mistral7b-mdmp-lora

LoRA adapter for MDMP staff-planning coaching — public U.S. Army doctrine only, no proprietary algorithms or customer data.

Disclaimer: Unofficial educational tool. Not affiliated with the U.S. Army. Not a substitute for qualified staff planning or classified planning systems. Ground answers in FM 5-0 / ADP 5-0 and verify citations.

Model description

This is a parameter-efficient fine-tune (LoRA) of Mistral-7B-Instruct-v0.3 trained on leak-reviewed instruction pairs derived from open MDMP doctrine summaries and fictional scenarios.

• Base model: mistralai/Mistral-7B-Instruct-v0.3
• Method: 4-bit QLoRA via Unsloth / PEFT
• Training data: mdmp-staff-planning-pairs (324 reviewed pairs)
• GitHub: dlens/mdmp-assistant
• Apple Silicon sibling: mistral7b-mdmp-lora-mlx (mlx-lm; not interchangeable)

Training hyperparameters

Parameter	Value
LoRA rank (r)	16
LoRA alpha	32
LoRA dropout	0.05
Epochs	2
Learning rate	2e-4
Batch size	2 × 4 grad accum
Max sequence length	2048

Target modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj (all layers).

Evaluation

Held-out golden set (20 questions, separate from training data):

Run	Pass rate
Base Mistral-7B-Instruct-v0.3	1/20 (5%)
This adapter (Spark Unsloth v7)	14/20 (70%)

Eval script: eval/run_golden.py

Quick start (inference only, no training): hf-quick-start.md

Usage

Clone the GitHub repo, install requirements-ml.txt, download this adapter, then:

from train.inference import load_model, generate_answer

model, tokenizer = load_model(
    model_name="mistralai/Mistral-7B-Instruct-v0.3",
    adapter_path="path/to/mistral7b-mdmp-lora",
    load_in_4bit=True,
)
answer = generate_answer(model, tokenizer, "What MDMP step is war gaming?")
print(answer)

Prompt format: Mistral Instruct — <s>[INST] {question} [/INST] {answer}

Generation defaults (match golden eval): temperature=0.1, top_p=0.9, max_new_tokens=256.

CLI demo:

export TRITON_PTXAS_PATH=/usr/local/cuda/bin/ptxas
python demo/ask.py --adapter path/to/mistral7b-mdmp-lora

Hardware: ~5–8 GB VRAM for 4-bit inference with adapter.

Limitations

• Small golden eval set (n=20), English-only, U.S. MDMP framing
• Prompt fragility: high golden pass rate does not imply robustness to paraphrase. Informal checks show truncated or off-by-one step answers on near-miss wording (e.g. “What are the steps of MDMP?” vs the golden “seven steps … in order”; “What is step 5/7?” mapping to adjacent steps). A RAG control over the same public doctrine corpus is more stable for factual step lookup; this artifact is LoRA-only.
• Coaching assistant only — not operational planning or classified scenarios
• Automatic scoring can be brittle to near-synonyms (e.g. "synchronization" vs "synchronize")
• Users must verify operationally consequential answers against authoritative doctrine

License

Apache 2.0 for this adapter and training data. Base model subject to Mistral license.