Consolidate task-transfer assets 1-14

archive/task-transfer/openvla_bridge_cross_defendtheline_zero_deadly_zero_exp3/checkpoints/steps_2000_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:38e0a39116fe43c53068d76929d93fb391c669002fd43c359174ab267f609f3c
+size 9784896838

archive/task-transfer/openvla_bridge_cross_defendtheline_zero_deadly_zero_exp3/config.full.yaml

@@ -0,0 +1,368 @@
+framework:
+  name: QwenOFT
+  qwenvl:
+    base_vlm: /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/talha-aug26/playground/Pretrained_models/Qwen3-VL-4B-Instruct
+    attn_implementation: flash_attention_2
+    enable_gradient_checkpointing: true
+  action_model:
+    action_model_type: MLP
+    action_dim: 7
+    action_hidden_dim: 2560
+    future_action_window_size: 0
+    past_action_window_size: 0
+    loss_type: discrete_ce
+    state_dim: 7
+    action_horizon: 1
+    action_env_dim: 7
+  kv_memory:
+    enabled: false
+    window: 4
+    rollout_len: 8
+    packed_train: false
+    rebased_sink: true
+datasets:
+  vla_data:
+    dataset_py: lerobot_datasets
+    include_state: false
+    data_root_dir: /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/talha-aug26/playground/Datasets/rl_games
+    data_mix: cross__deadly_corridor_0_cross_train_openvla_bridge_cross_defendtheline_zero_deadly_zero_exp3__lat0__1000ep_per_lat__bridge__defend_the_line_0_cross_train_openvla_bridge_cross_defendtheline_zero_deadly_zero_exp3__lat0__1000ep_per_lat__bridge
+    eval_data_mix: cross__deadly_corridor_0_cross_train_openvla_bridge_cross_defendtheline_zero_deadly_zero_exp3__lat0__1000ep_per_lat__bridge__defend_the_line_0_cross_train_openvla_bridge_cross_defendtheline_zero_deadly_zero_exp3__lat0__1000ep_per_lat__bridge__val
+    custom_mixtures_path: /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/talha-aug26/playground/Datasets/rl_games/_generated_mixtures/cross__deadly_corridor_0_cross_train_openvla_bridge_cross_defendtheline_zero_deadly_zero_exp3__lat0__1000ep_per_lat__bridge__defend_the_line_0_cross_train_openvla_bridge_cross_defendtheline_zero_deadly_zero_exp3__lat0__1000ep_per_lat__bridge.json
+    action_type: discrete
+    sequential_step_sampling: true
+    eval_sequential_step_sampling: null
+    num_workers: 8
+    eval_num_workers: 8
+    prefetch_factor: 4
+    persistent_workers: true
+    pin_memory: true
+    shuffle: true
+    action_balance:
+      enabled: false
+      strategy: balanced_epoch
+      action_key: action_id
+      target_flap_fraction: 0.3
+      noop_id: 0
+      flap_id: 1
+    latency_curriculum:
+      enabled: false
+      strategy: exclusive
+      latencies: null
+      phase_steps: null
+      phase_distributions: null
+      new_latency_passes: 1.0
+      replay_passes: 0.25
+      target_total_passes: 2.0
+      final_equalization: true
+      step_budget_mode: auto
+      eval_at_phase_end: false
+      save_at_phase_end: false
+      computed_plan: null
+    per_device_batch_size: 4
+    load_all_data_for_training: true
+    num_obs_frames: 1
+    image_mode: single
+    prompt_mode: raw
+    stitch_grid:
+    - 2
+    - 2
+    obs_image_size: null
+    video_backend: torchvision_av
+dataset:
+  source_hf: ''
+  config_name: null
+  source_subdir: null
+  converted_name: cross_task_train
+  single_source_hf: ''
+  mixed_source_hf: ''
+  single_converted_name: flappy_train
+  mixed_converted_name: flappy_mixed_latency_train
+  single_latency_filter: null
+  mixed_latency_filter: null
+  force_download: false
+  setup_force: false
+  skip_verification: false
+  target_latency_unit: observation_steps
+  verify_rows: 200
+  max_episodes: null
+  episodes_per_latency: null
+  latency_filter: null
+  debug_subset:
+    enabled: false
+    max_episodes: 5
+    suffix: debug
+base_model:
+  repo_id: Qwen/Qwen3-VL-4B-Instruct
+initialization:
+  checkpoint_local_dir: playground/Pretrained_models/Qwen3VL-OFT-Bridge-RT-1
+  checkpoint_hf_repo_id: StarVLA/Qwen3VL-OFT-Bridge-RT-1
+  checkpoint_filename: checkpoints/steps_5000_pytorch_model.pt
+trainer:
+  max_train_steps: 2000
+  num_warmup_steps: 0
+  save_interval: 500
+  eval_interval: 400
+  eval_num_batches: 100
+  per_latency_eval_num_batches: 5
+  eval_action_classification: true
+  eval_action_classification_interval: null
+  cc_f1_tolerance: 1
+  learning_rate:
+    base: 2.0e-05
+    qwen_vl_interface: 1.0e-05
+    action_model: 0.0001
+  lr_scheduler_type: cosine_with_min_lr
+  scheduler_specific_kwargs:
+    min_lr: 1.0e-06
+  freeze_modules: ''
+  freeze_vit: false
+  freeze_tied_embedding: false
+  freeze_llm_layers: []
+  loss_scale:
+    vla: 1.0
+    vlm: 0.1
+  max_grad_norm: 1.0
+  weight_decay: 0.0
+  logging_frequency: 1
+  profile_timing:
+    enabled: false
+    log_interval: 10
+  gradient_clipping: 1.0
+  gradient_accumulation_steps: 4
+  distributed_backend: deepspeed
+  is_resume: false
+  pretrained_checkpoint: null
+  resume_step: 0
+  reload_modules: null
+  optimizer:
+    name: AdamW
+    betas:
+    - 0.9
+    - 0.95
+    eps: 1.0e-08
+    weight_decay: 1.0e-08
+    fused: true
+  save_format: pt
+workspace_dir: /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/talha-aug26
+run_root_dir: /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/talha-aug26/results/Checkpoints
+seed: 42
+wandb_entity: zihanwang-ai-northwestern-university
+wandb_project: starVLA_rl_games
+auth:
+  env_file: null
+  hf_token_env: HF_TOKEN
+  wandb_api_key_env: WANDB_API_KEY
+paths:
+  run_root_dir: results/Checkpoints
+  dataset_local_dir: playground/Datasets/rl_games
+  dataset_cache_dir: null
+  base_model_dir: playground/Pretrained_models/Qwen3-VL-4B-Instruct
+  accelerate_config: starVLA/config/deepseeds/deepspeed_zero2.yaml
+launch:
+  use_accelerate: true
+  gpus: null
+  num_processes: 1
+  dry_run: false
+conda:
+  enabled: true
+  env_name: null
+rl_games:
+  model_alias: openvla
+  task: cross_task
+  env_eval:
+    image_size: 224
+    frameskip: 1
+    deadly:
+      action_layout: multibinary_7
+      multibinary_threshold: 0.0
+    enabled: true
+    eval_backend: eval_core
+    distributed_mode: rank_sharded
+    vectorized:
+      enabled: false
+      batch_size: 1
+    latency:
+      prompt_map_path: null
+      mode: mixed
+      values:
+      - 0
+      - 1
+      - 2
+      - 3
+      - 4
+      - 5
+    mid_train:
+      enabled: true
+      interval_steps: 500
+      latencies:
+      - 0
+      - 1
+      - 2
+      - 3
+      - 4
+      - 5
+      num_episodes: 5
+      max_steps_per_episode: 3600
+    post_train:
+      enabled: true
+      latencies:
+      - 0
+      - 1
+      - 2
+      - 3
+      - 4
+      num_episodes: 5
+      max_steps_per_episode: 3600
+  cross_task:
+    train_tasks:
+    - name: defend_the_line
+      role: A
+      train_source_hf: latency-sensitive-bench/memory-rollouts
+      prompt_source_hf: latency-sensitive-bench/memory-rollouts
+      train_source_subdir: defend_the_line_fixed_latency_0_1000ep_7k2steps
+      prompt_source_subdir: defend_the_line_fixed_latency_0_1000ep_7k2steps
+      converted_name: defend_the_line_0_cross_train_openvla_bridge_cross_defendtheline_zero_deadly_zero_exp3
+      train_latency_filter:
+      - 0
+      eval_latency_filter:
+      - 0
+      episodes_per_latency: 1000
+      max_episodes: null
+      weight: 1.0
+    - name: deadly_corridor
+      role: B
+      train_source_hf: latency-sensitive-bench/deadly_1000ep
+      prompt_source_hf: latency-sensitive-bench/deadly_1000ep
+      converted_name: deadly_corridor_0_cross_train_openvla_bridge_cross_defendtheline_zero_deadly_zero_exp3
+      train_latency_filter:
+      - 0
+      eval_latency_filter:
+      - 0
+      - 2
+      - 4
+      episodes_per_latency: 1000
+      max_episodes: null
+      action_layout: multibinary_7
+      weight: 1.0
+    loss_by_task:
+      flappy: discrete_ce
+      demon_attack: discrete_ce
+      defend_the_line: discrete_ce
+      deadly_corridor: multibinary_bce
+    loss_weight_by_task:
+      flappy: 1.0
+      demon_attack: 1.0
+      defend_the_line: 1.0
+      deadly_corridor: 1.0
+    eval_tasks:
+      flappy:
+        prompt_map_path: null
+        frameskip: 1
+        image_size: 224
+        task_description: null
+        mid_train:
+          enabled: true
+          latencies: null
+          num_episodes: null
+          max_steps_per_episode: null
+        post_train:
+          enabled: true
+          latencies: null
+          num_episodes: null
+          max_steps_per_episode: null
+      demon_attack:
+        prompt_map_path: null
+        frameskip: 4
+        image_size: 224
+        task_description: null
+        mid_train:
+          enabled: true
+          latencies: null
+          num_episodes: null
+          max_steps_per_episode: null
+        post_train:
+          enabled: true
+          latencies: null
+          num_episodes: null
+          max_steps_per_episode: null
+      defend_the_line:
+        prompt_map_path: /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/talha-aug26/playground/Datasets/rl_games/_prompt_maps/defend_the_line_0_cross_train_openvla_bridge_cross_defendtheline_zero_deadly_zero_exp3__lat0__1000ep_per_lat__bridge/eval_latency_prompt_map.json
+        frameskip: 4
+        image_size: 224
+        task_description: 'You are playing Defend the Line in VizDoom. Your goal is
+          to survive by turning toward enemies and shooting them before they reach
+          you. The available actions are: TURN_LEFT, TURN_RIGHT, and ATTACK. Choose
+          the best next action from the current visual observation.'
+        mid_train:
+          enabled: false
+          latencies:
+          - 0
+          num_episodes: 5
+          max_steps_per_episode: 3600
+        post_train:
+          enabled: true
+          latencies:
+          - 0
+          - 2
+          - 4
+          num_episodes: 20
+          max_steps_per_episode: 3600
+      deadly_corridor:
+        prompt_map_path: /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/talha-aug26/playground/Datasets/rl_games/_prompt_maps/deadly_corridor_0_cross_train_openvla_bridge_cross_defendtheline_zero_deadly_zero_exp3__lat0__1000ep_per_lat__bridge/eval_latency_prompt_map.json
+        frameskip: 4
+        image_size: 224
+        task_description: You are playing Deadly Corridor in VizDoom. Choose actions
+          from MOVE_FORWARD, MOVE_BACKWARD, MOVE_LEFT, MOVE_RIGHT, TURN_LEFT, TURN_RIGHT,
+          ATTACK.
+        mid_train:
+          enabled: false
+          latencies:
+          - 0
+          num_episodes: 5
+          max_steps_per_episode: 3600
+        post_train:
+          enabled: true
+          latencies:
+          - 0
+          - 2
+          - 4
+          num_episodes: 20
+          max_steps_per_episode: 3600
+  initialization_mode: bridge
+  action_carrier: bridge
+model: openvla
+env: cross_task
+init: bridge
+bridge_base_model:
+  repo_id:
+    openvla: Qwen/Qwen3-VL-4B-Instruct
+    pi0: StarVLA/Qwen2.5-VL-3B-Instruct-Action
+    pi05: Qwen/Qwen3-VL-4B-Instruct
+    gr00t: Qwen/Qwen3-VL-4B-Instruct
+  local_dir:
+    openvla: playground/Pretrained_models/Qwen3-VL-4B-Instruct
+    pi0: playground/Pretrained_models/Qwen2.5-VL-3B-Instruct-Action
+    pi05: playground/Pretrained_models/Qwen3-VL-4B-Instruct
+    gr00t: playground/Pretrained_models/Qwen3-VL-4B-Instruct
+mode: cross_task
+checkpoint:
+  load: auto
+  hf_repo_id: latency-sensitive-bench/openvla_bridge_cross_defendtheline_zero_deadly_zero_exp3
+  save_best_model: false
+  save_final_model: true
+  save_pt_file: false
+  save_training_state: true
+  save_safetensors_file: true
+  local:
+    keep_last_n: 1
+  sync:
+    enabled: false
+    repo_id: latency-sensitive-bench/openvla_bridge_cross_defendtheline_zero_deadly_zero_exp3
+    keep_last_n: 0
+    sync_every_n_checkpoints: 1
+    resume_policy: local_latest
+run_id: openvla_bridge_cross_defendtheline_zero_deadly_zero_exp3
+output_dir: /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/talha-aug26/results/Checkpoints/openvla_bridge_cross_defendtheline_zero_deadly_zero_exp3
+config_yaml: null
+is_debug: false
+version_id: '0.21'

archive/task-transfer/openvla_bridge_cross_defendtheline_zero_deadly_zero_exp3/config.yaml

@@ -0,0 +1,155 @@
+checkpoint:
+  local:
+    keep_last_n: 1
+  save_best_model: false
+  save_final_model: true
+  save_pt_file: false
+  save_safetensors_file: true
+  save_training_state: true
+  sync:
+    enabled: false
+    keep_last_n: 0
+    repo_id: latency-sensitive-bench/openvla_bridge_cross_defendtheline_zero_deadly_zero_exp3
+datasets:
+  vla_data:
+    data_mix: cross__deadly_corridor_0_cross_train_openvla_bridge_cross_defendtheline_zero_deadly_zero_exp3__lat0__1000ep_per_lat__bridge__defend_the_line_0_cross_train_openvla_bridge_cross_defendtheline_zero_deadly_zero_exp3__lat0__1000ep_per_lat__bridge
+    dataset_py: lerobot_datasets
+    eval_data_mix: cross__deadly_corridor_0_cross_train_openvla_bridge_cross_defendtheline_zero_deadly_zero_exp3__lat0__1000ep_per_lat__bridge__defend_the_line_0_cross_train_openvla_bridge_cross_defendtheline_zero_deadly_zero_exp3__lat0__1000ep_per_lat__bridge__val
+    eval_num_workers: 8
+    include_state: false
+    latency_curriculum:
+      enabled: false
+    obs_image_size: null
+    per_device_batch_size: 4
+    persistent_workers: true
+    pin_memory: true
+    prefetch_factor: 4
+framework:
+  action_model:
+    action_dim: 7
+    action_env_dim: 7
+    action_hidden_dim: 2560
+    action_horizon: 1
+    action_model_type: MLP
+    loss_type: discrete_ce
+    state_dim: 7
+  kv_memory:
+    enabled: false
+    packed_train: false
+    rebased_sink: true
+    rollout_len: 8
+    window: 4
+  name: QwenOFT
+  qwenvl:
+    attn_implementation: flash_attention_2
+    base_vlm: /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/talha-aug26/playground/Pretrained_models/Qwen3-VL-4B-Instruct
+    enable_gradient_checkpointing: true
+output_dir: /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/talha-aug26/results/Checkpoints/openvla_bridge_cross_defendtheline_zero_deadly_zero_exp3
+rl_games:
+  cross_task:
+    eval_tasks:
+      deadly_corridor:
+        mid_train:
+          enabled: false
+        prompt_map_path: /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/talha-aug26/playground/Datasets/rl_games/_prompt_maps/deadly_corridor_0_cross_train_openvla_bridge_cross_defendtheline_zero_deadly_zero_exp3__lat0__1000ep_per_lat__bridge/eval_latency_prompt_map.json
+      defend_the_line:
+        mid_train:
+          enabled: false
+        prompt_map_path: /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/talha-aug26/playground/Datasets/rl_games/_prompt_maps/defend_the_line_0_cross_train_openvla_bridge_cross_defendtheline_zero_deadly_zero_exp3__lat0__1000ep_per_lat__bridge/eval_latency_prompt_map.json
+      demon_attack:
+        frameskip: 4
+        image_size: 224
+        mid_train:
+          enabled: true
+          latencies: null
+          max_steps_per_episode: null
+          num_episodes: null
+        prompt_map_path: null
+        task_description: null
+      flappy:
+        frameskip: 1
+        image_size: 224
+        mid_train:
+          enabled: true
+          latencies: null
+          max_steps_per_episode: null
+          num_episodes: null
+        prompt_map_path: null
+        task_description: null
+    loss_by_task:
+      deadly_corridor: multibinary_bce
+      defend_the_line: discrete_ce
+      demon_attack: discrete_ce
+      flappy: discrete_ce
+    loss_weight_by_task:
+      deadly_corridor: 1.0
+      defend_the_line: 1.0
+      demon_attack: 1.0
+      flappy: 1.0
+  env_eval:
+    deadly:
+      action_layout: multibinary_7
+    distributed_mode: rank_sharded
+    enabled: true
+    eval_backend: eval_core
+    frameskip: 1
+    image_size: 224
+    latency:
+      prompt_map_path: null
+      values:
+      - 0
+      - 1
+      - 2
+      - 3
+      - 4
+      - 5
+    mid_train:
+      enabled: true
+      interval_steps: 500
+      max_steps_per_episode: 3600
+      num_episodes: 5
+    vectorized:
+      enabled: false
+  model_alias: openvla
+  task: cross_task
+run_id: openvla_bridge_cross_defendtheline_zero_deadly_zero_exp3
+run_root_dir: /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/talha-aug26/results/Checkpoints
+seed: 42
+trainer:
+  cc_f1_tolerance: 1
+  distributed_backend: deepspeed
+  eval_action_classification: true
+  eval_action_classification_interval: null
+  eval_interval: 400
+  eval_num_batches: 100
+  freeze_llm_layers: []
+  freeze_modules: ''
+  freeze_tied_embedding: false
+  freeze_vit: false
+  gradient_accumulation_steps: 4
+  is_resume: false
+  learning_rate:
+    action_model: 0.0001
+    base: 2.0e-05
+    qwen_vl_interface: 1.0e-05
+  logging_frequency: 1
+  lr_scheduler_type: cosine_with_min_lr
+  max_train_steps: 2000
+  num_warmup_steps: 0
+  optimizer:
+    betas:
+    - 0.9
+    - 0.95
+    eps: 1.0e-08
+    fused: true
+    weight_decay: 1.0e-08
+  per_latency_eval_num_batches: 5
+  pretrained_checkpoint: null
+  profile_timing:
+    enabled: false
+  reload_modules: null
+  save_interval: 500
+  scheduler_specific_kwargs:
+    min_lr: 1.0e-06
+wandb_entity: zihanwang-ai-northwestern-university
+wandb_project: starVLA_rl_games

archive/task-transfer/openvla_bridge_cross_defendtheline_zero_deadly_zero_exp3/dataset_statistics.json

