Fix #6575 (#6596)

docs/source/preprocessing.rst

@@ -284,6 +284,12 @@ The tokenizer also accept pre-tokenized inputs. This is particularly useful when
 predictions in `named entity recognition (NER) <https://en.wikipedia.org/wiki/Named-entity_recognition>`__ or
 `part-of-speech tagging (POS tagging) <https://en.wikipedia.org/wiki/Part-of-speech_tagging>`__.
 
+.. warning::
+
+    Pre-tokenized does not mean your inputs are already tokenized (you wouldn't need to pass them though the tokenizer
+    if that was the case) but just split into words (which is often the first step in subword tokenization algorithms
+    like BPE).
+
 If you want to use pre-tokenized inputs, just set :obj:`is_pretokenized=True` when passing your inputs to the
 tokenizer. For instance, we have:
 

src/transformers/tokenization_utils_base.py

@@ -1088,7 +1088,8 @@ ENCODE_KWARGS_DOCSTRING = r"""
                 returned to provide some overlap between truncated and overflowing sequences. The value of this
                 argument defines the number of overlapping tokens.
             is_pretokenized (:obj:`bool`, `optional`, defaults to :obj:`False`):
-                Whether or not the input is already tokenized.
+                Whether or not the input is already pre-tokenized (e.g., split into words), in which case the tokenizer
+                will skip the pre-tokenization step. This is useful for NER or token classification.
             pad_to_multiple_of (:obj:`int`, `optional`):
                 If set will pad the sequence to a multiple of the provided value. This is especially useful to enable
                 the use of Tensor Cores on NVIDIA hardware with compute capability >= 7.5 (Volta).