# Changelog All notable changes to **`nvidia/PhysicalAI-VANTAGE-Bench`** on Hugging Face. ## 2026-05-27 - **`data/pointing/` migrated to JSONL as the canonical annotation format.** `data/pointing/VANTAGE_2DPointing.jsonl` (1,005 lines, one sample object per line, 8 fields: `index, question_id, image_path, question, A, B, C, D`) replaces `VANTAGE_2DPointing.tsv`, which has been removed. The JSONL is a lossless conversion of the prior TSV — same 1,005 samples, same field values, same row order, with `index` widened from string to integer. No sample content changed and no ground-truth fields (`answer`, `target_point`) were introduced. The Hugging Face Dataset Viewer `pointing` config now resolves cleanly alongside the other JSON/JSONL configs (the mixed `.tsv`/`.jsonl` `configs:` block was triggering HF's JSON builder on the TSV and failing with an `ArrowInvalid: JSON parse error`). - **`data/event_verification/data_jsons/annotations/*.json` unwrapped to top-level lists.** Each of the four files (`VANTAGE_EventVerification.json` — 67 entries, `tailgating_location_a.json` — 28, `tailgating_location_b.json` — 22, `warehouse_near_miss.json` — 46; 163 total) was rewritten from `{"bcq": [...]}` to `[...]`, matching the `vqa/data_jsons/annotations/` and `temporal_localization/data_jsons/annotations/` layouts. Every sample object (`{id, video, system_prompt, question}`) is preserved byte-for-byte and entry order is unchanged. This lets the Hugging Face Dataset Viewer row-expand the files to 163 (instead of collapsing each top-level object to a single row, which produced 4 rows in the viewer). - **`data/pointing/VANTAGE_2DPointing.tsv` updated.** MCQ option coordinates (columns `A`–`D`) are now expressed in the normalized `0–1000` coordinate system (each cell is an `x,y` pair, with both components in `[0, 1000]` relative to the image dimensions). The public TSV ships only the question side: columns are `index, question_id, image_path, question, A, B, C, D`; the ground-truth fields (`answer`, `target_point`) are held server-side and are not included in the released file. - **`data/event_verification/` flattened** to match the `vqa/` and `temporal_localization/` layout. The `filtered/.../{metropolis,tailgating,warehouse_near_miss}` subtree was removed; all 163 videos now live directly under `data/event_verification/videos/`, and the four annotation JSONs were moved + renamed to `data/event_verification/data_jsons/annotations/{metropolis_event_verification,tailgating_location_a,tailgating_location_b,warehouse_near_miss}.json`. Each `bcq[].video` is now the basename (e.g. `example.mp4`) and each `bcq[].id` is the stem (e.g. `example`); all other fields, entry order, and counts (163 total) are preserved. The `configs:` glob in the top-level README is updated accordingly. - Renamed remaining active Metropolis-named annotation JSON artifacts to VANTAGE naming for public dataset consistency: `data/event_verification/data_jsons/annotations/metropolis_event_verification.json` → `data/event_verification/data_jsons/annotations/VANTAGE_EventVerification.json`; `data/vqa/data_jsons/annotations/Metropolis_VQA_Verification_Final_ITS_Data.json` → `data/vqa/data_jsons/annotations/VANTAGE_VQA_Verification_Final_ITS_Data.json`. File contents are unchanged; only filenames moved. The `configs:` globs in the top-level README already match the new filenames. `data/README.md` updated to reference the new event-verification filename. - **`data/vqa/data_jsons/annotations/*.json` reduced to inference-oriented fields.** Each entry now carries exactly `{q_uid, question, options}`. The metadata fields `industry`, `event_type`, `start_time`, `end_time`, `video_duration`, `task_type`, and `dimension` were removed across all five files. Entry count (1,195) and values of the retained fields are unchanged. Smoke-tested against VLMEvalKit's `VANTAGE_VQA` inference preparation: TSV regeneration, prompt building, and video resolution all pass. - **`data/temporal_localization/data_jsons/annotations/*.json` reduced to inference-oriented fields.** Each entry now carries exactly `{vid, question_id, question}` (key order preserved). The metadata fields `industry`, `event_type`, `task_type`, and `duration` were removed across all three files. Entry count (1,067) and values of the retained fields are unchanged; the 1,067 `question_id`s remain globally unique. Smoke-tested against VLMEvalKit's `VANTAGE_Temporal` inference preparation: TSV regeneration, prompt building, and video resolution all pass. - Event Verification annotations were left unchanged in this pass; the current `data/event_verification/data_jsons/annotations/*.json` schema is treated as already inference-appropriate. ## 2026-05-19 - README YAML updated with a `configs:` block so the HF dataset viewer exposes one tab per task (`vqa`, `temporal_localization`, `event_verification`, `referring`, `pointing`, `tracking`, `2dbbox`, `dense_captioning`). - Added `data/2dbbox/metadata.jsonl` and `data/dense_captioning/metadata.jsonl` so the image / video tasks are browsable in the viewer. ## 2026-05-15 - Per-task `prompt.json` added under `data/2dbbox/` and `data/dense_captioning/` (schema: `{"prompt": ""}`). - Question-only annotation JSONs added under `data/vqa/data_jsons/annotations/`, `data/temporal_localization/data_jsons/annotations/`, `data/event_verification/filtered/**`, and `data/referring/refdrone_test_llava.json`. Answer fields (`gt`, `gt_option`, `answer`, GT bbox coordinates, etc.) are stripped — only the question side ships. - `Metropolis2DPointing` references renamed to `Vantage2DPointing`. - New `data/README.md` summarizing the dataset layout and listing the per-task prompts. ## 2026-05-09 - **`data/` restructured.** Top-level task directories renamed: `dense_captioning/`, `event_verification/`, `pointing/`, `referring/`, `temporal_localization/`, `vqa/`. The legacy `Spatial/` directory was dropped. ## 2026-05-04 - **Initial release**: full release uploaded to `nvidia/PhysicalAI-VANTAGE-Bench` under `data/` (PR #2).