@@ -0,0 +1,127 @@
+{
+  "new_embodiment": {
+    "action": {
+      "mean": [
+        0.4634477236540988,
+        0.3839454741682857,
+        0.08175837993621826,
+        0.3876986764371395,
+        0.11705999199693906,
+        0.11682328209280968,
+        0.27798372507095337
+      ],
+      "std": [
+        0.4986501006596913,
+        0.4862513963960781,
+        0.2740502521082896,
+        0.48729818194014035,
+        0.32147348539970855,
+        0.3209976104148355,
+        0.4480408052663252
+      ],
+      "max": [
+        1.0,
+        1.0,
+        1.0,
+        1.0,
+        1.0,
+        1.0,
+        1.0
+      ],
+      "min": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "q01": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "q99": [
+        1.0,
+        1.0,
+        1.0,
+        1.0,
+        1.0,
+        1.0,
+        1.0
+      ],
+      "mask": [
+        true,
+        true,
+        true,
+        true,
+        true,
+        true,
+        true
+      ]
+    },
+    "state": {
+      "mean": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "std": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "max": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "min": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "q01": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "q99": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ]
+    },
+    "num_transitions": 325456,
+    "num_trajectories": 1890
+  }
+}

archive/task-transfer/openvla_bridge_cross_defendtheline_zero_deadly_zero_exp3/dataset_statistics_eval.json

@@ -0,0 +1,127 @@
+{
+  "new_embodiment": {
+    "action": {
+      "mean": [
+        0.4792165304534137,
+        0.38396821450442076,
+        0.14564630389213562,
+        0.3346198499202728,
+        0.11296086758375168,
+        0.17345838993787766,
+        0.37078356742858887
+      ],
+      "std": [
+        0.4995725429693537,
+        0.4863533629302698,
+        0.35275196406561327,
+        0.4718558228919714,
+        0.31653974339283536,
+        0.3786538218191283,
+        0.48301901353259197
+      ],
+      "max": [
+        1.0,
+        1.0,
+        1.0,
+        1.0,
+        1.0,
+        1.0,
+        1.0
+      ],
+      "min": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "q01": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "q99": [
+        1.0,
+        1.0,
+        1.0,
+        1.0,
+        1.0,
+        1.0,
+        1.0
+      ],
+      "mask": [
+        true,
+        true,
+        true,
+        true,
+        true,
+        true,
+        true
+      ]
+    },
+    "state": {
+      "mean": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "std": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "max": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "min": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "q01": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "q99": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ]
+    },
+    "num_transitions": 14317,
+    "num_trajectories": 310
+  }
+}

archive/task-transfer/openvla_bridge_cross_mixed_all_024_5env200ep_deadly1000ep_exp1/checkpoints/steps_15000_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ccab5fb797b96ccde78035c1f1d9e83ebe5f2a4bc9e5b8134f0682bc4e979bdd
+size 9138230516

archive/task-transfer/openvla_bridge_cross_mixed_all_024_5env200ep_deadly1000ep_exp1/config.full.yaml

@@ -0,0 +1,503 @@
+framework:
+  name: QwenOFT
+  qwenvl:
+    base_vlm: /mnt/local/talha/playground/Pretrained_models/Qwen3-VL-4B-Instruct
+    attn_implementation: flash_attention_2
+    enable_gradient_checkpointing: true
+  action_model:
+    action_model_type: MLP
+    action_dim: 7
+    state_dim: 7
+    action_hidden_dim: 2560
+    future_action_window_size: 0
+    past_action_window_size: 0
+    loss_type: discrete_ce
+    state_encoding: discretized_text
+    task_objective: null
+    action_horizon: 1
+    action_env_dim: 7
+    action_layout: bridge_cross_task_7
+  kv_memory:
+    enabled: false
+    window: 4
+    rollout_len: 8
+    packed_train: false
+    rebased_sink: true
+datasets:
+  vla_data:
+    dataset_py: lerobot_datasets
+    include_state: false
+    data_root_dir: /mnt/local/talha/playground/Datasets/rl_games
+    data_mix: cross__asterix_mixed_024_200ep_cross_train__lat0_2_4__200ep__bridge__atlantis_mixed_024_200ep_cross_train__lat0_2_4__200ep__bridge__deadly_corridor_mixed_024_1000ep_cross_train__lat0_2_4__1000ep__bridge__defend_the_line_mixed_024_200ep_cross_train__lat0_2_4__200ep__bridge__demon_attack_mixed_024_200ep_cross_train__lat0_2_4__200ep__bridge__flappy_mixed_024_200ep_cross_train__lat0_2_4__200ep__bridge
+    eval_data_mix: cross__asterix_mixed_024_200ep_cross_train__lat0_2_4__200ep__bridge__atlantis_mixed_024_200ep_cross_train__lat0_2_4__200ep__bridge__deadly_corridor_mixed_024_1000ep_cross_train__lat0_2_4__1000ep__bridge__defend_the_line_mixed_024_200ep_cross_train__lat0_2_4__200ep__bridge__demon_attack_mixed_024_200ep_cross_train__lat0_2_4__200ep__bridge__flappy_mixed_024_200ep_cross_train__lat0_2_4__200ep__bridge__val
+    custom_mixtures_path: /mnt/local/talha/playground/Datasets/rl_games/_generated_mixtures/cross_09b5517c91a12cca.json
+    action_type: discrete
+    sequential_step_sampling: true
+    eval_sequential_step_sampling: null
+    num_workers: 8
+    eval_num_workers: 8
+    prefetch_factor: 4
+    persistent_workers: true
+    pin_memory: true
+    shuffle: true
+    action_balance:
+      enabled: false
+      strategy: balanced_epoch
+      action_key: action_id
+      target_flap_fraction: 0.3
+      noop_id: 0
+      flap_id: 1
+    latency_curriculum:
+      enabled: false
+      strategy: exclusive
+      latencies: null
+      phase_steps: null
+      phase_distributions: null
+      new_latency_passes: 1.0
+      replay_passes: 0.25
+      target_total_passes: 2.0
+      final_equalization: true
+      step_budget_mode: auto
+      eval_at_phase_end: false
+      save_at_phase_end: false
+      computed_plan: null
+    per_device_batch_size: 32
+    load_all_data_for_training: true
+    num_obs_frames: 1
+    image_mode: single
+    prompt_mode: raw
+    stitch_grid:
+    - 2
+    - 2
+    obs_image_size: null
+    video_backend: torchvision_av
+dataset:
+  source_hf: ''
+  config_name: null
+  source_subdir: null
+  converted_name: cross_task_train
+  single_source_hf: ''
+  mixed_source_hf: ''
+  single_converted_name: flappy_train
+  mixed_converted_name: flappy_mixed_latency_train
+  single_latency_filter: null
+  mixed_latency_filter: null
+  force_download: false
+  setup_force: false
+  skip_verification: false
+  target_latency_unit: observation_steps
+  verify_rows: 200
+  max_episodes: null
+  episodes_per_latency: null
+  latency_filter: null
+  debug_subset:
+    enabled: false
+    max_episodes: 5
+    suffix: debug
+base_model:
+  repo_id: Qwen/Qwen3-VL-4B-Instruct
+initialization:
+  checkpoint_local_dir: playground/Pretrained_models/Qwen3VL-OFT-Bridge-RT-1
+  checkpoint_hf_repo_id: StarVLA/Qwen3VL-OFT-Bridge-RT-1
+  checkpoint_filename: checkpoints/steps_5000_pytorch_model.pt
+trainer:
+  max_train_steps: 15000
+  num_warmup_steps: 0
+  save_interval: 15000
+  eval_interval: 15000
+  eval_num_batches: 100
+  per_latency_eval_num_batches: 5
+  eval_action_classification: true
+  eval_action_classification_interval: null
+  cc_f1_tolerance: 1
+  learning_rate:
+    base: 2.0e-05
+    qwen_vl_interface: 1.0e-05
+    action_model: 0.0001
+  lr_scheduler_type: cosine_with_min_lr
+  scheduler_specific_kwargs:
+    min_lr: 1.0e-06
+  freeze_modules: ''
+  freeze_vit: false
+  freeze_tied_embedding: false
+  freeze_llm_layers: []
+  loss_scale:
+    vla: 1.0
+    vlm: 0.1
+  max_grad_norm: 1.0
+  weight_decay: 0.0
+  logging_frequency: 1
+  profile_timing:
+    enabled: false
+    log_interval: 10
+  gradient_clipping: 1.0
+  gradient_accumulation_steps: 8
+  distributed_backend: none
+  is_resume: false
+  pretrained_checkpoint: null
+  resume_step: 0
+  reload_modules: null
+  optimizer:
+    name: AdamW
+    betas:
+    - 0.9
+    - 0.95
+    eps: 1.0e-08
+    weight_decay: 1.0e-08
+    fused: true
+  save_format: pt
+workspace_dir: /mnt/local/talha
+run_root_dir: /mnt/local/talha/results/Checkpoints
+seed: 42
+wandb_entity: talha1503
+wandb_project: starVLA_rl_games
+auth:
+  env_file: null
+  hf_token_env: HF_TOKEN
+  wandb_api_key_env: WANDB_API_KEY
+paths:
+  run_root_dir: /mnt/local/talha/results/Checkpoints
+  dataset_local_dir: /mnt/local/talha/playground/Datasets/rl_games
+  dataset_cache_dir: null
+  base_model_dir: /mnt/local/talha/playground/Pretrained_models/Qwen3-VL-4B-Instruct
+  accelerate_config: starVLA/config/deepseeds/deepspeed_zero2.yaml
+launch:
+  use_accelerate: true
+  gpus: null
+  num_processes: 1
+  dry_run: false
+conda:
+  enabled: true
+  env_name: null
+rl_games:
+  model_alias: openvla
+  task: cross_task
+  env_eval:
+    image_size: 224
+    frameskip: 1
+    deadly:
+      action_layout: multibinary_7
+      multibinary_threshold: 0.0
+    enabled: true
+    eval_backend: eval_core
+    distributed_mode: none
+    vectorized:
+      enabled: false
+      batch_size: 1
+    latency:
+      prompt_map_path: null
+      mode: mixed
+      values:
+      - 0
+      - 1
+      - 2
+      - 3
+      - 4
+      - 5
+    mid_train:
+      enabled: false
+      interval_steps: 1000
+      latencies:
+      - 0
+      - 1
+      - 2
+      - 3
+      - 4
+      - 5
+      num_episodes: 5
+      max_steps_per_episode: 3600
+    post_train:
+      enabled: true
+      latencies:
+      - 0
+      - 2
+      - 4
+      latencies_from_prompt_map: false
+      num_episodes: 5
+      max_steps_per_episode: 3600
+    asterix:
+      action_layout: factorized_6
+  cross_task:
+    train_tasks:
+    - name: flappy
+      role: A
+      train_source_hf: latency-sensitive-bench/flappy_200ep
+      prompt_source_hf: latency-sensitive-bench/flappy_200ep
+      train_source_subdir: flappy_fix_latency_0_200ep
+      prompt_source_subdir: flappy_fix_latency_0_200ep
+      converted_name: flappy_mixed_024_200ep_cross_train
+      train_latency_filter:
+      - 0
+      - 2
+      - 4
+      eval_latency_filter:
+      - 0
+      - 2
+      - 4
+      episodes_per_latency: null
+      max_episodes: 200
+      weight: 1.0
+    - name: demon_attack
+      role: B
+      train_source_hf: latency-sensitive-bench/demon_attack_200ep
+      prompt_source_hf: latency-sensitive-bench/demon_attack_200ep
+      train_source_subdir: demon_attack_fix_latency_0_200ep
+      prompt_source_subdir: demon_attack_fix_latency_0_200ep
+      converted_name: demon_attack_mixed_024_200ep_cross_train
+      train_latency_filter:
+      - 0
+      - 2
+      - 4
+      eval_latency_filter:
+      - 0
+      - 2
+      - 4
+      episodes_per_latency: null
+      max_episodes: 200
+      weight: 1.0
+    - name: deadly_corridor
+      role: C
+      train_source_hf: latency-sensitive-bench/deadly_1000ep
+      prompt_source_hf: latency-sensitive-bench/deadly_1000ep
+      train_source_subdir: deadly_corridor_fix_latency_0_1000ep
+      prompt_source_subdir: deadly_corridor_fix_latency_0_1000ep
+      converted_name: deadly_corridor_mixed_024_1000ep_cross_train
+      train_latency_filter:
+      - 0
+      - 2
+      - 4
+      eval_latency_filter:
+      - 0
+      - 2
+      - 4
+      episodes_per_latency: null
+      max_episodes: 1000
+      action_layout: multibinary_7
+      weight: 1.0
+    - name: asterix
+      role: D
+      train_source_hf: latency-sensitive-bench/asterix_200ep
+      prompt_source_hf: latency-sensitive-bench/asterix_200ep
+      train_source_subdir: asterix_fixed_latency_0_200ep_7k2steps
+      prompt_source_subdir: asterix_fixed_latency_0_200ep_7k2steps
+      converted_name: asterix_mixed_024_200ep_cross_train
+      train_latency_filter:
+      - 0
+      - 2
+      - 4
+      eval_latency_filter:
+      - 0
+      - 2
+      - 4
+      episodes_per_latency: null
+      max_episodes: 200
+      action_layout: factorized_6
+      weight: 1.0
+    - name: atlantis
+      role: E
+      train_source_hf: latency-sensitive-bench/atlantis_200ep
+      prompt_source_hf: latency-sensitive-bench/atlantis_200ep
+      train_source_subdir: atlantis_fixed_latency_0_200ep_7k2steps
+      prompt_source_subdir: atlantis_fixed_latency_0_200ep_7k2steps
+      converted_name: atlantis_mixed_024_200ep_cross_train
+      train_latency_filter:
+      - 0
+      - 2
+      - 4
+      eval_latency_filter:
+      - 0
+      - 2
+      - 4
+      episodes_per_latency: null
+      max_episodes: 200
+      weight: 1.0
+    - name: defend_the_line
+      role: F
+      train_source_hf: latency-sensitive-bench/defend_the_line_200ep
+      prompt_source_hf: latency-sensitive-bench/defend_the_line_200ep
+      train_source_subdir: defend_the_line_fixed_latency_0_200ep_7k2steps
+      prompt_source_subdir: defend_the_line_fixed_latency_0_200ep_7k2steps
+      converted_name: defend_the_line_mixed_024_200ep_cross_train
+      train_latency_filter:
+      - 0
+      - 2
+      - 4
+      eval_latency_filter:
+      - 0
+      - 2
+      - 4
+      episodes_per_latency: null
+      max_episodes: 200
+      weight: 1.0
+    loss_by_task:
+      flappy: discrete_ce
+      demon_attack: discrete_ce
+      defend_the_line: discrete_ce
+      deadly_corridor: multibinary_bce
+      asterix: asterix_factorized_ce
+      atlantis: discrete_ce
+    loss_weight_by_task:
+      flappy: 1.0
+      demon_attack: 1.0
+      defend_the_line: 1.0
+      deadly_corridor: 1.0
+      asterix: 1.0
+      atlantis: 1.0
+    eval_tasks:
+      flappy:
+        prompt_map_path: /mnt/local/talha/playground/Datasets/rl_games/_prompt_maps/flappy_mixed_024_200ep_cross_train__lat0_2_4__200ep__bridge/eval_latency_prompt_map.json
+        frameskip: 1
+        image_size: 224
+        task_description: 'You are playing Flappy Bird. Pass through the pipe gaps
+          and stay alive. Choose the action: NOOP, FLAP.'
+        mid_train:
+          enabled: false
+          latencies:
+          - 0
+          - 2
+          - 4
+          num_episodes: 5
+          max_steps_per_episode: 3600
+        post_train:
+          enabled: true
+          latencies:
+          - 0
+          - 2
+          - 4
+          num_episodes: 20
+          max_steps_per_episode: 3600
+      demon_attack:
+        prompt_map_path: /mnt/local/talha/playground/Datasets/rl_games/_prompt_maps/demon_attack_mixed_024_200ep_cross_train__lat0_2_4__200ep__bridge/eval_latency_prompt_map.json
+        frameskip: 4
+        image_size: 224
+        task_description: 'You are playing Demon Attack from a single game image.
+          Choose exactly one action from: NOOP, FIRE, RIGHT, LEFT, RIGHTFIRE, LEFTFIRE.'
+        mid_train:
+          enabled: false
+          latencies:
+          - 0
+          - 2
+          - 4
+          num_episodes: 5
+          max_steps_per_episode: 3600
+        post_train:
+          enabled: true
+          latencies:
+          - 0
+          - 2
+          - 4
+          num_episodes: 20
+          max_steps_per_episode: 3600
+      defend_the_line:
+        prompt_map_path: null
+        frameskip: 4
+        image_size: 224
+        task_description: null
+        mid_train:
+          enabled: false
+          latencies: null
+          num_episodes: null
+          max_steps_per_episode: null
+        post_train:
+          enabled: false
+          latencies: null
+          num_episodes: null
+          max_steps_per_episode: null
+      deadly_corridor:
+        prompt_map_path: /mnt/local/talha/playground/Datasets/rl_games/_prompt_maps/deadly_corridor_mixed_024_1000ep_cross_train__lat0_2_4__1000ep__bridge/eval_latency_prompt_map.json
+        frameskip: 4
+        image_size: 224
+        task_description: You are playing Deadly Corridor in VizDoom. Choose actions
+          from MOVE_FORWARD, MOVE_BACKWARD, MOVE_LEFT, MOVE_RIGHT, TURN_LEFT, TURN_RIGHT,
+          ATTACK.
+        mid_train:
+          enabled: false
+          latencies:
+          - 0
+          - 2
+          - 4
+          num_episodes: 5
+          max_steps_per_episode: 3600
+        post_train:
+          enabled: true
+          latencies:
+          - 0
+          - 2
+          - 4
+          num_episodes: 20
+          max_steps_per_episode: 3600
+      asterix:
+        prompt_map_path: null
+        frameskip: 4
+        image_size: 224
+        task_description: null
+        mid_train:
+          enabled: false
+          latencies: null
+          num_episodes: null
+          max_steps_per_episode: null
+        post_train:
+          enabled: false
+          latencies: null
+          num_episodes: null
+          max_steps_per_episode: null
+      atlantis:
+        prompt_map_path: null
+        frameskip: 4
+        image_size: 224
+        task_description: null
+        mid_train:
+          enabled: false
+          latencies: null
+          num_episodes: null
+          max_steps_per_episode: null
+        post_train:
+          enabled: false
+          latencies: null
+          num_episodes: null
+          max_steps_per_episode: null
+  initialization_mode: bridge
+  action_carrier: bridge
+model: openvla
+env: cross_task
+init: bridge
+bridge_base_model:
+  repo_id:
+    openvla: Qwen/Qwen3-VL-4B-Instruct
+    pi0: StarVLA/Qwen2.5-VL-3B-Instruct-Action
+    pi05: Qwen/Qwen3-VL-4B-Instruct
+    gr00t: Qwen/Qwen3-VL-4B-Instruct
+  local_dir:
+    openvla: playground/Pretrained_models/Qwen3-VL-4B-Instruct
+    pi0: playground/Pretrained_models/Qwen2.5-VL-3B-Instruct-Action
+    pi05: playground/Pretrained_models/Qwen3-VL-4B-Instruct
+    gr00t: playground/Pretrained_models/Qwen3-VL-4B-Instruct
+mode: cross_task
+checkpoint:
+  load: auto
+  hf_repo_id: latency-sensitive-bench/openvla_bridge_cross_mixed_all_024_5env200ep_deadly1000ep_exp1
+  save_best_model: false
+  save_final_model: true
+  save_pt_file: false
+  save_training_state: true
+  save_safetensors_file: false
+  local:
+    keep_last_n: 1
+  sync:
+    enabled: true
+    repo_id: latency-sensitive-bench/openvla_bridge_cross_mixed_all_024_5env200ep_deadly1000ep_exp1
+    keep_last_n: 0
+    sync_every_n_checkpoints: 1
+    resume_policy: local_latest
+run_id: openvla_bridge_cross_mixed_all_024_5env200ep_deadly1000ep_exp1
+output_dir: /mnt/local/talha/results/Checkpoints/openvla_bridge_cross_mixed_all_024_5env200ep_deadly1000ep_exp1
+config_yaml: null
+is_debug: false
+version_id: '0.21'
+wandb_tags:
+- final_vla_profiling_training
+- demon_attack
+- openvla
+- rtx3090_profile
+training_latency_condition: rtx3090_profile
+wandb_group: final_vla_profiling_training

