jylin001206/SII-Decision-27B-Full
InnerJev-27B-Full by jylin001206 is a 27 billion parameter decision model, adapted from Qwen/Qwen3.8-27B, designed to return probability distributions over user-supplied options with thinking disabled. It uses full-rank updates on language linear layers and is optimized for decision distributions rather than general chat. This model excels at Choice, binary/probabilistic Noul, and ordered Score tasks, achieving 80.1734/100 on a fixed 58-benchmark decision suite.
Loading preview...
InnerJev-27B-Full: A Specialized Decision Model
InnerJev-27B-Full, developed by jylin001206, is a 27 billion parameter model specifically designed for decision-making tasks. Adapted from the Qwen/Qwen3.8-27B architecture, this model focuses on providing probability distributions for user-defined options, with its 'thinking' mechanism disabled to streamline decision outputs. Unlike general-purpose chat models, InnerJev-27B-Full is not supervised for conversational abilities but is instead optimized for precise decision distributions.
Key Capabilities
- Decision Distribution Output: Returns a probability distribution over user-supplied options.
- Specialized Task Handling: Excels in Choice, binary/probabilistic Noul, and ordered Score tasks, all sharing a unified answer-slot interface.
- Optimized for Decisions: Training specifically targets decision distributions, not general chat or SFT supervision.
- High Evaluation Score: Achieves 80.1734 / 100 on a fixed 58-benchmark decision suite (BF16 checkpoint).
- Full-Rank Adaptation: Utilizes full-rank updates on language linear layers, ensuring comprehensive adaptation while retaining original vision components (though decision training uses text inputs).
Good For
- Automated Decision Systems: Ideal for applications requiring probabilistic choices or scoring based on predefined options.
- Research in Decision Modeling: Useful for studying and implementing models focused purely on decision outcomes.
- Integration with Custom Interfaces: Designed to work with specific decision scoring interfaces, as demonstrated by the provided inference scripts.
- Resource-Conscious Deployment: Can be loaded in BF16 (requiring substantial RAM/GPU) or quantized (NF4/INT8) for deployment on GPUs like the RTX 4090, including sharded configurations for larger models.