.6888101954261779,
+      "vbench_overall_consistency": 0.24728918850421905,
+      "vbench_core5_mean": 0.819218663365942
+    },
+    {
+      "method": "C1-r10",
+      "threshold_log_c1": -3.523588657,
+      "fallback_rate": 0.1438095238095238,
+      "p1_fallback_rate": 0.0,
+      "p2_fallback_rate": 0.0,
+      "p3_fallback_rate": 0.43142857142857144,
+      "mean_end_to_end_s": 84.42822845273069,
+      "frame_psnr_mean_db": 22.60256892630426,
+      "lpips_alex": 0.16236638777270912,
+      "pixel_mae": 0.05034142703045158,
+      "pixel_mse": 0.010999858650310487,
+      "pixel_rmse": 0.09900532903010674,
+      "psnr_db": 20.673972463574263,
+      "ssim": 0.7189267579293257,
+      "temporal_delta_mse": 0.008911931206093869,
+      "vbench_subject_consistency": 0.9403002772648489,
+      "vbench_background_consistency": 0.9388887466922883,
+      "vbench_temporal_flickering": 0.9618239460903055,
+      "vbench_motion_smoothness": 0.9887010740619526,
+      "vbench_dynamic_degree": 0.82,
+      "vbench_aesthetic_quality": 0.5869080236554146,
+      "vbench_imaging_quality": 0.7173659597297669,
+      "vbench_overall_consistency": 0.2491347920894623,
+      "vbench_core5_mean": 0.8344328162808543
+    },
+    {
+      "method": "C1-r25",
+      "threshold_log_c1": -3.915276289,
+      "fallback_rate": 0.2861904761904762,
+      "p1_fallback_rate": 0.0,
+      "p2_fallback_rate": 0.0,
+      "p3_fallback_rate": 0.8585714285714285,
+      "mean_end_to_end_s": 83.51112565651128,
+      "frame_psnr_mean_db": 23.00035589281393,
+      "lpips_alex": 0.13290133690908554,
+      "pixel_mae": 0.04774270830776542,
+      "pixel_mse": 0.010374391180268139,
+      "pixel_rmse": 0.095749054033939,
+      "psnr_db": 21.025889159217662,
+      "ssim": 0.7327034662091734,
+      "temporal_delta_mse": 0.009165997090021846,
+      "vbench_subject_consistency": 0.9460808533273878,
+      "vbench_background_consistency": 0.9421683231476814,
+      "vbench_temporal_flickering": 0.9604834818723156,
+      "vbench_motion_smoothness": 0.9894952394128942,
+      "vbench_dynamic_degree": 0.8,
+      "vbench_aesthetic_quality": 0.5959329095482826,
+      "vbench_imaging_quality": 0.7276483651573181,
+      "vbench_overall_consistency": 0.2491569009423256,
+      "vbench_core5_mean": 0.8402651381187128
+    },
+    {
+      "method": "C1-r33",
+      "threshold_log_c1": -4.329442501,
+      "fallback_rate": 0.3380952380952381,
+      "p1_fallback_rate": 0.014285714285714285,
+      "p2_fallback_rate": 0.05285714285714286,
+      "p3_fallback_rate": 0.9471428571428572,
+      "mean_end_to_end_s": 84.58286115615047,
+      "frame_psnr_mean_db": 23.192352492844282,
+      "lpips_alex": 0.12620077199377122,
+      "pixel_mae": 0.046917045371290295,
+      "pixel_mse": 0.010171981018946102,
+      "pixel_rmse": 0.09456364497563306,
+      "psnr_db": 21.183387303768146,
+      "ssim": 0.7375190385472777,
+      "temporal_delta_mse": 0.009169250550476135,
+      "vbench_subject_consistency": 0.9475698670841033,
+      "vbench_background_consistency": 0.9426687129851311,
+      "vbench_temporal_flickering": 0.9599090755009183,
+      "vbench_motion_smoothness": 0.9895824586437182,
+      "vbench_dynamic_degree": 0.8,
+      "vbench_aesthetic_quality": 0.5974819004535675,
+      "vbench_imaging_quality": 0.729475989491272,
+      "vbench_overall_consistency": 0.2490800651907921,
+      "vbench_core5_mean": 0.8413557857315584
+    },
+    {
+      "method": "C1-r50",
+      "threshold_log_c1": -4.79892683,
+      "fallback_rate": 0.5171428571428571,
+      "p1_fallback_rate": 0.13142857142857142,
+      "p2_fallback_rate": 0.42,
+      "p3_fallback_rate": 1.0,
+      "mean_end_to_end_s": 99.25920022401726,
+      "frame_psnr_mean_db": 24.12363936617357,
+      "lpips_alex": 0.1050749322929606,
+      "pixel_mae": 0.040959788803271946,
+      "pixel_mse": 0.008328448020131038,
+      "pixel_rmse": 0.08546495607525889,
+      "psnr_db": 22.11610699774378,
+      "ssim": 0.7683776850652694,
+      "temporal_delta_mse": 0.008148231709055835,
+      "vbench_subject_consistency": 0.9477616070198916,
+      "vbench_background_consistency": 0.9430070643271169,
+      "vbench_temporal_flickering": 0.9589383629805901,
+      "vbench_motion_smoothness": 0.9895042519327572,
+      "vbench_dynamic_degree": 0.8,
+      "vbench_aesthetic_quality": 0.5991387969255447,
+      "vbench_imaging_quality": 0.7326260545639037,
+      "vbench_overall_consistency": 0.24868253156542777,
+      "vbench_core5_mean": 0.8424075549538428
+    }
+  ],
+  "comparisons": {
+    "C1-r33_vs_FPPF": {
+      "psnr_db": -0.4099588926624804,
+      "ssim": -0.009165007499456257,
+      "lpips_alex": 0.00520488655427008,
+      "pixel_mse": 0.00048042507441620544,
+      "temporal_delta_mse": 0.00034564203218906235,
+      "vbench_subject_consistency": 0.00010932957453113623,
+      "vbench_background_consistency": -0.0011537219632056228,
+      "vbench_temporal_flickering": -0.0005416798100752462,
+      "vbench_motion_smoothness": -0.00010019194109278651,
+      "vbench_dynamic_degree": 0.0,
+      "vbench_aesthetic_quality": -0.000733357965946202,
+      "vbench_imaging_quality": -0.0014540961441039668,
+      "vbench_overall_consistency": -9.74409282207267e-05,
+      "vbench_core5_mean": -0.000666407687963444
+    },
+    "C1-r50_vs_FPPF": {
+      "psnr_db": 0.5227608013131544,
+      "ssim": 0.021693639018535382,
+      "lpips_alex": -0.015920953146540545,
+      "pixel_mse": -0.0013631079243988584,
+      "temporal_delta_mse": -0.0006753768092312377,
+      "vbench_subject_consistency": 0.000301069510319496,
+      "vbench_background_consistency": -0.0008153706212198353,
+      "vbench_temporal_flickering": -0.0015123923304034559,
+      "vbench_motion_smoothness": -0.00017839865205382388,
+      "vbench_dynamic_degree": 0.0,
+      "vbench_aesthetic_quality": 0.0009235385060309964,
+      "vbench_imaging_quality": 0.0016959689285277468,
+      "vbench_overall_consistency": -0.000494974553585048,
+      "vbench_core5_mean": 0.0003853615343208938
+    }
+  }
+}