archive/task-transfer/openvla_bridge_cross_mixed_all_024_5env200ep_deadly1000ep_exp1/config.yaml

@@ -0,0 +1,121 @@
+checkpoint:
+  local:
+    keep_last_n: 1
+  save_best_model: false
+  save_final_model: true
+  save_pt_file: false
+  save_safetensors_file: false
+  save_training_state: true
+  sync:
+    enabled: true
+    keep_last_n: 0
+    repo_id: latency-sensitive-bench/openvla_bridge_cross_mixed_all_024_5env200ep_deadly1000ep_exp1
+datasets:
+  vla_data:
+    data_mix: cross__asterix_mixed_024_200ep_cross_train__lat0_2_4__200ep__bridge__atlantis_mixed_024_200ep_cross_train__lat0_2_4__200ep__bridge__deadly_corridor_mixed_024_1000ep_cross_train__lat0_2_4__1000ep__bridge__defend_the_line_mixed_024_200ep_cross_train__lat0_2_4__200ep__bridge__demon_attack_mixed_024_200ep_cross_train__lat0_2_4__200ep__bridge__flappy_mixed_024_200ep_cross_train__lat0_2_4__200ep__bridge
+    dataset_py: lerobot_datasets
+    eval_data_mix: cross__asterix_mixed_024_200ep_cross_train__lat0_2_4__200ep__bridge__atlantis_mixed_024_200ep_cross_train__lat0_2_4__200ep__bridge__deadly_corridor_mixed_024_1000ep_cross_train__lat0_2_4__1000ep__bridge__defend_the_line_mixed_024_200ep_cross_train__lat0_2_4__200ep__bridge__demon_attack_mixed_024_200ep_cross_train__lat0_2_4__200ep__bridge__flappy_mixed_024_200ep_cross_train__lat0_2_4__200ep__bridge__val
+    eval_num_workers: 8
+    latency_curriculum:
+      enabled: false
+    obs_image_size: null
+    per_device_batch_size: 32
+    persistent_workers: true
+    pin_memory: true
+    prefetch_factor: 4
+framework:
+  action_model:
+    action_dim: 7
+    action_env_dim: 7
+    action_hidden_dim: 2560
+    action_horizon: 1
+    action_model_type: MLP
+    loss_type: discrete_ce
+    state_encoding: discretized_text
+    task_objective: null
+  kv_memory:
+    enabled: false
+    packed_train: false
+    rebased_sink: true
+    rollout_len: 8
+    window: 4
+  name: QwenOFT
+  qwenvl:
+    attn_implementation: flash_attention_2
+    base_vlm: /mnt/local/talha/playground/Pretrained_models/Qwen3-VL-4B-Instruct
+    enable_gradient_checkpointing: true
+output_dir: /mnt/local/talha/results/Checkpoints/openvla_bridge_cross_mixed_all_024_5env200ep_deadly1000ep_exp1
+rl_games:
+  cross_task:
+    loss_by_task:
+      asterix: asterix_factorized_ce
+      atlantis: discrete_ce
+      deadly_corridor: multibinary_bce
+      defend_the_line: discrete_ce
+      demon_attack: discrete_ce
+      flappy: discrete_ce
+    loss_weight_by_task:
+      asterix: 1.0
+      atlantis: 1.0
+      deadly_corridor: 1.0
+      defend_the_line: 1.0
+      demon_attack: 1.0
+      flappy: 1.0
+  env_eval:
+    deadly:
+      action_layout: multibinary_7
+    enabled: true
+    eval_backend: eval_core
+    mid_train:
+      enabled: false
+      interval_steps: 1000
+  task: cross_task
+run_id: openvla_bridge_cross_mixed_all_024_5env200ep_deadly1000ep_exp1
+run_root_dir: /mnt/local/talha/results/Checkpoints
+seed: 42
+trainer:
+  cc_f1_tolerance: 1
+  distributed_backend: none
+  eval_action_classification: true
+  eval_action_classification_interval: null
+  eval_interval: 15000
+  eval_num_batches: 100
+  freeze_llm_layers: []
+  freeze_modules: ''
+  freeze_tied_embedding: false
+  freeze_vit: false
+  gradient_accumulation_steps: 8
+  gradient_clipping: 1.0
+  is_resume: false
+  learning_rate:
+    action_model: 0.0001
+    base: 2.0e-05
+    qwen_vl_interface: 1.0e-05
+  logging_frequency: 1
+  lr_scheduler_type: cosine_with_min_lr
+  max_train_steps: 15000
+  num_warmup_steps: 0
+  optimizer:
+    betas:
+    - 0.9
+    - 0.95
+    eps: 1.0e-08
+    fused: true
+    weight_decay: 1.0e-08
+  per_latency_eval_num_batches: 5
+  pretrained_checkpoint: null
+  profile_timing:
+    enabled: false
+  reload_modules: null
+  save_interval: 15000
+  scheduler_specific_kwargs:
+    min_lr: 1.0e-06
+training_latency_condition: rtx3090_profile
+wandb_entity: talha1503
+wandb_group: final_vla_profiling_training
+wandb_project: starVLA_rl_games
+wandb_tags:
+- final_vla_profiling_training
+- demon_attack
+- openvla
+- rtx3090_profile

archive/task-transfer/openvla_bridge_cross_mixed_all_024_5env200ep_deadly1000ep_exp1/dataset_statistics.json

@@ -0,0 +1,127 @@
+{
+  "new_embodiment": {
+    "action": {
+      "mean": [
+        0.40483391160766286,
+        0.28502383750552934,
+        0.2153583544616898,
+        0.1819893904030323,
+        0.16567657360186183,
+        0.12371420363585156,
+        0.12493560711542767
+      ],
+      "std": [
+        0.49074092958869836,
+        0.4513986874197357,
+        0.41100783371886684,
+        0.38589290819222594,
+        0.37179515437093547,
+        0.3293991390251945,
+        0.33063914047949866
+      ],
+      "max": [
+        1.0,
+        1.0,
+        1.0,
+        1.0,
+        1.0,
+        1.0,
+        1.0
+      ],
+      "min": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "q01": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "q99": [
+        1.0,
+        1.0,
+        1.0,
+        1.0,
+        1.0,
+        1.0,
+        1.0
+      ],
+      "mask": [
+        true,
+        true,
+        true,
+        true,
+        true,
+        true,
+        true
+      ]
+    },
+    "state": {
+      "mean": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "std": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "max": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "min": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "q01": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "q99": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ]
+    },
+    "num_transitions": 1472485,
+    "num_trajectories": 2000
+  }
+}

archive/task-transfer/openvla_bridge_cross_mixed_all_024_5env200ep_deadly1000ep_exp1/dataset_statistics_eval.json

@@ -0,0 +1,127 @@
+{
+  "new_embodiment": {
+    "action": {
+      "mean": [
+        0.43306747389336436,
+        0.2607775665819645,
+        0.21196447925952577,
+        0.18366811672846478,
+        0.16187932109460237,
+        0.12526643524567288,
+        0.1235945224761963
+      ],
+      "std": [
+        0.4954584738193594,
+        0.43907057726691345,
+        0.4086948653424082,
+        0.3871752274341832,
+        0.3682965818338933,
+        0.3309705670576461,
+        0.3291175408484206
+      ],
+      "max": [
+        1.0,
+        1.0,
+        1.0,
+        1.0,
+        1.0,
+        1.0,
+        1.0
+      ],
+      "min": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "q01": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "q99": [
+        1.0,
+        1.0,
+        1.0,
+        1.0,
+        1.0,
+        1.0,
+        1.0
+      ],
+      "mask": [
+        true,
+        true,
+        true,
+        true,
+        true,
+        true,
+        true
+      ]
+    },
+    "state": {
+      "mean": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "std": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "max": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "min": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "q01": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "q99": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ]
+    },
+    "num_transitions": 428869,
+    "num_trajectories": 438
+  }
+}

archive/task-transfer/openvla_bridge_cross_mixed_all_024_5env200ep_deadly1000ep_exp1/hydra/.hydra/config.yaml

