| idm_model: |
| action_decoder: |
| embed_dim: 256 |
| n_kv_head: 8 |
| n_q_head: 8 |
| n_transformer_layers: 3 |
| attention_history_len: |
| - 100 |
| - 100 |
| - 100 |
| - 100 |
| mask_block_size: null |
| model_type: dense |
| n_kv_head: 8 |
| n_kv_sink_tokens: 1 |
| n_q_head: 8 |
| n_thinking_tokens: 1 |
| n_transformer_layers: 4 |
| sparse_moe: |
| experts_per_token: 4 |
| lb_loss_weight: 0.01 |
| num_experts: 16 |
| rz_loss_weight: 0.001 |
| top_p: null |
| transformer_dim: 128 |
| z_loss_weight: 0.0001 |
| inference: |
| checkpoint_path: null |
| mouse_sampling_approach: truncated_normal |
| sampling_temperature: 1.0 |
| launch_all: |
| env: [] |
| job_name: 300M_1e_4_100_1e_4-20251110232545-all |
| run_cmd: uv run python elefant/lapo/train.py --stage all |
| --config={config_path} |
| vm_flavor: n3-H100x8-NVLink |
| launch_stage23: |
| env: [] |
| job_name: 300M_1e_4_100_1e_4-20251110232545-stage23 |
| run_cmd: uv run python elefant/lapo/train.py --stage stage23 |
| --config={config_path} |
| vm_flavor: n3-H100x8-NVLink |
| policy_model: |
| action_decoder: |
| embed_dim: 1024 |
| n_kv_head: 8 |
| n_q_head: 8 |
| n_transformer_layers: 3 |
| attention_history_len: |
| - 200 |
| - 200 |
| - 200 |
| - 200 |
| - 200 |
| - 200 |
| - 200 |
| - 200 |
| - 200 |
| - 200 |
| - 200 |
| - 200 |
| - 200 |
| - 200 |
| - 200 |
| - 200 |
| - 200 |
| - 200 |
| - 200 |
| - 200 |
| mask_block_size: 128 |
| model_type: dense |
| n_kv_head: 16 |
| n_kv_sink_tokens: 0 |
| n_q_head: 16 |
| n_thinking_tokens: 1 |
| n_transformer_layers: 20 |
| sparse_moe: |
| experts_per_token: 4 |
| lb_loss_weight: 0.01 |
| num_experts: 16 |
| rz_loss_weight: 0.001 |
| top_p: null |
| transformer_dim: 1024 |
| z_loss_weight: 0.0001 |
| shared: |
| action_mapping: |
| max_keys: 4 |
| max_mouse_keys: 2 |
| fast_dev_run: false |
| frame_height: 192 |
| frame_width: 192 |
| n_seq_timesteps: 200 |
| output_path: yueyugua/lapo/dev/1/20251110232545/300M_1e_4_100_1e_4 |
| precision: bf16-mixed |
| text_tokenizer_config: |
| text_annotation_model_version: |
| - gemini-2.5-flash |
| - gemini-2.5-flash-thinking-0905 |
| text_embedding_shape: |
| - 1 |
| - 768 |
| text_tokenizer_name: gemma |
| tokenizer: |
| conv_tokenizer_config: |
| num_tokens: 1 |
| magavit_v2_tokenizer_config: |
| checkpoint_path: |
| /data/magvit/checkpoints/roblox-all-elefant-run6/epoch=0-step=23000.ckpt |
| compression_rate: 8 |
| config_path: |
| /root/ml-playground/3rdparty/SEED-Voken/configs/Open-MAGVIT2/gpu/roblox_all_elefant_run6.yaml |
| type: conv |
| vit_tokenizer_config: |
| patch_size: 16 |
| stage1_idm: |
| accumulate_grad_batches: 1 |
| launch: |
| env: [] |
| job_name: 300M_1e_4_100_1e_4-20251110232545-st1 |
| run_cmd: uv run python elefant/lapo/train.py --stage stage1_idm |
| --config={config_path} |
| vm_flavor: n3-H100x8-NVLink |
| n_training_steps: 1000 |
| n_validation_steps: 10 |
| optim: |
| beta_1: 0.95 |
| beta_2: 0.999 |
| learning_rate: 0.0003 |
| weight_decay: 0.01 |
| save_every_n_steps: 1000 |
| training_dataset: |
| always_labelled: false |
| batch_size: 128 |
| dataset_worker_num_workers_per_gpu: 1 |
| dataset_worker_prefetch_factor: 1 |
| downloaded_files_queue_size_per_gpu: 16 |
| n_parallel_downloads_per_gpu: 8 |
