Update README.md

README.md

@@ -1,3 +1,151 @@
----
-license: agpl-3.0
----
+---
+license: agpl-3.0
+---
+# CommandNet Military Science Dataset (1900-1999)
+
+This repository contains a synthetic, instruction-tuning dataset generator and a generated dataset built for **Unsloth fine-tuning** workflows.
+
+The dataset is designed around a strict persona voice (ruthless, militaristic, tactical, arrogant, analytical) and focuses on **historical/doctrinal military analysis** with explicit:
+
+- Causal Analysis
+- Counterfactual Analysis
+
+It includes both conventional warfare and asymmetric contexts (insurgencies, militias, and counterinsurgency vs larger conventional forces), constrained to the historical window **1900-1999**.
+
+---
+
+## Dataset Goals
+
+- Produce **1,000+ samples** (default: 1,400)
+- Maintain a consistent strategic persona style
+- Cover core military science dimensions in depth:
+  - operational art
+  - logistics and sustainment
+  - command and control (C2)
+  - intelligence cycle and ISR pressure
+  - deception and disruption
+  - maneuver vs attrition dynamics
+  - civil-military and legitimacy constraints
+  - tempo and decision-cycle pressure
+- Include **asymmetric warfare** in substantial volume
+- Keep all samples anchored to historical doctrine periods between 1900 and 1999
+
+---
+
+## Output Schema (Default: ShareGPT)
+
+The default output (`--schema sharegpt`) is directly suitable for common Unsloth chat fine-tuning pipelines.
+
+Each line is a JSON object like:
+
+```json
+{
+  "id": "skynet-000001",
+  "conversations": [
+    {"from": "system", "value": "..."},
+    {"from": "human", "value": "..."},
+    {"from": "gpt", "value": "..."}
+  ],
+  "metadata": {
+    "year": 1966,
+    "decade": "1960s",
+    "warfare_type": "conventional",
+    "doctrine_family": "Amphibious Operational Sequencing",
+    "military_science_tags": ["..."]
+  }
+}
+```
+
+## Record Design (Raw Schema)
+
+- `id`
+- `year`
+- `decade`
+- `year_range`
+- `warfare_type` (`conventional`, `insurgency`, `counterinsurgency`, `hybrid`)
+- `doctrine_family`
+- `force_asymmetry_index`
+- `scenario_context`
+  - `terrain`
+  - `weather`
+  - `objective`
+  - `adversary_posture`
+  - `constraint`
+- `prompt`
+- `response`
+- `causal_analysis`
+- `counterfactual_analysis`
+- `style_tags`
+- `military_science_tags`
+- `source_refs`
+- `quality_flags`
+
+---
+
+## Doctrine Coverage
+
+The generator samples from doctrine families including:
+
+- Industrial Attrition and Trench Penetration
+- Infiltration and Decentralized Assault Groups
+- Deep Operation
+- Blitz and Mobile Combined Arms
+- Amphibious Operational Sequencing
+- Protracted People's War
+- Population-Centric Counterinsurgency
+- Maneuver Warfare and Decision-Cycle Pressure (OODA-linked)
+- AirLand Battle
+- Deterrence and Escalation Management
+
+Each doctrine has active year bounds and compatible warfare types.
+
+---
+
+## Quality Controls
+
+Built-in checks include:
+
+- Section presence checks in responses:
+  - `Causal Analysis:`
+  - `Counterfactual Analysis:`
+  - `Command Verdict:`
+- Historical bound checks (`1900-1999`)
+- Minimum response-length check
+- Duplicate fingerprint suppression using scenario keys:
+  - year
+  - warfare type
+  - doctrine
+  - terrain
+  - objective
+  - adversary posture
+
+If duplicates become excessive, generation fails fast with a clear error.
+
+---
+
+## Unsloth Integration Notes
+
+For chat SFT with Unsloth, default `sharegpt` output is the intended path.
+
+Typical flow:
+
+1. Load JSONL in your training notebook/script.
+2. Map `conversations` according to your model chat template.
+3. Keep the system prompt if you want style-locking behavior.
+4. Train with your normal Unsloth SFT trainer config.
+
+If your pipeline expects `instruction/input/output`, use `--schema alpaca`.
+
+---
+
+## Safety and Scope
+
+This dataset is for **historical and doctrinal analysis simulation**. It is not intended to provide real-world modern operational attack guidance.
+
+Design constraints intentionally emphasize:
+
+- historical framing
+- abstract operational reasoning
+- doctrinal critique and counterfactuals
+
+---