@@ -0,0 +1,500 @@
+framework:
+  qwenvl:
+    base_vlm: /mnt/local/talha/playground/Pretrained_models/Qwen3-VL-4B-Instruct
+    attn_implementation: flash_attention_2
+    enable_gradient_checkpointing: true
+  action_model:
+    state_dim: 7
+    loss_type: discrete_ce
+    action_horizon: 1
+    future_action_window_size: 0
+    past_action_window_size: 0
+    action_dim: 7
+    action_env_dim: 7
+    state_encoding: discretized_text
+    action_layout: bridge_cross_task_7
+  kv_memory:
+    enabled: false
+    window: 4
+    rollout_len: 8
+    packed_train: false
+    rebased_sink: true
+  name: QwenOFT
+datasets:
+  vla_data:
+    dataset_py: lerobot_datasets
+    include_state: false
+    data_root_dir: /mnt/local/talha/playground/Datasets/rl_games
+    data_mix: cross__asterix_mixed_024_200ep_cross_train__lat0_2_4__200ep__bridge__atlantis_mixed_024_200ep_cross_train__lat0_2_4__200ep__bridge__deadly_corridor_mixed_024_1000ep_cross_train__lat0_2_4__1000ep__bridge__defend_the_line_mixed_024_200ep_cross_train__lat0_2_4__200ep__bridge__demon_attack_mixed_024_200ep_cross_train__lat0_2_4__200ep__bridge__flappy_mixed_024_200ep_cross_train__lat0_2_4__200ep__bridge
+    eval_data_mix: cross__asterix_mixed_024_200ep_cross_train__lat0_2_4__200ep__bridge__atlantis_mixed_024_200ep_cross_train__lat0_2_4__200ep__bridge__deadly_corridor_mixed_024_1000ep_cross_train__lat0_2_4__1000ep__bridge__defend_the_line_mixed_024_200ep_cross_train__lat0_2_4__200ep__bridge__demon_attack_mixed_024_200ep_cross_train__lat0_2_4__200ep__bridge__flappy_mixed_024_200ep_cross_train__lat0_2_4__200ep__bridge__val
+    custom_mixtures_path: /mnt/local/talha/playground/Datasets/rl_games/_generated_mixtures/cross_09b5517c91a12cca.json
+    action_type: discrete
+    sequential_step_sampling: true
+    eval_sequential_step_sampling: null
+    num_workers: 8
+    eval_num_workers: 8
+    prefetch_factor: 4
+    persistent_workers: true
+    pin_memory: true
+    shuffle: true
+    action_balance:
+      enabled: false
+      strategy: balanced_epoch
+      action_key: action_id
+      target_flap_fraction: 0.3
+      noop_id: 0
+      flap_id: 1
+    latency_curriculum:
+      enabled: false
+      strategy: exclusive
+      latencies: null
+      phase_steps: null
+      phase_distributions: null
+      new_latency_passes: 1.0
+      replay_passes: 0.25
+      target_total_passes: 2.0
+      final_equalization: true
+      step_budget_mode: auto
+      eval_at_phase_end: false
+      save_at_phase_end: false
+      computed_plan: null
+    per_device_batch_size: 32
+    load_all_data_for_training: true
+    num_obs_frames: 1
+    image_mode: single
+    prompt_mode: raw
+    stitch_grid:
+    - 2
+    - 2
+    obs_image_size: null
+    video_backend: torchvision_av
+dataset:
+  source_hf: ''
+  config_name: null
+  source_subdir: null
+  converted_name: cross_task_train
+  single_source_hf: ''
+  mixed_source_hf: ''
+  single_converted_name: flappy_train
+  mixed_converted_name: flappy_mixed_latency_train
+  single_latency_filter: null
+  mixed_latency_filter: null
+  force_download: false
+  setup_force: false
+  skip_verification: false
+  target_latency_unit: observation_steps
+  verify_rows: 200
+  max_episodes: null
+  episodes_per_latency: null
+  latency_filter: null
+  debug_subset:
+    enabled: false
+    max_episodes: 5
+    suffix: debug
+base_model:
+  repo_id: Qwen/Qwen3-VL-4B-Instruct
+initialization:
+  checkpoint_local_dir: playground/Pretrained_models/Qwen3VL-OFT-Bridge-RT-1
+  checkpoint_hf_repo_id: StarVLA/Qwen3VL-OFT-Bridge-RT-1
+  checkpoint_filename: checkpoints/steps_5000_pytorch_model.pt
+trainer:
+  max_train_steps: 15000
+  num_warmup_steps: 0
+  save_interval: 15000
+  eval_interval: 15000
+  eval_num_batches: 100
+  per_latency_eval_num_batches: 5
+  eval_action_classification: true
+  eval_action_classification_interval: null
+  cc_f1_tolerance: 1
+  learning_rate:
+    base: 2.0e-05
+    qwen_vl_interface: 1.0e-05
+    action_model: 0.0001
+  lr_scheduler_type: cosine_with_min_lr
+  scheduler_specific_kwargs:
+    min_lr: 1.0e-06
+  freeze_modules: ''
+  freeze_vit: false
+  freeze_tied_embedding: false
+  freeze_llm_layers: []
+  loss_scale:
+    vla: 1.0
+    vlm: 0.1
+  max_grad_norm: 1.0
+  weight_decay: 0.0
+  logging_frequency: 1
+  profile_timing:
+    enabled: false
+    log_interval: 10
+  gradient_clipping: 1.0
+  gradient_accumulation_steps: 8
+  distributed_backend: none
+  is_resume: false
+  pretrained_checkpoint: /mnt/local/talha/results/Checkpoints/openvla_bridge_cross_mixed_all_024_5env200ep_deadly1000ep_exp1/checkpoints/_initialization/StarVLA_Qwen3VL-OFT-Bridge-RT-1/checkpoints/steps_5000_pytorch_model.pt
+  resume_step: 0
+  reload_modules: null
+  optimizer:
+    name: AdamW
+    betas:
+    - 0.9
+    - 0.95
+    eps: 1.0e-08
+    weight_decay: 1.0e-08
+    fused: true
+  save_format: pt
+workspace_dir: /mnt/local/talha
+run_root_dir: /mnt/local/talha/results/Checkpoints
+seed: 42
+wandb_entity: ${oc.env:WANDB_ENTITY}
+wandb_project: ${oc.env:WANDB_PROJECT,starVLA_rl_games}
+auth:
+  env_file: null
+  hf_token_env: HF_TOKEN
+  wandb_api_key_env: WANDB_API_KEY
+paths:
+  run_root_dir: /mnt/local/talha/results/Checkpoints
+  dataset_local_dir: /mnt/local/talha/playground/Datasets/rl_games
+  dataset_cache_dir: null
+  base_model_dir: /mnt/local/talha/playground/Pretrained_models/Qwen3-VL-4B-Instruct
+  accelerate_config: starVLA/config/deepseeds/deepspeed_zero2.yaml
+launch:
+  use_accelerate: true
+  gpus: null
+  num_processes: 1
+  dry_run: false
+conda:
+  enabled: true
+  env_name: null
+rl_games:
+  model_alias: openvla
+  task: cross_task
+  env_eval:
+    image_size: 224
+    frameskip: 1
+    deadly:
+      action_layout: multibinary_7
+      multibinary_threshold: 0.0
+    enabled: true
+    eval_backend: eval_core
+    distributed_mode: none
+    vectorized:
+      enabled: false
+      batch_size: 1
+    latency:
+      prompt_map_path: null
+      mode: mixed
+      values:
+      - 0
+      - 1
+      - 2
+      - 3
+      - 4
+      - 5
+    mid_train:
+      enabled: false
+      interval_steps: 1000
+      latencies:
+      - 0
+      - 1
+      - 2
+      - 3
+      - 4
+      - 5
+      num_episodes: 5
+      max_steps_per_episode: 3600
+    post_train:
+      enabled: true
+      latencies:
+      - 0
+      - 2
+      - 4
+      latencies_from_prompt_map: false
+      num_episodes: 5
+      max_steps_per_episode: 3600
+    asterix:
+      action_layout: factorized_6
+  cross_task:
+    train_tasks:
+    - name: flappy
+      role: A
+      train_source_hf: latency-sensitive-bench/flappy_200ep
+      prompt_source_hf: latency-sensitive-bench/flappy_200ep
+      train_source_subdir: flappy_fix_latency_0_200ep
+      prompt_source_subdir: flappy_fix_latency_0_200ep
+      converted_name: flappy_mixed_024_200ep_cross_train
+      train_latency_filter:
+      - 0
+      - 2
+      - 4
+      eval_latency_filter:
+      - 0
+      - 2
+      - 4
+      episodes_per_latency: null
+      max_episodes: 200
+      weight: 1.0
+    - name: demon_attack
+      role: B
+      train_source_hf: latency-sensitive-bench/demon_attack_200ep
+      prompt_source_hf: latency-sensitive-bench/demon_attack_200ep
+      train_source_subdir: demon_attack_fix_latency_0_200ep
+      prompt_source_subdir: demon_attack_fix_latency_0_200ep
+      converted_name: demon_attack_mixed_024_200ep_cross_train
+      train_latency_filter:
+      - 0
+      - 2
+      - 4
+      eval_latency_filter:
+      - 0
+      - 2
+      - 4
+      episodes_per_latency: null
+      max_episodes: 200
+      weight: 1.0
+    - name: deadly_corridor
+      role: C
+      train_source_hf: latency-sensitive-bench/deadly_1000ep
+      prompt_source_hf: latency-sensitive-bench/deadly_1000ep
+      train_source_subdir: deadly_corridor_fix_latency_0_1000ep
+      prompt_source_subdir: deadly_corridor_fix_latency_0_1000ep
+      converted_name: deadly_corridor_mixed_024_1000ep_cross_train
+      train_latency_filter:
+      - 0
+      - 2
+      - 4
+      eval_latency_filter:
+      - 0
+      - 2
+      - 4
+      episodes_per_latency: null
+      max_episodes: 1000
+      action_layout: multibinary_7
+      weight: 1.0
+    - name: asterix
+      role: D
+      train_source_hf: latency-sensitive-bench/asterix_200ep
+      prompt_source_hf: latency-sensitive-bench/asterix_200ep
+      train_source_subdir: asterix_fixed_latency_0_200ep_7k2steps
+      prompt_source_subdir: asterix_fixed_latency_0_200ep_7k2steps
+      converted_name: asterix_mixed_024_200ep_cross_train
+      train_latency_filter:
+      - 0
+      - 2
+      - 4
+      eval_latency_filter:
+      - 0
+      - 2
+      - 4
+      episodes_per_latency: null
+      max_episodes: 200
+      action_layout: factorized_6
+      weight: 1.0
+    - name: atlantis
+      role: E
+      train_source_hf: latency-sensitive-bench/atlantis_200ep
+      prompt_source_hf: latency-sensitive-bench/atlantis_200ep
+      train_source_subdir: atlantis_fixed_latency_0_200ep_7k2steps
+      prompt_source_subdir: atlantis_fixed_latency_0_200ep_7k2steps
+      converted_name: atlantis_mixed_024_200ep_cross_train
+      train_latency_filter:
+      - 0
+      - 2
+      - 4
+      eval_latency_filter:
+      - 0
+      - 2
+      - 4
+      episodes_per_latency: null
+      max_episodes: 200
+      weight: 1.0
+    - name: defend_the_line
+      role: F
+      train_source_hf: latency-sensitive-bench/defend_the_line_200ep
+      prompt_source_hf: latency-sensitive-bench/defend_the_line_200ep
+      train_source_subdir: defend_the_line_fixed_latency_0_200ep_7k2steps
+      prompt_source_subdir: defend_the_line_fixed_latency_0_200ep_7k2steps
+      converted_name: defend_the_line_mixed_024_200ep_cross_train
+      train_latency_filter:
+      - 0
+      - 2
+      - 4
+      eval_latency_filter:
+      - 0
+      - 2
+      - 4
+      episodes_per_latency: null
+      max_episodes: 200
+      weight: 1.0
+    loss_by_task:
+      flappy: discrete_ce
+      demon_attack: discrete_ce
+      defend_the_line: discrete_ce
+      deadly_corridor: multibinary_bce
+      asterix: asterix_factorized_ce
+      atlantis: discrete_ce
+    loss_weight_by_task:
+      flappy: 1.0
+      demon_attack: 1.0
+      defend_the_line: 1.0
+      deadly_corridor: 1.0
+      asterix: 1.0
+      atlantis: 1.0
+    eval_tasks:
+      flappy:
+        prompt_map_path: /mnt/local/talha/playground/Datasets/rl_games/_prompt_maps/flappy_mixed_024_200ep_cross_train__lat0_2_4__200ep__bridge/eval_latency_prompt_map.json
+        frameskip: 1
+        image_size: 224
+        task_description: 'You are playing Flappy Bird. Pass through the pipe gaps
+          and stay alive. Choose the action: NOOP, FLAP.'
+        mid_train:
+          enabled: false
+          latencies:
+          - 0
+          - 2
+          - 4
+          num_episodes: 5
+          max_steps_per_episode: 3600
+        post_train:
+          enabled: true
+          latencies:
+          - 0
+          - 2
+          - 4
+          num_episodes: 20
+          max_steps_per_episode: 3600
+      demon_attack:
+        prompt_map_path: /mnt/local/talha/playground/Datasets/rl_games/_prompt_maps/demon_attack_mixed_024_200ep_cross_train__lat0_2_4__200ep__bridge/eval_latency_prompt_map.json
+        frameskip: 4
+        image_size: 224
+        task_description: 'You are playing Demon Attack from a single game image.
+          Choose exactly one action from: NOOP, FIRE, RIGHT, LEFT, RIGHTFIRE, LEFTFIRE.'
+        mid_train:
+          enabled: false
+          latencies:
+          - 0
+          - 2
+          - 4
+          num_episodes: 5
+          max_steps_per_episode: 3600
+        post_train:
+          enabled: true
+          latencies:
+          - 0
+          - 2
+          - 4
+          num_episodes: 20
+          max_steps_per_episode: 3600
+      defend_the_line:
+        prompt_map_path: null
+        frameskip: 4
+        image_size: 224
+        task_description: null
+        mid_train:
+          enabled: false
+          latencies: null
+          num_episodes: null
+          max_steps_per_episode: null
+        post_train:
+          enabled: false
+          latencies: null
+          num_episodes: null
+          max_steps_per_episode: null
+      deadly_corridor:
+        prompt_map_path: /mnt/local/talha/playground/Datasets/rl_games/_prompt_maps/deadly_corridor_mixed_024_1000ep_cross_train__lat0_2_4__1000ep__bridge/eval_latency_prompt_map.json
+        frameskip: 4
+        image_size: 224
+        task_description: You are playing Deadly Corridor in VizDoom. Choose actions
+          from MOVE_FORWARD, MOVE_BACKWARD, MOVE_LEFT, MOVE_RIGHT, TURN_LEFT, TURN_RIGHT,
+          ATTACK.
+        mid_train:
+          enabled: false
+          latencies:
+          - 0
+          - 2
+          - 4
+          num_episodes: 5
+          max_steps_per_episode: 3600
+        post_train:
+          enabled: true
+          latencies:
+          - 0
+          - 2
+          - 4
+          num_episodes: 20
+          max_steps_per_episode: 3600
+      asterix:
+        prompt_map_path: null
+        frameskip: 4
+        image_size: 224
+        task_description: null
+        mid_train:
+          enabled: false
+          latencies: null
+          num_episodes: null
+          max_steps_per_episode: null
+        post_train:
+          enabled: false
+          latencies: null
+          num_episodes: null
+          max_steps_per_episode: null
+      atlantis:
+        prompt_map_path: null
+        frameskip: 4
+        image_size: 224
+        task_description: null
+        mid_train:
+          enabled: false
+          latencies: null
+          num_episodes: null
+          max_steps_per_episode: null
+        post_train:
+          enabled: false
+          latencies: null
+          num_episodes: null
+          max_steps_per_episode: null
+  initialization_mode: bridge
+  action_carrier: bridge
+model: openvla
+env: cross_task
+init: bridge
+bridge_base_model:
+  repo_id:
+    openvla: Qwen/Qwen3-VL-4B-Instruct
+    pi0: StarVLA/Qwen2.5-VL-3B-Instruct-Action
+    pi05: Qwen/Qwen3-VL-4B-Instruct
+    gr00t: Qwen/Qwen3-VL-4B-Instruct
+  local_dir:
+    openvla: playground/Pretrained_models/Qwen3-VL-4B-Instruct
+    pi0: playground/Pretrained_models/Qwen2.5-VL-3B-Instruct-Action
+    pi05: playground/Pretrained_models/Qwen3-VL-4B-Instruct
+    gr00t: playground/Pretrained_models/Qwen3-VL-4B-Instruct
+mode: cross_task
+checkpoint:
+  load: auto
+  hf_repo_id: latency-sensitive-bench/openvla_bridge_cross_mixed_all_024_5env200ep_deadly1000ep_exp1
+  save_best_model: false
+  save_final_model: true
+  save_pt_file: false
+  save_training_state: true
+  save_safetensors_file: false
+  local:
+    keep_last_n: 1
+  sync:
+    enabled: true
+    repo_id: latency-sensitive-bench/openvla_bridge_cross_mixed_all_024_5env200ep_deadly1000ep_exp1
+    keep_last_n: 0
+    sync_every_n_checkpoints: 1
+    resume_policy: local_latest
+run_id: openvla_bridge_cross_mixed_all_024_5env200ep_deadly1000ep_exp1
+output_dir: null
+config_yaml: null
+is_debug: false
+version_id: 0.21
+wandb_tags:
+- final_vla_profiling_training
+- demon_attack
+- openvla
+- rtx3090_profile
+training_latency_condition: rtx3090_profile
+wandb_group: final_vla_profiling_training

archive/task-transfer/openvla_bridge_cross_mixed_all_024_5env200ep_deadly1000ep_exp1/hydra/.hydra/hydra.yaml

