| { |
| "same_20_samples": true, |
| "same_captions": true, |
| "only_request_change": "body.model", |
| "same_prompt_sha256": "f4f2c370eab558f31c5e195fe1953a6232ee5bab77e1c89e5fed96768dd56213", |
| "same_260_image_inputs": true, |
| "same_schema_and_reasoning_settings": true, |
| "sol_statistics": { |
| "requested_original_videos": 20, |
| "returned_annotations": 20, |
| "api_or_parse_failures": 0, |
| "distinct_returned_original_videos": 20, |
| "structurally_valid_samples": 20, |
| "complete_correspondence_candidates": 0, |
| "partial_or_needs_review_samples": 20, |
| "total_text_groups": 107, |
| "total_text_mentions": 188, |
| "groups_by_grounding_status": { |
| "matched": 24, |
| "missing_mask": 52, |
| "not_visible_in_reference": 21, |
| "mask_quality_issue": 7, |
| "ambiguous": 2, |
| "requires_mask_union": 1 |
| }, |
| "matched_groups_exported": 24, |
| "groups_by_entity_kind": { |
| "thing": 66, |
| "stuff": 14, |
| "collective": 18, |
| "body_part": 9 |
| }, |
| "source_masks": 32, |
| "all_captions_preserved": true, |
| "structural_errors": [], |
| "local_adjustments_by_type": { |
| "unconfirmed_quality_mask_to_candidate": 3, |
| "unique_whole_token_occurrence_resolution": 2 |
| }, |
| "usage": { |
| "input_tokens": 326855, |
| "output_tokens": 23152, |
| "total_tokens": 350007, |
| "reasoning_tokens": 4239, |
| "cached_input_tokens": 180500 |
| }, |
| "human_reviewed_samples": 0, |
| "batch_id": "batch_6ae4c59512d08190928be89ef2e23f08", |
| "prompt_sha256": "f4f2c370eab558f31c5e195fe1953a6232ee5bab77e1c89e5fed96768dd56213", |
| "requests_sha256": "93fe0de6dbb058a13286c72dca789cee3b7c6a696f64656d3d02ebd1e6ff343a", |
| "no_claim_of_full_video_mask_tracks": true, |
| "completion_scope": "All 20 model annotations are retained. Complete-candidate and partial/review exports are separate. Semantic accuracy is not certified by structural checks." |
| }, |
| "luna_statistics": { |
| "requested_original_videos": 20, |
| "returned_annotations": 20, |
| "api_or_parse_failures": 0, |
| "distinct_returned_original_videos": 20, |
| "structurally_valid_samples": 17, |
| "complete_correspondence_candidates": 0, |
| "partial_or_needs_review_samples": 20, |
| "total_text_groups": 109, |
| "total_text_mentions": 187, |
| "groups_by_grounding_status": { |
| "matched": 23, |
| "missing_mask": 51, |
| "not_visible_in_reference": 24, |
| "mask_quality_issue": 9, |
| "ambiguous": 1, |
| "requires_mask_union": 1 |
| }, |
| "matched_groups_exported": 20, |
| "groups_by_entity_kind": { |
| "thing": 70, |
| "stuff": 12, |
| "collective": 17, |
| "body_part": 10 |
| }, |
| "source_masks": 32, |
| "all_captions_preserved": true, |
| "structural_errors": [ |
| { |
| "sample_id": "v_xbA294b3Z5c____1", |
| "errors": [ |
| "g5:mention_inside_word:'it':1" |
| ] |
| }, |
| { |
| "sample_id": "v_iksziMLKsBQ____2", |
| "errors": [ |
| "g4:within_group_overlap:[77, 91]:[86, 91]" |
| ] |
| }, |
| { |
| "sample_id": "v_XLspZ5bof7o____1", |
| "errors": [ |
| "g2:within_group_overlap:[58, 79]:[71, 79]" |
| ] |
| } |
| ], |
| "local_adjustments_by_type": { |
| "unique_whole_token_occurrence_resolution": 3 |
| }, |
| "usage": { |
| "input_tokens": 326855, |
| "output_tokens": 52595, |
| "total_tokens": 379450, |
| "reasoning_tokens": 35163, |
| "cached_input_tokens": 171000 |
| }, |
| "human_reviewed_samples": 0, |
| "run_id": "gcg-object20-luna-direct-27164e5005fc45de97fc9bf2296f9f74", |
| "api_transport": "direct_responses_api", |
| "prompt_sha256": "f4f2c370eab558f31c5e195fe1953a6232ee5bab77e1c89e5fed96768dd56213", |
| "requests_sha256": "4211af1e82619dd38c6dbc110e1c1665b604a5a753232cf7675ef593b61c20a7", |
| "no_claim_of_full_video_mask_tracks": true, |
| "completion_scope": "All 20 model annotations are retained. Complete-candidate and partial/review exports are separate. Semantic accuracy is not certified by structural checks." |
| }, |
| "note": "Different entity grouping and counts are model outputs, not an accuracy benchmark. Neither run is human certified." |
| } |
