GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
decisionlens
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Run this model inference with full control and performance in your environment.
Model Details
Model Provider
decisionlens
Model Tree
mistralai/Mistral-7B-Instruct-v0.3
Input Modalities
Output Modalities
Supported Functionality
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.
mistralai/Mistral-7B-Instruct-v0.3| 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).
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
Clone the GitHub repo, install requirements-ml.txt, download this adapter, then:
python
from train.inference import load_model, generate_answermodel, 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:
bash
export TRITON_PTXAS_PATH=/usr/local/cuda/bin/ptxaspython demo/ask.py --adapter path/to/mistral7b-mdmp-lora
Hardware: ~5–8 GB VRAM for 4-bit inference with adapter.
Apache 2.0 for this adapter and training data. Base model subject to Mistral license.