@@ -0,0 +1,463 @@
+hydra:
+  run:
+    dir: ${run_root_dir}/${run_id}/hydra
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - model=openvla
+    - env=cross_task
+    - init=bridge
+    - mode=cross_task
+    - ++framework.qwenvl.base_vlm=/mnt/local/talha/playground/Pretrained_models/Qwen3-VL-4B-Instruct
+    - ++framework.qwenvl.attn_implementation=flash_attention_2
+    - ++framework.qwenvl.enable_gradient_checkpointing=true
+    - ++framework.action_model.state_dim=7
+    - ++framework.action_model.loss_type=discrete_ce
+    - ++framework.action_model.action_horizon=1
+    - ++framework.action_model.future_action_window_size=0
+    - ++framework.action_model.past_action_window_size=0
+    - ++framework.action_model.action_dim=7
+    - ++framework.action_model.action_env_dim=7
+    - ++framework.action_model.state_encoding=discretized_text
+    - ++framework.action_model.action_layout=bridge_cross_task_7
+    - ++framework.kv_memory.enabled=false
+    - ++framework.kv_memory.window=4
+    - ++framework.kv_memory.rollout_len=8
+    - ++framework.kv_memory.packed_train=false
+    - ++framework.kv_memory.rebased_sink=true
+    - ++framework.name=QwenOFT
+    - ++datasets.vla_data.dataset_py=lerobot_datasets
+    - ++datasets.vla_data.include_state=false
+    - ++datasets.vla_data.data_root_dir=playground/Datasets/rl_games
+    - ++datasets.vla_data.data_mix=cross_task_train
+    - ++datasets.vla_data.eval_data_mix=null
+    - ++datasets.vla_data.custom_mixtures_path=null
+    - ++datasets.vla_data.action_type=discrete
+    - ++datasets.vla_data.sequential_step_sampling=true
+    - ++datasets.vla_data.eval_sequential_step_sampling=null
+    - ++datasets.vla_data.num_workers=8
+    - ++datasets.vla_data.eval_num_workers=8
+    - ++datasets.vla_data.prefetch_factor=4
+    - ++datasets.vla_data.persistent_workers=true
+    - ++datasets.vla_data.pin_memory=true
+    - ++datasets.vla_data.shuffle=true
+    - ++datasets.vla_data.action_balance.enabled=false
+    - ++datasets.vla_data.action_balance.strategy=balanced_epoch
+    - ++datasets.vla_data.action_balance.action_key=action_id
+    - ++datasets.vla_data.action_balance.target_flap_fraction=0.3
+    - ++datasets.vla_data.action_balance.noop_id=0
+    - ++datasets.vla_data.action_balance.flap_id=1
+    - ++datasets.vla_data.latency_curriculum.enabled=false
+    - ++datasets.vla_data.latency_curriculum.strategy=exclusive
+    - ++datasets.vla_data.latency_curriculum.latencies=null
+    - ++datasets.vla_data.latency_curriculum.phase_steps=null
+    - ++datasets.vla_data.latency_curriculum.phase_distributions=null
+    - ++datasets.vla_data.latency_curriculum.new_latency_passes=1.0
+    - ++datasets.vla_data.latency_curriculum.replay_passes=0.25
+    - ++datasets.vla_data.latency_curriculum.target_total_passes=2.0
+    - ++datasets.vla_data.latency_curriculum.final_equalization=true
+    - ++datasets.vla_data.latency_curriculum.step_budget_mode=auto
+    - ++datasets.vla_data.latency_curriculum.eval_at_phase_end=false
+    - ++datasets.vla_data.latency_curriculum.save_at_phase_end=false
+    - ++datasets.vla_data.latency_curriculum.computed_plan=null
+    - ++datasets.vla_data.per_device_batch_size=32
+    - ++datasets.vla_data.load_all_data_for_training=true
+    - ++datasets.vla_data.num_obs_frames=1
+    - ++datasets.vla_data.image_mode=single
+    - ++datasets.vla_data.prompt_mode=raw
+    - ++datasets.vla_data.stitch_grid=[2,2]
+    - ++datasets.vla_data.obs_image_size=null
+    - ++datasets.vla_data.video_backend=torchvision_av
+    - ++dataset.source_hf=
+    - ++dataset.config_name=null
+    - ++dataset.source_subdir=null
+    - ++dataset.converted_name=cross_task_train
+    - ++dataset.single_source_hf=
+    - ++dataset.mixed_source_hf=
+    - ++dataset.single_converted_name=flappy_train
+    - ++dataset.mixed_converted_name=flappy_mixed_latency_train
+    - ++dataset.single_latency_filter=null
+    - ++dataset.mixed_latency_filter=null
+    - ++dataset.force_download=false
+    - ++dataset.setup_force=false
+    - ++dataset.skip_verification=false
+    - ++dataset.target_latency_unit=observation_steps
+    - ++dataset.verify_rows=200
+    - ++dataset.max_episodes=null
+    - ++dataset.episodes_per_latency=null
+    - ++dataset.latency_filter=null
+    - ++dataset.debug_subset.enabled=false
+    - ++dataset.debug_subset.max_episodes=5
+    - ++dataset.debug_subset.suffix=debug
+    - ++base_model.repo_id=Qwen/Qwen3-VL-4B-Instruct
+    - ++initialization.checkpoint_local_dir=playground/Pretrained_models/Qwen3VL-OFT-Bridge-RT-1
+    - ++initialization.checkpoint_hf_repo_id=StarVLA/Qwen3VL-OFT-Bridge-RT-1
+    - ++initialization.checkpoint_filename=checkpoints/steps_5000_pytorch_model.pt
+    - ++trainer.max_train_steps=15000
+    - ++trainer.num_warmup_steps=0
+    - ++trainer.save_interval=15000
+    - ++trainer.eval_interval=15000
+    - ++trainer.eval_num_batches=100
+    - ++trainer.per_latency_eval_num_batches=5
+    - ++trainer.eval_action_classification=true
+    - ++trainer.eval_action_classification_interval=null
+    - ++trainer.cc_f1_tolerance=1
+    - ++trainer.learning_rate.base=2e-05
+    - ++trainer.learning_rate.qwen_vl_interface=1e-05
+    - ++trainer.learning_rate.action_model=0.0001
+    - ++trainer.lr_scheduler_type=cosine_with_min_lr
+    - ++trainer.scheduler_specific_kwargs.min_lr=1e-06
+    - ++trainer.freeze_modules=
+    - ++trainer.freeze_vit=false
+    - ++trainer.freeze_tied_embedding=false
+    - ++trainer.freeze_llm_layers=[]
+    - ++trainer.loss_scale.vla=1.0
+    - ++trainer.loss_scale.vlm=0.1
+    - ++trainer.max_grad_norm=1.0
+    - ++trainer.weight_decay=0.0
+    - ++trainer.logging_frequency=1
+    - ++trainer.profile_timing.enabled=false
+    - ++trainer.profile_timing.log_interval=10
+    - ++trainer.gradient_clipping=1.0
+    - ++trainer.gradient_accumulation_steps=8
+    - ++trainer.distributed_backend=none
+    - ++trainer.is_resume=false
+    - ++trainer.pretrained_checkpoint=null
+    - ++trainer.resume_step=0
+    - ++trainer.reload_modules=null
+    - ++trainer.optimizer.name=AdamW
+    - ++trainer.optimizer.betas=[0.9,0.95]
+    - ++trainer.optimizer.eps=1e-08
+    - ++trainer.optimizer.weight_decay=1e-08
+    - ++trainer.optimizer.fused=true
+    - ++trainer.save_format=pt
+    - ++workspace_dir=/mnt/local/talha
+    - ++run_root_dir=results/Checkpoints
+    - ++seed=42
+    - ++auth.env_file=null
+    - ++auth.hf_token_env=HF_TOKEN
+    - ++auth.wandb_api_key_env=WANDB_API_KEY
+    - ++paths.run_root_dir=/mnt/local/talha/results/Checkpoints
+    - ++paths.dataset_local_dir=/mnt/local/talha/playground/Datasets/rl_games
+    - ++paths.dataset_cache_dir=null
+    - ++paths.base_model_dir=/mnt/local/talha/playground/Pretrained_models/Qwen3-VL-4B-Instruct
+    - ++paths.accelerate_config=starVLA/config/deepseeds/deepspeed_zero2.yaml
+    - ++rl_games.model_alias=openvla
+    - ++rl_games.task=cross_task
+    - ++rl_games.env_eval.image_size=224
+    - ++rl_games.env_eval.frameskip=1
+    - ++rl_games.env_eval.deadly.action_layout=multibinary_7
+    - ++rl_games.env_eval.deadly.multibinary_threshold=0.0
+    - ++rl_games.env_eval.enabled=true
+    - ++rl_games.env_eval.eval_backend=eval_core
+    - ++rl_games.env_eval.distributed_mode=none
+    - ++rl_games.env_eval.vectorized.enabled=false
+    - ++rl_games.env_eval.vectorized.batch_size=1
+    - ++rl_games.env_eval.latency.prompt_map_path=null
+    - ++rl_games.env_eval.latency.mode=mixed
+    - ++rl_games.env_eval.latency.values=[0,1,2,3,4,5]
+    - ++rl_games.env_eval.mid_train.enabled=false
+    - ++rl_games.env_eval.mid_train.interval_steps=1000
+    - ++rl_games.env_eval.mid_train.latencies=[0,1,2,3,4,5]
+    - ++rl_games.env_eval.mid_train.num_episodes=5
+    - ++rl_games.env_eval.mid_train.max_steps_per_episode=3600
+    - ++rl_games.env_eval.post_train.enabled=true
+    - ++rl_games.env_eval.post_train.latencies=[0,2,4]
+    - ++rl_games.env_eval.post_train.latencies_from_prompt_map=false
+    - ++rl_games.env_eval.post_train.num_episodes=5
+    - ++rl_games.env_eval.post_train.max_steps_per_episode=3600
+    - ++rl_games.env_eval.asterix.action_layout=factorized_6
+    - ++rl_games.cross_task.train_tasks=[{name:flappy,role:A,train_source_hf:latency-sensitive-bench/flappy_200ep,prompt_source_hf:latency-sensitive-bench/flappy_200ep,train_source_subdir:flappy_fix_latency_0_200ep,prompt_source_subdir:flappy_fix_latency_0_200ep,converted_name:flappy_mixed_024_200ep_cross_train,train_latency_filter:[0,2,4],eval_latency_filter:[0,2,4],episodes_per_latency:null,max_episodes:200,weight:1.0},{name:demon_attack,role:B,train_source_hf:latency-sensitive-bench/demon_attack_200ep,prompt_source_hf:latency-sensitive-bench/demon_attack_200ep,train_source_subdir:demon_attack_fix_latency_0_200ep,prompt_source_subdir:demon_attack_fix_latency_0_200ep,converted_name:demon_attack_mixed_024_200ep_cross_train,train_latency_filter:[0,2,4],eval_latency_filter:[0,2,4],episodes_per_latency:null,max_episodes:200,weight:1.0},{name:deadly_corridor,role:C,train_source_hf:latency-sensitive-bench/deadly_1000ep,prompt_source_hf:latency-sensitive-bench/deadly_1000ep,train_source_subdir:deadly_corridor_fix_latency_0_1000ep,prompt_source_subdir:deadly_corridor_fix_latency_0_1000ep,converted_name:deadly_corridor_mixed_024_1000ep_cross_train,train_latency_filter:[0,2,4],eval_latency_filter:[0,2,4],episodes_per_latency:null,max_episodes:1000,action_layout:multibinary_7,weight:1.0},{name:asterix,role:D,train_source_hf:latency-sensitive-bench/asterix_200ep,prompt_source_hf:latency-sensitive-bench/asterix_200ep,train_source_subdir:asterix_fixed_latency_0_200ep_7k2steps,prompt_source_subdir:asterix_fixed_latency_0_200ep_7k2steps,converted_name:asterix_mixed_024_200ep_cross_train,train_latency_filter:[0,2,4],eval_latency_filter:[0,2,4],episodes_per_latency:null,max_episodes:200,action_layout:factorized_6,weight:1.0},{name:atlantis,role:E,train_source_hf:latency-sensitive-bench/atlantis_200ep,prompt_source_hf:latency-sensitive-bench/atlantis_200ep,train_source_subdir:atlantis_fixed_latency_0_200ep_7k2steps,prompt_source_subdir:atlantis_fixed_latency_0_200ep_7k2steps,converted_name:atlantis_mixed_024_200ep_cross_train,train_latency_filter:[0,2,4],eval_latency_filter:[0,2,4],episodes_per_latency:null,max_episodes:200,weight:1.0},{name:defend_the_line,role:F,train_source_hf:latency-sensitive-bench/defend_the_line_200ep,prompt_source_hf:latency-sensitive-bench/defend_the_line_200ep,train_source_subdir:defend_the_line_fixed_latency_0_200ep_7k2steps,prompt_source_subdir:defend_the_line_fixed_latency_0_200ep_7k2steps,converted_name:defend_the_line_mixed_024_200ep_cross_train,train_latency_filter:[0,2,4],eval_latency_filter:[0,2,4],episodes_per_latency:null,max_episodes:200,weight:1.0}]
+    - ++rl_games.cross_task.loss_by_task.flappy=discrete_ce
+    - ++rl_games.cross_task.loss_by_task.demon_attack=discrete_ce
+    - ++rl_games.cross_task.loss_by_task.defend_the_line=discrete_ce
+    - ++rl_games.cross_task.loss_by_task.deadly_corridor=multibinary_bce
+    - ++rl_games.cross_task.loss_by_task.asterix=asterix_factorized_ce
+    - ++rl_games.cross_task.loss_by_task.atlantis=discrete_ce
+    - ++rl_games.cross_task.loss_weight_by_task.flappy=1.0
+    - ++rl_games.cross_task.loss_weight_by_task.demon_attack=1.0
+    - ++rl_games.cross_task.loss_weight_by_task.defend_the_line=1.0
+    - ++rl_games.cross_task.loss_weight_by_task.deadly_corridor=1.0
+    - ++rl_games.cross_task.loss_weight_by_task.asterix=1.0
+    - ++rl_games.cross_task.loss_weight_by_task.atlantis=1.0
+    - ++rl_games.cross_task.eval_tasks.flappy.prompt_map_path=null
+    - ++rl_games.cross_task.eval_tasks.flappy.frameskip=1
+    - ++rl_games.cross_task.eval_tasks.flappy.image_size=224
+    - '++rl_games.cross_task.eval_tasks.flappy.task_description=''You are playing
+      Flappy Bird. Pass through the pipe gaps and stay alive. Choose the action: NOOP,
+      FLAP.'''
+    - ++rl_games.cross_task.eval_tasks.flappy.mid_train.enabled=false
+    - ++rl_games.cross_task.eval_tasks.flappy.mid_train.latencies=[0,2,4]
+    - ++rl_games.cross_task.eval_tasks.flappy.mid_train.num_episodes=5
+    - ++rl_games.cross_task.eval_tasks.flappy.mid_train.max_steps_per_episode=3600
+    - ++rl_games.cross_task.eval_tasks.flappy.post_train.enabled=true
+    - ++rl_games.cross_task.eval_tasks.flappy.post_train.latencies=[0,2,4]
+    - ++rl_games.cross_task.eval_tasks.flappy.post_train.num_episodes=20
+    - ++rl_games.cross_task.eval_tasks.flappy.post_train.max_steps_per_episode=3600
+    - ++rl_games.cross_task.eval_tasks.demon_attack.prompt_map_path=null
+    - ++rl_games.cross_task.eval_tasks.demon_attack.frameskip=4
+    - ++rl_games.cross_task.eval_tasks.demon_attack.image_size=224
+    - '++rl_games.cross_task.eval_tasks.demon_attack.task_description=''You are playing
+      Demon Attack from a single game image. Choose exactly one action from: NOOP,
+      FIRE, RIGHT, LEFT, RIGHTFIRE, LEFTFIRE.'''
+    - ++rl_games.cross_task.eval_tasks.demon_attack.mid_train.enabled=false
+    - ++rl_games.cross_task.eval_tasks.demon_attack.mid_train.latencies=[0,2,4]
+    - ++rl_games.cross_task.eval_tasks.demon_attack.mid_train.num_episodes=5
+    - ++rl_games.cross_task.eval_tasks.demon_attack.mid_train.max_steps_per_episode=3600
+    - ++rl_games.cross_task.eval_tasks.demon_attack.post_train.enabled=true
+    - ++rl_games.cross_task.eval_tasks.demon_attack.post_train.latencies=[0,2,4]
+    - ++rl_games.cross_task.eval_tasks.demon_attack.post_train.num_episodes=20
+    - ++rl_games.cross_task.eval_tasks.demon_attack.post_train.max_steps_per_episode=3600
+    - ++rl_games.cross_task.eval_tasks.defend_the_line.prompt_map_path=null
+    - ++rl_games.cross_task.eval_tasks.defend_the_line.frameskip=4
+    - ++rl_games.cross_task.eval_tasks.defend_the_line.image_size=224
+    - ++rl_games.cross_task.eval_tasks.defend_the_line.task_description=null
