mistralops-ministral-8b-agent-orchestrator

MistralOps: Ministral 8B Agent Orchestrator

Fine-tuned LoRA adapter for multi-step AI agent orchestration planning.

Model Description

This is a QLoRA adapter for Ministral-8B-Instruct-2410 fine-tuned to generate structured multi-step execution plans for AI agent orchestration.

Given a natural language user request, the model generates a JSON execution plan with:

• 46 available tools across 12 categories (email, calendar, GitHub, Slack, smart home, browser, shell, etc.)
• Step dependencies and parallel execution groups
• Error recovery strategies (retry, fallback, abort) per step
• Data flow between steps using output references

Training Details

Evaluation Results

The base model scores 0.000 across all metrics (wrong output schema — it doesn't know the plan format). Fine-tuning enables structured agent planning from scratch.

Usage

from mlx_lm import load, generate
from mlx_lm.sample_utils import make_sampler

model, tokenizer = load(
    "mlx-community/Ministral-8B-Instruct-2410-4bit",
    adapter_path="path/to/adapters"
)
sampler = make_sampler(temp=0.1)

messages = [
    {"role": "system", "content": "You are an AI agent orchestrator. Generate a multi-step execution plan as JSON."},
    {"role": "user", "content": "Check my emails, create tasks for action items, and send a summary to Slack"},
]
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
response = generate(model, tokenizer, prompt=prompt, max_tokens=2048, sampler=sampler)

Project

Built for the Mistral AI Worldwide Hackathon — Tokyo Edition (Feb 28 - Mar 1, 2026).

• W&B Project: mistralops
• Track: Fine-Tuning by W&B