Access to SVI-Bench is provided only for approved research and educational use. Access requests must be submitted using an institutional .edu email address. Please describe your intended use. By requesting access, you agree to the SVI-Bench Dataset Access and Non-Disclosure Terms in the dataset card.

SVI-Bench

SVI-Bench is a benchmark for Strategic Video Intelligence: evaluating whether multimodal models can move from seeing what happens in complex multi-agent sports video to explaining why it happens, simulating alternatives, and autonomously integrating evidence for higher-level analysis.

SVI-Bench spans basketball, hockey, and soccer, combining broadcast video with aligned play-by-play logs, expert commentary, game reports, and structured statistics. The benchmark organizes nine tasks across four capability pillars:

Perception: parse short video clips into structured descriptions, answer fine-grained action questions, and retrieve clips from compositional natural-language queries.
Reasoning: explain strategic events, forecast future outcomes from partial game context, and synthesize long-form narratives from full-game evidence.
Simulation: generate plausible future play sequences from motion context and produce goal-conditioned actions that satisfy strategic objectives.
Agency: autonomously gather and integrate evidence across large-scale multimodal sports corpora—including video, event logs, commentary, reports, and statistics—to answer complex analytical questions.

SVI-Bench is designed for research on video-language models, sports analytics, long-horizon video understanding, video retrieval, controllable video generation, and agentic reasoning over multimodal sports corpora.

Links

Resource	Link
Project page	https://svi-bench.github.io/
Code & Github Page	https://github.com/Texaser/SVI-Bench
Arxiv	http://arxiv.org/abs/2605.31529
Extended paper	https://svi-bench.github.io/svi_bench_extended.pdf

Dataset Summary

SVI-Bench provides task-specific data, videos, annotations, embeddings, checkpoints, tracker weights, questions, and evaluation assets for a multi-task sports video benchmark. Unlike single-task video QA or captioning datasets, SVI-Bench is structured to test a progression of model abilities: from localized perception and action recognition, to strategic understanding, to future prediction and action generation, to agentic reasoning across a large sports-video corpus.

The benchmark is organized into nine task directories (T1–T9). Each task includes sport-specific data or evaluation assets where applicable. Some tasks focus on classic video-language understanding, while others require retrieval, forecasting, generation, or tool/agent-style reasoning.

Supported Tasks

ID	Task	Sports	Pillar
T1	Structured play-by-play description	Basketball, hockey, soccer	Perception
T2	Fine-grained action QA	Basketball, hockey, soccer	Perception
T3	Compositional video retrieval	Basketball, hockey, soccer	Perception
T4	Strategic reasoning QA	Basketball, hockey, soccer	Reasoning
T5	Outcome forecasting	Basketball, soccer	Reasoning
T6	Long-form narrative synthesis	Basketball, hockey, soccer	Reasoning
T7	Motion-conditioned generation	Basketball, soccer	Simulation
T8	Goal-conditioned action generation	Basketball	Simulation
T9	Cross-corpus agentic reasoning	Basketball, hockey, soccer	Agency

Repository Layout

T1/{basketball,hockey,soccer,captions}/        Structured play-by-play description
T2/{basketball,hockey,soccer,data}/            Fine-grained action QA
T3/{clips,ckpts,embeds,compositions,data}/     Compositional video retrieval
T4/{basketball,hockey,soccer}/                 Strategic reasoning QA
T5/{basketball,soccer}/                        Outcome forecasting
T6/soccer/                                     Long-form narrative synthesis
T7/{basketball,soccer}/                        Motion-conditioned generation
T7/tracker_weights/                            YOLOX + MixFormer-ViT tracker weights shared with T8
T8/basketball/                                 Goal-conditioned action generation
T8/llava_qa_checkpoint/                        Fine-tuned LLaVA-Qwen QA model
T8/tracker_weights/                            YOLOX + MixFormer-ViT tracker weights
T9/{data,ckpts,embeds,questions,storage}/      Cross-corpus agentic reasoning

Intended Uses

SVI-Bench is intended for:

evaluating multimodal and video-language models on sports video understanding;
benchmarking long-horizon temporal reasoning, action recognition, retrieval, and forecasting;
studying strategic understanding in dynamic multi-agent sports environments;
developing reproducible evaluation pipelines for video QA, captioning, retrieval, generation, and agentic reasoning;
non-commercial research and educational use.

Out-of-Scope Uses

SVI-Bench is not intended for:

commercial products, services, or model training without separate written approval;
redistribution, mirroring, or re-hosting of source videos or restricted assets;
surveillance, biometric identification, athlete profiling, betting, gambling, or other harmful profiling applications;
attempts to reconstruct restricted source content outside the approved dataset access flow;
use that violates the dataset license or access terms.

Access

This is a gated dataset. To request access, use the Hugging Face access form using your institutional .edu email account and provide your affiliation, intended use, and agreement to the dataset terms. Access is limited to approved research and educational use.

Because the repository is large and includes video assets, embeddings, checkpoints, tracker weights, and task-specific data, users should download only the task directories needed for their experiments.

Quick Start

1. Request dataset access

Complete the access request form on Hugging Face. We will manually review your information and approve access to the dataset.

2. Install dependencies

Access the code and scripts on GitHub. Use the task-specific README files and the code repository for the latest dependencies and evaluation instructions.

git clone https://github.com/Texaser/SVI-Bench.git
cd SVI-Bench

Install the dependencies specified in the repository or the per-task README files.

3. Download the data

After your Hugging Face access request is approved, you can download only the needed task directory. For example:

# Example: download T2 data repository after approval
from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="MVP-Group/SVI-Bench",
    repo_type="dataset",
    allow_patterns=[
        "T2/basketball/*",
    ],
    local_dir="SVI-Bench/T2/basketball",
)

For large files, consider using selective download patterns, huggingface_hub, or git lfs include/exclude rules so you do not download the full dataset unnecessarily.

4. Run evaluation

Evaluation scripts and task-specific instructions are maintained in the code repository:

https://github.com/Texaser/SVI-Bench

Each task may have its own expected input format, model output format, metrics, and preprocessing requirements. Refer to the corresponding task README before running experiments.

Licensing and Terms

SVI-Bench is released under CC BY-NC 4.0 for approved non-commercial research and educational use only. Redistribution is not permitted. Users must also follow the gated access terms shown on the Hugging Face request form.

Ethical Considerations

Sports video datasets can involve athletes, spectators, and broadcast or third-party footage. Users should respect privacy, licensing, and data-use restrictions. Do not use this dataset for surveillance, biometric identification, harmful profiling, gambling/betting systems, or unauthorized commercial deployments. Report security or access incidents to the maintainers.

Citation

If you use SVI-Bench, please cite the benchmark:

@misc{pan2026svibenchdynamicmicroworldstrategic,
      title={SVI-Bench: A Dynamic Microworld for Strategic Video Intelligence}, 
      author={Yulu Pan and Han Yi and Seongsu Ha and Md Mohaiminul Islam and Benjamin Zhang and Lorenzo Torresani and Gedas Bertasius},
      year={2026},
      eprint={2605.31529},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2605.31529}, 
}

Contact

For questions about access, licensing, evaluation, or benchmark maintenance, please open an issue in the code repository or contact the SVI-Bench maintainers through the project page.

Downloads last month: 618

Total file size:

15.2 TB

Paper for MVP-Group/SVI-Bench

SVI-Bench: A Dynamic Microworld for Strategic Video Intelligence

Paper • 2605.31529 • Published 5 days ago • 4