stage3.yaml

datasets:
# text
  - json_path: ./datasets/annotations/texts/evol_instruct_gpt_turbo.jsonl # 143k * 0.5
    sampling_strategy: "first:2%"

# image
  - json_path: ./datasets/annotations/images/allava_instruct_laion4v.jsonl # 49817
    sampling_strategy: "first:5%"
    data_root: ./datasets/images/allava_instruct_laion4v/

  - json_path: ./datasets/annotations/images/allava_instruct_vflan4v.jsonl # 19884
    sampling_strategy: "first:5%"
    data_root: ./datasets/images/allava_instruct_vflan4v/

  - json_path: ./datasets/annotations/images/sharegpt4o.jsonl # 129950
    sampling_strategy: "first:5%"
    data_root: ./datasets/images/sharegpt4o/

  - json_path: ./datasets/annotations/images/llava_instruct_processed.jsonl # 369400 * 0.5
    sampling_strategy: "first:2%"
    data_root: ./datasets/images/llava_instruct/

  - json_path: ./datasets/annotations/images/textcaps.jsonl # 21877
    sampling_strategy: "first:5%"
    data_root: ./datasets/images/textcaps/

  - json_path: ./datasets/annotations/images/textocr_gpt4v.jsonl # 25105
    sampling_strategy: "first:5%"
    data_root: ./datasets/images/textocr_gpt4v/

  - json_path: ./datasets/annotations/images/rendered_text.jsonl # 18921
    sampling_strategy: "first:5%"
    data_root: ./datasets/images/rendered_text/

  - json_path: ./datasets/annotations/images/iam.jsonl # 5782
    sampling_strategy: "first:5%"
    data_root: ./datasets/images/iam/

  - json_path: ./datasets/annotations/images/llavar.jsonl # 19796
    sampling_strategy: "first:5%"
    data_root: ./datasets/images/llavar/

  - json_path: ./datasets/annotations/images/st_vqa.jsonl # 32307
    sampling_strategy: "first:5%"
    data_root: ./datasets/images/st_vqa/

  - json_path: ./datasets/annotations/images/tallyqa_processed.jsonl # 50k * 0.2
    sampling_strategy: "first:1%"
    data_root: ./datasets/images/tallyqa/

  - json_path: ./datasets/annotations/images/m4_instruct_annotations_processed.json # 600k
    sampling_strategy: "first:5%"
    data_root: ./datasets/images/M4-Instruct-Data/

# video
  # caption, 186893
  # all mp4, or jpeg/jpg, only has id, video, conversations. and len(conversations) == 2, num_images = 1
  - json_path: ./datasets/annotations/videos/caption_sharegptvideo_300k-sharegptvideo-train_300k_302k_first20.json # frame / jpeg / 20% 60000
    sampling_strategy: "first:5%"
    data_root: ./datasets/videos/ShareGPTVideo60k/

  - json_path: ./datasets/annotations/videos/caption_vidln_kinetics-vidln-kinetics_train_28k_filtered.json # 27215
    sampling_strategy: "first:5%"
    data_root: ./datasets/videos/caption_vidln_kinetics-vidln-kinetics_train_28k/
  - json_path: ./datasets/annotations/videos/caption_vidln_oops-vidln-oops_train_11k.json # 10697
    sampling_strategy: "first:5%"
    data_root: ./datasets/videos/caption_vidln_oops-vidln-oops_train_11k/
  - json_path: ./datasets/annotations/videos/caption_vidln_ovis-vidln-ovis_train_1k.json # frame / jpg 610
    sampling_strategy: "first:5%"
    data_root: ./datasets/videos/caption_vidln_ovis-vidln-ovis_train_1k/
  - json_path: ./datasets/annotations/videos/caption_vidln_uvo_dense-vidln-uvo_dense_train_1k.json # 563
    sampling_strategy: "first:5%"
    data_root: ./datasets/videos/caption_vidln_uvo_dense-vidln-uvo_dense_train_1k/
  - json_path: ./datasets/annotations/videos/caption_vidln_uvo_sparse-vidln-uvo_sparse_train_6k.json # 5702
    sampling_strategy: "first:5%"
    data_root: ./datasets/videos/caption_vidln_uvo_sparse-vidln-uvo_sparse_train_6k/

