run_name: multitask_train model: model_name: molmo llm: d_model: 4096 n_heads: 32 n_kv_heads: 8 head_dim: null qkv_bias: false clip_qkv: null n_layers: 36 mlp_ratio: 4 mlp_hidden_size: 24576 activation_type: swiglu block_type: sequential rope: true rope_full_precision: true rope_theta: 1000000.0 rope_type: default rope_factor: null rope_high_freq_factor: null rope_low_freq_factor: null rope_original_max_position_embeddings: null attention_type: sdpa float32_attention: true attention_dropout: 0.0 attention_layer_norm: true attention_layer_norm_type: qwen3 residual_dropout: 0.1 response_residual_dropout: 0.0 layer_norm_type: rms layer_norm_with_affine: true layer_norm_eps: 1.0e-06 attention_layer_norm_with_affine: true max_sequence_length: 4096 max_position_embeddings: null include_bias: false bias_for_layer_norm: null norm_after: false moe_num_experts: 8 moe_top_k: 2 moe_mlp_impl: sparse moe_log_expert_assignment: false moe_shared_expert: false moe_lbl_in_fp32: false moe_interleave: false moe_loss_weight: 0.1 moe_zloss_weight: null moe_dropless: true moe_capacity_factor: 1.25 embedding_dropout: 0.0 scale_logits: false vocab_size: 151936 additional_vocab_size: 128 weight_tying: false embedding_size: 151936 use_position_ids: true tokenizer: identifier: Qwen/Qwen3-8B tokenizer_dir: null init_path: /weka/oe-training-default/mm-olmo/pretrained_llms/qwen3-8b.pt init_incremental: null new_embedding_init_range: 0.02 initializer_range: 0.02 normalize_input_embeds: false activation_checkpoint: whole_layer compile: blocks fix_pad_tokenizer: false init_std: 0.02 init_fn: normal init_cutoff_factor: null vision_backbone: vit: image_model_type: siglip image_default_input_size: - 378 - 378 image_patch_size: 14 image_pos_patch_size: 14 image_emb_dim: 1152 image_num_heads: 16 image_num_key_value_heads: 16 image_num_layers: 27 image_head_dim: 72 image_mlp_dim: 4304 image_mlp_activations: gelu_pytorch_tanh image_dropout_rate: 0.0 image_num_pos: 729 image_norm_eps: 1.0e-06 attention_dropout: 0.0 residual_dropout: 0.0 initializer_range: 0.02 float32_attention: true attention_type: sdpa activation_checkpointing: true init_path: /weka/oe-training-default/mm-olmo/pretrained_image_encoders/siglip2-so400m-14-384.pt resize_mode: siglip pad_value: 0.0 normalize: siglip image_pooling_2d: attention_meanq pooling_attention_mask: false image_projector: mlp image_padding_embed: null vit_layers: - -3 - -9 skip_unused_layers: true image_feature_dropout: 0.0 connector_activation_checkpointing: true compile_vit: blocks compile_connector: dynamic normalize_on_gpu: false data_formatter: prompt_templates: uber_model message_format: role system_prompt: demo_or_style always_start_with_space: false default_inference_len: 65 select_answer: best debug: false image_last: false format_message_list: null p_one_message: 0.0 eval_system_prompt_mapping: null timestamp_mode: 50-percent-seconds p_choice_content_in_mc: 1.0 mm_preprocessor: crop_mode: overlap-and-resize-c2 use_col_tokens: true max_crops: 8 pooling_w: 2 pooling_h: 2 overlap_margins: - 4 - 4 max_images: null max_multi_image_crops: 4 max_answer_len: null last_message_loss_only: false loss_token_weighting: root_subsegments max_text_tokens: null image_padding_mask: false legacy_image_mask: false bi_directional_attn: null seed: 6198 epoch: null dry_run: false ft_llm: true ft_vit: true ft_connector: true ft_embedding: lm_head optimizer: name: adamw learning_rate: 0.0001 weight_decay: 0.01 betas: - 0.9 - 0.95 eps: 1.0e-05 connector_learning_rate: 5.0e-06 vit_learning_rate: 5.0e-06 llm_learning_rate: 