BenchFlow

BenchFlow is an Open-source Benchmark Hub and Eval Infra for AI production and benchmark developers.

Overview

BenchFlow.Demo.2.0-720p-50fps.mp4

Within the dashed box, you will find the interfaces (BaseAgent, BenchClient) provided by BenchFlow. For benchmark users, you are required to extend and implement the BaseAgent interface to interact with the benchmark. The call_api method supplies a step_input which provides the input for each step of a task (a task may have one or more steps).

Quick Start For Benchmark Users

Before you start, please get your BF_TOKEN on BenchFlow.ai.

1. Install BenchFlow

   git clone https://github.com/benchflow-ai/benchflow.git
   cd benchflow
   pip install -e .

2. Browse Benchmarks

   Find benchmarks tailored to your needs on our Benchmark Hub.

3. Implement Your Agent

   Extend the BaseAgent interface:

   def call_api(self, task_step_inputs: Dict[str, Any]) -> str:
       pass

   Optional: You can include a requirements.txt file to install additional dependencies, such as openai and requests.

4. Test Your Agent

   Here is a quick example to run your agent:

   import os
   from benchflow import load_benchmark
   from benchflow.agents.webarena_openai import WebarenaAgent
   
   # The benchmark name follows the schema: org_name/benchmark_name.
   # You can obtain the benchmark name from the Benchmark Hub.
   bench = load_benchmark(benchmark_name="benchflow/webarena", bf_token=os.getenv("BF_TOKEN"))
   
   your_agents = WebarenaAgent()
   
   run_ids = bench.run(
       task_ids=[0],
       agents=your_agents,
       api={"provider": "openai", "model": "gpt-4o-mini", "OPENAI_API_KEY": os.getenv("OPENAI_API_KEY")},
       requirements_txt="webarena_requirements.txt",
       args={}
   )
   
   results = bench.get_results(run_ids)

Quick Start for Benchmark Developers

1. Install BenchFlow

   Install BenchFlow via pip:

   pip install benchflow

2. Embed BenchClient into Your Benchmark Evaluation Scripts

   Refer to this example for how MMLU-Pro integrates BenchClient.

3. Containerize Your Benchmark and Upload the Image to Dockerhub

   Ensure your benchmark can run in a single container without any additional steps. Below is an example Dockerfile for MMLU-Pro:

   FROM python:3.11-slim
   
   COPY . /app
   WORKDIR /app
   COPY scripts/entrypoint.sh /app/entrypoint.sh
   
   RUN chmod +x /app/entrypoint.sh
   RUN pip install -r requirements.txt
   
   ENTRYPOINT ["/app/entrypoint.sh"]

4. Extend BaseBench to Run Your Benchmarks

   See this example for how MMLU-Pro extends BaseBench

5. Upload Your Benchmark into BenchFlow

   Go to the Benchmark Hub and click on +new benchmarks to upload your benchmark Git repository. Make sure you place the benchflow_interface.py file at the root of your project.

---

License

This project is licensed under the MIT License.