src/lighteval/models/nanotron/nanotron_model.py (7 lines):
	- line 187: # TODO @nouamane: this is only needed for training, can we just mark params as NanotronParameter instead?
	- line 256: TODO: Remove these conditionals once HuggingFace supports a way to
	- line 323: # TODO: Merge `tok_encode_batch` here.
	- line 573: ),  # [padding_length - seq] #TODO: padding_token not always 0
	- line 584: ),  # [padding_length - seq] #TODO: padding_token not always 0
	- line 954: False  # if we unpad in modeling, it doesn't matter if left or right # TODO: not supported yet
	- line 1013: )  # [1, seq, voc] #TODO: this assumes padding on the right


src/lighteval/metrics/harness_compatibility/truthful_qa.py (1 line):
	- line 57: # TODO: This completely ignores any normalization, but keeping it as is


src/lighteval/metrics/metrics_sample.py (1 line):
	- line 788: TODO: Will have to move this to sacrebleu.