  - json_path: ./datasets/annotations/videos/caption_sharegpt4o-sharegpt4o_3k.json # 2111
    sampling_strategy: "first:5%"
    data_root: ./datasets/videos/caption_sharegpt4o-sharegpt4o_3k/
  - json_path: ./datasets/annotations/videos/caption_youcook2-youcook2-train_debug_9k.json # 8700
    sampling_strategy: "first:5%"
    data_root: ./datasets/videos/caption_youcook2-youcook2-train_debug_9k/
  - json_path: ./datasets/annotations/videos/caption_textvr-textvr-train_40k.json # 39648
    sampling_strategy: "first:5%"
    data_root: ./datasets/videos/TextVR/

  - json_path: ./datasets/annotations/videos/vript_short_videos_en_20240911_fix.jsonl # 8776
    sampling_strategy: "first:5%"
    data_root: ./datasets/videos/Vript/vript_short_videos_clips/
  - json_path: ./datasets/annotations/videos/vript_long_videos_en_20240911_split_30_60s_processed.jsonl # 16137
    sampling_strategy: "first:5%"
    data_root: ./datasets/videos/Vript/vript_long_videos_clips/30_60s/
  - json_path: ./datasets/annotations/videos/vript_long_videos_en_20240911_split_60_120s_processed.jsonl # 3644
    sampling_strategy: "all"
    data_root: ./datasets/videos/Vript/vript_long_videos_clips/60_120s/
  - json_path: ./datasets/annotations/videos/vript_long_videos_en_20240911_split_120_180s_processed.jsonl # 773
    sampling_strategy: "all"
    data_root: ./datasets/videos/Vript/vript_long_videos_clips/120_180s/
  - json_path: ./datasets/annotations/videos/vript_long_videos_en_20240911_split_180s_plus_processed.jsonl # 727
    sampling_strategy: "all"
    data_root: ./datasets/videos/Vript/vript_long_videos_clips/180s_plus/ # 0.84

  - json_path: ./datasets/annotations/videos/moviechat1k_caption-MovieChat-train_caption_1k_filtered.json # 795
    sampling_strategy: "all"
    data_root: ./datasets/videos/MovieChat-1K_train_fixed/ # 4.01 min
  - json_path: ./datasets/annotations/videos/moviechat1k_global-MovieChat-train_global_1k_filtered.jsonl # muti-turn, num_image =1 795
    sampling_strategy: "all"
    data_root: ./datasets/videos/MovieChat-1K_train_fixed/

  - json_path: ./datasets/annotations/videos/caption_sharegemini_k400_223k.json # 223k
    sampling_strategy: "first:5%"
    data_root: ./datasets/videos/Kinetics-400/

  # vqa, 412243
  # all mp4, or avi or gif. and len(conversations) == 2, num_images = 1, star dataset has start, end
  - json_path: ./datasets/annotations/videos/vqa_sharegptvideo_240k-sharegptvideo-train_240k_240k_first25.json # 60000
    sampling_strategy: "first:5%"
    data_root: ./datasets/videos/ShareGPTVideo60k/

  # reasoning
  - json_path: ./datasets/annotations/videos/reasoning_star-star-train_46k.json # with start, end # 45731
    sampling_strategy: "first:5%"
    data_root: ./datasets/videos/reasoning_star-star-train_46k/
  - json_path: ./datasets/annotations/videos/reasoning_next_qa_oe-openend_qa_annos-next_qa_train_openend_35k.jsonl # 34132
    sampling_strategy: "first:5%"
    data_root: ./datasets/videos/NExT-QA/
  - json_path: ./datasets/annotations/videos/reasoning_next_qa-next_qa-train_35k.jsonl # 34132
    sampling_strategy: "first:5%"
    data_root: ./datasets/videos/NExT-QA/
  - json_path: ./datasets/annotations/videos/reasoning_clevrer_mc-clevrer_mc-train_43k_debug_43k.jsonl # 42620
    sampling_strategy: "first:5%"
    data_root: ./datasets/videos/clevrer/
  - json_path: ./datasets/annotations/videos/reasoning_clevrer_qa-clevrer_qa-train_mc_40k.jsonl # 40000
    sampling_strategy: "first:5%"
    data_root: ./datasets/videos/clevrer/

