min1321 commited on
Commit
bca8fe4
·
verified ·
1 Parent(s): 72816ee

Update dataset card (Introduction + Results + figures)

Browse files
Files changed (1) hide show
  1. README.md +189 -33
README.md CHANGED
@@ -1,35 +1,191 @@
1
  ---
2
- dataset_info:
3
- features:
4
- - name: qid
5
- dtype: string
6
- - name: video_id
7
- dtype: string
8
- - name: source_corpus
9
- dtype: string
10
- - name: question
11
- dtype: string
12
- - name: correct_interval
13
- list: float64
14
- - name: duration_s
15
- dtype: float64
16
- - name: video_duration_s
17
- dtype: float64
18
- - name: event_summary
19
- dtype: string
20
- - name: category
21
- dtype: string
22
- - name: youtube_url
23
- dtype: string
24
- splits:
25
- - name: test
26
- num_bytes: 813749
27
- num_examples: 2273
28
- download_size: 365218
29
- dataset_size: 813749
30
- configs:
31
- - config_name: default
32
- data_files:
33
- - split: test
34
- path: data/test-*
35
  ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
  ---
2
+ license: cc-by-4.0
3
+ task_categories:
4
+ - video-text-to-text
5
+ language:
6
+ - en
7
+ size_categories:
8
+ - 1K<n<10K
9
+ pretty_name: ExtremeWhenBench
10
+ tags:
11
+ - video
12
+ - temporal-grounding
13
+ - long-video
14
+ - benchmark
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
15
  ---
16
+
17
+ # ExtremeWhenBench
18
+
19
+ Hour-scale natural-language temporal grounding benchmark.
20
+
21
+ <p align="center">
22
+ <img src="https://raw.githubusercontent.com/naver-ai/ExtremeWhenbench/main/assets/teaser.png" alt="Short-clip vs hour-long temporal grounding" width="640">
23
+ </p>
24
+
25
+ **2,273 open-form natural-language questions** over **194 hour-long
26
+ videos** (mean 75.7 min, max 9 hr) sourced from LVBench, MLVU, and
27
+ VideoMME. The median GT event is 9 s — matched to Charades-STA's 7.1 s
28
+ — so the same event grain now sits inside a search space ~**153× larger**.
29
+
30
+ Companion to *Natural-Language Temporal Grounding in Hour-Long Videos is a
31
+ Search Problem: A Benchmark and Empirical Decomposition* (under review).
32
+ Code, evaluation script, and lmms-eval task plugin live at
33
+ **<https://github.com/naver-ai/ExtremeWhenbench>**.
34
+
35
+ ## Introduction
36
+
37
+ At hour-scale, the binding constraint is **search, not recognition**.
38
+ Video-LLMs are bottlenecked not by localizing a nearby event, but —
39
+ given a natural-language query — by finding the right region of a long
40
+ video. Short-video benchmarks can't distinguish these failure modes;
41
+ hour-scale grounding makes the distinction central and lets the task
42
+ decompose into a **search** stage and a **localize** stage —
43
+ structurally the *retrieve-then-read* split that reshaped open-domain QA.
44
+
45
+ ## Results
46
+
47
+ Best per-model mIoU on Charades-STA vs. ExtremeWhenBench (same task).
48
+ Open Video-LLMs collapse 5–120×, and a frame-level CLIP retriever
49
+ overtakes every open Video-LLM under our compute budget:
50
+
51
+ | Model | Charades-STA | ExtremeWhenBench | Ratio |
52
+ | ------------------------------ | -----------: | ---------------: | ----: |
53
+ | **Qwen3.5-9B** | **0.579** | **0.110** | 5.3× |
54
+ | InternVL3.5-8B | 0.359 | 0.003 | 120× |
55
+ | LLaVA-OneVision-7B | 0.226 | 0.003 | 75× |
56
+ | LLaVA-NextVideo-7B | 0.089 | 0.001 | 89× |
57
+ | GPT-5.4 (64f) | 0.299 | 0.013 | 23× |
58
+ | Gemini-2.5-flash (1k f) | 0.308 | 0.053 | 5.8× |
59
+ | Gemini-3.5-flash (auto-fps) | 0.466 | 0.115 | 4.1× |
60
+ | **CLIP ViT-L/14-336** (retrieval) | 0.332 | **0.269** | 1.2× |
61
+
62
+ CLIP outperforms every open Video-LLM on ours while sitting in the
63
+ middle of the pack on Charades-STA — the order flips when search
64
+ dominates. A failure taxonomy attributes **85%** of Video-LLM errors to
65
+ search; a retrieve-then-ground hybrid recovers **6.7×** over the
66
+ monolithic Video-LLM.
67
+
68
+ Scaling frames alone doesn't close the gap — Qwen3.5-9B keeps climbing
69
+ out to N=2,048 frames but remains far below the retrieval baseline:
70
+
71
+ <p align="center">
72
+ <img src="https://raw.githubusercontent.com/naver-ai/ExtremeWhenbench/main/assets/figure3_sweep.png" alt="Frame-count sweep, Qwen3.5-9B" width="560">
73
+ </p>
74
+
75
+ Queries are **open-form natural language**, not template-bound:
76
+ MATTR 0.78 vs. 0.60 (TVBench) and 0.54 (Charades-STA); 1,578 unique
77
+ 4-gram stems (~25× per-question gap over TVBench).
