title: VideoGLaMM ActivityNet Object Review - GPT-6 Luna
emoji: 🌙
colorFrom: indigo
colorTo: blue
sdk: static
app_file: index.html
VideoGLaMM ActivityNet object text — same 20 samples, GPT-6 Luna
A new independent Space for gpt-6-luna. The previous GPT-6.1 Sol Space and the original VIPOSeg Space are unchanged.
All 20 requests are deep-equal to the previous submitted requests except body.model, which is changed from gpt-6.1-sol to gpt-6-luna. Exact same local custom IDs, 20 samples in the same order, refined captions, complete prompt, 260 image input blocks in the same order, JSON schema, medium reasoning, and 16000 output token cap. Each request contains 10 clean RGB frames plus the source ID overlay, identity sheet, and actual binary-mask sheet. Each direct HTTP body equals the prepared Luna request body, and raw API responses and request audits are retained locally. The earlier waiting Batch was cancellation-requested at the user's instruction; its input verification and cancellation receipt are retained separately. See data/request_audit.json, data/direct_api_receipt.json, and data/batch_cancel_receipt.json.
Prompt SHA256: f4f2c370eab558f31c5e195fe1953a6232ee5bab77e1c89e5fed96768dd56213. The user annotation rule is unchanged, with the full original text preserved in data/TEXT_ANNOTATION.original.md and embedded verbatim in data/annotation_prompt.txt.
20/20 direct Responses API outputs returned; 17 passed structural checks. 109 entity groups, 187 mentions, 23 model mask-match candidates and 86 other pending groups; 0 complete correspondence candidates. All statuses and any validation errors are displayed. Human reviewed: 0. Differences from the previous run are model outputs, not measured accuracy differences. See data/model_comparison.json.
Original supplied media: ten sampled frames per caption segment, not a full source video. 32 original masks across the 20 samples; masks and sparse bboxes exist on a single reference frame per sample, not on the other nine. No new masks or tracks generated. Time labels are estimates from the source sampling procedure. Playback is an inspection slideshow at two images per second, not original video FPS (unknown).
All text groups are retained, including entities without masks, body parts and collective references. Missing spatial supervision is not a background negative. Model claims about visibility, caption hallucinations and mask quality still require human review. Original JSON and unmodified GPT outputs are downloadable per sample, separately from deterministic postprocessing adjustments and span validation.
Review status and notes are browser-local only, isolated from the Sol Space with storage key videoglamm-activitynet-object20-luna-review-v1. Export/import JSON to save or aggregate reviews; exports bind to the Luna dataset hash. Marking a sample reviewed does not supply its missing masks.
Direct Responses API run gcg-object20-luna-direct-27164e5005fc45de97fc9bf2296f9f74. Actual API usage: 326855 input tokens, 52595 output tokens. No bill estimate is presented as an invoice.
Source: VideoGLaMM Dataset.md, ActivityNet Entities GCG. Media rights remain with original data owners. Research annotation preview. Layout based on the prior review Space.