  # general vqa
  - json_path: ./datasets/annotations/videos/vcg-plus_112K_clean_97k.json # 9645
    sampling_strategy: "first:1%"
    data_root: ./datasets/videos/vcg-plus_112K_clean_97k/
  - json_path: ./datasets/annotations/videos/vqa_rgbd-nturgbd_clean_110k_filtered.json # avi 27354
    sampling_strategy: "first:5%"
    data_root: ./datasets/videos/vqa_rgbd-nturgbd_clean_110k/
  - json_path: ./datasets/annotations/videos/vqa_perception_train-mc_question_train_forchoice_8k.json # 7392
    sampling_strategy: "first:5%"
    data_root: ./datasets/videos/perception_test/
  - json_path: ./datasets/annotations/videos/vqa_ego_qa-ego_qa-train_8k_filtered.jsonl # 7791
    sampling_strategy: "first:5%"
    data_root: ./datasets/videos/vqa_ego_qa-ego_qa-train_8k/
  - json_path: ./datasets/annotations/videos/vqa_tvqa-tvqa_123k_filtered.jsonl # 123k
    sampling_strategy: "first:5%"
    data_root: ./datasets/videos/tvqa/
  - json_path: ./datasets/annotations/videos/guiworld_en_20241029_fix.jsonl # 94998
    sampling_strategy: "first:5%"
    data_root: ./datasets/videos/GUI-World/

  # gif
  - json_path: ./datasets/annotations/videos/vqa_tgif_action-openend_qa_train_openend_20471.jsonl # 20471
    sampling_strategy: "first:5%"
    data_root: ./datasets/videos/tgif-qa/
  - json_path: ./datasets/annotations/videos/vqa_tgif_count-openend_qa_train_openend_26839.jsonl # 26839
    sampling_strategy: "first:5%"
    data_root: ./datasets/videos/tgif-qa/
  - json_path: ./datasets/annotations/videos/vqa_tgif_frame_qa-tgif_frame_qa-train_40k.jsonl # 9788
    sampling_strategy: "first:2%"
    data_root: ./datasets/videos/tgif-qa/
  - json_path: ./datasets/annotations/videos/vqa_tgif_transition_qa_openend-openend_qa_annos-tgif_transition_qa_train_openend_53k.jsonl # 13174
    sampling_strategy: "first:2%"
    data_root: ./datasets/videos/tgif-qa/
  - json_path: ./datasets/annotations/videos/vqa_tgif_transition_qa-tgif_transition_qa-train_53k.jsonl # 13174
    sampling_strategy: "first:2%"
    data_root: ./datasets/videos/tgif-qa/

  # classification, unzip in 0.43 min
  - json_path: ./datasets/annotations/videos/classification_k710-k710-train_40k_filtered.jsonl # 10000
    sampling_strategy: "first:5%"
    data_root: ./datasets/videos/classification_k710-k710-train_40k/
  - json_path: ./datasets/annotations/videos/classification_ssv2-ssv2-train_40k_filtered.jsonl # webm 10000
    sampling_strategy: "first:5%"
    data_root: ./datasets/videos/classification_ssv2-ssv2-train_40k/

  # lsmdc, unzip in 10.26 min, 74232
  - json_path: ./datasets/annotations/videos/lsmdc-lsmdc_297k_filtered.json # avi 74232
    sampling_strategy: "first:5%"
    data_root: ./datasets/videos/lsmdc-lsmdc_297k/

  # Grounding, 152416
  # all mp4. and len(conversations) == 2, num_images = 1
  - json_path: ./datasets/annotations/videos/grounding_ANetHL-ANet-HL-ANet_HL2_11k.json # 10340
    sampling_strategy: "all"
    data_root: ./datasets/videos/activity-net/
  - json_path: ./datasets/annotations/videos/grounding_ANetRTL-ActivityNet-RTL-ANet_RTL_34k.json # 33557
    sampling_strategy: "all"
    data_root: ./datasets/videos/activity-net/
  - json_path: ./datasets/annotations/videos/timeit_ANet-TimeIT-Activitynet_Captions_11k_filtered.json # 10009
    sampling_strategy: "all"
    data_root: ./datasets/videos/activity-net/

