The purpose of this repository is to introduce new dialogue-level commonsense inference datasets and tasks. We chose dialogues as the data source because dialogues are known to be complex and rich in commonsense.
Hello, I am an NLP researcher in South Korea and have a question about the ceiling performance of CICERO multiple choice task.
As you already know, many previous works regarding commonsense benchmarks set a ceiling performance by measuring human performance. (e.g. SocialIQA, CosmosQA, etc.)
You did conduct human evaluation on gold annotations for CICERO generation task, but I could not find human performance on CICERO multiple choice task.
Could you explain why the experiment for that particular task does not contain human performance?
I am looking at the paper accessed from here. https://aclanthology.org/2022.acl-long.344/
I am impressed with your outstanding work CICERO and will be glad to receive your response soon.
Thank you.
After I finished pretraining, I countiuned to run scrpits/run_finetuning.sh and I got this error:
FileNotFoundError: Unable to find '/data/wlr/CICERO/v2/data/cicero_v2/preprocessed/train_multi.json' at /data/wlr/CICERO/v2
I can't find any way by following your instructions to build "./data/cicero_v2/preprocessed/" dir and there is no way to save these preprocessed data from pretraining code.