Training puzzles are too computerized

I tried the new training feature and my impression is that even if we humanly solve the puzzle, i.e. play the wining variation (win a queen or get mating position), we often end up failing the puzzle because we didn't play the best computer move, which shortens the game by one or two moves sometimes. Is the idea behind training to make people think like computers? Or is it that these puzzles need to be fixed?

I've tried pretty hard to be lenient in the determination of whether a line is a win, retry, or loss. The current rules are for mate in N puzzles:
-All lines that are equally as long are wins
-All lines that are 1 or 2 moves longer are retries
-Anything else is a loss

I wanted to allow longer mate sequences to be playable, but the problem was the program became stuck on repeating positions by continually going for longer sequences, so I decided to make the trade-off that you see.

For material advantage lines, the rules are:
-Moves that are within 5% of the top move's computer evaluation that also gain material advantage in adequate time are also wins.
-Moves that are within 10% of the top move's computer evaluation are retries.
-Anything else is a loss.

In material advantage puzzles, you are expected to play on if:
-You are in check
-There is a high likelihood that you will lose a minor/major piece without accurate counterplay
-You have a high likelihood of being mated without accurate counterplay

The evaluations and solutions are generated by a computer, so naturally they may seem computerised. There's no way to generate the 10,000+ puzzles on here by hand.

Hey Clarkey, I would suggest going through previously generated puzzles (http://www.yacpdb.org/) and using the computer solutions from these initial positions. There tends to be only one unique solution for each puzzle.

I've encountered quite a few curious puzzles. In one, I had a rook and five pawns against an opponent's lone king and was expected to find the 10-move sequence that leads to mate. I'm not entirely sure if these sorts of puzzles are that helpful as they favor a computerized approach over a human one.

http://en.lichess.org/training/5009

That is an example of what Matetricks is saying.

Those puzzles aren't generated anymore, but old ones of that format ( under ~#4000. There are more than 10k puzzles in the db now ) will remain in the db.

I've considered sourcing yacpdb for chess puzzles, but there's issues with compatibility that are almost too tricky to bother with. All we'd be using is their starting positions and determining the rest from there -- there's already millions of decent starting positions in the lichess db.

Main issues are:
-no link to a source game
-positions are determined from FEN instead of PGN (this causes more issues than you think)
-solutions in yacpdb are algebraic notation instead of long algebraic notation or smith notation - PITA to work with.

I understand - it's possible to simplify the notation using chess.js, but it might be more trouble than it's worth. I haven't run into many issues with using FENs for puzzles instead of PGNs aside from the obvious problem of being unable to source to a game.

Also, a quick suggestion - it would be easy to write a script to parse the FEN string of each puzzle to check for unhelpful puzzles where white has many more pieces than black. If the number of uppercase letters significantly exceeds that of the lowercase letters (i.e, in my aforementioned example), you can just delete them from the database so no one else will run into them.

It'd be more accurate to run SF on the starting position for a quarter of a second and see if it matches the new criteria.

...or how about wangling the doodah flip and run the hyperbole on top of the doobery-drive?