confidence_validation/checkpoints/confidence_c0.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:45097451bb0c7600e17ba42b12d66091378a34012ee5b62e2dd16dc82bbb73ba
+size 2152902

confidence_validation/checkpoints/confidence_c1.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9db356a091c6d1ab6767ccd27e82ecf5a797a7cd5cee9055d8e08ec770a92dc3
+size 2152902

confidence_validation/checkpoints/confidence_c2.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bc814e92e5e671ea1e2fd41acf9c0d32f011966b6795c955a188a3781e60f64d
+size 2152902

confidence_validation/features/train/features_rank_00.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4f4eb5ee6d955f62b1a67da5ea05280a7495cdfa5be6d33cfc26a556fa1ef655
+size 8670311

confidence_validation/features/train/features_rank_01.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:34fc3f348d2a510a9995fd9313c961f2992cad96cdf1eac68f83e8ee6b55a7c0
+size 8670311

confidence_validation/features/train/features_rank_02.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:06ae94db76ef0cd565c767d1ff4751b98a613720e6422224d52584091285e051
+size 8670311

confidence_validation/features/train/features_rank_03.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:57d382233c680657bff91fe1ea33a3623665c0daaf21cb5c917597c6f2207441
+size 8670311

confidence_validation/features/train/features_rank_04.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5e207abe3bf0ed3bd210b548bacba9cc5f766730584ee3af16f72b2024ccbd3a
+size 8670311

confidence_validation/features/train/features_rank_05.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c9a592ec52cd23d68d508f0a1e7cd92dbd1bbb6f5b6cc9113d8c3e3b362bd92a
+size 8670311

confidence_validation/features/train/features_rank_06.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a47996c3e99d6ec2d19fa03356fac4a4ce51e9fff3e3d1472ae19518fa537fec
+size 8670311

confidence_validation/features/train/features_rank_07.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e601719384dd980f33dcf6fdb5d1b5be9e95307d38bca1549a6f36a65f0704d2
+size 8670311

confidence_validation/features/validation/features_rank_00.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:78e749aeb56ba195882ed7a8f15e4bfdc14e9e89fb467c3c517fb6b01fc351ca
+size 2256359

confidence_validation/features/validation/features_rank_01.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:925cc3fc119cfd513b695188c90440a3f213127bf18704720685ce33e29b9a9c
+size 2256359

