SEASON Reproduction

Reproduction of SEASON: Mitigating Temporal Hallucination in Video Large Language Models via Self-Diagnostic Contrastive Decoding (Wu et al., 2025), with room for extension experiments.

1. Setup

# Option A: conda
conda env create -f environment.yml
conda activate season

# Option B: pip
pip install -r requirements.txt

2. Quick Sanity Check

python smoke_test.py

3. Run VidHalluc Evaluation

python -m eval.vidhalluc --data_root /path/to/vidhalluc

4. Repository Structure

SEASON- experiment/
├── PROJECT.md
├── README.md
├── environment.yml
├── requirements.txt
├── smoke_test.py
├── season/
│   ├── config.py
│   ├── model.py
│   ├── negatives.py
│   ├── diagnostic.py
│   ├── decoding.py
│   └── utils/video.py
└── eval/
    └── vidhalluc.py

5. Default Hyperparameters

6. Notes

• data/ and model/ are intentionally excluded from Git tracking.
• Spatial negative noise (sigma) is set to 0.1 by default due to paper ambiguity.
• If decoder attentions are unavailable, diagnostic weights fall back to equal weighting.

7. Next Extensions

• Add VideoHallucer and EventHallusion evaluators.
• Compare alternative negative constructions.
• Benchmark additional VideoLLM backbones.