Real_toxicity_Prompts

  • شرح :

RealToxicityPrompts: ارزیابی انحطاط سمی عصبی در مدل‌های زبان

RealToxicityPrompts مجموعه‌ای از اعلان‌های انگلیسی (به‌ویژه آغاز جملات) با سمیت متفاوت است. این اعلان‌ها به‌عنوان زمینه‌ای برای یک مدل زبان خودرگرسیون (یعنی GPT-2) ارائه می‌شوند و برای ایجاد تکمیل‌ها استفاده می‌شوند.

جزئیات بیشتر در مقاله اصلی ارائه شده است.

شکاف مثال ها
'train' 99,442
  • ساختار ویژگی :
FeaturesDict({
   
'begin': int32,
   
'challenging': bool,
   
'continuation': FeaturesDict({
       
'flirtation': float32,
       
'identity_attack': float32,
       
'insult': float32,
       
'profanity': float32,
       
'severe_toxicity': float32,
       
'sexually_explicit': float32,
       
'text': Text(shape=(), dtype=string),
       
'threat': float32,
       
'toxicity': float32,
   
}),
   
'end': int32,
   
'filename': Text(shape=(), dtype=string),
   
'prompt': FeaturesDict({
       
'flirtation': float32,
       
'identity_attack': float32,
       
'insult': float32,
       
'profanity': float32,
       
'severe_toxicity': float32,
       
'sexually_explicit': float32,
       
'text': Text(shape=(), dtype=string),
       
'threat': float32,
       
'toxicity': float32,
   
}),
})
  • مستندات ویژگی :
ویژگی کلاس شکل نوع D شرح
FeaturesDict
شروع تانسور int32
چالش برانگیز تانسور بوول
ادامه FeaturesDict
ادامه/ معاشقه تانسور float32
Continuation/identity_attack تانسور float32
ادامه / توهین تانسور float32
ادامه / فحاشی تانسور float32
ادامه/سمیت_شدید تانسور float32
ادامه/جنسی_صریح تانسور float32
ادامه / متن متن رشته
ادامه/تهدید تانسور float32
ادامه / سمیت تانسور float32
پایان تانسور int32
نام فایل متن رشته
سریع FeaturesDict
سریع / معاشقه تانسور float32
prompt/identity_attack تانسور float32
سریع / توهین تانسور float32
سریع / ناسزا تانسور float32
prompt/severe_toxicity تانسور float32
prompt/sexually_explicit تانسور float32
اعلان / متن متن رشته
اعلان / تهدید تانسور float32
سریع / سمیت تانسور float32
  • نقل قول :
@article{gehman2020realtoxicityprompts,
  title
={Realtoxicityprompts: Evaluating neural toxic degeneration in language models},
  author
={Gehman, Samuel and Gururangan, Suchin and Sap, Maarten and Choi, Yejin and Smith, Noah A},
  journal
={arXiv preprint arXiv:2009.11462},
  year
={2020}
}