confidence_validation/features/validation/features_rank_02.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5d3b0fad65997a4f2d0a652998399491503fab700f8684119b6d2526f1b2600a
+size 2256359

confidence_validation/features/validation/features_rank_03.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a42a824c1b56d438586b9964edd1f0d0eb8fe96216d90dccc9c63259298bb86d
+size 2256359

confidence_validation/features/validation/features_rank_04.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:425d4530620611b477067991a786bc728dac054802d5903ddcf05b46e36fa704
+size 2082791

confidence_validation/features/validation/features_rank_05.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0b14fb88ab25ef35d8e088b2c9bae36449ed1c7946230f20a988b8dc2218b4a2
+size 2082791

confidence_validation/features/validation/features_rank_06.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2c61034ddd0dc73baf40c5cc5f35d79ba56a7bcdc7fca1bbc36cd2024922d6f5
+size 2082791

confidence_validation/features/validation/features_rank_07.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:188a699a5c5599b8043ed3e99d280e8bbaf17d06a949cf487b86d8cae6fd7b77
+size 2082791

confidence_validation/runner.log

@@ -0,0 +1,114 @@
+{
+  "output": "/mnt/local_nvme/cz/projects/HY-WorldPlay-DEV-Predictor/datasets/hyworldplay_predictor_confidence_validation25_prefeature/cases.jsonl",
+  "cases": 25
+}
+{
+  "records": 800,
+  "train_records": 700
+}
+WARNING:__main__:
+*****************************************
+Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed. 
+*****************************************
+rank0 extracted 25/133 batches
+rank0 extracted 50/133 batches
+rank0 extracted 75/133 batches
+rank0 extracted 100/133 batches
+rank0 extracted 125/133 batches
+Saved 8 feature shards to /mnt/local_nvme/cz/projects/HY-WorldPlay-DEV-Predictor/confidence_validation/features/train
+WARNING:__main__:
+*****************************************
+Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed. 
+*****************************************
+rank0 extracted 25/35 batches
+Saved 8 feature shards to /mnt/local_nvme/cz/projects/HY-WorldPlay-DEV-Predictor/confidence_validation/features/validation
+WARNING:__main__:
+*****************************************
+Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed. 
+*****************************************
+[rank0]:[W819 05:29:45.641755152 ProcessGroupNCCL.cpp:4115] [PG ID 0 PG GUID 0 Rank 0]  using GPU 0 to perform barrier as devices used by this process are currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect.Specify device_ids in barrier() to force use of a particular device,or call init_process_group() with a device_id.
+[rank6]:[W819 05:29:45.820996456 ProcessGroupNCCL.cpp:4115] [PG ID 0 PG GUID 0 Rank 6]  using GPU 6 to perform barrier as devices used by this process are currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect.Specify device_ids in barrier() to force use of a particular device,or call init_process_group() with a device_id.
+[rank1]:[W819 05:29:45.868540764 ProcessGroupNCCL.cpp:4115] [PG ID 0 PG GUID 0 Rank 1]  using GPU 1 to perform barrier as devices used by this process are currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect.Specify device_ids in barrier() to force use of a particular device,or call init_process_group() with a device_id.
+[rank2]:[W819 05:29:46.322682030 ProcessGroupNCCL.cpp:4115] [PG ID 0 PG GUID 0 Rank 2]  using GPU 2 to perform barrier as devices used by this process are currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect.Specify device_ids in barrier() to force use of a particular device,or call init_process_group() with a device_id.
+[rank7]:[W819 05:29:46.386567270 ProcessGroupNCCL.cpp:4115] [PG ID 0 PG GUID 0 Rank 7]  using GPU 7 to perform barrier as devices used by this process are currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect.Specify device_ids in barrier() to force use of a particular device,or call init_process_group() with a device_id.
+[rank5]:[W819 05:29:46.398470481 ProcessGroupNCCL.cpp:4115] [PG ID 0 PG GUID 0 Rank 5]  using GPU 5 to perform barrier as devices used by this process are currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect.Specify device_ids in barrier() to force use of a particular device,or call init_process_group() with a device_id.
+[rank4]:[W819 05:29:46.426298692 ProcessGroupNCCL.cpp:4115] [PG ID 0 PG GUID 0 Rank 4]  using GPU 4 to perform barrier as devices used by this process are currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect.Specify device_ids in barrier() to force use of a particular device,or call init_process_group() with a device_id.
+[rank3]:[W819 05:29:46.471978908 ProcessGroupNCCL.cpp:4115] [PG ID 0 PG GUID 0 Rank 3]  using GPU 3 to perform barrier as devices used by this process are currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect.Specify device_ids in barrier() to force use of a particular device,or call init_process_group() with a device_id.
+c0 epoch 1/40 loss=1.951573
+c0 epoch 10/40 loss=0.075422
+c0 epoch 20/40 loss=0.035323
+c0 epoch 30/40 loss=0.025954
+c0 epoch 40/40 loss=0.020488
+c1 epoch 1/40 loss=3.383299
+c1 epoch 10/40 loss=0.149253
+c1 epoch 20/40 loss=0.046591
+c1 epoch 30/40 loss=0.029699
+c1 epoch 40/40 loss=0.022912
+c2 epoch 1/40 loss=2.855274
+c2 epoch 10/40 loss=0.109034
+c2 epoch 20/40 loss=0.040014
+c2 epoch 30/40 loss=0.029119
+c2 epoch 40/40 loss=0.022536
+{
+  "protocol": "WorldPlay fulltrain100 (8400 transitions); independent validation25 (2100 transitions); no validation tuning",
+  "train_samples": 8400,
+  "validation_samples": 2100,
+  "c1_validation_metrics": {
+    "overall": {
+      "model": "Confidence C1",
+      "split": "validation",
+      "step": "overall",
+      "spearman": 0.9529179491248462,
+      "pearson": 0.96539306640625,
+      "log_mae": 0.15999889373779297
+    },
+    "P1": {
+      "model": "Confidence C1",
+      "split": "validation",
+      "step": "P1",
+      "spearman": 0.8590718873187204,
+      "pearson": 0.8829622268676758,
+      "log_mae": 0.17691564559936523
+    },
+    "P2": {
+      "model": "Confidence C1",
+      "split": "validation",
+      "step": "P2",
+      "spearman": 0.8631690385373965,
+      "pearson": 0.9074602127075195,
+      "log_mae": 0.14503394067287445
+    },
+    "P3": {
+      "model": "Confidence C1",
+      "split": "validation",
+      "step": "P3",
+      "spearman": 0.8700319930976419,
+      "pearson": 0.8872363567352295,
+      "log_mae": 0.15804709494113922
+    }
+  },
+  "remaining_risk_at_33pct": {
+    "Random": 0.009558672141283751,
+    "Timestep-only": 0.005005753133445978,
+    "Confidence C0": 0.008430036716163158,
+    "Confidence C1": 0.004878982435911894,
+    "Confidence C2": 0.004925357177853584,
+    "Oracle": 0.004840569104999304
+  },
+  "gain_vs_random": 0.48957529207015615,
+  "oracle_gap": 0.007935705508866102,
+  "routing_at_33pct": {
+    "P1": {
+      "selection_fraction": 0.012857142857142857,
+      "within_step_fallback_rate": 0.012857142857142857
+    },
+    "P2": {
+      "selection_fraction": 0.05714285714285714,
+      "within_step_fallback_rate": 0.05714285714285714
+    },
+    "P3": {
+      "selection_fraction": 0.93,
+      "within_step_fallback_rate": 0.93
+    }
+  }
+}