  - json_path: ./datasets/annotations/videos/grounding_didemo-didemo-train_66k.json # 65566
    sampling_strategy: "first:5%"
    data_root: ./datasets/videos/didemo/
  - json_path: ./datasets/annotations/videos/timeit_DiDeMo-TimeIT-DiDeMo_33k_filtered.json # 32944
    sampling_strategy: "first:5%"
    data_root: ./datasets/videos/didemo/

  # More data, 23549
  # all mp4 (except vitt). and len(conversations) == 2, num_images = 1
  - json_path: ./datasets/annotations/videos/timeit_COIN-TimeIT-COIN_10k_filtered.json # 9026
    sampling_strategy: "all"
    data_root: ./datasets/videos/timeit_COIN-TimeIT-COIN_10k/
  - json_path: ./datasets/annotations/videos/timeit_HiREST-TimeIT-HiREST_1k_filtered.json # 842
    sampling_strategy: "all"
    data_root: ./datasets/videos/timeit_HiREST-TimeIT-HiREST_1k/
  - json_path: ./datasets/annotations/videos/timeit_QuerYD-TimeIT-QuerYD_15k_filtered.json # 13791
    sampling_strategy: "all"
    data_root: ./datasets/videos/timeit_QuerYD-TimeIT-QuerYD_15k/
  - json_path: ./datasets/annotations/videos/timeit_ViTT-TimeIT-ViTT_6k_filtered.json # mp4, mkv, webm, slow decode 5085
    sampling_strategy: "all"
    data_root: ./datasets/videos/timeit_ViTT-TimeIT-ViTT_6k/

# longvid, 3870
  - json_path: ./datasets/annotations/videos/htstep_eventcount-longvideo_annos-htstep_eventcount_2k_2k.json # 1933
    sampling_strategy: "all"
    data_root: ./datasets/videos/htstep_eventcount_2k/ # 4.36 min
  - json_path: ./datasets/annotations/videos/htstep_eventrelationship-longvideo_annos-htstep_eventrelationship_1k_1k.json # 965
    sampling_strategy: "all"
    data_root: ./datasets/videos/htstep_eventrelationship_1k/ # 1.31 min
  - json_path: ./datasets/annotations/videos/htstep_eventunderstanding-longvideo_annos-htstep_eventunderstanding_1k_1k.json # 972
    sampling_strategy: "all"
    data_root: ./datasets/videos/htstep_eventunderstanding_1k/ # 1.22 min, has start, end time
  - json_path: ./datasets/annotations/videos/ego4dhcap_eventunderstanding-longvideo_annos-ego4dhcap_eventunderstanding_2k_2k_filtered.json # 2k
    sampling_strategy: "all"
    data_root: ./datasets/videos/ego4dhcap_eventunderstanding_2k/

