Shepherd-Alpha/README.md

---
license: apache-2.0
language:
- en
tags:
- tactical-reasoning
- military
- defense-ai
- bicell-dispersal
- sft
- dual-perspective
- shepherd
- convergentintel
- qwen
- ai
base_model: Qwen/Qwen3-1.7B
datasets:
- ZennyKenny/tactical-military-reasoning-v.1.0
library_name: transformers
pipeline_tag: text-generation
---

# Shepherd-Alpha

**The first defense AI reasoning model on Hugging Face.**

Shepherd-Alpha is a tactical reasoning model fine-tuned on dual-perspective military scenario analysis using BiCell Depth Dispersal — a novel training methodology that partitions transformer layers by abstraction depth and trains them asymmetrically to separate representation encoding from task-specific reasoning.

Developed by [Convergent Intelligence LLC: Research Division](https://convergentintel.com)

## What This Model Does

Given a tactical scenario, Shepherd-Alpha produces structured dual-perspective analysis:
- **Attack reasoning** — how an adversary would exploit the situation
- **Defense reasoning** — how to counter, mitigate, and survive

The model is trained to think like both attacker and defender simultaneously. A model that understands how to attack becomes a defender that anticipates.

## Training Methodology: BiCell Depth Dispersal

Standard fine-tuning updates all layers jointly, allowing co-adaptation that can mask shallow learning. BiCell Depth Dispersal forces genuine specialization:

| Phase | Frozen | Training | Purpose |
|-------|--------|----------|---------|
| 1 | Upper layers (14-27) | Lower layers (0-13) | Foundations encode before specialization exists |
| 2 | Lower layers (0-13) | Upper layers (14-27) | Reasoning learns over frozen representations |
| 3 | None | All layers | Joint integration of asymmetric gradient history |

All three backward passes accumulate gradients before a single optimizer step. The asymmetric gradient history forces each depth zone to develop independently before integration.

**Key finding during training:** Lower layers consistently produce ~1.7x the gradient magnitude of upper layers during domain adaptation. The pretrained upper layers already possess sufficient reasoning capacity — the primary adaptation is teaching lower layers to encode tactical domain structure. This suggests that for domain-specific SFT, representation layers (not reasoning layers) are the bottleneck.

### Training Details

- **Base model:** Qwen/Qwen3-1.7B (28 layers, all full attention)
- **Dataset:** [ZennyKenny/tactical-military-reasoning-v.1.0](https://huggingface.co/datasets/ZennyKenny/tactical-military-reasoning-v.1.0) — 150 dual-perspective tactical scenarios with attack and defense chain-of-thought reasoning (MIT licensed)
- **Architecture:** 28 transformer layers split at depth 14 — Zone Lo (layers 0-13) and Zone Hi (layers 14-27)
- **Hardware:** NVIDIA A100
- **Epochs:** 3
- **Batch size:** 2
- **Learning rate:** 2e-5 (AdamW, weight decay 0.01)
- **Precision:** bfloat16
- **Label masking:** Loss computed only on assistant (reasoning) tokens, not scenario prompts

## Usage

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("reaperdoesntknow/Shepherd-Alpha")
tokenizer = AutoTokenizer.from_pretrained("reaperdoesntknow/Shepherd-Alpha")

messages = [
    {
        "role": "user",
        "content": "Analyze this tactical scenario.\n\nScenario: A mechanized platoon advancing through urban terrain detects a coordinated drone swarm from the northeast. Limited anti-air capability. Civilian structures restrict fields of fire."
    }
]

inputs = tokenizer.apply_chat_template(
    messages,
    add_generation_prompt=True,
    tokenize=True,
    return_dict=True,
    return_tensors="pt",
)

output = model.generate(
    **inputs,
    max_new_tokens=512,
    temperature=0.7,
    top_p=0.9,
    do_sample=True,
)

generated = output[0][inputs["input_ids"].shape[1]:]
print(tokenizer.decode(generated, skip_special_tokens=True))
```

## The Shepherd Program

Shepherd-Alpha is the first public model in the Shepherd family — an ongoing research program developing AI systems for autonomous defense applications. The program spans:

- **Shepherd Doctrine** — a comprehensive counter-swarm and area defense blueprint covering 28+ subsystems across five concentric engagement layers
- **Shepherd AI** — tactical reasoning models trained on dual-perspective analysis (this model)
- **BiCell Dispersal** — a training methodology based on the B_i Cell Dispersal framework for stochastic layer partitioning during fine-tuning

## Limitations

- **Alpha release** — this is a research checkpoint, not a production system
- **Small training set** — 150 scenarios provides format and domain grounding but limited tactical depth. Future versions will incorporate augmented datasets with multi-model generated reasoning
- **Base model thinking mode** — Qwen3's pretrained `<think>` generation pattern can override the structured output format. Use `enable_thinking=False` in generation config for cleaner output
- **Not a weapon system** — this model performs analysis and reasoning. It does not control, target, or actuate anything

## Citation

```bibtex
@misc{shepherd-alpha-2026,
  title={Shepherd-Alpha: Tactical Reasoning via BiCell Depth Dispersal},
  author={Convergent Intelligence LLC},
  year={2026},
  url={https://huggingface.co/reaperdoesntknow/Shepherd-Alpha}
}
```

## Related Work

- [Structure Over Scale](https://doi.org/10.57967/hf/5165) — Foundation paper on structure-first training methodologies
- [DualMind Methodology](https://doi.org/10.57967/hf/5184) — Dual-cognitive-mode SFT using EXPLORE/EXAMINE tokens
- [Discrepancy Calculus](https://doi.org/10.57967/hf/5194) — Mathematical framework grounding BiCell dispersal theory
- [B_i Cell Dispersal Framework](https://convergentintel.com) — Stochastic layer freezing grounded in DISC measure theory

---

*Convergent Intelligence LLC: Research Division*
*"Structure beats scale. Collaboration beats hierarchy. Observation beats theory."*
<!-- cix-keeper-ts:2026-06-12T13:16:55Z -->