confidence_validation/summary.json

@@ -0,0 +1,63 @@
+{
+  "c1_validation_metrics": {
+    "P1": {
+      "log_mae": 0.17691564559936523,
+      "model": "Confidence C1",
+      "pearson": 0.8829622268676758,
+      "spearman": 0.8590718873187204,
+      "split": "validation",
+      "step": "P1"
+    },
+    "P2": {
+      "log_mae": 0.14503394067287445,
+      "model": "Confidence C1",
+      "pearson": 0.9074602127075195,
+      "spearman": 0.8631690385373965,
+      "split": "validation",
+      "step": "P2"
+    },
+    "P3": {
+      "log_mae": 0.15804709494113922,
+      "model": "Confidence C1",
+      "pearson": 0.8872363567352295,
+      "spearman": 0.8700319930976419,
+      "split": "validation",
+      "step": "P3"
+    },
+    "overall": {
+      "log_mae": 0.15999889373779297,
+      "model": "Confidence C1",
+      "pearson": 0.96539306640625,
+      "spearman": 0.9529179491248462,
+      "split": "validation",
+      "step": "overall"
+    }
+  },
+  "gain_vs_random": 0.48957529207015615,
+  "oracle_gap": 0.007935705508866102,
+  "protocol": "WorldPlay fulltrain100 (8400 transitions); independent validation25 (2100 transitions); no validation tuning",
+  "remaining_risk_at_33pct": {
+    "Confidence C0": 0.008430036716163158,
+    "Confidence C1": 0.004878982435911894,
+    "Confidence C2": 0.004925357177853584,
+    "Oracle": 0.004840569104999304,
+    "Random": 0.009558672141283751,
+    "Timestep-only": 0.005005753133445978
+  },
+  "routing_at_33pct": {
+    "P1": {
+      "selection_fraction": 0.012857142857142857,
+      "within_step_fallback_rate": 0.012857142857142857
+    },
+    "P2": {
+      "selection_fraction": 0.05714285714285714,
+      "within_step_fallback_rate": 0.05714285714285714
+    },
+    "P3": {
+      "selection_fraction": 0.93,
+      "within_step_fallback_rate": 0.93
+    }
+  },
+  "train_samples": 8400,
+  "validation_samples": 2100
+}

