reaperdoesntknow/Shepherd-Alpha

Shepherd-Alpha: Dual-Perspective Tactical AI Reasoning

Shepherd-Alpha, developed by Convergent Intelligence LLC, is the first defense AI reasoning model publicly available. Built on the Qwen3-1.7B base model, it specializes in generating structured dual-perspective analysis for military scenarios, providing both adversary attack reasoning and defensive counter-strategies.

Key Capabilities & Innovations

• Dual-Perspective Reasoning: Trained to simultaneously analyze tactical situations from both attacker and defender viewpoints, anticipating threats and formulating responses.
• BiCell Depth Dispersal: A novel training methodology that partitions transformer layers by abstraction depth (lower layers 0-13, upper layers 14-27) and trains them asymmetrically. This forces genuine specialization, with lower layers encoding domain structure and upper layers performing reasoning.
• Targeted Domain Adaptation: Training revealed that for domain-specific fine-tuning, representation layers (lower layers) are the primary bottleneck, consistently showing ~1.7x higher gradient magnitude during adaptation.
• Structured Output: Produces clear attack and defense chain-of-thought reasoning for given tactical scenarios.

Training Details

Shepherd-Alpha was fine-tuned on the ZennyKenny/tactical-military-reasoning-v.1.0 dataset, comprising 150 dual-perspective scenarios. The BiCell methodology involved three phases: training lower layers, then upper layers, followed by joint integration, ensuring specialized learning before full model adaptation.

Limitations

As an alpha release, it has a small training set (150 scenarios), limiting tactical depth. Users should be aware that the base model's <think> generation pattern might occasionally override the structured output, which can be mitigated using enable_thinking=False in generation config. This model is for analysis and reasoning only, not for control or actuation of weapon systems.