| n_preprocess_threads_per_gpu: 16 |
| n_seq_timesteps: 3 |
| preprocessed_chunks_queue_size_per_gpu: 4096 |
| rand_augmentation: |
| fraction_augmented: 0.0 |
| shuffle: true |
| shuffle_buffer_size_per_gpu: 1024 |
| shuffled_chunks_queue_size_per_gpu: 16 |
| sql_query: SELECT filepath FROM video_dataset_metadata |
| text_annotation_model_version: |
| - gemini-2.5-flash |
| - gemini-2.5-flash-thinking-0905 |
| tokenizer: |
| conv_tokenizer_config: |
| num_tokens: 1 |
| magavit_v2_tokenizer_config: null |
| type: vit |
| vit_tokenizer_config: |
| patch_size: 16 |
| warn_on_starvation: false |
| validation_datasets: [] |
| validation_step_interval: 100 |
| stage2_pretrain: |
| accumulate_grad_batches: 1 |
| freeze_transformer_layers_for_steps: 0 |
| launch: |
| env: [] |
| job_name: lapo3-labelled-bc |
| run_cmd: uv run python elefant/lapo/train.py --stage |
| stage2_pretrain_with_idm --config={config_path} |
| vm_flavor: n3-H100x1 |
| n_training_steps: 100000 |
| n_validation_steps: 10 |
| optim: |
| beta_1: 0.95 |
| beta_2: 0.999 |
| learning_rate: 0.0003 |
| weight_decay: 0.01 |
| save_every_n_steps: 1000 |
| supervised_idm_model_path: null |
| training_dataset: |
| always_labelled: false |
| batch_size: 128 |
| dataset_worker_num_workers_per_gpu: 1 |
| dataset_worker_prefetch_factor: 1 |
| downloaded_files_queue_size_per_gpu: 16 |
| n_parallel_downloads_per_gpu: 8 |
| n_preprocess_threads_per_gpu: 16 |
| n_seq_timesteps: 3 |
| preprocessed_chunks_queue_size_per_gpu: 4096 |
| rand_augmentation: |
| fraction_augmented: 0.0 |
| shuffle: true |
| shuffle_buffer_size_per_gpu: 1024 |
| shuffled_chunks_queue_size_per_gpu: 16 |
| sql_query: SELECT filepath FROM video_dataset_metadata |
| text_annotation_model_version: |
| - gemini-2.5-flash |
| - gemini-2.5-flash-thinking-0905 |
| tokenizer: |
| conv_tokenizer_config: |
| num_tokens: 1 |
| magavit_v2_tokenizer_config: null |
| type: vit |
| vit_tokenizer_config: |
| patch_size: 16 |
| warn_on_starvation: false |
| validation_datasets: [] |
| validation_step_interval: 100 |
| stage3_finetune: |
| accumulate_grad_batches: 1 |
| freeze_transformer_layers_for_steps: 0 |
| init: |
| random: true |
| stage2_model_path: null |
| stage3_model_path: null |
| launch: |
| env: [] |
| job_name: 300M_1e_4_100_1e_4-20251110232545-st3 |
| run_cmd: uv run python elefant/lapo/train.py --stage stage3_finetune |
| --config={config_path} |
| vm_flavor: n3-H100x8-NVLink |
| n_training_steps: 500000 |
| n_validation_steps: 200 |
| optim: |
| beta_1: 0.9 |
| beta_2: 0.999 |
| learning_rate: 0.0001 |
| weight_decay: 0.0001 |
| save_every_n_steps: 10000 |
| training_dataset: |
| always_labelled: true |
| batch_size: 8 |
| dataset_worker_num_workers_per_gpu: 1 |
| dataset_worker_prefetch_factor: 16 |
| downloaded_files_queue_size_per_gpu: 8 |
| n_parallel_downloads_per_gpu: 16 |
| n_preprocess_threads_per_gpu: 32 |
| n_seq_timesteps: 3 |
| preprocessed_chunks_queue_size_per_gpu: 2 |
| rand_augmentation: |
| fraction_augmented: 1.0 |
| shuffle: true |
| shuffle_buffer_size_per_gpu: 1900 |
| shuffled_chunks_queue_size_per_gpu: 16 |
| sql_query: "SELECT filepath FROM video_dataset_metadata\nWHERE \n(\nfilepath LIKE |