+    - ++rl_games.cross_task.eval_tasks.defend_the_line.mid_train.enabled=false
+    - ++rl_games.cross_task.eval_tasks.defend_the_line.mid_train.latencies=null
+    - ++rl_games.cross_task.eval_tasks.defend_the_line.mid_train.num_episodes=null
+    - ++rl_games.cross_task.eval_tasks.defend_the_line.mid_train.max_steps_per_episode=null
+    - ++rl_games.cross_task.eval_tasks.defend_the_line.post_train.enabled=false
+    - ++rl_games.cross_task.eval_tasks.defend_the_line.post_train.latencies=null
+    - ++rl_games.cross_task.eval_tasks.defend_the_line.post_train.num_episodes=null
+    - ++rl_games.cross_task.eval_tasks.defend_the_line.post_train.max_steps_per_episode=null
+    - ++rl_games.cross_task.eval_tasks.deadly_corridor.prompt_map_path=null
+    - ++rl_games.cross_task.eval_tasks.deadly_corridor.frameskip=4
+    - ++rl_games.cross_task.eval_tasks.deadly_corridor.image_size=224
+    - ++rl_games.cross_task.eval_tasks.deadly_corridor.task_description='You are playing
+      Deadly Corridor in VizDoom. Choose actions from MOVE_FORWARD, MOVE_BACKWARD,
+      MOVE_LEFT, MOVE_RIGHT, TURN_LEFT, TURN_RIGHT, ATTACK.'
+    - ++rl_games.cross_task.eval_tasks.deadly_corridor.mid_train.enabled=false
+    - ++rl_games.cross_task.eval_tasks.deadly_corridor.mid_train.latencies=[0,2,4]
+    - ++rl_games.cross_task.eval_tasks.deadly_corridor.mid_train.num_episodes=5
+    - ++rl_games.cross_task.eval_tasks.deadly_corridor.mid_train.max_steps_per_episode=3600
+    - ++rl_games.cross_task.eval_tasks.deadly_corridor.post_train.enabled=true
+    - ++rl_games.cross_task.eval_tasks.deadly_corridor.post_train.latencies=[0,2,4]
+    - ++rl_games.cross_task.eval_tasks.deadly_corridor.post_train.num_episodes=20
+    - ++rl_games.cross_task.eval_tasks.deadly_corridor.post_train.max_steps_per_episode=3600
+    - ++rl_games.cross_task.eval_tasks.asterix.prompt_map_path=null
+    - ++rl_games.cross_task.eval_tasks.asterix.frameskip=4
+    - ++rl_games.cross_task.eval_tasks.asterix.image_size=224
+    - ++rl_games.cross_task.eval_tasks.asterix.task_description=null
+    - ++rl_games.cross_task.eval_tasks.asterix.mid_train.enabled=false
+    - ++rl_games.cross_task.eval_tasks.asterix.mid_train.latencies=null
+    - ++rl_games.cross_task.eval_tasks.asterix.mid_train.num_episodes=null
+    - ++rl_games.cross_task.eval_tasks.asterix.mid_train.max_steps_per_episode=null
+    - ++rl_games.cross_task.eval_tasks.asterix.post_train.enabled=false
+    - ++rl_games.cross_task.eval_tasks.asterix.post_train.latencies=null
+    - ++rl_games.cross_task.eval_tasks.asterix.post_train.num_episodes=null
+    - ++rl_games.cross_task.eval_tasks.asterix.post_train.max_steps_per_episode=null
+    - ++rl_games.cross_task.eval_tasks.atlantis.prompt_map_path=null
+    - ++rl_games.cross_task.eval_tasks.atlantis.frameskip=4
+    - ++rl_games.cross_task.eval_tasks.atlantis.image_size=224
+    - ++rl_games.cross_task.eval_tasks.atlantis.task_description=null
+    - ++rl_games.cross_task.eval_tasks.atlantis.mid_train.enabled=false
+    - ++rl_games.cross_task.eval_tasks.atlantis.mid_train.latencies=null
+    - ++rl_games.cross_task.eval_tasks.atlantis.mid_train.num_episodes=null
+    - ++rl_games.cross_task.eval_tasks.atlantis.mid_train.max_steps_per_episode=null
+    - ++rl_games.cross_task.eval_tasks.atlantis.post_train.enabled=false
+    - ++rl_games.cross_task.eval_tasks.atlantis.post_train.latencies=null
+    - ++rl_games.cross_task.eval_tasks.atlantis.post_train.num_episodes=null
+    - ++rl_games.cross_task.eval_tasks.atlantis.post_train.max_steps_per_episode=null
+    - ++rl_games.initialization_mode=bridge
+    - ++rl_games.action_carrier=bridge
+    - ++bridge_base_model.repo_id.openvla=Qwen/Qwen3-VL-4B-Instruct
+    - ++bridge_base_model.repo_id.pi0=StarVLA/Qwen2.5-VL-3B-Instruct-Action
+    - ++bridge_base_model.repo_id.pi05=Qwen/Qwen3-VL-4B-Instruct
+    - ++bridge_base_model.repo_id.gr00t=Qwen/Qwen3-VL-4B-Instruct
+    - ++bridge_base_model.local_dir.openvla=playground/Pretrained_models/Qwen3-VL-4B-Instruct
+    - ++bridge_base_model.local_dir.pi0=playground/Pretrained_models/Qwen2.5-VL-3B-Instruct-Action
+    - ++bridge_base_model.local_dir.pi05=playground/Pretrained_models/Qwen3-VL-4B-Instruct
+    - ++bridge_base_model.local_dir.gr00t=playground/Pretrained_models/Qwen3-VL-4B-Instruct
+    - ++checkpoint.load=auto
+    - ++checkpoint.hf_repo_id=latency-sensitive-bench/openvla_bridge_cross_mixed_all_024_5env200ep_deadly1000ep_exp1
+    - ++checkpoint.save_best_model=false
+    - ++checkpoint.save_final_model=true
+    - ++checkpoint.save_pt_file=false
+    - ++checkpoint.save_training_state=true
+    - ++checkpoint.save_safetensors_file=false
+    - ++checkpoint.local.keep_last_n=1
+    - ++checkpoint.sync.enabled=true
+    - ++checkpoint.sync.repo_id=latency-sensitive-bench/openvla_bridge_cross_mixed_all_024_5env200ep_deadly1000ep_exp1
+    - ++checkpoint.sync.keep_last_n=0
+    - ++checkpoint.sync.sync_every_n_checkpoints=1
+    - ++checkpoint.sync.resume_policy=local_latest
+    - ++run_id=openvla_bridge_cross_mixed_all_024_5env200ep_deadly1000ep_exp1
+    - ++output_dir=null
+    - ++config_yaml=null
+    - ++is_debug=false
+    - ++version_id=0.21
+    - ++wandb_tags=[final_vla_profiling_training,demon_attack,openvla,rtx3090_profile]
+    - ++training_latency_condition=rtx3090_profile
+    - ++wandb_group=final_vla_profiling_training
+    - ++run_root_dir=/mnt/local/talha/results/Checkpoints
+    - ++trainer.is_resume=false
+    - ++trainer.pretrained_checkpoint=/mnt/local/talha/results/Checkpoints/openvla_bridge_cross_mixed_all_024_5env200ep_deadly1000ep_exp1/checkpoints/_initialization/StarVLA_Qwen3VL-OFT-Bridge-RT-1/checkpoints/steps_5000_pytorch_model.pt
+    - ++trainer.resume_step=0
+    - ++datasets.vla_data.data_root_dir=/mnt/local/talha/playground/Datasets/rl_games
+    - ++datasets.vla_data.data_mix=cross__asterix_mixed_024_200ep_cross_train__lat0_2_4__200ep__bridge__atlantis_mixed_024_200ep_cross_train__lat0_2_4__200ep__bridge__deadly_corridor_mixed_024_1000ep_cross_train__lat0_2_4__1000ep__bridge__defend_the_line_mixed_024_200ep_cross_train__lat0_2_4__200ep__bridge__demon_attack_mixed_024_200ep_cross_train__lat0_2_4__200ep__bridge__flappy_mixed_024_200ep_cross_train__lat0_2_4__200ep__bridge
+    - ++datasets.vla_data.eval_data_mix=cross__asterix_mixed_024_200ep_cross_train__lat0_2_4__200ep__bridge__atlantis_mixed_024_200ep_cross_train__lat0_2_4__200ep__bridge__deadly_corridor_mixed_024_1000ep_cross_train__lat0_2_4__1000ep__bridge__defend_the_line_mixed_024_200ep_cross_train__lat0_2_4__200ep__bridge__demon_attack_mixed_024_200ep_cross_train__lat0_2_4__200ep__bridge__flappy_mixed_024_200ep_cross_train__lat0_2_4__200ep__bridge__val
+    - ++datasets.vla_data.custom_mixtures_path=/mnt/local/talha/playground/Datasets/rl_games/_generated_mixtures/cross_09b5517c91a12cca.json
+    - ++framework.qwenvl.base_vlm=/mnt/local/talha/playground/Pretrained_models/Qwen3-VL-4B-Instruct
+    - ++rl_games.cross_task.eval_tasks.flappy.prompt_map_path=/mnt/local/talha/playground/Datasets/rl_games/_prompt_maps/flappy_mixed_024_200ep_cross_train__lat0_2_4__200ep__bridge/eval_latency_prompt_map.json
+    - ++rl_games.cross_task.eval_tasks.demon_attack.prompt_map_path=/mnt/local/talha/playground/Datasets/rl_games/_prompt_maps/demon_attack_mixed_024_200ep_cross_train__lat0_2_4__200ep__bridge/eval_latency_prompt_map.json
+    - ++rl_games.cross_task.eval_tasks.deadly_corridor.prompt_map_path=/mnt/local/talha/playground/Datasets/rl_games/_prompt_maps/deadly_corridor_mixed_024_1000ep_cross_train__lat0_2_4__1000ep__bridge/eval_latency_prompt_map.json
+  job:
+    name: train_starvla_hydra
+    chdir: false
+    override_dirname: '++auth.env_file=null,++auth.hf_token_env=HF_TOKEN,++auth.wandb_api_key_env=WANDB_API_KEY,++base_model.repo_id=Qwen/Qwen3-VL-4B-Instruct,++bridge_base_model.local_dir.gr00t=playground/Pretrained_models/Qwen3-VL-4B-Instruct,++bridge_base_model.local_dir.openvla=playground/Pretrained_models/Qwen3-VL-4B-Instruct,++bridge_base_model.local_dir.pi05=playground/Pretrained_models/Qwen3-VL-4B-Instruct,++bridge_base_model.local_dir.pi0=playground/Pretrained_models/Qwen2.5-VL-3B-Instruct-Action,++bridge_base_model.repo_id.gr00t=Qwen/Qwen3-VL-4B-Instruct,++bridge_base_model.repo_id.openvla=Qwen/Qwen3-VL-4B-Instruct,++bridge_base_model.repo_id.pi05=Qwen/Qwen3-VL-4B-Instruct,++bridge_base_model.repo_id.pi0=StarVLA/Qwen2.5-VL-3B-Instruct-Action,++checkpoint.hf_repo_id=latency-sensitive-bench/openvla_bridge_cross_mixed_all_024_5env200ep_deadly1000ep_exp1,++checkpoint.load=auto,++checkpoint.local.keep_last_n=1,++checkpoint.save_best_model=false,++checkpoint.save_final_model=true,++checkpoint.save_pt_file=false,++checkpoint.save_safetensors_file=false,++checkpoint.save_training_state=true,++checkpoint.sync.enabled=true,++checkpoint.sync.keep_last_n=0,++checkpoint.sync.repo_id=latency-sensitive-bench/openvla_bridge_cross_mixed_all_024_5env200ep_deadly1000ep_exp1,++checkpoint.sync.resume_policy=local_latest,++checkpoint.sync.sync_every_n_checkpoints=1,++config_yaml=null,++dataset.config_name=null,++dataset.converted_name=cross_task_train,++dataset.debug_subset.enabled=false,++dataset.debug_subset.max_episodes=5,++dataset.debug_subset.suffix=debug,++dataset.episodes_per_latency=null,++dataset.force_download=false,++dataset.latency_filter=null,++dataset.max_episodes=null,++dataset.mixed_converted_name=flappy_mixed_latency_train,++dataset.mixed_latency_filter=null,++dataset.mixed_source_hf=,++dataset.setup_force=false,++dataset.single_converted_name=flappy_train,++dataset.single_latency_filter=null,++dataset.single_source_hf=,++dataset.skip_verification=false,++dataset.source_hf=,++dataset.source_subdir=null,++dataset.target_latency_unit=observation_steps,++dataset.verify_rows=200,++datasets.vla_data.action_balance.action_key=action_id,++datasets.vla_data.action_balance.enabled=false,++datasets.vla_data.action_balance.flap_id=1,++datasets.vla_data.action_balance.noop_id=0,++datasets.vla_data.action_balance.strategy=balanced_epoch,++datasets.vla_data.action_balance.target_flap_fraction=0.3,++datasets.vla_data.action_type=discrete,++datasets.vla_data.custom_mixtures_path=/mnt/local/talha/playground/Datasets/rl_games/_generated_mixtures/cross_09b5517c91a12cca.json,++datasets.vla_data.custom_mixtures_path=null,++datasets.vla_data.data_mix=cross__asterix_mixed_024_200ep_cross_train__lat0_2_4__200ep__bridge__atlantis_mixed_024_200ep_cross_train__lat0_2_4__200ep__bridge__deadly_corridor_mixed_024_1000ep_cross_train__lat0_2_4__1000ep__bridge__defend_the_line_mixed_024_200ep_cross_train__lat0_2_4__200ep__bridge__demon_attack_mixed_024_200ep_cross_train__lat0_2_4__200ep__bridge__flappy_mixed_024_200ep_cross_train__lat0_2_4__200ep__bridge,++datasets.vla_data.data_mix=cross_task_train,++datasets.vla_data.data_root_dir=/mnt/local/talha/playground/Datasets/rl_games,++datasets.vla_data.data_root_dir=playground/Datasets/rl_games,++datasets.vla_data.dataset_py=lerobot_datasets,++datasets.vla_data.eval_data_mix=cross__asterix_mixed_024_200ep_cross_train__lat0_2_4__200ep__bridge__atlantis_mixed_024_200ep_cross_train__lat0_2_4__200ep__bridge__deadly_corridor_mixed_024_1000ep_cross_train__lat0_2_4__1000ep__bridge__defend_the_line_mixed_024_200ep_cross_train__lat0_2_4__200ep__bridge__demon_attack_mixed_024_200ep_cross_train__lat0_2_4__200ep__bridge__flappy_mixed_024_200ep_cross_train__lat0_2_4__200ep__bridge__val,++datasets.vla_data.eval_data_mix=null,++datasets.vla_data.eval_num_workers=8,++datasets.vla_data.eval_sequential_step_sampling=null,++datasets.vla_data.image_mode=single,++datasets.vla_data.include_state=false,++datasets.vla_data.latency_curriculum.computed_plan=null,++datasets.vla_data.latency_curriculum.enabled=false,++datasets.vla_data.latency_curriculum.eval_at_phase_end=false,++datasets.vla_data.latency_curriculum.final_equalization=true,++datasets.vla_data.latency_curriculum.latencies=null,++datasets.vla_data.latency_curriculum.new_latency_passes=1.0,++datasets.vla_data.latency_curriculum.phase_distributions=null,++datasets.vla_data.latency_curriculum.phase_steps=null,++datasets.vla_data.latency_curriculum.replay_passes=0.25,++datasets.vla_data.latency_curriculum.save_at_phase_end=false,++datasets.vla_data.latency_curriculum.step_budget_mode=auto,++datasets.vla_data.latency_curriculum.strategy=exclusive,++datasets.vla_data.latency_curriculum.target_total_passes=2.0,++datasets.vla_data.load_all_data_for_training=true,++datasets.vla_data.num_obs_frames=1,++datasets.vla_data.num_workers=8,++datasets.vla_data.obs_image_size=null,++datasets.vla_data.per_device_batch_size=32,++datasets.vla_data.persistent_workers=true,++datasets.vla_data.pin_memory=true,++datasets.vla_data.prefetch_factor=4,++datasets.vla_data.prompt_mode=raw,++datasets.vla_data.sequential_step_sampling=true,++datasets.vla_data.shuffle=true,++datasets.vla_data.stitch_grid=[2,2],++datasets.vla_data.video_backend=torchvision_av,++framework.action_model.action_dim=7,++framework.action_model.action_env_dim=7,++framework.action_model.action_horizon=1,++framework.action_model.action_layout=bridge_cross_task_7,++framework.action_model.future_action_window_size=0,++framework.action_model.loss_type=discrete_ce,++framework.action_model.past_action_window_size=0,++framework.action_model.state_dim=7,++framework.action_model.state_encoding=discretized_text,++framework.kv_memory.enabled=false,++framework.kv_memory.packed_train=false,++framework.kv_memory.rebased_sink=true,++framework.kv_memory.rollout_len=8,++framework.kv_memory.window=4,++framework.name=QwenOFT,++framework.qwenvl.attn_implementation=flash_attention_2,++framework.qwenvl.base_vlm=/mnt/local/talha/playground/Pretrained_models/Qwen3-VL-4B-Instruct,++framework.qwenvl.base_vlm=/mnt/local/talha/playground/Pretrained_models/Qwen3-VL-4B-Instruct,++framework.qwenvl.enable_gradient_checkpointing=true