confidence_validation/summary.md

@@ -0,0 +1,19 @@
+# HY-WorldPlay offline confidence validation
+
+WorldPlay fulltrain100 (8400 transitions); independent validation25 (2100 transitions); no validation tuning
+
+## C1 validation Spearman
+
+- overall: 0.9529
+- P1: 0.8591
+- P2: 0.8632
+- P3: 0.8700
+
+## 33.3% fallback remaining C1 risk
+
+- Random: 0.0095586721
+- Timestep-only: 0.0050057531
+- Confidence C0: 0.0084300367
+- Confidence C1: 0.0048789824
+- Confidence C2: 0.0049253572
+- Oracle: 0.0048405691

confidence_validation/tables/confidence_metrics.csv

@@ -0,0 +1,25 @@
+model,split,step,spearman,pearson,log_mae
+Confidence C0,train,overall,0.910420624592802,0.914232611656189,0.1561840921640396
+Confidence C0,train,P1,0.8801052591242132,0.8944251537322998,0.17658944427967072
+Confidence C0,train,P2,0.9063240986564994,0.9173239469528198,0.14159120619297028
+Confidence C0,train,P3,0.8942312332292026,0.9015886783599854,0.15037165582180023
+Confidence C1,train,overall,0.9599546928176265,0.9663121700286865,0.1623731553554535
+Confidence C1,train,P1,0.8680394642178618,0.8818591833114624,0.18546940386295319
+Confidence C1,train,P2,0.8950266639376704,0.9078938364982605,0.14816781878471375
+Confidence C1,train,P3,0.8855762887929935,0.8950272798538208,0.15348218381404877
+Confidence C2,train,overall,0.9524278955144021,0.9526013135910034,0.1651911735534668
+Confidence C2,train,P1,0.8683219754517324,0.8778232336044312,0.17134705185890198
+Confidence C2,train,P2,0.8974592285143828,0.9101301431655884,0.13969838619232178
+Confidence C2,train,P3,0.8512258048715202,0.8678408265113831,0.18452806770801544
+Confidence C0,validation,overall,0.9059050330545604,0.9138662815093994,0.15434052050113678
+Confidence C0,validation,P1,0.8688861580767074,0.8960813283920288,0.16730837523937225
+Confidence C0,validation,P2,0.8769245301666797,0.9138691425323486,0.1416863650083542
+Confidence C0,validation,P3,0.8817214962246278,0.8930692672729492,0.1540267914533615
+Confidence C1,validation,overall,0.9529179491248462,0.96539306640625,0.15999889373779297
+Confidence C1,validation,P1,0.8590718873187204,0.8829622268676758,0.17691564559936523
+Confidence C1,validation,P2,0.8631690385373965,0.9074602127075195,0.14503394067287445
+Confidence C1,validation,P3,0.8700319930976419,0.8872363567352295,0.15804709494113922
+Confidence C2,validation,overall,0.9472269125555032,0.9510321021080017,0.1618642508983612
+Confidence C2,validation,P1,0.8622886635628703,0.8770185708999634,0.1692851185798645
+Confidence C2,validation,P2,0.869344782919382,0.9097270965576172,0.13763289153575897
+Confidence C2,validation,P3,0.8291001643443586,0.8579860329627991,0.17867469787597656

