morriszjm/Tacit-1.7B
Tacit-1.7B is a 1.7 billion parameter model developed by morriszjm, built upon Qwen3-1.7B. It is specifically designed for making typed decisions (e.g., choosing options, yes/no, ordered levels) in a single forward pass, providing a probability distribution over the choices. This model utilizes AnyJev self-distillation, where the base model generated and answered its own decision problems, learning to provide direct answers without human labels or external datasets. It demonstrates improved accuracy on benchmarks like JevBench and bev-decision compared to its base model.
Loading preview...
Overview
Tacit-1.7B is a 1.7 billion parameter model, developed by morriszjm, that specializes in making typed decisions in a single forward pass. Built on the Qwen3-1.7B architecture, it returns a probability over given options, such as choosing from a list, answering yes/no, or selecting an ordered level.
Key Capabilities
- Single-pass Decision Making: Designed to output decisions directly in one forward pass, enhancing efficiency for specific tasks.
- Self-Distillation Training: Utilizes an innovative AnyJev self-distillation method where the base model generated its own decision problems and learned to solve them, eliminating the need for human labels or external datasets.
- Improved Accuracy: Demonstrates enhanced performance on decision-making benchmarks, showing a +6.9 point improvement on JevBench and +3.1 points on bev-decision compared to its Qwen3-1.7B base.
Good for
- Automated Decision Systems: Ideal for applications requiring rapid, probabilistic decisions from predefined options.
- Resource-Constrained Environments: Its 1.7 billion parameter size makes it suitable for deployment where computational resources are a consideration.
- Research in Self-Supervised Learning: Offers a practical example of self-distillation for specialized task learning without external supervision.