TempCore redistributes videos from MVBench, NExT-QA, MLVU and LongVideoBench. TempCore does not own these videos; all rights and copyright belong to their original creators and platforms. Each video stays under the license and terms of its source benchmark, and copyright holders may ask for a video to be removed. By requesting access you agree to the following terms.

TempCore: Are Video QA Benchmarks Temporally Grounded?

📖 arXiv • 🤗 Paper • 🧪 lmms-eval task • EMNLP 2026

TempCore is the temporally sensitive subset of six Video QA benchmarks: 2,560 questions that need the video and whose answer depends on which frames a model sees.

🚀 Evaluate with lmms-eval

TempCore is built into lmms-eval. This is the recommended way to evaluate on it: no access request to this repository is needed, since each task fetches its videos from the source benchmark.

git clone https://github.com/EvolvingLMMs-Lab/lmms-eval && cd lmms-eval
uv sync --extra video --extra qwen

uv run python -m lmms_eval --model qwen3_vl \
    --model_args pretrained=Qwen/Qwen3-VL-8B-Instruct \
    --tasks tempcore --batch_size 1

lmms-eval task	Source task	Questions
`tempcore`	all six below	2,560
`tempcore_mvbench`	`mvbench_*` (18 subtasks)	209
`tempcore_nextqa`	`nextqa_mc_test`	629
`tempcore_egoschema`	`egoschema_subset`	96
`tempcore_mlvu`	`mlvu_dev`	673
`tempcore_longvideobench`	`longvideobench_val_v`	261
`tempcore_videomme`	`videomme`	692

Each task reuses the matching lmms-eval benchmark task (prompt, frame sampling, scoring) and keeps only the TempCore questions, so a TempCore score is directly comparable with the full-benchmark score of the same task.

📌 Overview

Most Video QA questions can be answered no matter which frames a VLM sees. We measure this per sample with Frame Selection Sensitivity (FSS), the change in accuracy when the most relevant frames (SigLIP MaxProb) are swapped for the least relevant ones (MinProb), averaged over eight VLMs, and combine it with a Language Independence Score (LIS) from no-frame predictions. TempCore is the subset of each benchmark that is both vision-dependent (LIS > 0.5) and frame-sensitive (|FSS| > 0.15).

Config	Source benchmark	TempCore questions	Videos	Videos in this repo
`mvbench`	MVBench	209	203	✅
`nextqa`	NExT-QA (MC test)	629	421	✅
`egoschema`	EgoSchema (500-question subset)	96	96	❌ fetch from source
`mlvu`	MLVU (dev, MC)	673	488	✅
`longvideobench`	LongVideoBench (val)	261	219	✅
`videomme`	Video-MME	692	511	❌ fetch from source

Temporally grounded: FSS > 0.15, the most relevant frames help. Visual bias: FSS < −0.15, the frames an encoder ranks as most relevant mislead the models.

📦 Using this repository

This repository holds the TempCore questions in one schema across the six benchmarks, the videos of four of them, and per-sample scores. Request access above to download it.

from datasets import load_dataset

ds = load_dataset("HJOK/TempCore", "mlvu", split="test")
row = ds[0]
row["question"], row["options"], row["answer"], row["video_path"]

video_path is relative to the repository root. Download the videos for one benchmark (or videos/* for all hosted ones):

hf download HJOK/TempCore --repo-type dataset --local-dir TempCore \
    --include "data/*" "scripts/*" "videos/mlvu/*"

EgoSchema videos come from Ego4D, whose license does not allow redistribution, and Video-MME may not be redistributed without the authors' approval. scripts/fetch_videos.py pulls the needed videos from the upstream Hub repositories into the same layout (accept the upstream terms first; only the needed members are read from the upstream archives, not the whole archives):

cd TempCore
pip install huggingface_hub requests
python scripts/fetch_videos.py --benchmark egoschema videomme --out .

SHA256SUMS lists the checksum of every hosted video.