confidence_validation/tables/risk_coverage.csv

@@ -0,0 +1,67 @@
+method,fallback_ratio,remaining_risk,std_if_random
+Random,0.0,0.014347239397466183,0.0
+Timestep-only,0.0,0.014347239397466183,
+Confidence C0,0.0,0.014347239397466183,
+Confidence C1,0.0,0.014347239397466183,
+Confidence C2,0.0,0.014347239397466183,
+Oracle,0.0,0.014347239397466183,
+Random,0.05,0.013626513117924333,5.5243322704253306e-05
+Timestep-only,0.05,0.013086354359984398,
+Confidence C0,0.05,0.013266514986753464,
+Confidence C1,0.05,0.011860155500471592,
+Confidence C2,0.05,0.011852260679006577,
+Oracle,0.05,0.011713187210261822,
+Random,0.1,0.012901437673717738,7.81304187993621e-05
+Timestep-only,0.1,0.01186059694737196,
+Confidence C0,0.1,0.012261713854968548,
+Confidence C1,0.1,0.01016467995941639,
+Confidence C2,0.1,0.010183872655034065,
+Oracle,0.1,0.009989776648581028,
+Random,0.2,0.011489660385996104,0.00011283538260078906
+Timestep-only,0.2,0.009180759079754353,
+Confidence C0,0.2,0.010462735779583454,
+Confidence C1,0.2,0.007498461287468672,
+Confidence C2,0.2,0.007524735759943724,
+Oracle,0.2,0.0073388563469052315,
+Random,0.25,0.010771198542788624,0.00014086472540232944
+Timestep-only,0.25,0.008169419132173061,
+Confidence C0,0.25,0.009707297198474407,
+Confidence C1,0.25,0.0064151231199502945,
+Confidence C2,0.25,0.006412565242499113,
+Oracle,0.25,0.0062853023409843445,
+Random,0.3333333333333333,0.009558672141283751,0.00012621153655552298
+Timestep-only,0.3333333333333333,0.005005753133445978,
+Confidence C0,0.3333333333333333,0.008430036716163158,
+Confidence C1,0.3333333333333333,0.004878982435911894,
+Confidence C2,0.3333333333333333,0.004925357177853584,
+Oracle,0.3333333333333333,0.004840569104999304,
+Random,0.4,0.008614255879074335,0.000123129467857736
+Timestep-only,0.4,0.004496104083955288,
+Confidence C0,0.4,0.007375260349363089,
+Confidence C1,0.4,0.004071088042110205,
+Confidence C2,0.4,0.00408595334738493,
+Oracle,0.4,0.00398297468200326,
+Random,0.5,0.0071629173122346405,0.0001261506684215066
+Timestep-only,0.5,0.003779263701289892,
+Confidence C0,0.5,0.005708133336156607,
+Confidence C1,0.5,0.003076771041378379,
+Confidence C2,0.5,0.003123910864815116,
+Oracle,0.5,0.002952347509562969,
+Random,0.6,0.0057337418384850024,0.00013838311339076486
+Timestep-only,0.6,0.0027656827587634325,
+Confidence C0,0.6,0.004291518125683069,
+Confidence C1,0.6,0.0022384771145880222,
+Confidence C2,0.6,0.0022648375015705824,
+Oracle,0.6,0.002107207430526614,
+Random,0.75,0.003592234782408923,0.0001083880957844813
+Timestep-only,0.75,0.0016061057103797793,
+Confidence C0,0.75,0.0021801821421831846,
+Confidence C1,0.75,0.0011718758614733815,
+Confidence C2,0.75,0.0011971330968663096,
+Oracle,0.75,0.0010821536416187882,
+Random,1.0,0.0,0.0
+Timestep-only,1.0,0.0,
+Confidence C0,1.0,0.0,
+Confidence C1,1.0,0.0,
+Confidence C2,1.0,0.0,
+Oracle,1.0,0.0,

confidence_validation_wide4096x2_c1/checkpoints/confidence_c1.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:71cee71f42764bfccde6631802ac53d6b31d8b5cdcd03d046bcc5482de83886d
+size 101273674

