Input: Video + "What unfolded during the game in the video?"
Output: "During the game, there was a foul committed by a player from the yellow-jerseyed team, leading to a yellow card being issued..."

Architecture

Vision Encoder: DWT-VJEPA2-based video encoder (vit_giant, 1408 dim)
Text Model: LLaMA 3.2-3B with LoRA fine-tuning
Vision-Text Bridge: Learned projection layer (1408 → 2048 → 3072)
Specialization: Fine-tuned on soccer video QA data

Usage (Helper functions are in repo)

python

from soccer_qa_inference import SoccerQA

model = SoccerQA("/path/to/model")
answer = model.ask("video.mp4", "Was this a Foul?", max_tokens=45)
print(answer)

Model Details

Parameters: ~4B total
Input: Video files (16 frames, 256x256) + text questions
Output: Natural language answers
Domain: Soccer/football video analysis
Context: Handles complex game situations, player actions, fouls, etc.

Training Data

Soccer video clips with question-answer pairs
Covers various game situations: fouls, shots, saves, player actions
Annotated with detailed descriptions of game events

Limitations

Research use only, no commercial applications
Optimized specifically for soccer content
May not generalize well to other sports or video domains
Requires high-quality video input for best results

License

CC-BY-NC-4.0 - Research use only, no commercial applications permitted.

Citation

bibtex

@misc{soccer-qa-4b-2025,
  title={Soccer QA 4B: Video Question Answering for Soccer Analysis},
  author={Varun Kodathala, Sports Vision},
  year={2025},
  note={Research model for soccer video understanding}
}

sportsvision/soccer-qa-4b

Soccer QA 4B - Soccer Video Question Answering Model

Model Description

Architecture

Usage (Helper functions are in repo)

Model Details

Training Data

Limitations

License

Citation

sportsvision/
soccer-qa-4b