| 'employee_labelled/%'\n)\nAND filepath_random_num >= 0.0 AND filepath_random_num |
| < 0.95\nAND \"user\" NOT IN ('charlie_morry', 'Irakli')\nAND timestamp <= epoch(strptime('2025-11-01', |
| '%Y-%m-%d')) * 1000;\n" |
| text_annotation_model_version: |
| - gemini-2.5-flash |
| - gemini-2.5-flash-thinking-0905 |
| tokenizer: |
| conv_tokenizer_config: |
| num_tokens: 1 |
| magavit_v2_tokenizer_config: null |
| type: vit |
| vit_tokenizer_config: |
| patch_size: 16 |
| warn_on_starvation: false |
| validation_datasets: |
| - always_labelled: true |
| batch_size: 8 |
| dataset_worker_num_workers_per_gpu: 1 |
| dataset_worker_prefetch_factor: 8 |
| downloaded_files_queue_size_per_gpu: 16 |
| n_parallel_downloads_per_gpu: 1 |
| n_preprocess_threads_per_gpu: 8 |
| n_seq_timesteps: 3 |
| preprocessed_chunks_queue_size_per_gpu: 1 |
| rand_augmentation: |
| fraction_augmented: 0.0 |
| shuffle: true |
| shuffle_buffer_size_per_gpu: 1024 |
| shuffled_chunks_queue_size_per_gpu: 1 |
| sql_query: "SELECT filepath FROM video_dataset_metadata\nWHERE \n(\nfilepath LIKE |
| 'employee_labelled/%'\n)\nAND filepath_random_num >= 0.95 AND filepath_random_num |
| < 1.0\nAND env_name = 'elefant-test' AND env_sub_type = 'hovercraft'\nAND timestamp |
| <= epoch(strptime('2025-11-01', '%Y-%m-%d')) * 1000\nORDER BY filepath_random_num |
| ASC\n" |
| text_annotation_model_version: |
| - gemini-2.5-flash |
| - gemini-2.5-flash-thinking-0905 |
| tokenizer: |
| conv_tokenizer_config: |
| num_tokens: 1 |
| magavit_v2_tokenizer_config: null |
| type: vit |
| vit_tokenizer_config: |
| patch_size: 16 |
| validation_name: elefant-test-racing |
| warn_on_starvation: false |
| - always_labelled: true |
| batch_size: 8 |
| dataset_worker_num_workers_per_gpu: 1 |
| dataset_worker_prefetch_factor: 8 |
| downloaded_files_queue_size_per_gpu: 16 |
| n_parallel_downloads_per_gpu: 1 |
| n_preprocess_threads_per_gpu: 8 |
| n_seq_timesteps: 3 |
| preprocessed_chunks_queue_size_per_gpu: 1 |
| rand_augmentation: |
| fraction_augmented: 0.0 |
| shuffle: true |
| shuffle_buffer_size_per_gpu: 1024 |
| shuffled_chunks_queue_size_per_gpu: 1 |
| sql_query: "SELECT filepath FROM video_dataset_metadata\nWHERE \n(\nfilepath LIKE |
| 'employee_labelled/%'\n)\nAND filepath_random_num >= 0.95 AND filepath_random_num |
| < 1.0\nAND env_name = 'elefant-test' AND (env_sub_type = 'basic-fps' OR env_sub_type |
| = 'fps-basic')\nORDER BY filepath_random_num ASC\n" |
| text_annotation_model_version: |
| - gemini-2.5-flash |
| - gemini-2.5-flash-thinking-0905 |
| tokenizer: |
| conv_tokenizer_config: |
| num_tokens: 1 |
| magavit_v2_tokenizer_config: null |
| type: vit |
| vit_tokenizer_config: |
| patch_size: 16 |
| validation_name: elefant-test-fps |
| warn_on_starvation: false |
| - always_labelled: true |
| batch_size: 8 |
| dataset_worker_num_workers_per_gpu: 1 |
| dataset_worker_prefetch_factor: 8 |
| downloaded_files_queue_size_per_gpu: 16 |
| n_parallel_downloads_per_gpu: 1 |
| n_preprocess_threads_per_gpu: 8 |
| n_seq_timesteps: 3 |
| preprocessed_chunks_queue_size_per_gpu: 1 |
| rand_augmentation: |
| fraction_augmented: 0.0 |
| shuffle: true |
| shuffle_buffer_size_per_gpu: 1024 |
| shuffled_chunks_queue_size_per_gpu: 1 |
