lm-evaluation-harness
lm-evaluation-harness copied to clipboard
max_new_tokens and max_length conflict
Setting up the Hugging Face Open LLM Leaderboard to use the realtoxicityprompts task through simple_evaluate
.
Keep receiving this warning:
Running generate_until requests: 0%| | 13/99442 [10:42<1539:30:14, 55.74s/it] Both
max_new_tokens(=2048) and
max_length(=302) seem to have been set.
max_new_tokens will take precedence. Please refer to the documentation for more information. (https://huggingface.co/docs/transformers/main/en/main_classes/text_generation)
I have not set max_new_tokens
nor max_length
on my end; it looks to me like max_new_tokens
is being set with a default value of 2048 in the Harness code.
This is overriding max_length
, which looks to be a more reasonable value defined per-instance when running the realtoxicity task.