🗂️ Fields

Field	Description
`sample_id`	Stable id within the source benchmark (MVBench `<task>/<index_in_task>`, NExT-QA `<video>/<row>`, EgoSchema `<q_uid>`, MLVU `<json_stem>/<index_in_file>`, LongVideoBench `<id>`, Video-MME `<question_id>`)
`benchmark`	Config name
`video_path`	Video file, relative to the repository root
`video_hosted`	Whether the video is stored in this repository
`video_source`	Upstream location: `repo_id`, `files` (archive or file names), `format` (`zip` / `tar` / `file`), `member`
`question`	Question text, as in the source benchmark
`options`	Answer options, with any leading `A.`-style letters removed
`answer` / `answer_index`	Correct option as a letter (A–E) and a 0-based index
`fss`, `lis`	Frame Selection Sensitivity and Language Independence Score
`fss_class`	`key` (FSS > 0.15), `antikey` (FSS < −0.15) or `nonkey`
`category`	`temporally_grounded` (key) or `visual_bias` (antikey); every TempCore row is one of these two
`tempcore`	Always `true` here
`n_models_fss`, `n_models_lis`	Number of models behind each score (8)
`per_model`	Per-model correctness (0/1) under MaxProb, MinProb and no-frame input
`meta`	JSON string with the remaining source fields (e.g. MVBench `task`, `start`/`end`; MLVU `question_type`, `duration`; LongVideoBench categories and subtitle file name; Video-MME `duration`, `domain`, `task_type`)

Scores come from eight VLMs: SmolVLM2-256M/500M/2.2B, LLaVA-OneVision-0.5B/7B, VideoLLaMA3-2B, Qwen3-VL-8B and InternVL3.5-8B, with K=32 frames scored by SigLIP2-so400m on question and answer text.

annotations/ holds the scores for every sample of the six benchmarks (not only TempCore), so thresholds can be changed without re-running models, and annotations/human_eval/ the 100 human temporal-necessity ratings from the paper. These files carry ids and text hashes only.

🎬 Video sources and processing

Config	Upstream repository	Upstream license / terms
`mvbench`	`OpenGVLab/MVBench`	MIT; videos for academic research only, and some sources carry their own agreements
`nextqa`	`lmms-lab/NExTQA` (videos from VidOR)	NExT-QA: MIT, cite and attach the source
`mlvu`	`sy1998/MLVU` (files identical by SHA-256 to `MLVU/MVLU`)	CC-BY-NC-SA 4.0, academic research only
`longvideobench`	`longvideobench/LongVideoBench`	CC-BY-NC-SA 4.0, non-commercial only
`egoschema`	`lmms-lab/egoschema`	Ego4D license
`videomme`	`lmms-lab/Video-MME`	Academic research only; no redistribution without approval

Videos are stored byte-for-byte as extracted from the upstream files, with one exception: the 23 MVBench episodic_reasoning questions use TVQA clips that MVBench distributes as 3 fps JPEG frames. Those are encoded to H.264 mp4 at 3 fps (CRF 18), keeping one video frame per source frame. MVBench start/end clip bounds are kept in meta; the paper's scores use the whole video file.

⚖️ License

The questions, annotations and scores are released under CC-BY-NC-SA 4.0. The videos remain the property of their creators and are covered by the terms of their source benchmark listed above. If you own a video in this dataset and want it removed, open a discussion on this repository.

✒️ Citation

@inproceedings{ok2026tempcore,
  title     = {TempCore: Are Video QA Benchmarks Temporally Grounded?},
  author    = {Ok, Hyunjong and Lee, Jaeho},
  booktitle = {Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing},
  year      = {2026}
}

Please also cite the source benchmarks you use.

Downloads last month: 23

Total file size:

101 GB

Paper for HJOK/TempCore

TempCore: Are Video QA Benchmarks Temporally Grounded?

Paper • 2509.01167 • Published Sep 20