Skip to content

Latest commit

 

History

History
71 lines (60 loc) · 3.72 KB

how_to_use_your_own_eval_code.md

File metadata and controls

71 lines (60 loc) · 3.72 KB
Raw

Guidelines for Custom Evaluation

Feel free to use your own evaluation code to evaluate with MixEval data. To help you smoothly finish your evaluation and ensure fairness at the same time, we suggest some protocols to follow and provide example model output formats.

❗Protocols

  1. You can evaluate models on either MixEval or MixEval-hard (or both of them). Each of them contains two files: free-form.json and multiple-choice.json, indicating two different kinds of problems, and the final score is the accuracy over all samples of the two files.

    └── data
    └── mixeval-<version>
        │
        ├── mixeval
        │   ├──free-form.json
        │   └──multiple-choice.json
        │
        └── mixeval-hard
            ├──free-form.json
            └──multiple-choice.json

    Here <version> denotes the dynamic benchmark version indicated at the top of README.md. You can also load the data from the 🤗 huggingface repository.

  2. To reduce the impact of prompt formatting, we provided fixed input prompts. In your custom interence code, you should use the mixeval.evaluation_prompts.construct_prompt_freeform function to format free-form entries and the mixeval.evaluation_prompts.construct_prompt_multichoice function to format multiple-choice entries.

    You can do so by simply wraping the function over each entry of the data. E.g., construct_prompt_freeform(entry), where entry is a dictionary.


    When evaluating base models, besides the above prompt format, the formated inputs should be additionally prefixed with the provided 5-shot examples. Each free-form entry should be prefixed with mixeval.evaluation_prompts.FIVE_SHOT_PREFIX_FREEFORM and each multiple-choice entry should be prefixed with mixeval.evaluation_prompts.FIVE_SHOT_PREFIX_MULTIPLECHOICE.


  3. We highy recommend you to use our parsing pipeline to compute scores to ensure fairness, which is very fast and accurate. However, you can also use your own parsing logic.


Model Output Format (if using our parsing pipeline)

Free-form

[
    {
        "id": "0",
        "problem_type": "free-form", 
        "context": null, 
        "prompt": "What does a manometer measure?", 
        "target": ["Manometric unit", "Absolute Pressure"], 
        "benchmark_name": "TriviaQA", 
        "formated_input": "Question: What does a manometer measure?\nAnswer the question shortly.", 
        "response": "A manometer measures pressure."
    },
    ...
]

Multiple-choice

[
    {
        "id": "0",
        "problem_type": "single-choice", 
        "context": null, 
        "prompt": "Which solution is correct?", 
        "options": ["can be ruined if they get wet by a mop ", "can be ruined if they get wet by a jar "], 
        "target": [0], 
        "benchmark_name": "PIQA", 
        "formated_input": "magazine\nWhich solution is correct?\nA. can be ruined if they get wet by a mop \nB. can be ruined if they get wet by a jar \nAnswer with the option letter from the given choices directly.", 
        "response": "A."
    },
    ...
]
  1. You should prepare the model response files under mix_eval/data/model_responses/. As an example, you can refer to the file structure of mix_eval/data/model_responses/gemma_11_7b_instruct/.
  2. The difference between model input (the benchmark data) and output (the model response file) is just the "response" field, i.e., each entry in your output file should keep all key-value pairs (including the 'id') of the input entry, with an additional "response" field representing the model's output.