,++initialization.checkpoint_filename=checkpoints/steps_5000_pytorch_model.pt,++initialization.checkpoint_hf_repo_id=StarVLA/Qwen3VL-OFT-Bridge-RT-1,++initialization.checkpoint_local_dir=playground/Pretrained_models/Qwen3VL-OFT-Bridge-RT-1,++is_debug=false,++output_dir=null,++paths.accelerate_config=starVLA/config/deepseeds/deepspeed_zero2.yaml,++paths.base_model_dir=/mnt/local/talha/playground/Pretrained_models/Qwen3-VL-4B-Instruct,++paths.dataset_cache_dir=null,++paths.dataset_local_dir=/mnt/local/talha/playground/Datasets/rl_games,++paths.run_root_dir=/mnt/local/talha/results/Checkpoints,++rl_games.action_carrier=bridge,++rl_games.cross_task.eval_tasks.asterix.frameskip=4,++rl_games.cross_task.eval_tasks.asterix.image_size=224,++rl_games.cross_task.eval_tasks.asterix.mid_train.enabled=false,++rl_games.cross_task.eval_tasks.asterix.mid_train.latencies=null,++rl_games.cross_task.eval_tasks.asterix.mid_train.max_steps_per_episode=null,++rl_games.cross_task.eval_tasks.asterix.mid_train.num_episodes=null,++rl_games.cross_task.eval_tasks.asterix.post_train.enabled=false,++rl_games.cross_task.eval_tasks.asterix.post_train.latencies=null,++rl_games.cross_task.eval_tasks.asterix.post_train.max_steps_per_episode=null,++rl_games.cross_task.eval_tasks.asterix.post_train.num_episodes=null,++rl_games.cross_task.eval_tasks.asterix.prompt_map_path=null,++rl_games.cross_task.eval_tasks.asterix.task_description=null,++rl_games.cross_task.eval_tasks.atlantis.frameskip=4,++rl_games.cross_task.eval_tasks.atlantis.image_size=224,++rl_games.cross_task.eval_tasks.atlantis.mid_train.enabled=false,++rl_games.cross_task.eval_tasks.atlantis.mid_train.latencies=null,++rl_games.cross_task.eval_tasks.atlantis.mid_train.max_steps_per_episode=null,++rl_games.cross_task.eval_tasks.atlantis.mid_train.num_episodes=null,++rl_games.cross_task.eval_tasks.atlantis.post_train.enabled=false,++rl_games.cross_task.eval_tasks.atlantis.post_train.latencies=null,++rl_games.cross_task.eval_tasks.atlantis.post_train.max_steps_per_episode=null,++rl_games.cross_task.eval_tasks.atlantis.post_train.num_episodes=null,++rl_games.cross_task.eval_tasks.atlantis.prompt_map_path=null,++rl_games.cross_task.eval_tasks.atlantis.task_description=null,++rl_games.cross_task.eval_tasks.deadly_corridor.frameskip=4,++rl_games.cross_task.eval_tasks.deadly_corridor.image_size=224,++rl_games.cross_task.eval_tasks.deadly_corridor.mid_train.enabled=false,++rl_games.cross_task.eval_tasks.deadly_corridor.mid_train.latencies=[0,2,4],++rl_games.cross_task.eval_tasks.deadly_corridor.mid_train.max_steps_per_episode=3600,++rl_games.cross_task.eval_tasks.deadly_corridor.mid_train.num_episodes=5,++rl_games.cross_task.eval_tasks.deadly_corridor.post_train.enabled=true,++rl_games.cross_task.eval_tasks.deadly_corridor.post_train.latencies=[0,2,4],++rl_games.cross_task.eval_tasks.deadly_corridor.post_train.max_steps_per_episode=3600,++rl_games.cross_task.eval_tasks.deadly_corridor.post_train.num_episodes=20,++rl_games.cross_task.eval_tasks.deadly_corridor.prompt_map_path=/mnt/local/talha/playground/Datasets/rl_games/_prompt_maps/deadly_corridor_mixed_024_1000ep_cross_train__lat0_2_4__1000ep__bridge/eval_latency_prompt_map.json,++rl_games.cross_task.eval_tasks.deadly_corridor.prompt_map_path=null,++rl_games.cross_task.eval_tasks.deadly_corridor.task_description=''You
+      are playing Deadly Corridor in VizDoom. Choose actions from MOVE_FORWARD, MOVE_BACKWARD,
+      MOVE_LEFT, MOVE_RIGHT, TURN_LEFT, TURN_RIGHT, ATTACK.'',++rl_games.cross_task.eval_tasks.defend_the_line.frameskip=4,++rl_games.cross_task.eval_tasks.defend_the_line.image_size=224,++rl_games.cross_task.eval_tasks.defend_the_line.mid_train.enabled=false,++rl_games.cross_task.eval_tasks.defend_the_line.mid_train.latencies=null,++rl_games.cross_task.eval_tasks.defend_the_line.mid_train.max_steps_per_episode=null,++rl_games.cross_task.eval_tasks.defend_the_line.mid_train.num_episodes=null,++rl_games.cross_task.eval_tasks.defend_the_line.post_train.enabled=false,++rl_games.cross_task.eval_tasks.defend_the_line.post_train.latencies=null,++rl_games.cross_task.eval_tasks.defend_the_line.post_train.max_steps_per_episode=null,++rl_games.cross_task.eval_tasks.defend_the_line.post_train.num_episodes=null,++rl_games.cross_task.eval_tasks.defend_the_line.prompt_map_path=null,++rl_games.cross_task.eval_tasks.defend_the_line.task_description=null,++rl_games.cross_task.eval_tasks.demon_attack.frameskip=4,++rl_games.cross_task.eval_tasks.demon_attack.image_size=224,++rl_games.cross_task.eval_tasks.demon_attack.mid_train.enabled=false,++rl_games.cross_task.eval_tasks.demon_attack.mid_train.latencies=[0,2,4],++rl_games.cross_task.eval_tasks.demon_attack.mid_train.max_steps_per_episode=3600,++rl_games.cross_task.eval_tasks.demon_attack.mid_train.num_episodes=5,++rl_games.cross_task.eval_tasks.demon_attack.post_train.enabled=true,++rl_games.cross_task.eval_tasks.demon_attack.post_train.latencies=[0,2,4],++rl_games.cross_task.eval_tasks.demon_attack.post_train.max_steps_per_episode=3600,++rl_games.cross_task.eval_tasks.demon_attack.post_train.num_episodes=20,++rl_games.cross_task.eval_tasks.demon_attack.prompt_map_path=/mnt/local/talha/playground/Datasets/rl_games/_prompt_maps/demon_attack_mixed_024_200ep_cross_train__lat0_2_4__200ep__bridge/eval_latency_prompt_map.json,++rl_games.cross_task.eval_tasks.demon_attack.prompt_map_path=null,++rl_games.cross_task.eval_tasks.demon_attack.task_description=''You
+      are playing Demon Attack from a single game image. Choose exactly one action
+      from: NOOP, FIRE, RIGHT, LEFT, RIGHTFIRE, LEFTFIRE.'',++rl_games.cross_task.eval_tasks.flappy.frameskip=1,++rl_games.cross_task.eval_tasks.flappy.image_size=224,++rl_games.cross_task.eval_tasks.flappy.mid_train.enabled=false,++rl_games.cross_task.eval_tasks.flappy.mid_train.latencies=[0,2,4],++rl_games.cross_task.eval_tasks.flappy.mid_train.max_steps_per_episode=3600,++rl_games.cross_task.eval_tasks.flappy.mid_train.num_episodes=5,++rl_games.cross_task.eval_tasks.flappy.post_train.enabled=true,++rl_games.cross_task.eval_tasks.flappy.post_train.latencies=[0,2,4],++rl_games.cross_task.eval_tasks.flappy.post_train.max_steps_per_episode=3600,++rl_games.cross_task.eval_tasks.flappy.post_train.num_episodes=20,++rl_games.cross_task.eval_tasks.flappy.prompt_map_path=/mnt/local/talha/playground/Datasets/rl_games/_prompt_maps/flappy_mixed_024_200ep_cross_train__lat0_2_4__200ep__bridge/eval_latency_prompt_map.json,++rl_games.cross_task.eval_tasks.flappy.prompt_map_path=null,++rl_games.cross_task.eval_tasks.flappy.task_description=''You
+      are playing Flappy Bird. Pass through the pipe gaps and stay alive. Choose the
+      action: NOOP, FLAP.'',++rl_games.cross_task.loss_by_task.asterix=asterix_factorized_ce,++rl_games.cross_task.loss_by_task.atlantis=discrete_ce,++rl_games.cross_task.loss_by_task.deadly_corridor=multibinary_bce,++rl_games.cross_task.loss_by_task.defend_the_line=discrete_ce,++rl_games.cross_task.loss_by_task.demon_attack=discrete_ce,++rl_games.cross_task.loss_by_task.flappy=discrete_ce,++rl_games.cross_task.loss_weight_by_task.asterix=1.0,++rl_games.cross_task.loss_weight_by_task.atlantis=1.0,++rl_games.cross_task.loss_weight_by_task.deadly_corridor=1.0,++rl_games.cross_task.loss_weight_by_task.defend_the_line=1.0,++rl_games.cross_task.loss_weight_by_task.demon_attack=1.0,++rl_games.cross_task.loss_weight_by_task.flappy=1.0,++rl_games.cross_task.train_tasks=[{name:flappy,role:A,train_source_hf:latency-sensitive-bench/flappy_200ep,prompt_source_hf:latency-sensitive-bench/flappy_200ep,train_source_subdir:flappy_fix_latency_0_200ep,prompt_source_subdir:flappy_fix_latency_0_200ep,converted_name:flappy_mixed_024_200ep_cross_train,train_latency_filter:[0,2,4],eval_latency_filter:[0,2,4],episodes_per_latency:null,max_episodes:200,weight:1.0},{name:demon_attack,role:B,train_source_hf:latency-sensitive-bench/demon_attack_200ep,prompt_source_hf:latency-sensitive-bench/demon_attack_200ep,train_source_subdir:demon_attack_fix_latency_0_200ep,prompt_source_subdir:demon_attack_fix_latency_0_200ep,converted_name:demon_attack_mixed_024_200ep_cross_train,train_latency_filter:[0,2,4],eval_latency_filter:[0,2,4],episodes_per_latency:null,max_episodes:200,weight:1.0},{name:deadly_corridor,role:C,train_source_hf:latency-sensitive-bench/deadly_1000ep,prompt_source_hf:latency-sensitive-bench/deadly_1000ep,train_source_subdir:deadly_corridor_fix_latency_0_1000ep,prompt_source_subdir:deadly_corridor_fix_latency_0_1000ep,converted_name:deadly_corridor_mixed_024_1000ep_cross_train,train_latency_filter:[0,2,4],eval_latency_filter:[0,2,4],episodes_per_latency:null,max_episodes:1000,action_layout:multibinary_7,weight:1.0},{name:asterix,role:D,train_source_hf:latency-sensitive-bench/asterix_200ep,prompt_source_hf:latency-sensitive-bench/asterix_200ep,train_source_subdir:asterix_fixed_latency_0_200ep_7k2steps,prompt_source_subdir:asterix_fixed_latency_0_200ep_7k2steps,converted_name:asterix_mixed_024_200ep_cross_train,train_latency_filter:[0,2,4],eval_latency_filter:[0,2,4],episodes_per_latency:null,max_episodes:200,action_layout:factorized_6,weight:1.0},{name:atlantis,role:E,train_source_hf:latency-sensitive-bench/atlantis_200ep,prompt_source_hf:latency-sensitive-bench/atlantis_200ep,train_source_subdir:atlantis_fixed_latency_0_200ep_7k2steps,prompt_source_subdir:atlantis_fixed_latency_0_200ep_7k2steps,converted_name:atlantis_mixed_024_200ep_cross_train,train_latency_filter:[0,2,4],eval_latency_filter:[0,2,4],episodes_per_latency:null,max_episodes:200,weight:1.0},{name:defend_the_line,role:F,train_source_hf:latency-sensitive-bench/defend_the_line_200ep,prompt_source_hf:latency-sensitive-bench/defend_the_line_200ep,train_source_subdir:defend_the_line_fixed_latency_0_200ep_7k2steps,prompt_source_subdir:defend_the_line_fixed_latency_0_200ep_7k2steps,converted_name:defend_the_line_mixed_024_200ep_cross_train,train_latency_filter:[0,2,4],eval_latency_filter:[0,2,4],episodes_per_latency:null,max_episodes:200,weight:1.0}],++rl_games.env_eval.asterix.action_layout=factorized_6,++rl_games.env_eval.deadly.action_layout=multibinary_7,++rl_games.env_eval.deadly.multibinary_threshold=0.0,++rl_games.env_eval.distributed_mode=none,++rl_games.env_eval.enabled=true,++rl_games.env_eval.eval_backend=eval_core,++rl_games.env_eval.frameskip=1,++rl_games.env_eval.image_size=224,++rl_games.env_eval.latency.mode=mixed,++rl_games.env_eval.latency.prompt_map_path=null,++rl_games.env_eval.latency.values=[0,1,2,3,4,5],++rl_games.env_eval.mid_train.enabled=false,++rl_games.env_eval.mid_train.interval_steps=1000,++rl_games.env_eval.mid_train.latencies=[0,1,2,3,4,5],++rl_games.env_eval.mid_train.max_steps_per_episode=3600,++rl_games.env_eval.mid_train.num_episodes=5,++rl_games.env_eval.post_train.enabled=true,++rl_games.env_eval.post_train.latencies=[0,2,4],++rl_games.env_eval.post_train.latencies_from_prompt_map=false,++rl_games.env_eval.post_train.max_steps_per_episode=3600,++rl_games.env_eval.post_train.num_episodes=5,++rl_games.env_eval.vectorized.batch_size=1,++rl_games.env_eval.vectorized.enabled=false,++rl_games.initialization_mode=bridge,++rl_games.model_alias=openvla,++rl_games.task=cross_task,++run_id=openvla_bridge_cross_mixed_all_024_5env200ep_deadly1000ep_exp1,++run_root_dir=/mnt/local/talha/results/Checkpoints,++run_root_dir=results/Checkpoints,++seed=42,++trainer.cc_f1_tolerance=1,++trainer.distributed_backend=none,++trainer.eval_action_classification=true,++trainer.eval_action_classification_interval=null,++trainer.eval_interval=15000,++trainer.eval_num_batches=100,++trainer.freeze_llm_layers=[],++trainer.freeze_modules=,++trainer.freeze_tied_embedding=false,++trainer.freeze_vit=false,++trainer.gradient_accumulation_steps=8,++trainer.gradient_clipping=1.0,++trainer.is_resume=false,++trainer.is_resume=false,++trainer.learning_rate.action_model=0.0001,++trainer.learning_rate.base=2e-05,++trainer.learning_rate.qwen_vl_interface=1e-05,++trainer.logging_frequency=1,++trainer.loss_scale.vla=1.0,++trainer.loss_scale.vlm=0.1,++trainer.lr_scheduler_type=cosine_with_min_lr,++trainer.max_grad_norm=1.0,++trainer.max_train_steps=15000,++trainer.num_warmup_steps=0,++trainer.optimizer.betas=[0.9,0.95],++trainer.optimizer.eps=1e-08,++trainer.optimizer.fused=true,++trainer.optimizer.name=AdamW,++trainer.optimizer.weight_decay=1e-08,++trainer.per_latency_eval_num_batches=5,++trainer.pretrained_checkpoint=/mnt/local/talha/results/Checkpoints/openvla_bridge_cross_mixed_all_024_5env200ep_deadly1000ep_exp1/checkpoints/_initialization/StarVLA_Qwen3VL-OFT-Bridge-RT-1/checkpoints/steps_5000_pytorch_model.pt,++trainer.pretrained_checkpoint=null,++trainer.profile_timing.enabled=false,++trainer.profile_timing.log_interval=10,++trainer.reload_modules=null,++trainer.resume_step=0,++trainer.resume_step=0,++trainer.save_format=pt,++trainer.save_interval=15000,++trainer.scheduler_specific_kwargs.min_lr=1e-06,++trainer.weight_decay=0.0,++training_latency_condition=rtx3090_profile,++version_id=0.21,++wandb_group=final_vla_profiling_training,++wandb_tags=[final_vla_profiling_training,demon_attack,openvla,rtx3090_profile],++workspace_dir=/mnt/local/talha,env=cross_task,init=bridge,mode=cross_task,model=openvla'
+    id: ???
+    num: ???
+    config_name: train
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.7
+    version_base: '1.1'
+    cwd: /home/ubuntu/talha/starVLA
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /home/ubuntu/talha/starVLA/examples/rl_games/config
+      schema: file
+      provider: main
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/local/talha/results/Checkpoints/openvla_bridge_cross_mixed_all_024_5env200ep_deadly1000ep_exp1/hydra
+    choices:
+      cross_task_setup: null
+      checkpoint: default
+      mode: cross_task
+      init: bridge
+      env: cross_task
+      model: openvla
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false

