dipta007/VCInspector-3B
VCInspector-3B by dipta007 is a lightweight 3 billion parameter large multimodal model (LMM) fine-tuned from Qwen2.5-VL-3B-Instruct. It specializes in reference-free evaluation of video captions, focusing on detecting factual errors in objects and actions. This model provides interpretable quality scores and natural language explanations, making it suitable for resource-constrained environments and cross-domain generalization for both video and image caption evaluation.
Loading preview...
VC-Inspector-3B: Lightweight Factual Video Caption Evaluation
VC-Inspector-3B is a 3 billion parameter large multimodal model (LMM) developed by dipta007, specifically designed for the reference-free evaluation of video captions. Fine-tuned from Qwen2.5-VL-3B-Instruct, this model excels at identifying factual inaccuracies within captions, particularly concerning objects and actions depicted in videos. It provides both a numerical quality score (1-5) and a natural language explanation for its assessment, offering interpretable feedback.
Key Capabilities
- Reference-free Evaluation: Assesses video captions without needing ground-truth references.
- Factual Grounding: Pinpoints factual errors in objects and actions within captions.
- Interpretable Outputs: Generates quality scores (1-5) accompanied by natural language explanations.
- Cross-domain Generalization: Effective for evaluating both video and image captions.
- Lightweight & Efficient: At 3B parameters, it's suitable for resource-constrained environments and offers fast inference (0.30 seconds per video clip on A100 GPU).
Good for
- Resource-constrained deployments: Ideal for on-device or environments with limited computational power.
- Batch processing: Efficiently evaluates large datasets of video captions.
- Automated quality control: Automatically identifies factual errors in generated video descriptions.
- Applications requiring interpretable feedback: Provides clear explanations for caption quality scores.