feat: add palindrome_generation

[joenorton]

Adds Palindrome Generation, to the Algorithmic section.
I made the decision to base this on made-up words. I think this will isolate the algorithmic challenge of creating a palindrome without interference from linguistic factors.

Prompt example:

q="Rearrange these letters to form a palindrome (a word, phrase, or sequence that remains the same in reverse): c, c, f

Example format:
racecar

What is your palindrome?", a="cfc"

Object example:
metadata: {'letters': ['c', 'c', 'f'], 'generated_palindrome': 'cfc'}

Can change min & max length as part of config.

[andreaskoepf]

Can you please try to rebase to current main head? I first forgot to add the pytz dependency in pyproject.yaml.

[joenorton]

Hopefully did that correctly. If not I'll circle back on Sunday.

[andreaskoepf]

OK, now formatting & unit tests pass. Since the result in not unique we need a score_answer() function which check whether the right letter-set was used and that the submitted answer is a palindrome.

[Miserlou]

max_length: int = 10 seems like a low limit? Would love to see this done with 32K tokens, for instance..

[andreaskoepf]

@Miserlou I suggest we keep the default config very simple .. in ~1 week we'll have some experience of "curriculum engineering" and can then give some guidelines for "calibrating" the default/baseline complexity which can then be adjusted in a similar way for all tasks scalar value.

[joenorton]

I'm fine with changing the default config to whatever is reasonable.

[andreaskoepf]

Perfect! :-D I like the 0.0 - 0.05 reward leader .. keeps a tangible delta between right and wrong ;-)

        # Check if the answer is a palindrome
        if answer != answer[::-1]:
            return 0.02

        # Check if answer contains the same letters as provided (ignoring order)
        if sorted(answer) != sorted(expected_letters):
            return 0.05