| sql_query: "SELECT filepath FROM video_dataset_metadata\nWHERE \n(\nfilepath LIKE |
| 'employee_labelled/%'\n)\nAND filepath_random_num >= 0.95 AND filepath_random_num |
| < 1.0\nAND env_name = 'roblox' AND (env_sub_type = 'blade-ball' OR env_sub_type |
| = 'blade ball')\nAND timestamp <= epoch(strptime('2025-11-01', '%Y-%m-%d')) |
| * 1000\nORDER BY filepath_random_num ASC\n" |
| text_annotation_model_version: |
| - gemini-2.5-flash |
| - gemini-2.5-flash-thinking-0905 |
| tokenizer: |
| conv_tokenizer_config: |
| num_tokens: 1 |
| magavit_v2_tokenizer_config: null |
| type: vit |
| vit_tokenizer_config: |
| patch_size: 16 |
| validation_name: blade-ball |
| warn_on_starvation: false |
| - always_labelled: true |
| batch_size: 8 |
| dataset_worker_num_workers_per_gpu: 1 |
| dataset_worker_prefetch_factor: 8 |
| downloaded_files_queue_size_per_gpu: 16 |
| n_parallel_downloads_per_gpu: 1 |
| n_preprocess_threads_per_gpu: 8 |
| n_seq_timesteps: 3 |
| preprocessed_chunks_queue_size_per_gpu: 1 |
| rand_augmentation: |
| fraction_augmented: 0.0 |
| shuffle: true |
| shuffle_buffer_size_per_gpu: 1024 |
| shuffled_chunks_queue_size_per_gpu: 1 |
| sql_query: "SELECT filepath FROM video_dataset_metadata\nWHERE \n(\nfilepath LIKE |
| 'employee_labelled/%'\n)\nAND filepath_random_num >= 0.95 AND filepath_random_num |
| < 1.0\nAND ((env_name = 'msdos' AND env_sub_type = 'quake')\n OR (env_name |
| = 'quake')\n OR (env_name = 'quake2')\n OR (env_name = 'doom')\n OR (env_name |
| = 'call-of-duty-mobile')\n OR (env_name LIKE '%left-4%'))\nAND timestamp <= |
| epoch(strptime('2025-11-01', '%Y-%m-%d')) * 1000\nORDER BY filepath_random_num |
| ASC\n" |
| text_annotation_model_version: |
| - gemini-2.5-flash |
| - gemini-2.5-flash-thinking-0905 |
| tokenizer: |
| conv_tokenizer_config: |
| num_tokens: 1 |
| magavit_v2_tokenizer_config: null |
| type: vit |
| vit_tokenizer_config: |
| patch_size: 16 |
| validation_name: simple-fps |
| warn_on_starvation: false |
| - always_labelled: true |
| batch_size: 8 |
| dataset_worker_num_workers_per_gpu: 1 |
| dataset_worker_prefetch_factor: 8 |
| downloaded_files_queue_size_per_gpu: 16 |
| n_parallel_downloads_per_gpu: 1 |
| n_preprocess_threads_per_gpu: 8 |
| n_seq_timesteps: 3 |
| preprocessed_chunks_queue_size_per_gpu: 1 |
| rand_augmentation: |
| fraction_augmented: 0.0 |
| shuffle: true |
| shuffle_buffer_size_per_gpu: 1024 |
| shuffled_chunks_queue_size_per_gpu: 1 |
| sql_query: "SELECT filepath FROM video_dataset_metadata\nWHERE \n(\nfilepath LIKE |
| 'employee_labelled/%'\n)\nAND filepath_random_num >= 0.95 AND filepath_random_num |
| < 1.0\nAND \"user\" NOT IN ('charlie_morry', 'Irakli')\nAND timestamp <= epoch(strptime('2025-11-01', |
| '%Y-%m-%d')) * 1000\nORDER BY filepath_random_num ASC\n" |
| text_annotation_model_version: |
| - gemini-2.5-flash |
| - gemini-2.5-flash-thinking-0905 |
| tokenizer: |
| conv_tokenizer_config: |
| num_tokens: 1 |
| magavit_v2_tokenizer_config: null |
| type: vit |
| vit_tokenizer_config: |
| patch_size: 16 |
| validation_name: overall |
| warn_on_starvation: false |
| validation_step_interval: 200000000 |
| wandb: |
| enabled: true |
| exp_name: 300M_1e_4_100_1e_4-20251110232545 |
| project: scaling-law |
| run_id: ujoc4pyf |
| tags: [] |
|
|
