Datasets:

---

FutureMa

/

EvasionBench

like 89

Hi @SaylorTwift , thank you for setting this up!

I noticed a small issue in the field_spec: the task 'evasion detection' requires the model to see both the question and the answer fields to make a judgment. The current config only passes the question.

Could we update the input to include both? Or should I provide a pre-formatted 'prompt' column in the dataset to make it easier for the evaluator?

Also, the labels are direct, intermediate, and fully_evasive. The rest of the config looks great!

Hi @FutureMa ! Thanks for your input this is still a first draft , we use inspect-ai as a backend so we can build the prompt on the fly without the need to have it in the dataset !

However, thre is one thing that would make the implementation much easier, if i understand correctly, your benchmark is a multichoice benchmark. therefore having a choices field with: ["direct", "intermediate", "fully_evasive"] for each sample as well as the letter for the target (A, B, or C) would be super helpful.

The eval.yaml file would look liek this:

tasks:
  - id: evasion_bench
    config: default
    split: train

    epochs: 1

    shuffle_choices: true

    field_spec:
      input: question
      target: "eva4b_label_letter"
      choices: "choices"
      metadata: ["answer"]

    solvers:
      - name: prompt_template
        template: |
          Question: {prompt}

          Answer: {answer}
      - name: multiple_choice
        template: |
          You are a financial analyst. Your task is to Detect Evasive Answers in Financial Q&A.
          The entire content of your response should be of the following format: 'ANSWER: $LETTER' (without quotes) where LETTER is one of {letters}.

          {question}

          {choices}

    scorers:
      - name: choice

Here is an example of what the prompt would look like using inspect: https://huggingface.co/spaces/SaylorTwift/evasive-bench

Hi @SaylorTwift ,

I've added the fields you suggested:

• choices: ["direct", "intermediate", "fully_evasive"] for each sample
• eva4b_label_letter: Letter representation (A/B/C) of the labels

The dataset now has 6 columns total. The updated parquet file has been uploaded.

Please let me know if anything else needs to be adjusted!