# llava video
  - json_path: ./datasets/annotations/videos/llava_video/llava-video_2_3_m_academic_mc_v0_1_qa_processed_6901_with_duration.jsonl
    sampling_strategy: "first:25%"
    data_root: ./datasets/videos/LLaVA-Video-178K/
  - json_path: ./datasets/annotations/videos/llava_video/llava-video_2_3_m_nextqa_oe_qa_processed_61_with_duration.jsonl
    sampling_strategy: "first:25%"
    data_root: ./datasets/videos/LLaVA-Video-178K/
  - json_path: ./datasets/annotations/videos/llava_video/llava-video_0_30_s_youtube_oe_v0_1_qa_processed_420200_with_duration.jsonl
    sampling_strategy: "first:5%"
    data_root: ./datasets/videos/LLaVA-Video-178K/
  - json_path: ./datasets/annotations/videos/llava_video/llava-video_1_2_m_academic_oe_v0_1_qa_processed_26302_with_duration.jsonl
    sampling_strategy: "first:25%"
    data_root: ./datasets/videos/LLaVA-Video-178K/
  - json_path: ./datasets/annotations/videos/llava_video/llava-video_1_2_m_youtube_mc_v0_1_qa_processed_39710_with_duration.jsonl # 420188
    sampling_strategy: "first:25%"
    data_root: ./datasets/videos/LLaVA-Video-178K/
  - json_path: ./datasets/annotations/videos/llava_video/llava-video_30_60_s_nextqa_oe_qa_processed_6843_with_duration.jsonl
    sampling_strategy: "first:5%"
    data_root: ./datasets/videos/LLaVA-Video-178K/
  - json_path: ./datasets/annotations/videos/llava_video/llava-video_2_3_m_youtube_mc_v0_1_qa_processed_39967_with_duration.jsonl # 39951
    sampling_strategy: "first:25%"
    data_root: ./datasets/videos/LLaVA-Video-178K/
  - json_path: ./datasets/annotations/videos/llava_video/llava-video_2_3_m_academic_v0_1_cap_processed_3124_with_duration.jsonl
    sampling_strategy: "first:25%"
    data_root: ./datasets/videos/LLaVA-Video-178K/
  - json_path: ./datasets/annotations/videos/llava_video/llava-video_30_60_s_academic_oe_v0_1_qa_processed_57924_with_duration.jsonl
    sampling_strategy: "first:5%"
    data_root: ./datasets/videos/LLaVA-Video-178K/
  - json_path: ./datasets/annotations/videos/llava_video/llava-video_2_3_m_youtube_v0_1_cap_processed_24685_with_duration.jsonl
    sampling_strategy: "first:25%"
    data_root: ./datasets/videos/LLaVA-Video-178K/
  - json_path: ./datasets/annotations/videos/llava_video/llava-video_30_60_s_youtube_mc_v0_1_qa_processed_39927_with_duration.jsonl
    sampling_strategy: "first:5%"
    data_root: ./datasets/videos/LLaVA-Video-178K/
  - json_path: ./datasets/annotations/videos/llava_video/llava-video_0_30_s_activitynetqa_oe_qa_processed_2950_with_duration.jsonl
    sampling_strategy: "first:5%"
    data_root: ./datasets/videos/LLaVA-Video-178K/
  - json_path: ./datasets/annotations/videos/llava_video/llava-video_1_2_m_nextqa_oe_qa_processed_4694_with_duration.jsonl
    sampling_strategy: "first:25%"
    data_root: ./datasets/videos/LLaVA-Video-178K/
  - json_path: ./datasets/annotations/videos/llava_video/llava-video_30_60_s_youtube_oe_v0_1_qa_processed_110624_with_duration.jsonl
    sampling_strategy: "first:5%"
    data_root: ./datasets/videos/LLaVA-Video-178K/
  - json_path: ./datasets/annotations/videos/llava_video/llava-video_1_2_m_academic_mc_v0_1_qa_processed_4241_with_duration.jsonl
    sampling_strategy: "first:25%"
    data_root: ./datasets/videos/LLaVA-Video-178K/
  - json_path: ./datasets/annotations/videos/llava_video/llava-video_0_30_s_youtube_mc_v0_1_qa_processed_39353_with_duration.jsonl
    sampling_strategy: "first:5%"
    data_root: ./datasets/videos/LLaVA-Video-178K/
  - json_path: ./datasets/annotations/videos/llava_video/llava-video_30_60_s_activitynetqa_oe_qa_processed_4530_with_duration.jsonl
    sampling_strategy: "first:5%"
    data_root: ./datasets/videos/LLaVA-Video-178K/
  - json_path: ./datasets/annotations/videos/llava_video/llava-video_1_2_m_youtube_oe_v0_1_qa_processed_137645_with_duration.jsonl
    sampling_strategy: "first:25%"
    data_root: ./datasets/videos/LLaVA-Video-178K/
  - json_path: ./datasets/annotations/videos/llava_video/llava-video_30_60_s_academic_mc_v0_1_qa_processed_20346_with_duration.jsonl
    sampling_strategy: "first:5%"
    data_root: ./datasets/videos/LLaVA-Video-178K/
  - json_path: ./datasets/annotations/videos/llava_video/llava-video_30_60_s_youtube_v0_1_cap_processed_19995_with_duration.jsonl
    sampling_strategy: "first:5%"
    data_root: ./datasets/videos/LLaVA-Video-178K/
  - json_path: ./datasets/annotations/videos/llava_video/llava-video_0_30_s_nextqa_mc_qa_processed_5496_with_duration.jsonl
    sampling_strategy: "first:5%"
    data_root: ./datasets/videos/LLaVA-Video-178K/
  - json_path: ./datasets/annotations/videos/llava_video/llava-video_0_30_s_academic_mc_v0_1_qa_processed_5753_with_duration.jsonl
    sampling_strategy: "first:5%"
    data_root: ./datasets/videos/LLaVA-Video-178K/
  - json_path: ./datasets/annotations/videos/llava_video/llava-video_2_3_m_youtube_oe_v0_1_qa_processed_141495_with_duration.jsonl
    sampling_strategy: "first:25%"
    data_root: ./datasets/videos/LLaVA-Video-178K/
  - json_path: ./datasets/annotations/videos/llava_video/llava-video_1_2_m_nextqa_mc_qa_processed_4633_with_duration.jsonl
    sampling_strategy: "first:25%"
    data_root: ./datasets/videos/LLaVA-Video-178K/
  - json_path: ./datasets/annotations/videos/llava_video/llava-video_2_3_m_activitynetqa_oe_qa_processed_7460_with_duration.jsonl
    sampling_strategy: "first:25%"
    data_root: ./datasets/videos/LLaVA-Video-178K/
  - json_path: ./datasets/annotations/videos/llava_video/llava-video_2_3_m_nextqa_mc_qa_processed_52_with_duration.jsonl
    sampling_strategy: "first:25%"
    data_root: ./datasets/videos/LLaVA-Video-178K/
  - json_path: ./datasets/annotations/videos/llava_video/llava-video_1_2_m_activitynetqa_oe_qa_processed_8590_with_duration.jsonl
    sampling_strategy: "first:25%"
    data_root: ./datasets/videos/LLaVA-Video-178K/
  - json_path: ./datasets/annotations/videos/llava_video/llava-video_1_2_m_academic_v0_1_cap_processed_4627_with_duration.jsonl
    sampling_strategy: "first:25%"
    data_root: ./datasets/videos/LLaVA-Video-178K/
  - json_path: ./datasets/annotations/videos/llava_video/llava-video_30_60_s_academic_v0_1_cap_processed_10514_with_duration.jsonl
    sampling_strategy: "first:5%"
    data_root: ./datasets/videos/LLaVA-Video-178K/

