mistral7b-mdmp-lora-mlx

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

This is not a drop-in for the Unsloth/PEFT adapter. Use decisionlens/mistral7b-mdmp-lora on NVIDIA GPUs. Adapters are not interchangeable.

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

Parameter-efficient fine-tune (LoRA) of mlx-community/Mistral-7B-Instruct-v0.3-4bit via mlx-lm, trained on leak-reviewed instruction pairs from open MDMP doctrine summaries and fictional scenarios.

• Base model: mlx-community/Mistral-7B-Instruct-v0.3-4bit
• Method: 4-bit QLoRA via mlx-lm (Mac sidecar / mlx-v4)
• Training data: mdmp-staff-planning-pairs (324 reviewed pairs)
• GPU sibling: mistral7b-mdmp-lora (Unsloth / PEFT)
• GitHub: dlens/mdmp-assistant

Training hyperparameters

Parameter	Value
LoRA rank	16
LoRA scale	2.0 (mlx-lm multiplies the update by scale; PEFT equivalent is alpha/rank = 32/16)
LoRA dropout	0.05
Iters	600
Learning rate	5e-5 (cosine decay, 20-step warmup)
Batch	1 × 4 grad accum
Max sequence length	2048
Layers	all (num_layers: -1)

Target modules: self_attn.q_proj, k_proj, v_proj, o_proj, mlp.gate_proj, up_proj, down_proj.

Evaluation

Held-out golden set (20 questions, separate from training data). Score below is this uploaded artifact (eval/run_golden_mlx.py, 2026-09-01), not a copy of the Unsloth 14/20 result or an older Mac run.

Run	Pass rate
Base MLX Mistral-7B-Instruct-v0.3-4bit	(see repo eval reports)
This adapter (mlx-v4, pre-publish)	18/20 (90%)

Eval script: eval/run_golden_mlx.py

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

Usage

git clone https://github.com/dlens/mdmp-assistant.git
cd mdmp-assistant
python -m venv .venv-mlx && source .venv-mlx/bin/activate
pip install -r requirements-mlx.txt huggingface_hub

hf download decisionlens/mistral7b-mdmp-lora-mlx --local-dir outputs/mlx-mistral7b-mdmp-lora-v4
python demo/ask.py --backend mlx

Or from Python:

from train.mlx_inference import load_mlx, generate_answer

model, tokenizer = load_mlx(
    model_name="mlx-community/Mistral-7B-Instruct-v0.3-4bit",
    adapter_path="outputs/mlx-mistral7b-mdmp-lora-v4",
)
print(generate_answer(model, tokenizer, "What MDMP step is war gaming?"))

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

Hardware: Apple Silicon. First run also downloads the 4-bit base weights from mlx-community.

Limitations

• Small golden eval set (n=20), English-only, U.S. MDMP framing
• Coaching assistant only — not operational planning or classified scenarios
• Not interchangeable with the Unsloth adapter on Hugging Face
• Users must verify operationally consequential answers against authoritative doctrine

License

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