confidence_validation_wide4096x2_c1/runner.log

@@ -0,0 +1,85 @@
+WARNING:__main__:
+*****************************************
+Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed. 
+*****************************************
+[rank6]:[W820 12:16:16.962510208 ProcessGroupNCCL.cpp:4115] [PG ID 0 PG GUID 0 Rank 6]  using GPU 6 to perform barrier as devices used by this process are currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect.Specify device_ids in barrier() to force use of a particular device,or call init_process_group() with a device_id.
+[rank7]:[W820 12:16:16.989267769 ProcessGroupNCCL.cpp:4115] [PG ID 0 PG GUID 0 Rank 7]  using GPU 7 to perform barrier as devices used by this process are currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect.Specify device_ids in barrier() to force use of a particular device,or call init_process_group() with a device_id.
+[rank3]:[W820 12:16:16.993386558 ProcessGroupNCCL.cpp:4115] [PG ID 0 PG GUID 0 Rank 3]  using GPU 3 to perform barrier as devices used by this process are currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect.Specify device_ids in barrier() to force use of a particular device,or call init_process_group() with a device_id.
+[rank5]:[W820 12:16:17.073358047 ProcessGroupNCCL.cpp:4115] [PG ID 0 PG GUID 0 Rank 5]  using GPU 5 to perform barrier as devices used by this process are currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect.Specify device_ids in barrier() to force use of a particular device,or call init_process_group() with a device_id.
+[rank0]:[W820 12:16:17.201143716 ProcessGroupNCCL.cpp:4115] [PG ID 0 PG GUID 0 Rank 0]  using GPU 0 to perform barrier as devices used by this process are currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect.Specify device_ids in barrier() to force use of a particular device,or call init_process_group() with a device_id.
+[rank1]:[W820 12:16:17.363060718 ProcessGroupNCCL.cpp:4115] [PG ID 0 PG GUID 0 Rank 1]  using GPU 1 to perform barrier as devices used by this process are currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect.Specify device_ids in barrier() to force use of a particular device,or call init_process_group() with a device_id.
+[rank2]:[W820 12:16:17.411170725 ProcessGroupNCCL.cpp:4115] [PG ID 0 PG GUID 0 Rank 2]  using GPU 2 to perform barrier as devices used by this process are currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect.Specify device_ids in barrier() to force use of a particular device,or call init_process_group() with a device_id.
+[rank4]:[W820 12:16:17.432620994 ProcessGroupNCCL.cpp:4115] [PG ID 0 PG GUID 0 Rank 4]  using GPU 4 to perform barrier as devices used by this process are currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect.Specify device_ids in barrier() to force use of a particular device,or call init_process_group() with a device_id.
+c1 epoch 1/40 loss=1.998238
+c1 epoch 10/40 loss=0.041588
+c1 epoch 20/40 loss=0.014896
+c1 epoch 30/40 loss=0.012734
+c1 epoch 40/40 loss=0.011734
+{
+  "protocol": "WorldPlay fulltrain100 (8400 transitions); independent validation25 (2100 transitions); no validation tuning",
+  "targets": [
+    "c1"
+  ],
+  "mlp_dims": [
+    4096,
+    4096
+  ],
+  "train_samples": 8400,
+  "validation_samples": 2100,
+  "c1_validation_metrics": {
+    "overall": {
+      "model": "Confidence C1",
+      "split": "validation",
+      "step": "overall",
+      "spearman": 0.9707018864565339,
+      "pearson": 0.9785629510879517,
+      "log_mae": 0.13104061782360077
+    },
+    "P1": {
+      "model": "Confidence C1",
+      "split": "validation",
+      "step": "P1",
+      "spearman": 0.9057680759115251,
+      "pearson": 0.9183362126350403,
+      "log_mae": 0.1512252390384674
+    },
+    "P2": {
+      "model": "Confidence C1",
+      "split": "validation",
+      "step": "P2",
+      "spearman": 0.9236297800024663,
+      "pearson": 0.9456437230110168,
+      "log_mae": 0.12534597516059875
+    },
+    "P3": {
+      "model": "Confidence C1",
+      "split": "validation",
+      "step": "P3",
+      "spearman": 0.947315764201268,
+      "pearson": 0.947665274143219,
+      "log_mae": 0.11655063927173615
+    }
+  },
+  "remaining_risk_at_33pct": {
+    "Random": 0.009558672141283751,
+    "Timestep-only": 0.005005753133445978,
+    "Confidence C1": 0.004887485411018133,
+    "Oracle": 0.004840569104999304
+  },
+  "gain_vs_random": 0.48868573597067294,
+  "oracle_gap": 0.009692311999094184,
+  "routing_at_33pct": {
+    "P1": {
+      "selection_fraction": 0.02857142857142857,
+      "within_step_fallback_rate": 0.02857142857142857
+    },
+    "P2": {
+      "selection_fraction": 0.06857142857142857,
+      "within_step_fallback_rate": 0.06857142857142857
+    },
+    "P3": {
+      "selection_fraction": 0.9028571428571428,
+      "within_step_fallback_rate": 0.9028571428571428
+    }
+  }
+}

confidence_validation_wide4096x2_c1/summary.json

@@ -0,0 +1,68 @@
+{
+  "c1_validation_metrics": {
+    "P1": {
+      "log_mae": 0.1512252390384674,
+      "model": "Confidence C1",
+      "pearson": 0.9183362126350403,
+      "spearman": 0.9057680759115251,
+      "split": "validation",
+      "step": "P1"
+    },
+    "P2": {
+      "log_mae": 0.12534597516059875,
+      "model": "Confidence C1",
+      "pearson": 0.9456437230110168,
+      "spearman": 0.9236297800024663,
+      "split": "validation",
+      "step": "P2"
+    },
+    "P3": {
+      "log_mae": 0.11655063927173615,
+      "model": "Confidence C1",
+      "pearson": 0.947665274143219,
+      "spearman": 0.947315764201268,
+      "split": "validation",
+      "step": "P3"
+    },
+    "overall": {
+      "log_mae": 0.13104061782360077,
+      "model": "Confidence C1",
+      "pearson": 0.9785629510879517,
+      "spearman": 0.9707018864565339,
+      "split": "validation",
+      "step": "overall"
+    }
+  },
+  "gain_vs_random": 0.48868573597067294,
+  "mlp_dims": [
+    4096,
+    4096
+  ],
+  "oracle_gap": 0.009692311999094184,
+  "protocol": "WorldPlay fulltrain100 (8400 transitions); independent validation25 (2100 transitions); no validation tuning",
+  "remaining_risk_at_33pct": {
+    "Confidence C1": 0.004887485411018133,
+    "Oracle": 0.004840569104999304,
+    "Random": 0.009558672141283751,
+    "Timestep-only": 0.005005753133445978
+  },
+  "routing_at_33pct": {
+    "P1": {
+      "selection_fraction": 0.02857142857142857,
+      "within_step_fallback_rate": 0.02857142857142857
+    },
+    "P2": {
+      "selection_fraction": 0.06857142857142857,
+      "within_step_fallback_rate": 0.06857142857142857
+    },
+    "P3": {
+      "selection_fraction": 0.9028571428571428,
+      "within_step_fallback_rate": 0.9028571428571428
+    }
+  },
+  "targets": [
+    "c1"
+  ],
+  "train_samples": 8400,
+  "validation_samples": 2100
+}