  - json_path: ./datasets/annotations/videos/llava_video/llava-video_1_2_m_youtube_v0_1_cap_processed_24234_with_duration.jsonl
    sampling_strategy: "first:25%"
    data_root: ./datasets/videos/LLaVA-Video-178K/
  - json_path: ./datasets/annotations/videos/llava_video/llava-video_30_60_s_nextqa_mc_qa_processed_6843_with_duration.jsonl
    sampling_strategy: "first:5%"
    data_root: ./datasets/videos/LLaVA-Video-178K/
  - json_path: ./datasets/annotations/videos/llava_video/llava-video_0_30_s_nextqa_oe_qa_processed_5492_with_duration.jsonl
    sampling_strategy: "first:5%"
    data_root: ./datasets/videos/LLaVA-Video-178K/
  - json_path: ./datasets/annotations/videos/llava_video/llava-video_0_30_s_academic_oe_v0_1_qa_processed_48468_with_duration.jsonl
    sampling_strategy: "first:5%"
    data_root: ./datasets/videos/LLaVA-Video-178K/
  - json_path: ./datasets/annotations/videos/llava_video/llava-video_0_30_s_youtube_v0_1_cap_processed_79346_with_duration.jsonl
    sampling_strategy: "first:5%"
    data_root: ./datasets/videos/LLaVA-Video-178K/
  - json_path: ./datasets/annotations/videos/llava_video/llava-video_2_3_m_academic_oe_v0_1_qa_processed_18134_with_duration.jsonl
    sampling_strategy: "first:25%"
    data_root: ./datasets/videos/LLaVA-Video-178K/
  - json_path: ./datasets/annotations/videos/llava_video/llava-video_0_30_s_perceptiontest_mc_qa_processed_1785_with_duration.jsonl
    sampling_strategy: "first:5%"
    data_root: ./datasets/videos/LLaVA-Video-178K/
  - json_path: ./datasets/annotations/videos/llava_video/llava-video_30_60_s_perceptiontest_mc_qa_processed_618_with_duration.jsonl
    sampling_strategy: "first:5%"
    data_root: ./datasets/videos/LLaVA-Video-178K/
  - json_path: ./datasets/annotations/videos/llava_video/llava-video_0_30_s_academic_v0_1_cap_processed_11985_with_duration.jsonl
    sampling_strategy: "first:5%"
    data_root: ./datasets/videos/LLaVA-Video-178K/