78
+
79
+ ## Loading
80
+
81
+ ```python
82
+ from datasets import load_dataset
83
+ ds = load_dataset("min1321/extreme-when-bench", split="test")
84
+ print(ds[0])
85
+ ```
86
+
87
+ ## Schema
88
+
89
+ | Column | Type | Description |
90
+ | ------------------ | -------------- | ---------------------------------------------- |
91
+ | `qid` | string | Stable question id |
92
+ | `video_id` | string | Source-corpus video id |
93
+ | `source_corpus` | string | `LVBench` / `MLVU` / `VideoMME` |
94
+ | `question` | string | The grounding question |
95
+ | `correct_interval` | list\[float, float] | Ground-truth `[start_s, end_s]` |
96
+ | `duration_s` | float | Length of the GT interval (s) |
97
+ | `video_duration_s` | float | Length of the full source video (s) |
98
+ | `event_summary` | string | Short paraphrase of the event |
99
+ | `category` | string | Coarse event category |
100
+ | `youtube_url` | string | YouTube URL (empty for MLVU) |
101
+
102
+ Predictions are `[start, end]` intervals in seconds. Metrics: **mIoU**,
103
+ **R@0.3 / R@0.5 / R@0.7**, **parse-failure rate** (parse failures count
104
+ as IoU = 0).
105
+
106
+ ## Source videos
107
+
108
+ This dataset ships annotations only — download the videos from their
109
+ source corpora:
110
+
111
+ | Corpus | Videos | Where to get it |
112
+ | --------- | -----: | ------------------------------------------------- |
113
+ | VideoMME | 89 | <https://video-mme.github.io/> |
114
+ | LVBench | 67 | <https://lvbench.github.io/> |
115
+ | MLVU | 38 | <https://github.com/JUNJIE99/MLVU> |
116
+
117
+ The `source_corpus`, `youtube_url`, and `video_duration_s` columns give
118
+ you the per-video lookup.
119
+
120
+ ## Evaluation
121
+
122
+ Two paths, both in the companion repo
123
+ [naver-ai/ExtremeWhenbench](https://github.com/naver-ai/ExtremeWhenbench).
124
+
125
+ ### A. Reference script (`evaluation/eval_qwen35_vllm.py`)
126
+
127
+ vLLM + async OpenAI client. Loads this dataset directly:
128
+
129
+ ```bash
130
+ vllm serve <Qwen3.5-9B-checkpoint> --port 8000 \
131
+ --served-model-name qwen3.5-9b \
132
+ --tensor-parallel-size 8 --reasoning-parser qwen3 \
133
+ --allowed-local-media-path ./videos
134
+
135
+ python evaluation/eval_qwen35_vllm.py \
136
+ --video-dir ./videos \
137
+ --num-frames 768 \
138
+ --out qwen35_f768.jsonl
139
+ ```
140
+
141
+ ### B. lmms-eval drop-in
142
+
143
+ Copy the task plugin and the small adapter patch from the companion repo
144
+ into your local [lmms-eval](https://github.com/EvolvingLMMs-Lab/lmms-eval) clone:
145
+
146
+ ```bash
147
+ git clone https://github.com/EvolvingLMMs-Lab/lmms-eval
148
+ cp -r ExtremeWhenbench/lmms-eval/extremewhenbench lmms-eval/lmms_eval/tasks/
149
+ cp ExtremeWhenbench/lmms-eval/patches/chat_openai.py \
150
+ lmms-eval/lmms_eval/models/chat/openai.py
151
+ pip install -e lmms-eval
152
+
153
+ python -m lmms_eval --model openai \
154
+ --model_args "model=qwen3.5-9b,base_url=http://localhost:8000/v1,api_key=EMPTY,\
155
+ pass_video_url=True,max_frames_num=768,enable_thinking_kwarg=False" \
156
+ --gen_kwargs "max_new_tokens=128,temperature=0,top_p=0.8" \
157
+ --tasks extremewhenbench --batch_size 1 \
158
+ --output_path ./logs/ewb --log_samples
159
+ ```
160
+
161
+ The two-flag adapter patch lets the OpenAI-compatible adapter send video
162
+ as a URL so vLLM can decode server-side; without it, mIoU on hour-long
163
+ inputs drops ~10× because client-side frame extraction strips
164
+ absolute-time signals. We plan to upstream both pieces as a PR to
165
+ `EvolvingLMMs-Lab/lmms-eval`; until that lands, the drop-in above is the
166
+ way to use them.
167
+
168
+ Reported numbers (Qwen3.5-9B, `num_frames=768`, `enable_thinking=False`):
169
+
170
+ | Path | mIoU |
171
+ | ------------------------------------------ | ------ |
172
+ | Paper Table 4 (reported) | 0.053 |
173
+ | `evaluation/eval_qwen35_vllm.py` | 0.0469 |
174
+ | `lmms-eval` task | 0.0485 |
175
+
176
+ ## Citation
177
+
178
+ ```bibtex
179
+ @inproceedings{seo2026extremewhenbench,
180
+ title = {Natural-Language Temporal Grounding in Hour-Long Videos
181
+ is a Search Problem: A Benchmark and Empirical Decomposition},
182
+ author = {Seo, Sukmin and Kim, Geewook},
183
+ year = {2026}
184
+ }
185
+ ```
186
+
187
+ ## License
188
+
189
+ Annotations: CC-BY-4.0. Source videos are governed by the licenses of
190
+ LVBench, MLVU, and VideoMME respectively; this dataset does not
191
+ redistribute them.