timoschick/pet
A semi-supervised training procedure for few-shot NLP tasks that reformulates inputs as cloze-style phrases.
Not currently ranked — collecting fresh signals.
star history
Pattern-Exploiting Training (PET) is a semi-supervised training procedure for language models that reformulates input examples as cloze-style phrases. It outperforms regular supervised training and even GPT-3 in low-resource settings while requiring 99.9% fewer parameters. The iterative variant iPET trains multiple generations of models and can operate without any training data, making it effective for few-shot text classification and natural language inference tasks.
Frequently asked
- What is timoschick/pet?
- A semi-supervised training procedure for few-shot NLP tasks that reformulates inputs as cloze-style phrases.
- Is pet open source?
- Yes — timoschick/pet is open source, released under the Apache-2.0 license.
- What language is pet written in?
- timoschick/pet is primarily written in Python.
- How popular is pet?
- timoschick/pet has 1.6k stars on GitHub.
- Where can I find pet?
- timoschick/pet is on GitHub at https://github.com/timoschick/pet.