search

INK-USC / CrossFit

Code for paper "CrossFit :weight_lifting:: A Few-shot Learning Challenge for Cross-task Generalization in NLP" (https://arxiv.org/abs/2104.08835)
102 stars 6 forks source link

Absolute Scores for Random Multi-task Learning #8

Open JonathanJao opened 2 years ago

JonathanJao commented 2 years ago

Hi, our team would like to cite your paper and compare our approach with your scores for multi-task learning random setting. The Relative scores are displayed in your paper in Figure 4(a), would you be able to share the

  1. multi-task learning absolute scores and
  2. direct fine tuning base line scores.

Specifically if we could get absolute scores for the following 20 tasks:

["quoref", "wiki_split", "ethos-disability", "yelp_polarity", "superglue-rte", "glue-cola", "ethos-sexual_orientation", "blimp-sentential_negation_npi_scope", "ai2_arc", "amazon_polarity", "race-high", "blimp-sentential_negation_npi_licensor_present", "tweet_eval-irony", "break-QDMR", "crawl_domain", "freebase_qa", "glue-qnli", "hatexplain", "ag_news", "circa"]

that would be great!

Thanks!

cherry979988 commented 2 years ago

Hi @JonathanJao

Thanks for raising this issue. Here they are: https://docs.google.com/spreadsheets/d/1ICX825URGOd0G3DFn44in0GlufKQTbDbAlw-gC-0x_Y/edit?usp=sharing