1.0e-05 frame_selector_learning_rate: 0.0001 temporal_token_scorer_learning_rate: 0.0001 connector_weight_decay: 0.0 vit_weight_decay: 0.0 llm_weight_decay: 0.0 frame_selector_weight_decay: 0.01 temporal_token_scorer_weight_decay: 0.01 connector_betas: - 0.9 - 0.95 vit_betas: - 0.9 - 0.95 llm_betas: - 0.9 - 0.95 frame_selector_betas: - 0.9 - 0.95 temporal_token_scorer_betas: - 0.9 - 0.95 connector_eps: 1.0e-06 vit_eps: 1.0e-06 llm_eps: 1.0e-06 frame_selector_eps: 1.0e-06 temporal_token_scorer_eps: 1.0e-06 metrics_log_interval: -1 scheduler: name: multimodal units: steps t_warmup: 100 t_max: null alpha_f: 0.1 connector_t_warmup: 200 vit_t_warmup: 200 llm_t_warmup: 200 frame_selector_t_warmup: 200 temporal_token_scorer_t_warmup: 200 grad_clip_warmup_steps: null grad_clip_warmup_factor: null warmup_min_lr: 0.0 data: dataset: null mixture: null root_size_mixture: - rate: 0.15 mixture: pixmo_ask_model_anything: null pixmo_cap: 50000.0 pixmo_cap_qa_as_user_qa: null pixmo_pointing_explanations: null - rate: 0.5 mixture: coco_2014_vqa_multi: null text_vqa: null okvqa: null chart_qa_weighted: null doc_qa: null info_qa: null ai2_diagram_v2_mix_transparent: null a_okvqa_mc: null a_okvqa_da: null android_control: null science_qa_img: null tabwmp_da: null st_qa: null tally_qa: null pixmo_clocks: 250000.0 dv_qa: 10000.0 figure_qa: 10000.0 plot_qa: 20000.0 cosyn_chart_exp: null cosyn_chemical_exp: null cosyn_diagram_exp: null cosyn_document: null cosyn_math_exp: null cosyn_music_exp: null cosyn_table_exp: null - rate: 0.35 mixture: pixmo_points_train: null pixmo_count_train: null pixmo_points_high_freq_train: null cosyn_point: null kwargs_mixture: null split: train seed: 50189 pad: to_max sequence_length: 2304 max_text_seq_len: null shuffle: true start_index: 0 packing: null num_workers: 2 drop_last: true pin_memory: true prefetch_factor: null persistent_workers: false timeout: 0 restore_dataloader: true fast_forward_batches: null evaluators: [] eval_interval: 2000 inf_evaluators: - label: chart_qa data: dataset: chart_qa mixture: null root_size_mixture: null kwargs_mixture: null split: validation seed: 691203 pad: to_max sequence_length: 1792 max_text_seq_len: null shuffle: true start_index: 0 packing: null num_workers: 2 drop_last: true pin_memory: true prefetch_factor: null persistent_workers: true timeout: 0 evaluator: n_to_log: 0 num_wandb_examples: 32 save_predictions: null save_tokens: false vqa_eval: relaxed_correctness,scifi_relaxed_correctness,em pointing_eval: false count_eval: false point_count_eval: false android_eval: false clock_eval: false clock_bench_eval: false math_vista_eval: false temp_compass_eval: '' temp_compass_disable_api: false video_mme_eval: '' mlvu_gen_eval: false long_video_bench_eval: false plm_fgqa_eval: false long_video_bench_caption_eval: false vinoground_eval: false vixmo_caption_eval: false refexp_eval: false coco_caption_eval: false qv_highlights_eval: false tomato: false temporal_bench: false max_new_tokens: 12 device_batch_size: 4 subset_num_batches: null max_examples: 2048 console_log_interval: 20 include_image: false - label: chart_qa_exp data: dataset: chart_qa_exp mixture: null root_size_mixture: null kwargs_mixture: null split: validation seed: 691203 pad: to_max sequence_length: 1792 max_text_seq_len: null shuffle: true start_index: 0 packing: null num_workers: 2 drop_last: true pin_memory: true prefetch_factor: null persistent_workers: true timeout: 0 evaluator: n_to_log: 0 num_wandb_examples: 32 save_predictions: null save_tokens: false vqa_eval: relaxed_correctness,scifi_relaxed_correctness,em pointing_eval: false count_eval: false point_count_eval: false android_eval: false clock_eval: false clock_bench_eval: false math_vista_eval: false temp_compass_eval: '' temp_compass_disable_api: false video_mme_eval: '' mlvu_gen_eval: false long_video_bench_eval: false plm_fgqa_eval: false long_video_bench_caption_eval: false vinoground_eval: false vixmo_caption_eval: false refexp_eval: false coco_caption_eval: false qv_highlights_eval: false tomato: false temporal_bench: false max_new_tokens: 256 device_batch_size: 4 subset_num_batches: null max_examples: 2048 console_log_interval: 20 include_image: false - label: info_qa data: dataset: info_qa mixture: null root_size_mixture: null kwargs_mixture: null split: validation seed: 691203 pad: to_max sequence_length: 1792 max_text_seq_len: null shuffle: true start_index: 0 packing: null num_workers: 2 drop_last: true pin_memory: true prefetch_factor: null persistent_workers: true timeout: 0 evaluator: n_to_log: 0 num_wandb_examples: 32 save_predictions: null save_tokens: false vqa_eval: ansl,em pointing_eval: false count_eval: false point_count_eval: false android_eval: false clock_eval: false clock_bench_eval: false math_vista_eval: false temp_compass_eval: '' temp_compass_disable_api: false video_mme_eval: '' mlvu_gen_eval: false long_video_bench_eval: false plm_fgqa_eval: false long_video_bench_caption_eval: false vinoground_eval: false vixmo_caption_eval: false refexp_eval: false coco_caption_eval: false qv_highlights_eval: false tomato: false temporal_bench: false max_new_tokens: 12 device_batch_size: 4 subset_num_batches: null max_examples: 2048 console_log_interval: 20 include_image: false - label: doc_qa data: dataset: doc_qa mixture: null root_size_mixture: null kwargs_mixture: null split: validation seed: 691203 pad: to_max sequence_length: 1792 max_text_seq_len: null shuffle: true start_index: 0 packing: null num_workers: 2 drop_last: true pin_memory: true prefetch_factor: null persistent_workers: true timeout: 0 evaluator: n_to_log: 0 num_wandb_examples: 32 save_predictions: null save_tokens: false vqa_eval: ansl,em pointing_eval: false count_eval: false point_count_eval: false android_eval: false clock_eval: false clock_bench_eval: false math_vista_eval: false temp_compass_eval: '' temp_compass_disable_api: false video_mme_eval: '' mlvu_gen_eval: false long_video_bench_eval: false plm_fgqa_eval: false long_video_bench_caption_eval: false vinoground_eval: false vixmo_caption_eval: false refexp_eval: false coco_caption_eval: false qv_highlights_eval: false tomato: false temporal_bench: false max_new_tokens: 12 device_batch_size: 4 subset_num_batches: null max_examples: 2048 console_log_interval: 20 include_image: false - label: ai2_diagram data: dataset: ai2_diagram_v2_mix_transparent mixture: null root_size_mixture: null kwargs_mixture: null split: validation seed: 691203 pad: to_max sequence_length: 1792 max_text_seq_len: null shuffle: true start_index: 0 packing: null num_workers: 2 drop_last: true pin_memory: true prefetch_factor: null persistent_workers: true timeout: 0 evaluator: n_to_log: 0 num_wandb_examples: 32 save_predictions: null save_tokens: false vqa_eval: mc_ai2d_opaque,mc_ai2d_transparent pointing_eval: false count_eval: false point_count_eval: false android_eval: false clock_eval: false clock_bench_eval: false math_vista_eval: false temp_compass_eval: '' temp_compass_disable_api: false video_mme_eval: '' mlvu_gen_eval: false long_video_bench_eval: false plm_fgqa_eval: false