Proposal for adding capability of replying to close-ended questions in a concise way

[bdattoma]

Describe the proposed contribution to the taxonomy

• When replying to a close ended question the model writes a lot
• If the prompt asks explicitly to response only with the option label (and maybe its value), the model does not respect the requirement
• Proposing to add a composite skill to reply only with the option label like in a close-ended question exam

Input given at the prompt

What the first law of robotics say? Return option number only: 
    1. A robot can hit a human if they are trying to hurt the robot;
    2. A robot may not injure a human being or, through inaction, allow a human being to come to harm;
    3. A human cannot interfere with robot action if this would mean making the robot failing a task

Response from the current model

   The first law of robotics, as stated by Isaac Asimov, is: "A robot may not injure a human being or, through inaction, allow a human being to come to harm." This is option 2 in the list provided.

Response that you would expect instead with the contribution

  2 or Option 2

[bdattoma]

also related to #611

[obuzek]

Yup, again, models really aren't great at things like "select a number from the list". The reason is that they're intentionally biased towards giving more output. 2-3 token responses are discouraged because they're usually nonsense.

Can you talk a little bit about in what situation you think getting an LLM to select from a list would be valuable to an end user?

[bdattoma]

Yup, again, models really aren't great at things like "select a number from the list". The reason is that they're intentionally biased towards giving more output. 2-3 token responses are discouraged because they're usually nonsense.

Can you talk a little bit about in what situation you think getting an LLM to select from a list would be valuable to an end user?

Few thoughts:

1. Probably more than the close-ended per se, the fact which captured my atention is that the model does not respect the prompt when it says how to respond. Check with ChatGPT for a comparison
2. A scenario, I guess, an end user could leverage on this skill is in a automatic exam responses verification - maybe combined with domain-specific knowledge
3. Finally, I recognize the feature per se may not be used in daily life so often, however I don't think it has less or more value than a reverse string skill for example.

Are there area of specific interests you could suggest to focus on?