confidence_validation_wide4096x2_c1/summary.md

@@ -0,0 +1,17 @@
+# HY-WorldPlay offline confidence validation
+
+WorldPlay fulltrain100 (8400 transitions); independent validation25 (2100 transitions); no validation tuning
+
+## C1 validation Spearman
+
+- overall: 0.9707
+- P1: 0.9058
+- P2: 0.9236
+- P3: 0.9473
+
+## 33.3% fallback remaining C1 risk
+
+- Random: 0.0095586721
+- Timestep-only: 0.0050057531
+- Confidence C1: 0.0048874854
+- Oracle: 0.0048405691

confidence_validation_wide4096x2_c1/tables/confidence_metrics.csv

@@ -0,0 +1,9 @@
+model,split,step,spearman,pearson,log_mae
+Confidence C1,train,overall,0.9786988309040863,0.9830387234687805,0.11529678106307983
+Confidence C1,train,P1,0.9239389073837476,0.9279435873031616,0.1396360695362091
+Confidence C1,train,P2,0.948503091905051,0.9554183483123779,0.10605843365192413
+Confidence C1,train,P3,0.9556997813787577,0.9599348306655884,0.10019586980342865
+Confidence C1,validation,overall,0.9707018864565339,0.9785629510879517,0.13104061782360077
+Confidence C1,validation,P1,0.9057680759115251,0.9183362126350403,0.1512252390384674
+Confidence C1,validation,P2,0.9236297800024663,0.9456437230110168,0.12534597516059875
+Confidence C1,validation,P3,0.947315764201268,0.947665274143219,0.11655063927173615

confidence_validation_wide4096x2_c1/tables/risk_coverage.csv

@@ -0,0 +1,45 @@
+method,fallback_ratio,remaining_risk,std_if_random
+Random,0.0,0.014347239397466183,0.0
+Timestep-only,0.0,0.014347239397466183,
+Confidence C1,0.0,0.014347239397466183,
+Oracle,0.0,0.014347239397466183,
+Random,0.05,0.013626513117924333,5.5243322704253306e-05
+Timestep-only,0.05,0.013086354359984398,
+Confidence C1,0.05,0.011780897155404091,
+Oracle,0.05,0.011713187210261822,
+Random,0.1,0.012901437673717738,7.81304187993621e-05
+Timestep-only,0.1,0.01186059694737196,
+Confidence C1,0.1,0.010083409957587719,
+Oracle,0.1,0.009989776648581028,
+Random,0.2,0.011489660385996104,0.00011283538260078906
+Timestep-only,0.2,0.009180759079754353,
+Confidence C1,0.2,0.007394451182335615,
+Oracle,0.2,0.0073388563469052315,
+Random,0.25,0.010771198542788624,0.00014086472540232944
+Timestep-only,0.25,0.008169419132173061,
+Confidence C1,0.25,0.006334020756185055,
+Oracle,0.25,0.0062853023409843445,
+Random,0.3333333333333333,0.009558672141283751,0.00012621153655552298
+Timestep-only,0.3333333333333333,0.005005753133445978,
+Confidence C1,0.3333333333333333,0.004887485411018133,
+Oracle,0.3333333333333333,0.004840569104999304,
+Random,0.4,0.008614255879074335,0.000123129467857736
+Timestep-only,0.4,0.004496104083955288,
+Confidence C1,0.4,0.004047052469104528,
+Oracle,0.4,0.00398297468200326,
+Random,0.5,0.0071629173122346405,0.0001261506684215066
+Timestep-only,0.5,0.003779263701289892,
+Confidence C1,0.5,0.003037535585463047,
+Oracle,0.5,0.002952347509562969,
+Random,0.6,0.0057337418384850024,0.00013838311339076486
+Timestep-only,0.6,0.0027656827587634325,
+Confidence C1,0.6,0.0021919310092926025,
+Oracle,0.6,0.002107207430526614,
+Random,0.75,0.003592234782408923,0.0001083880957844813
+Timestep-only,0.75,0.0016061057103797793,
+Confidence C1,0.75,0.0011335358722135425,
+Oracle,0.75,0.0010821536416187882,
+Random,1.0,0.0,0.0
+Timestep-only,1.0,0.0,
+Confidence C1,1.0,0.0,
+Oracle,1.0,0.0,