morriszjm/Tacit-2B

Tacit-2B: Efficient Typed Decision-Making

Tacit-2B is a 2.3 billion parameter language model, derived from Qwen/Qwen3.5-2B, engineered for rapid, typed decision-making. Unlike general-purpose LLMs, Tacit-2B specializes in outputting a probability distribution over predefined options (e.g., choosing from a list, answering yes/no, or selecting an ordered level) in a single forward pass. This specialization is achieved through a unique AnyJev self-distillation process, where the model autonomously generated and learned from its own decision problems, without relying on human labels or external datasets.

Key Capabilities

• Single-Pass Decision Output: Provides probabilistic answers for typed decisions (choice, yes/no, score) in one forward pass, optimizing for speed and efficiency.
• Adaptive Reasoning: Supports an optional adaptive mode where low-confidence decisions can be escalated to the model's own reasoning process (Chain-of-Thought) for improved accuracy, while still providing probabilities.
• Self-Distilled Training: Leverages a novel self-distillation method, allowing the model to learn robust decision-making directly from its own generated data.
• vLLM Integration: Designed for high-throughput serving with vLLM, supporting both in-process and server-based deployments.

Good For

• Automated Customer Support: Quickly categorizing customer inquiries or determining next steps based on user input.
• Interactive Systems: Powering applications that require fast, structured responses to user choices or questions.
• Decision Automation: Implementing systems where a model needs to select from a fixed set of options with high confidence and speed.
• Resource-Constrained Environments: Its 2.3B parameter size makes it efficient for deployment where computational resources are a consideration, while still offering specialized performance.