info_pool_byHJ / README.md
Heeyewon's picture
Upload folder using huggingface_hub
752aed8 verified
|
Raw
History Blame Contribute Delete
2.44 kB
metadata
configs:
  - config_name: ON_SCREEN_TEXT_SCOREBOARD_READOUT_READ
    data_files:
      - split: test
        path: ON_SCREEN_TEXT_SCOREBOARD_READOUT_READ/test-*.parquet
  - config_name: PERSON_ON_SCREEN_CLOSEUP_IN_BLOCK_RETRIEVE
    data_files:
      - split: test
        path: PERSON_ON_SCREEN_CLOSEUP_IN_BLOCK_RETRIEVE/test-*.parquet
  - config_name: SCORING_PLAY_LINK
    data_files:
      - split: test
        path: SCORING_PLAY_LINK/test-*.parquet
  - config_name: PITCH_IN_BASE_OUT_STATE_RETRIEVE
    data_files:
      - split: test
        path: PITCH_IN_BASE_OUT_STATE_RETRIEVE/test-*.parquet
  - config_name: BROADCAST_STRUCTURE_INNING_FLOW_SEGMENT
    data_files:
      - split: test
        path: BROADCAST_STRUCTURE_INNING_FLOW_SEGMENT/test-*.parquet

SME eval information pool -- staging additions

Row schema matches twelvelabs/sme_eval_information_pool exactly (id / media / messages / metadata.media_metadata / metadata.sample_metadata[0] with domain, specific_domain, segment_dict, user_query_segment, chapters, metadata_schema, annotation_info, derivation).

Stored as test-*.parquet per config, matching the TwelveLabs convention: id is a string, media is list<struct<media_path, region, type>>, and messages/metadata are JSON-encoded strings (not native structs) -- json.loads() them after reading the parquet.

Config names mirror <Segment ID, upper-cased, dots -> underscores>_<Category>, the same convention used by the target dataset's existing configs (e.g. ON_SCREEN_TEXT_SCOREBOARD_READOUT_READ).

config Segment ID rows
ON_SCREEN_TEXT_SCOREBOARD_READOUT_READ on_screen_text.scoreboard_readout 30
PERSON_ON_SCREEN_CLOSEUP_IN_BLOCK_RETRIEVE person_on_screen.closeup_in_block 20
SCORING_PLAY_LINK scoring_play 30
PITCH_IN_BASE_OUT_STATE_RETRIEVE pitch.in_base_out_state 35
BROADCAST_STRUCTURE_INNING_FLOW_SEGMENT broadcast_structure.inning_flow 7

ON_SCREEN_TEXT_SCOREBOARD_READOUT_READ is a config that already exists in the target dataset -- merging that one is a straight row-append. The other four configs are new and don't exist in the target dataset yet, so merging them means adding new configs, not appending to existing ones. Confirm the exact target config names with the eval team before merging, in case their internal naming differs from this best-effort mapping.

nothing_here.bases_empty (scenario 10) has no finalized JSON file yet and is not included here.