☁️ CloudEval-YAML

🤔 Considering integrating LLMs into your cloud application but unsure about their performance? Check this out 👇

📊 The table is generated with CloudEval-YAML, a practical Large Language Model (LLM) benchmark specifically designed for cloud-native application configuration generation. It comprises a comprehensive dataset of 1011 problems covering widely deployed applications including Kubernetes, Envoy and Istio. This benchmark is unique in its practical and realistic approach, featuring a combination of hand-written question contexts, reference YAML files, and unit test scripts for each problem. This benchmark can facilitate comparing different LLM models, as well as different sampling and prompt engineering techniques. 📄 Please refer to the following technical report for more detailed information: CloudEval-YAML: A Practical Benchmark for Cloud Native YAML Configuration Generation.

🚀 Quick Start

Benchmarking GPT-3.5 with all metrics but unit tests.

📋 Prerequisites:

Install the python prerequisites

pip install -r requirements.txt

🔑 Configure OpenAI API Key

export OPENAI_API_KEY='YOUR_API_KEY'

🏃 Run the Benchmark

python benchmark.py

Sample output from the terminal:

------ Problem: Kubernetes_pod_q1 ------
bleu score: 0.69 / 1
edit_distance score: 0.57 / 1
exact_match score: 0.00 / 1
kv_match score: 0.00 / 1
kv_wildcard score: 0.73 / 1

(more problems...)

------ Category: Kubernetes_pod ------
bleu score: 30.44 / 48
edit_distance score: 23.60 / 48
exact_match score: 4.00 / 48
kv_match score: 10.00 / 48
kv_wildcard score: 28.83 / 48

(more categories...)

------ Overall --- Model: gpt-3.5 ------
Dataset: original (337 problems)
bleu score: 211.66 / 337
edit_distance score: 174.34 / 337
exact_match score: 29.00 / 337
kv_match score: 58.00 / 337
kv_wildcard score: 212.39 / 337

(repeat above: more datasets...)

====== Overall === Model: gpt-3.5 ======
Dataset: original, simplified, translated (1011 problems)
bleu score: 0.631
edit_distance score: 0.522
exact_match score: 0.071
kv_match score: 0.154
kv_wildcard score: 0.605

All prompts, generated codes and scores by default will be saved in outputs/. You man use them as checkpoints for further analysis.

🎉 Congratulations! You have successfully completed the quickstart and benchmarked GPT-3.5 (other than unit tests).

🤖 Benchmarking Different Models

OpenAI GPT-3.5/4

• Configure OpenAI API key

  export OPENAI_API_KEY='YOUR_API_KEY'

• Include GPT-3.5/4 for benchmarking in config.json

  "models": [
    "gpt-3_5",
    "gpt-4"
  ]

• Model-specific query is defined in models/gpt.py. You may customize them if needed. Please refer to OpenAI API documentation for more details.

Google PaLM 2

• Configure PaLM API key

  export PALM_API_KEY='YOUR_API_KEY'

• Include PaLM 2 for benchmarking in config.json

  "models": [
    "palm-2"
  ]

• Model-specific query is defined in models/palm.py. You may customize them if needed. Please refer to Google API documentation for more details.

Replicate

Replicate is a platform that allows you to run open-source or your own models at scale in the cloud. It abstracts away the complexity of deploying models locally and provides a simple API for query and response.

• Configure Replicate API token

  export REPLICATE_API_TOKEN=='YOUR_API_TOKEN'

• Currently, a subset of supported open-source models are integrated into the benchmark. You may include them for benchmarking in config.json.

  "models": [
    "llama-2-70b-chat",
    "llama-2-13b-chat",
    "llama-2-7b-chat",
    "llama-13b-lora",
    "llama-7b",
    "wizardcoder-34b-v1.0",
    "wizardcoder-15b-v1.0",
    "codellama-13b-instruct",
    "codellama-7b-instruct"
  ]

• Model-specific query is defined in models/replicate.py. You may customize them, integrate more or even include your own models if needed. Please refer to Replicate API documentation for more details.

🏃 Run the Unit Tests

The easiest way to run unit tests is to launch an AWS EC2 instance using public AMI with cloudyaml_public as name and ami-012306c17129a3b71 as ID in us-west-1 region (N. California). Notice that this is going to take several hours to evaluate. Alternatively, one can create a pull request and we can help to run on a cluster.

After launching the EC2 instance, you can pull this repo and enable unit test in config.json

"metrics": {
  "unit_test": true
}

Then run the benchmark:

python benchmark.py 

Sample output from the terminal:

------ Problem: Kubernetes_pod_q1 ------
bleu score: 0.69 / 1
edit_distance score: 0.57 / 1
exact_match score: 0.00 / 1
kv_match score: 0.00 / 1
kv_wildcard score: 0.73 / 1
unit_test score: 1.00 / 1

(more problems, categories and datasets...)

====== Overall === Model: gpt-3.5 ======
Dataset: original, simplified, translated (1011 problems)
bleu score: 0.631
edit_distance score: 0.522
exact_match score: 0.071
kv_match score: 0.154
kv_wildcard score: 0.605
unit_test score: 0.418

Please refer to Advanced.md for how to run your own model and other advanced features.

📄 Citation

If you find this benchmark useful, please cite our paper:

@article{xu2024cloudeval,
  title={CloudEval-YAML: A Practical Benchmark for Cloud Configuration Generation},
  author={Xu, Yifei and Chen, Yuning and Zhang, Xumiao and Lin, Xianshang and Hu, Pan and Ma, Yunfei and Lu, Songwu and Du, Wan and Mao, Zhuoqing and Zhai, Ennan and others},
  journal={Proceedings of Machine Learning and Systems},
  volume={6},
  pages={173--195},
  year={2024}
}