long_video_bench_caption_eval: false vinoground_eval: false vixmo_caption_eval: false refexp_eval: false coco_caption_eval: false qv_highlights_eval: false tomato: false temporal_bench: false max_new_tokens: 12 device_batch_size: 4 subset_num_batches: null max_examples: 2048 console_log_interval: 20 include_image: false - label: coco_2014_vqa data: dataset: coco_2014_vqa mixture: null root_size_mixture: null kwargs_mixture: null split: validation seed: 691203 pad: to_max sequence_length: 1792 max_text_seq_len: null shuffle: true start_index: 0 packing: null num_workers: 2 drop_last: true pin_memory: true prefetch_factor: null persistent_workers: true timeout: 0 evaluator: n_to_log: 0 num_wandb_examples: 32 save_predictions: null save_tokens: false vqa_eval: vqa_score pointing_eval: false count_eval: false point_count_eval: false android_eval: false clock_eval: false clock_bench_eval: false math_vista_eval: false temp_compass_eval: '' temp_compass_disable_api: false video_mme_eval: '' mlvu_gen_eval: false long_video_bench_eval: false plm_fgqa_eval: false long_video_bench_caption_eval: false vinoground_eval: false vixmo_caption_eval: false refexp_eval: false coco_caption_eval: false qv_highlights_eval: false tomato: false temporal_bench: false max_new_tokens: 12 device_batch_size: 4 subset_num_batches: null max_examples: 2048 console_log_interval: 20 include_image: false - label: pixmo_clocks data: dataset: pixmo_clocks mixture: null root_size_mixture: null kwargs_mixture: null split: validation seed: 691203 pad: to_max sequence_length: 1792 max_text_seq_len: null shuffle: true start_index: 0 packing: null num_workers: 2 drop_last: true pin_memory: true prefetch_factor: null persistent_workers: true timeout: 0 evaluator: n_to_log: 0 num_wandb_examples: 32 save_predictions: null save_tokens: false vqa_eval: '' pointing_eval: false count_eval: false point_count_eval: false android_eval: false clock_eval: true clock_bench_eval: false math_vista_eval: false temp_compass_eval: '' temp_compass_disable_api: false video_mme_eval: '' mlvu_gen_eval: false long_video_bench_eval: false plm_fgqa_eval: false long_video_bench_caption_eval: false vinoground_eval: false vixmo_caption_eval: false refexp_eval: false coco_caption_eval: false qv_highlights_eval: false tomato: false temporal_bench: false max_new_tokens: 12 device_batch_size: 4 subset_num_batches: null max_examples: 2048 console_log_interval: 20 include_image: false - label: android_control_ll data: dataset: android_control_ll mixture: null root_size_mixture: null kwargs_mixture: null split: validation seed: 691203 pad: to_max sequence_length: 1792 max_text_seq_len: null shuffle: true start_index: 0 packing: null num_workers: 2 drop_last: true pin_memory: true prefetch_factor: null persistent_workers: true timeout: 0 evaluator: n_to_log: 0 num_wandb_examples: 32 save_predictions: null save_tokens: false vqa_eval: '' pointing_eval: false count_eval: false point_count_eval: false android_eval: true clock_eval: false clock_bench_eval: false math_vista_eval: false temp_compass_eval: '' temp_compass_disable_api: false video_mme_eval: '' mlvu_gen_eval: false long_video_bench_eval: false plm_fgqa_eval: false long_video_bench_caption_eval: false vinoground_eval: false vixmo_caption_eval: false refexp_eval: false coco_caption_eval: false qv_highlights_eval: false tomato: false temporal_bench: false max_new_tokens: 16 device_batch_size: 4 subset_num_batches: null max_examples: 2048 console_log_interval: 20 include_image: false - label: pointing_eval data: dataset: pointing_eval mixture: null root_size_mixture: null kwargs_mixture: null split: test seed: 691203 pad: to_max sequence_length: 1792 max_text_seq_len: null shuffle: true start_index: 0 packing: null num_workers: 2 drop_last: true pin_memory: true prefetch_factor: null persistent_workers: true timeout: 0 evaluator: n_to_log: 0 num_wandb_examples: 32 save_predictions: null save_tokens: false vqa_eval: '' pointing_eval: true count_eval: false point_count_eval: false android_eval: false clock_eval: false clock_bench_eval: false math_vista_eval: false temp_compass_eval: '' temp_compass_disable_api: false video_mme_eval: '' mlvu_gen_eval: false long_video_bench_eval: false plm_fgqa_eval: false long_video_bench_caption_eval: false vinoground_eval: false vixmo_caption_eval: false refexp_eval: false coco_caption_eval: false qv_highlights_eval: false tomato: false temporal_bench: false max_new_tokens: 192 device_batch_size: 4 subset_num_batches: null max_examples: 2048 console_log_interval: 20 include_image: false inf_eval_interval: 2000 eval_on_last_step: true eval_on_load: false save_folder: /weka/oe-training-default/sanghol/molmo/models/uber-v1/uber3.4-synthetic-siglip2-qwen3_8b checkpointer_config: save_thread_count: null load_thread_count: null pre_download: false work_dir: null throttle_uploads: false canceled_check_interval: 50 save_interval: 1000 save_at: null save_final_optim: true save_num_checkpoints_to_keep: 1 save_final_unsharded_checkpoint: false save_interval_ephemeral: null save_overwrite: true load_path: null reset_optimizer_state: false reset_trainer_state: false initial_model_checkpoint: /weka/oe-training-default/sanghol/molmo/models/dense-cap-v1/captioner-qwen3_8b/step22347 allow_resume: true max_duration: 30000 global_train_batch_size: 256 device_train_microbatch_size: 4 max_grad_norm: 1.0 multi_component_grad_norm: true batch_divisor: global_batch max_grad_norm_ratio: null precision: amp_bf16 wandb: project: molmo2-dev entity: prior-ai2 group: uber-v1 name: uber3.4-synthetic-siglip2-qwen3_8b tags: - watching log_artifacts: false rank_zero_only: true log_interval: 20 allow_resume: false beaker_log_interval: 50 speed_monitor: window_size: 20 gpu_flops_available: null console_log_interval: 20 gen1_gc_interval: 1 compile: mode: default fullgraph: false dynamic: false backend: inductor activation_checkpointing: true fsdp: fsdp2: true precision: float use_orig_params: true wrapping_strategy: by_block_and_size sharding_strategy: FULL_SHARD hybrid_sharding_num_model_replicas: null softmax_auxiliary_loss: true softmax_auxiliary_loss_scale: 0.0001 saliency_score_loss_wt: null frame_score_loss_wt: null frame_score_loss_type: mse frame_score_loss_target: 0.7 time_limit: null extra_steps_after_cancel: 10 python_profiling: false torch_profiling: false stop_at: 30000 stop_after: null fused_loss: false compile_loss: true runtime_data: args: /gantry-runtime/launch_scripts/train_multitask_model.py 3.4-synthetic /weka/oe-training-default/sanghol/molmo/models/dense-cap-v1/captioner-qwen3_8b/step22347 --save_overwrite --save_interval=1000 --wandb.group=uber-v1 --wandb.name=uber3.4-synthetic-siglip2-qwen3_8b --save_folder=/weka/oe-training-default/sanghol/molmo/models/uber-v1/uber3.4-synthetic-siglip2-qwen3_8b hostname: jupiter-cs-aus-109.reviz.ai2.in date: 09/13/2025, 07:25 world_size: 32 resuming_from: null beaker_experiment_id: 01K50Z8VKCVN9Z4X917WEEP9VQ beaker_experiment_url: https://beaker.org/ex/01K50Z8VKCVN9Z4X917WEEP9VQ wandb_id: xhs8xbpk wandb_url: https://wandb.ai/prior-ai2/molmo2-dev/runs/xhs8xbpk