Chinzhu/Lite-OmniJev
Chinzhu/Lite-OmniJev is a 7.9 billion parameter multimodal decision model built on a frozen Gemma 4 E4B base with a decision LoRA. It is designed for single-step multimodal decision-making, capable of processing text, images, screenshots, audio, and video inputs. The model efficiently reads out all option probabilities in a single forward pass for choice (2-20 options), score (2-10 levels), and yes/no questions, returning answer probabilities, confidence, and structured judgments via API. This model is optimized for scenarios requiring rapid, structured decision outputs across diverse modalities.
Loading preview...
Lite-OmniJev: Single-Step Multimodal Decision Model
Lite-OmniJev is a 7.9 billion parameter model developed by Chinzhu, specializing in single-step multimodal decision-making. It leverages a frozen Gemma 4 E4B base model enhanced with a decision-focused LoRA adapter. A key differentiator is its ability to read out all option probabilities for a given question in a single forward pass, significantly streamlining the decision process.
Key Capabilities
- Multimodal Input Support: Processes text, images, screenshots, audio, and video.
- Diverse Decision Types: Handles Choice (2-20 options), Score (2-10 levels), and Yes/No questions.
- Efficient Output: Provides direct API returns of answer probabilities, confidence scores, and structured judgments.
- Optimized Inference: Designed for efficiency, reading all option probabilities in one forward pass.
- Performance: Achieves 76.62% on JevBench231 and 48.5% on MVBench (nine tasks, 1800 samples), demonstrating strong performance in multimodal understanding and decision tasks.
Good For
- Applications requiring rapid, structured decision outputs from diverse input types.
- Scenarios where a single forward pass for probability readout is critical for efficiency.
- Systems needing to classify or score inputs based on multimodal content.
- Developers looking for a model that provides explicit probabilities and confidence for its decisions.