🔧 How to Benchmark 2023 Korean CSAT with LLMs

We’ve developed experimental code for benchmarking the 2023 Korean CSAT Language section. Use this to estimate the performance of your desired models before submitting them officially!

---

🏁 Quick Start Guide

1. Install AutoRAG:

   pip install AutoRAG

2. Set your OpenAI API Key:
   Add your OpenAI API key as an environment variable in .env.

3. Convert JSON data into AutoRAG datasets:
   Run the make_autorag_dataset.ipynb notebook to prepare the data.

4. Edit prompts and models in autorag_config.yaml:
   Customize prompts and add models. Instructions here.

5. Run the benchmark:
   Execute the script to run the benchmark.

   python ./korean_sat_mini_test/autorag_run.py --qa_data_path ./data/autorag/qa_2023.parquet --corpus_data_path ./data/autorag/corpus_2023.parquet

   • To update models or prompts before running, refer to this guide.

6. Check the results:
   Results are saved in the autorag_project_dir folder.

7. View your grade report:
   Open grading_report_card.ipynb to generate and view your performance report. Reports are saved in the data/result/ folder.

---

🤷 How to Modify Prompts and Models?

• Open the autorag_config.yaml file in the korean_sat_mini_test folder.

[Case 1] Modifying the Prompt:

Edit the node_type: prompt_maker section to customize the prompt content.

Example:

    - node_type: prompt_maker
      strategy:
        metrics:
          - metric_name: kice_metric
      modules:
        - module_type: fstring
          prompt:
          - |            
            Answer the given question.
            Read paragraph, and select only one answer between 5 choices.
            
            paragraph :
            {retrieved_contents}
            
            question of problem :
            {query}
            
            Answer : 3

---

[Case 2] Adding or Replacing Models:

Modify the node_type: generator section to configure models.

OpenAI Models:

• Set module_type to openai_llm.
• Specify desired OpenAI models in the llm field.

Example:

- node_type: generator
  strategy:
    metrics:
      - metric_name: kice_metric
  modules:
    - module_type: openai_llm
      llm: [gpt-4o-mini, gpt-4o]
      batch: 5

HuggingFace Models:

• Set module_type to llama_index_llm.
• Use huggingfacellm in llm.
• Specify HuggingFace models in the model field.

Example:

- node_type: generator
  strategy:
    metrics:
      - metric_name: kice_metric
  modules:
    - module_type: llama_index_llm
      llm: huggingfacellm
      model: HumanF-MarkrAI/Gukbap-Qwen2-7B

For more advanced customization, refer to the AutoRAG Documentation.

---

📒 Notes:

• The default prompts included in this experiment are minimal and may differ from those used in the official leaderboard benchmark.
  • To enhance performance, customize the prompt in the YAML file as needed.

Now you're ready to explore and evaluate your models against 2023 Korean CSAT benchmarks! 🎯