reaperdoesntknow/Shepherd-Alpha
Shepherd-Alpha is a 2 billion parameter tactical reasoning model developed by Convergent Intelligence LLC, fine-tuned on Qwen3-1.7B. It specializes in dual-perspective military scenario analysis, generating both attack and defense reasoning. This model utilizes a novel BiCell Depth Dispersal training methodology to separate representation encoding from task-specific reasoning, making it the first defense AI reasoning model on Hugging Face.
Loading preview...
Shepherd-Alpha: Tactical AI Reasoning
Shepherd-Alpha is a 2 billion parameter model developed by Convergent Intelligence LLC, designed for dual-perspective tactical scenario analysis. It is the first public model in the Shepherd program, focusing on AI systems for autonomous defense applications. Built on Qwen3-1.7B, it provides structured reasoning for both adversary exploitation (attack) and counter-measures (defense) within military contexts.
Key Capabilities & Differentiators
- Dual-Perspective Reasoning: Generates both attack and defense strategies for tactical scenarios, anticipating threats by understanding adversarial approaches.
- BiCell Depth Dispersal: Employs a novel training methodology that partitions transformer layers by abstraction depth, training them asymmetrically. This forces specialization, with lower layers encoding domain structure and upper layers focusing on reasoning.
- Specialized Domain Adaptation: Training insights revealed that for domain-specific fine-tuning, representation layers (lower layers) are the primary bottleneck, not reasoning layers.
- Structured Output: Produces clear, structured analysis for tactical situations.
Training Details
The model was fine-tuned on the ZennyKenny/tactical-military-reasoning-v.1.0 dataset, comprising 150 dual-perspective tactical scenarios. The BiCell Depth Dispersal method involves three phases of asymmetric layer freezing and training, accumulating gradients before a single optimizer step.
Limitations
As an alpha release, Shepherd-Alpha has a small training set (150 scenarios), limiting its tactical depth. Users should be aware that the base model's <think> generation pattern might override structured output, and enable_thinking=False is recommended for cleaner results. This model is for analysis and reasoning only, not for controlling or actuating systems.