Code for Monotonic paraphrasing improves generalization of language model prompting.
- To generate paraphrase with MonoPara, use the function
run_PPL_greedy
inmain_down_stream.py
. - The logits ensemble is at here. We manipulate the predicted logits at every decoding step.
- The larger
alpha
means more dependent onlogit_for_next_step_target
, which means assigning higher weight to the token that has the highest probability predicted by the target model, which is equivalent to lower perplexity wrt. target model. We setalpha=0.5
for all the experiments. - Ignore the
option
parameter and leave it as default. - We use model
mistralai/Mistral-7B-Instruct-v0.1
as both target model and paraphrase model.