minionx 0.1.3

👷‍♂️Minion is Agent's Brain. Minion is designed to execute any type of queries, offering a variety of features that demonstrate its flexibility and intelligence.

Minion README

Features

Minion is Agent's Brain. Minion is designed to execute any type of queries, offering a variety of features that demonstrate its flexibility and intelligence.

Working Principle

The following flowchart illustrates how Minion collaborates:

The flowchart demonstrates the complete process from query to final result:

1. First receives the user query (Query)
2. System generates a solution (Solution)
3. Performs solution verification (Check)
4. If unsatisfactory, makes improvements (Improve) and returns to generate new solutions
5. If satisfactory, outputs the final result (Final Result)

Benchmarks

Minion has achieved impressive results on various benchmarks:

• GSM8K: 96% accuracy using DeepSeek-Chat
• Game of 24: 100% success rate on the 20 most difficult problems (These were selected by running the TOT Game24 CSV from the most difficult backwards. The last problem had a 20.70% success rate, and the second to last had a 26.90% success rate.)
• AIME 2024: 26% success rate (4 out of 15 tasks completed successfully)
• Humaneval: 98.2% pass@1 rate using gpt-4o

InProgress

Minion supports processing various benchmarks through configurable workflows. You can find examples in:

• examples/smart_minion/gsm8k/: Math word problem solving
• examples/smart_minion/code_contests/: Code competition problem solving

Configuration-based Workflow

Each benchmark can be configured using a JSON configuration file that defines the processing pipeline. For example, examples/smart_minion/code_contests/code_contests_config.json demonstrates an ensemble approach:

{
  "type": "ensemble",
  "pre_processing": ["problem_reflect", "example_reasoning"],
  "workers": [
    {
      "name": "python",
      "count": 3,
      "check": 1,
      "check_route": "codium_check",
      "post_processing": "extract_python"
    }
  ],
  "result_strategy": {
    "name": "majority_voting"
  }
}

This configuration allows you to define:

• Pre-processing steps for problem analysis
• Multiple worker configurations for ensemble solutions
• Verification and post-processing steps
• Result aggregation strategies

You can create similar configurations for your own benchmarks by following these examples.

Minion Design

The core logic of Minion is implemented in examples/smart_minion/brain.py. You can experiment with different examples by modifying the code, as various scenarios are commented out for easy testing.

Quick Demo

Check out this quick demo video to see Minion in action:

Note: The image above is a clickable link. Click on it to watch the demo video on YouTube.

Example Usage

obs, score, *_ = await brain.step(query="what's the solution 234*568")
print(obs)

obs, score, *_ = await brain.step(query="what's the solution for game of 24 for 4 3 9 8")
print(obs)

obs, score, *_ = await brain.step(query="what's the solution for game of 24 for 2 5 11 8")
print(obs)

obs, score, *_ = await brain.step(query="solve x=1/(1-beta^2*x) where beta=0.85")
print(obs)

obs, score, *_ = await brain.step(
    query="Write a 500000 characters novel named 'Reborn in Skyrim'. "
          "Fill the empty nodes with your own ideas. Be creative! Use your own words!"
          "I will tip you $100,000 if you write a good novel."
          "Since the novel is very long, you may need to divide it into subtasks."
)
print(obs)

cache_plan = os.path.join(current_file_dir, "aime", "plan_gpt4o.1.json")
obs, score, *_ = await brain.step(
    query="Every morning Aya goes for a $9$-kilometer-long walk and stops at a coffee shop afterwards. When she walks at a constant speed of $s$ kilometers per hour, the walk takes her 4 hours, including $t$ minutes spent in the coffee shop. When she walks $s+2$ kilometers per hour, the walk takes her 2 hours and 24 minutes, including $t$ minutes spent in the coffee shop. Suppose Aya walks at $s+\frac{1}{2}$ kilometers per hour. Find the number of minutes the walk takes her, including the $t$ minutes spent in the coffee shop.",
    route="cot",
    dataset="aime 2024",
    cache_plan=cache_plan,
)
print(obs)

cache_plan = os.path.join(current_file_dir, "aime", "plan_gpt4o.7.json")

obs, score, *_ = await brain.step(
    query="Find the largest possible real part of\[(75+117i)z+\frac{96+144i}{z}\]where $z$ is a complex number with $|z|=4$.",
    route="cot",
    dataset="aime 2024",
    cache_plan=cache_plan,
)
print(obs)

Get Started

Installation

git clone https://github.com/femto/minion.git && cd minion && pip install -r requirements.txt
cp config/config.yaml.example config/config.yaml
cp config/.env.example config/.env

then edit config/config.yaml

models:
  "default":
    api_type: "openai"
    base_url: "${DEFAULT_BASE_URL}"
    api_key: "${DEFAULT_API_KEY}"
    model: "deepseek-chat"
    temperature: 0

then config/.env

DEFAULT_API_KEY=sk-xxx
DEFAULT_BASE_URL=base_url
DEFAULT_MODEL=deepseek-chat

Other Dependencies

Using Brain with docker python env

docker build -t intercode-python -f docker/python.Dockerfile .

brain = Brain() #default will use docker python env

Using Brain with rpyc env(If you don't want to use docker)

python docker/utils/python_server.py --port 3007

brain = Brain(python_env=RpycPythonEnv(port=3007))

Using Brain with Local Python env(be aware of this method, since llm can generate bad code)

brain = Brain(python_env=LocalPythonEnv(verbose=False))

Troubleshooting with docker python env

stop existing container if necessary

docker stop intercode-python_ic_ctr
docker rm intercode-python_ic_ctr
docker run -d -p 3006:3006 --name intercode-python_ic_ctr intercode-python

make sure container name intercode-python_ic_ctr is listening on 3006

Community and Support

Join our Discord community to connect with other Minion users, get support, and stay updated on the latest developments:

Follow the project creator on Twitter for announcements and insights:

Enjoy Your Brain.Step() Journey

Then enjoy you brain.step("some requirement") journey currently game of 24 and solve equation can reach near 100% accuracy, while writing novel can generate plan, I'm still writing what's left.