archive/task-transfer/openvla_bridge_cross_mixed_all_024_5env200ep_deadly1000ep_exp1/hydra/.hydra/overrides.yaml

@@ -0,0 +1,296 @@
+- model=openvla
+- env=cross_task
+- init=bridge
+- mode=cross_task
+- ++framework.qwenvl.base_vlm=/mnt/local/talha/playground/Pretrained_models/Qwen3-VL-4B-Instruct
+- ++framework.qwenvl.attn_implementation=flash_attention_2
+- ++framework.qwenvl.enable_gradient_checkpointing=true
+- ++framework.action_model.state_dim=7
+- ++framework.action_model.loss_type=discrete_ce
+- ++framework.action_model.action_horizon=1
+- ++framework.action_model.future_action_window_size=0
+- ++framework.action_model.past_action_window_size=0
+- ++framework.action_model.action_dim=7
+- ++framework.action_model.action_env_dim=7
+- ++framework.action_model.state_encoding=discretized_text
+- ++framework.action_model.action_layout=bridge_cross_task_7
+- ++framework.kv_memory.enabled=false
+- ++framework.kv_memory.window=4
+- ++framework.kv_memory.rollout_len=8
+- ++framework.kv_memory.packed_train=false
+- ++framework.kv_memory.rebased_sink=true
+- ++framework.name=QwenOFT
+- ++datasets.vla_data.dataset_py=lerobot_datasets
+- ++datasets.vla_data.include_state=false
+- ++datasets.vla_data.data_root_dir=playground/Datasets/rl_games
+- ++datasets.vla_data.data_mix=cross_task_train
+- ++datasets.vla_data.eval_data_mix=null
+- ++datasets.vla_data.custom_mixtures_path=null
+- ++datasets.vla_data.action_type=discrete
+- ++datasets.vla_data.sequential_step_sampling=true
+- ++datasets.vla_data.eval_sequential_step_sampling=null
+- ++datasets.vla_data.num_workers=8
+- ++datasets.vla_data.eval_num_workers=8
+- ++datasets.vla_data.prefetch_factor=4
+- ++datasets.vla_data.persistent_workers=true
+- ++datasets.vla_data.pin_memory=true
+- ++datasets.vla_data.shuffle=true
+- ++datasets.vla_data.action_balance.enabled=false
+- ++datasets.vla_data.action_balance.strategy=balanced_epoch
+- ++datasets.vla_data.action_balance.action_key=action_id
+- ++datasets.vla_data.action_balance.target_flap_fraction=0.3
+- ++datasets.vla_data.action_balance.noop_id=0
+- ++datasets.vla_data.action_balance.flap_id=1
+- ++datasets.vla_data.latency_curriculum.enabled=false
+- ++datasets.vla_data.latency_curriculum.strategy=exclusive
+- ++datasets.vla_data.latency_curriculum.latencies=null
+- ++datasets.vla_data.latency_curriculum.phase_steps=null
+- ++datasets.vla_data.latency_curriculum.phase_distributions=null
+- ++datasets.vla_data.latency_curriculum.new_latency_passes=1.0
+- ++datasets.vla_data.latency_curriculum.replay_passes=0.25
+- ++datasets.vla_data.latency_curriculum.target_total_passes=2.0
+- ++datasets.vla_data.latency_curriculum.final_equalization=true
+- ++datasets.vla_data.latency_curriculum.step_budget_mode=auto
+- ++datasets.vla_data.latency_curriculum.eval_at_phase_end=false
+- ++datasets.vla_data.latency_curriculum.save_at_phase_end=false
+- ++datasets.vla_data.latency_curriculum.computed_plan=null
+- ++datasets.vla_data.per_device_batch_size=32
+- ++datasets.vla_data.load_all_data_for_training=true
+- ++datasets.vla_data.num_obs_frames=1
+- ++datasets.vla_data.image_mode=single
+- ++datasets.vla_data.prompt_mode=raw
+- ++datasets.vla_data.stitch_grid=[2,2]
+- ++datasets.vla_data.obs_image_size=null
+- ++datasets.vla_data.video_backend=torchvision_av
+- ++dataset.source_hf=
+- ++dataset.config_name=null
+- ++dataset.source_subdir=null
+- ++dataset.converted_name=cross_task_train
+- ++dataset.single_source_hf=
+- ++dataset.mixed_source_hf=
+- ++dataset.single_converted_name=flappy_train
+- ++dataset.mixed_converted_name=flappy_mixed_latency_train
+- ++dataset.single_latency_filter=null
+- ++dataset.mixed_latency_filter=null
+- ++dataset.force_download=false
+- ++dataset.setup_force=false
+- ++dataset.skip_verification=false
+- ++dataset.target_latency_unit=observation_steps
+- ++dataset.verify_rows=200
+- ++dataset.max_episodes=null
+- ++dataset.episodes_per_latency=null
+- ++dataset.latency_filter=null
+- ++dataset.debug_subset.enabled=false
+- ++dataset.debug_subset.max_episodes=5
+- ++dataset.debug_subset.suffix=debug
+- ++base_model.repo_id=Qwen/Qwen3-VL-4B-Instruct
+- ++initialization.checkpoint_local_dir=playground/Pretrained_models/Qwen3VL-OFT-Bridge-RT-1
+- ++initialization.checkpoint_hf_repo_id=StarVLA/Qwen3VL-OFT-Bridge-RT-1
+- ++initialization.checkpoint_filename=checkpoints/steps_5000_pytorch_model.pt
+- ++trainer.max_train_steps=15000
+- ++trainer.num_warmup_steps=0
+- ++trainer.save_interval=15000
+- ++trainer.eval_interval=15000
+- ++trainer.eval_num_batches=100
+- ++trainer.per_latency_eval_num_batches=5
+- ++trainer.eval_action_classification=true
+- ++trainer.eval_action_classification_interval=null
+- ++trainer.cc_f1_tolerance=1
+- ++trainer.learning_rate.base=2e-05
+- ++trainer.learning_rate.qwen_vl_interface=1e-05
+- ++trainer.learning_rate.action_model=0.0001
+- ++trainer.lr_scheduler_type=cosine_with_min_lr
+- ++trainer.scheduler_specific_kwargs.min_lr=1e-06
+- ++trainer.freeze_modules=
+- ++trainer.freeze_vit=false
+- ++trainer.freeze_tied_embedding=false
+- ++trainer.freeze_llm_layers=[]
+- ++trainer.loss_scale.vla=1.0
+- ++trainer.loss_scale.vlm=0.1
+- ++trainer.max_grad_norm=1.0
+- ++trainer.weight_decay=0.0
+- ++trainer.logging_frequency=1
+- ++trainer.profile_timing.enabled=false
+- ++trainer.profile_timing.log_interval=10
+- ++trainer.gradient_clipping=1.0
+- ++trainer.gradient_accumulation_steps=8
+- ++trainer.distributed_backend=none
+- ++trainer.is_resume=false
+- ++trainer.pretrained_checkpoint=null
+- ++trainer.resume_step=0
+- ++trainer.reload_modules=null
+- ++trainer.optimizer.name=AdamW
+- ++trainer.optimizer.betas=[0.9,0.95]
+- ++trainer.optimizer.eps=1e-08
+- ++trainer.optimizer.weight_decay=1e-08
+- ++trainer.optimizer.fused=true
+- ++trainer.save_format=pt
+- ++workspace_dir=/mnt/local/talha
+- ++run_root_dir=results/Checkpoints
+- ++seed=42
+- ++auth.env_file=null
+- ++auth.hf_token_env=HF_TOKEN
+- ++auth.wandb_api_key_env=WANDB_API_KEY
+- ++paths.run_root_dir=/mnt/local/talha/results/Checkpoints
+- ++paths.dataset_local_dir=/mnt/local/talha/playground/Datasets/rl_games
+- ++paths.dataset_cache_dir=null
+- ++paths.base_model_dir=/mnt/local/talha/playground/Pretrained_models/Qwen3-VL-4B-Instruct
+- ++paths.accelerate_config=starVLA/config/deepseeds/deepspeed_zero2.yaml
+- ++rl_games.model_alias=openvla
+- ++rl_games.task=cross_task
+- ++rl_games.env_eval.image_size=224
+- ++rl_games.env_eval.frameskip=1
+- ++rl_games.env_eval.deadly.action_layout=multibinary_7
+- ++rl_games.env_eval.deadly.multibinary_threshold=0.0
+- ++rl_games.env_eval.enabled=true
+- ++rl_games.env_eval.eval_backend=eval_core
+- ++rl_games.env_eval.distributed_mode=none
+- ++rl_games.env_eval.vectorized.enabled=false
+- ++rl_games.env_eval.vectorized.batch_size=1
+- ++rl_games.env_eval.latency.prompt_map_path=null
+- ++rl_games.env_eval.latency.mode=mixed
+- ++rl_games.env_eval.latency.values=[0,1,2,3,4,5]
+- ++rl_games.env_eval.mid_train.enabled=false
+- ++rl_games.env_eval.mid_train.interval_steps=1000
+- ++rl_games.env_eval.mid_train.latencies=[0,1,2,3,4,5]
+- ++rl_games.env_eval.mid_train.num_episodes=5
+- ++rl_games.env_eval.mid_train.max_steps_per_episode=3600
+- ++rl_games.env_eval.post_train.enabled=true
+- ++rl_games.env_eval.post_train.latencies=[0,2,4]
+- ++rl_games.env_eval.post_train.latencies_from_prompt_map=false
+- ++rl_games.env_eval.post_train.num_episodes=5
+- ++rl_games.env_eval.post_train.max_steps_per_episode=3600
+- ++rl_games.env_eval.asterix.action_layout=factorized_6
+- ++rl_games.cross_task.train_tasks=[{name:flappy,role:A,train_source_hf:latency-sensitive-bench/flappy_200ep,prompt_source_hf:latency-sensitive-bench/flappy_200ep,train_source_subdir:flappy_fix_latency_0_200ep,prompt_source_subdir:flappy_fix_latency_0_200ep,converted_name:flappy_mixed_024_200ep_cross_train,train_latency_filter:[0,2,4],eval_latency_filter:[0,2,4],episodes_per_latency:null,max_episodes:200,weight:1.0},{name:demon_attack,role:B,train_source_hf:latency-sensitive-bench/demon_attack_200ep,prompt_source_hf:latency-sensitive-bench/demon_attack_200ep,train_source_subdir:demon_attack_fix_latency_0_200ep,prompt_source_subdir:demon_attack_fix_latency_0_200ep,converted_name:demon_attack_mixed_024_200ep_cross_train,train_latency_filter:[0,2,4],eval_latency_filter:[0,2,4],episodes_per_latency:null,max_episodes:200,weight:1.0},{name:deadly_corridor,role:C,train_source_hf:latency-sensitive-bench/deadly_1000ep,prompt_source_hf:latency-sensitive-bench/deadly_1000ep,train_source_subdir:deadly_corridor_fix_latency_0_1000ep,prompt_source_subdir:deadly_corridor_fix_latency_0_1000ep,converted_name:deadly_corridor_mixed_024_1000ep_cross_train,train_latency_filter:[0,2,4],eval_latency_filter:[0,2,4],episodes_per_latency:null,max_episodes:1000,action_layout:multibinary_7,weight:1.0},{name:asterix,role:D,train_source_hf:latency-sensitive-bench/asterix_200ep,prompt_source_hf:latency-sensitive-bench/asterix_200ep,train_source_subdir:asterix_fixed_latency_0_200ep_7k2steps,prompt_source_subdir:asterix_fixed_latency_0_200ep_7k2steps,converted_name:asterix_mixed_024_200ep_cross_train,train_latency_filter:[0,2,4],eval_latency_filter:[0,2,4],episodes_per_latency:null,max_episodes:200,action_layout:factorized_6,weight:1.0},{name:atlantis,role:E,train_source_hf:latency-sensitive-bench/atlantis_200ep,prompt_source_hf:latency-sensitive-bench/atlantis_200ep,train_source_subdir:atlantis_fixed_latency_0_200ep_7k2steps,prompt_source_subdir:atlantis_fixed_latency_0_200ep_7k2steps,converted_name:atlantis_mixed_024_200ep_cross_train,train_latency_filter:[0,2,4],eval_latency_filter:[0,2,4],episodes_per_latency:null,max_episodes:200,weight:1.0},{name:defend_the_line,role:F,train_source_hf:latency-sensitive-bench/defend_the_line_200ep,prompt_source_hf:latency-sensitive-bench/defend_the_line_200ep,train_source_subdir:defend_the_line_fixed_latency_0_200ep_7k2steps,prompt_source_subdir:defend_the_line_fixed_latency_0_200ep_7k2steps,converted_name:defend_the_line_mixed_024_200ep_cross_train,train_latency_filter:[0,2,4],eval_latency_filter:[0,2,4],episodes_per_latency:null,max_episodes:200,weight:1.0}]
+- ++rl_games.cross_task.loss_by_task.flappy=discrete_ce
+- ++rl_games.cross_task.loss_by_task.demon_attack=discrete_ce
+- ++rl_games.cross_task.loss_by_task.defend_the_line=discrete_ce
+- ++rl_games.cross_task.loss_by_task.deadly_corridor=multibinary_bce
+- ++rl_games.cross_task.loss_by_task.asterix=asterix_factorized_ce
+- ++rl_games.cross_task.loss_by_task.atlantis=discrete_ce
+- ++rl_games.cross_task.loss_weight_by_task.flappy=1.0
+- ++rl_games.cross_task.loss_weight_by_task.demon_attack=1.0
+- ++rl_games.cross_task.loss_weight_by_task.defend_the_line=1.0
+- ++rl_games.cross_task.loss_weight_by_task.deadly_corridor=1.0
+- ++rl_games.cross_task.loss_weight_by_task.asterix=1.0
+- ++rl_games.cross_task.loss_weight_by_task.atlantis=1.0
+- ++rl_games.cross_task.eval_tasks.flappy.prompt_map_path=null
+- ++rl_games.cross_task.eval_tasks.flappy.frameskip=1
+- ++rl_games.cross_task.eval_tasks.flappy.image_size=224
+- '++rl_games.cross_task.eval_tasks.flappy.task_description=''You are playing Flappy
+  Bird. Pass through the pipe gaps and stay alive. Choose the action: NOOP, FLAP.'''
+- ++rl_games.cross_task.eval_tasks.flappy.mid_train.enabled=false
+- ++rl_games.cross_task.eval_tasks.flappy.mid_train.latencies=[0,2,4]
+- ++rl_games.cross_task.eval_tasks.flappy.mid_train.num_episodes=5
+- ++rl_games.cross_task.eval_tasks.flappy.mid_train.max_steps_per_episode=3600
+- ++rl_games.cross_task.eval_tasks.flappy.post_train.enabled=true
+- ++rl_games.cross_task.eval_tasks.flappy.post_train.latencies=[0,2,4]
+- ++rl_games.cross_task.eval_tasks.flappy.post_train.num_episodes=20
+- ++rl_games.cross_task.eval_tasks.flappy.post_train.max_steps_per_episode=3600
+- ++rl_games.cross_task.eval_tasks.demon_attack.prompt_map_path=null
+- ++rl_games.cross_task.eval_tasks.demon_attack.frameskip=4
+- ++rl_games.cross_task.eval_tasks.demon_attack.image_size=224
+- '++rl_games.cross_task.eval_tasks.demon_attack.task_description=''You are playing
+  Demon Attack from a single game image. Choose exactly one action from: NOOP, FIRE,
+  RIGHT, LEFT, RIGHTFIRE, LEFTFIRE.'''
+- ++rl_games.cross_task.eval_tasks.demon_attack.mid_train.enabled=false
+- ++rl_games.cross_task.eval_tasks.demon_attack.mid_train.latencies=[0,2,4]
+- ++rl_games.cross_task.eval_tasks.demon_attack.mid_train.num_episodes=5
+- ++rl_games.cross_task.eval_tasks.demon_attack.mid_train.max_steps_per_episode=3600
+- ++rl_games.cross_task.eval_tasks.demon_attack.post_train.enabled=true
+- ++rl_games.cross_task.eval_tasks.demon_attack.post_train.latencies=[0,2,4]
+- ++rl_games.cross_task.eval_tasks.demon_attack.post_train.num_episodes=20
+- ++rl_games.cross_task.eval_tasks.demon_attack.post_train.max_steps_per_episode=3600
+- ++rl_games.cross_task.eval_tasks.defend_the_line.prompt_map_path=null
+- ++rl_games.cross_task.eval_tasks.defend_the_line.frameskip=4
+- ++rl_games.cross_task.eval_tasks.defend_the_line.image_size=224
+- ++rl_games.cross_task.eval_tasks.defend_the_line.task_description=null
+- ++rl_games.cross_task.eval_tasks.defend_the_line.mid_train.enabled=false
+- ++rl_games.cross_task.eval_tasks.defend_the_line.mid_train.latencies=null
+- ++rl_games.cross_task.eval_tasks.defend_the_line.mid_train.num_episodes=null
+- ++rl_games.cross_task.eval_tasks.defend_the_line.mid_train.max_steps_per_episode=null
+- ++rl_games.cross_task.eval_tasks.defend_the_line.post_train.enabled=false
+- ++rl_games.cross_task.eval_tasks.defend_the_line.post_train.latencies=null
+- ++rl_games.cross_task.eval_tasks.defend_the_line.post_train.num_episodes=null
+- ++rl_games.cross_task.eval_tasks.defend_the_line.post_train.max_steps_per_episode=null
+- ++rl_games.cross_task.eval_tasks.deadly_corridor.prompt_map_path=null
+- ++rl_games.cross_task.eval_tasks.deadly_corridor.frameskip=4
+- ++rl_games.cross_task.eval_tasks.deadly_corridor.image_size=224
+- ++rl_games.cross_task.eval_tasks.deadly_corridor.task_description='You are playing
+  Deadly Corridor in VizDoom. Choose actions from MOVE_FORWARD, MOVE_BACKWARD, MOVE_LEFT,
+  MOVE_RIGHT, TURN_LEFT, TURN_RIGHT, ATTACK.'
+- ++rl_games.cross_task.eval_tasks.deadly_corridor.mid_train.enabled=false
+- ++rl_games.cross_task.eval_tasks.deadly_corridor.mid_train.latencies=[0,2,4]
+- ++rl_games.cross_task.eval_tasks.deadly_corridor.mid_train.num_episodes=5
+- ++rl_games.cross_task.eval_tasks.deadly_corridor.mid_train.max_steps_per_episode=3600
+- ++rl_games.cross_task.eval_tasks.deadly_corridor.post_train.enabled=true
+- ++rl_games.cross_task.eval_tasks.deadly_corridor.post_train.latencies=[0,2,4]
+- ++rl_games.cross_task.eval_tasks.deadly_corridor.post_train.num_episodes=20
+- ++rl_games.cross_task.eval_tasks.deadly_corridor.post_train.max_steps_per_episode=3600
+- ++rl_games.cross_task.eval_tasks.asterix.prompt_map_path=null
+- ++rl_games.cross_task.eval_tasks.asterix.frameskip=4
+- ++rl_games.cross_task.eval_tasks.asterix.image_size=224
+- ++rl_games.cross_task.eval_tasks.asterix.task_description=null
+- ++rl_games.cross_task.eval_tasks.asterix.mid_train.enabled=false
+- ++rl_games.cross_task.eval_tasks.asterix.mid_train.latencies=null
+- ++rl_games.cross_task.eval_tasks.asterix.mid_train.num_episodes=null
+- ++rl_games.cross_task.eval_tasks.asterix.mid_train.max_steps_per_episode=null
+- ++rl_games.cross_task.eval_tasks.asterix.post_train.enabled=false
+- ++rl_games.cross_task.eval_tasks.asterix.post_train.latencies=null
+- ++rl_games.cross_task.eval_tasks.asterix.post_train.num_episodes=null
+- ++rl_games.cross_task.eval_tasks.asterix.post_train.max_steps_per_episode=null
+- ++rl_games.cross_task.eval_tasks.atlantis.prompt_map_path=null
+- ++rl_games.cross_task.eval_tasks.atlantis.frameskip=4
+- ++rl_games.cross_task.eval_tasks.atlantis.image_size=224
+- ++rl_games.cross_task.eval_tasks.atlantis.task_description=null
+- ++rl_games.cross_task.eval_tasks.atlantis.mid_train.enabled=false
+- ++rl_games.cross_task.eval_tasks.atlantis.mid_train.latencies=null
+- ++rl_games.cross_task.eval_tasks.atlantis.mid_train.num_episodes=null
+- ++rl_games.cross_task.eval_tasks.atlantis.mid_train.max_steps_per_episode=null
+- ++rl_games.cross_task.eval_tasks.atlantis.post_train.enabled=false
+- ++rl_games.cross_task.eval_tasks.atlantis.post_train.latencies=null
+- ++rl_games.cross_task.eval_tasks.atlantis.post_train.num_episodes=null
+- ++rl_games.cross_task.eval_tasks.atlantis.post_train.max_steps_per_episode=null
+- ++rl_games.initialization_mode=bridge
+- ++rl_games.action_carrier=bridge
+- ++bridge_base_model.repo_id.openvla=Qwen/Qwen3-VL-4B-Instruct
+- ++bridge_base_model.repo_id.pi0=StarVLA/Qwen2.5-VL-3B-Instruct-Action
+- ++bridge_base_model.repo_id.pi05=Qwen/Qwen3-VL-4B-Instruct
+- ++bridge_base_model.repo_id.gr00t=Qwen/Qwen3-VL-4B-Instruct
+- ++bridge_base_model.local_dir.openvla=playground/Pretrained_models/Qwen3-VL-4B-Instruct
+- ++bridge_base_model.local_dir.pi0=playground/Pretrained_models/Qwen2.5-VL-3B-Instruct-Action
+- ++bridge_base_model.local_dir.pi05=playground/Pretrained_models/Qwen3-VL-4B-Instruct
+- ++bridge_base_model.local_dir.gr00t=playground/Pretrained_models/Qwen3-VL-4B-Instruct
+- ++checkpoint.load=auto
+- ++checkpoint.hf_repo_id=latency-sensitive-bench/openvla_bridge_cross_mixed_all_024_5env200ep_deadly1000ep_exp1
+- ++checkpoint.save_best_model=false
+- ++checkpoint.save_final_model=true
+- ++checkpoint.save_pt_file=false
+- ++checkpoint.save_training_state=true
+- ++checkpoint.save_safetensors_file=false
+- ++checkpoint.local.keep_last_n=1
+- ++checkpoint.sync.enabled=true
+- ++checkpoint.sync.repo_id=latency-sensitive-bench/openvla_bridge_cross_mixed_all_024_5env200ep_deadly1000ep_exp1
+- ++checkpoint.sync.keep_last_n=0
+- ++checkpoint.sync.sync_every_n_checkpoints=1
+- ++checkpoint.sync.resume_policy=local_latest
+- ++run_id=openvla_bridge_cross_mixed_all_024_5env200ep_deadly1000ep_exp1
+- ++output_dir=null
+- ++config_yaml=null
+- ++is_debug=false
+- ++version_id=0.21
+- ++wandb_tags=[final_vla_profiling_training,demon_attack,openvla,rtx3090_profile]
+- ++training_latency_condition=rtx3090_profile
+- ++wandb_group=final_vla_profiling_training
+- ++run_root_dir=/mnt/local/talha/results/Checkpoints
+- ++trainer.is_resume=false
+- ++trainer.pretrained_checkpoint=/mnt/local/talha/results/Checkpoints/openvla_bridge_cross_mixed_all_024_5env200ep_deadly1000ep_exp1/checkpoints/_initialization/StarVLA_Qwen3VL-OFT-Bridge-RT-1/checkpoints/steps_5000_pytorch_model.pt
+- ++trainer.resume_step=0
+- ++datasets.vla_data.data_root_dir=/mnt/local/talha/playground/Datasets/rl_games
+- ++datasets.vla_data.data_mix=cross__asterix_mixed_024_200ep_cross_train__lat0_2_4__200ep__bridge__atlantis_mixed_024_200ep_cross_train__lat0_2_4__200ep__bridge__deadly_corridor_mixed_024_1000ep_cross_train__lat0_2_4__1000ep__bridge__defend_the_line_mixed_024_200ep_cross_train__lat0_2_4__200ep__bridge__demon_attack_mixed_024_200ep_cross_train__lat0_2_4__200ep__bridge__flappy_mixed_024_200ep_cross_train__lat0_2_4__200ep__bridge
+- ++datasets.vla_data.eval_data_mix=cross__asterix_mixed_024_200ep_cross_train__lat0_2_4__200ep__bridge__atlantis_mixed_024_200ep_cross_train__lat0_2_4__200ep__bridge__deadly_corridor_mixed_024_1000ep_cross_train__lat0_2_4__1000ep__bridge__defend_the_line_mixed_024_200ep_cross_train__lat0_2_4__200ep__bridge__demon_attack_mixed_024_200ep_cross_train__lat0_2_4__200ep__bridge__flappy_mixed_024_200ep_cross_train__lat0_2_4__200ep__bridge__val
+- ++datasets.vla_data.custom_mixtures_path=/mnt/local/talha/playground/Datasets/rl_games/_generated_mixtures/cross_09b5517c91a12cca.json
+- ++framework.qwenvl.base_vlm=/mnt/local/talha/playground/Pretrained_models/Qwen3-VL-4B-Instruct
+- ++rl_games.cross_task.eval_tasks.flappy.prompt_map_path=/mnt/local/talha/playground/Datasets/rl_games/_prompt_maps/flappy_mixed_024_200ep_cross_train__lat0_2_4__200ep__bridge/eval_latency_prompt_map.json
+- ++rl_games.cross_task.eval_tasks.demon_attack.prompt_map_path=/mnt/local/talha/playground/Datasets/rl_games/_prompt_maps/demon_attack_mixed_024_200ep_cross_train__lat0_2_4__200ep__bridge/eval_latency_prompt_map.json
+- ++rl_games.cross_task.eval_tasks.deadly_corridor.prompt_map_path=/mnt/local/talha/playground/Datasets/rl_games/_prompt_maps/deadly_corridor_mixed_024_1000ep_cross_train__lat0_2_4__1000ep__bridge/eval_latency_prompt_map.json

latency-aware/demon-attack/vla/starvla-qwenoft-h1/openvla_bridge_demon_attack_latency_mixed_024_exp1/checkpoints/steps_5000_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dc6f0c52cfc33ea2ae7d7c382abe762561693936045464b1c160b1d9fdf0f501
+size 9138230516