RewardConfig's max_length argument docstring should indicate that it filters out dataset, rather than truncating it

[Kallinteris-Andreas]

RewardConfig's max_length argument docstring should indicate that it filters out dataset entries that exceed the limit, rather than truncating them

maybe the current documentation is obvious, and I just misunderstood it on first glance

[qgallouedec]

Good point, given that for other trainers (like DPO), it's a truncation.

In fact, the best thing would be to have a common behavior for all trainers (truncation), but the urgent thing is to clarify the documentation.