Upload README.md with huggingface_hub

README.md

@@ -4,10 +4,18 @@ license: apache-2.0
 
 # SVI-Bench
 
-> The SVI-Bench artifacts are hosted as a **dataset**, not a model.
-> Please use the links below.
+**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.
 
-## Where to go
+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.
+
+## Resources
 
 - 📚 **Dataset:** [huggingface.co/datasets/MVP-Group/SVI-Bench](https://huggingface.co/datasets/MVP-Group/SVI-Bench)
 - 🌐 **Project page:** [svi-bench.github.io](https://svi-bench.github.io/)