morriszjm/Tacit-8B

Tacit-8B: Efficient Decision-Making Model

Tacit-8B is an 8 billion parameter model, based on Qwen/Qwen3-8B, specifically engineered for making typed decisions in a single forward pass. Unlike traditional LLMs that generate extensive reasoning, Tacit-8B directly outputs a probability distribution over predefined options (e.g., choice, yes/no, ordered levels).

Key Capabilities

• Single-Pass Decision-Making: Provides immediate probabilistic answers for classification tasks, reducing latency and computational cost.
• Self-Distillation Training: Developed using AnyJev self-distillation, where the base Qwen3-8B model generated its own decision problems and learned to answer them efficiently without human labels or external datasets.
• Improved Accuracy: Demonstrates enhanced performance on decision-making benchmarks, achieving +3.9 points on JevBench (public set) and +1.4 points on bev-decision (test split) compared to its Qwen3-8B base.
• Adaptive Decision Routing: Supports an adaptive mode where low-confidence decisions can be automatically escalated to the model's full reasoning capabilities (Chain-of-Thought) for increased accuracy, with configurable thresholds (tau) and limits (max_cot_share).
• vLLM Integration: Optimized for high-throughput inference with vLLM, allowing for efficient deployment and scaling.

Good for

• Automated Customer Support: Quickly classifying customer intent or routing queries.
• Content Moderation: Rapidly categorizing content (e.g., safe/unsafe, spam/not spam).
• Business Process Automation: Making quick, probabilistic decisions in workflows.
• Interactive Agents: Providing fast responses for choice-based interactions.