looking at the solution and your code I can only see two possible issues.

1. train_test_split has an optional random seed parameter. The seed needs to set for both the model and the split.

If you still fail the tests...

3. You set the learning_rate to 1. unless explicitly stated otherwise, the test cases assume default parameters for everything.

minor point, but enumerate accepts a starting number as an argument. In other words, you get enumerate to count from 1 rather than using n + 1

It is a 7th kyu Kata. So maybe people find this more clear than some of the one-liners?

fyi, I don't actually think this is a great solution; Index Error could be solved in a better way.

In the basic test cases you state:

it "possible numbers" $ do -- these, not coincidentally, have exactly one solution each
  solve 13 `shouldBe` Just "(((1) * 3) + 5) + 5"

Why doesn't Just "(1) + 3 + 3 + 3 + 3" count as a solution for the number 13 (ignoring parenthesis) ??

Either the text is wrong or the kata description needs to give an explanation for why the above would not be valid.

In terms of performance, I believe it it better to create a list and then call join add the very end.

This is because appending to a string is not a constant time operation whereas appending to a list is.

done