vtg-bench
Collection
2 items • Updated
qid string | model string | sheet string | video_path string | duration_s float64 | region list | question string | gt_segments list | gt_answer string | va string | k1 int64 | k2 int64 | w1 int64 | w2 int64 | pass1 dict | pass2 dict | final dict | iou float64 | iou_pass1 float64 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
wendi_misc-018 | us.openai.gpt-5.6-sol | wendi_misc | /home/greenland-user/vtg/videos/Pilot/Baby Wemby In Paris COULD NOT POST ME UP in a 1v1😈 #streetball #basketball.mp4 | 168.252 | [
0,
168.252
] | why is the hooper wearing the vest dancing and spinning? | [
[
20,
23
]
] | v | 96 | 96 | 512 | 640 | {
"n_frames": 96,
"step": 1.752625,
"first_ts": [
0.88
],
"last_ts": [
167.38
],
"raw": "{\"answer\":\"He is celebrating after beating his defender and finishing the final game with a dunk.\",\"start\":159.0,\"end\":168.3,\"reasoning\":\"He drives past the defender, dunks at about 160 seconds, the... | {
"n_frames": 38,
"step": 0.5,
"window": [
149,
168.252
],
"raw": "{\"answer\":\"He is using flashy footwork and a spin move to evade the defender, reach the basket, and dunk.\",\"start\":149.0,\"end\":161.8,\"in_window\":true,\"reasoning\":\"The sequence begins with his feints against the defender, c... | {
"answer": "He is using flashy footwork and a spin move to evade the defender, reach the basket, and dunk.",
"start": 149,
"end": 161.8,
"source": "pass2",
"in_window": true
} | 0 | 0 |
YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
Video temporal grounding with frontier models — frame-based eval (Sept 2026)
Can a frontier multimodal model, given a video and a question, return the answer and the time interval where it happens? Four models were tested with byte-identical inputs on 351 human-annotated questions over 178 YouTube videos (dashcam, SNL sketches, talk shows, vlogs, soccer and basketball highlights): Claude Fable 5.1, Claude Fable 5, GPT-6 Astra and GPT-5.6 Sol (all via Amazon Bedrock).
Neither model family accepts video input, so each video is turned into timestamped frames: 96 uniform frames
first, then a 2 fps zoom into the model's own coarse interval. Full protocol in code/README.md.
| model | mIoU | R@0.5 | answer correct % |
|---|---|---|---|
| GPT-6 Astra | 0.364 | 32.9 | 82.4 |
| Claude Fable 5.1 | 0.351 | 34.2 | 73.6 |
| Claude Fable 5 | 0.319 | 31.3 | 68.6 |
| GPT-5.6 Sol | 0.286 | 24.1 | 68.6 |
(clean subset, n = 319; answer accuracy on the 159 questions with a reference answer, judged by Claude Opus 5)
Layout
code/— the pipeline scripts, with a README describing each one and how to run them.data/— the labels (two annotation workbooks, normalisedqa_all.jsonl/qa_matched.jsonl), the Drive manifests and listings, durations. The videos and the frame cache are not in this repo: they are YouTube content and live as.tar.gzarchives in the separate media repo https://huggingface.co/datasets/Rakancorle1/vtg-frontier-eval-media (seedata/README.md).results/— every model's raw output for every question (both passes), the Opus 5 answer judgements, the aggregated report and a per-question spreadsheet.
- Downloads last month
- 141