A lot has happened in NLP from the time I released models.pratik.ai.
I received quite positive feedback and have decided to keep it updated.
Here's what got added recently 👻
Models suitable for mobile
MobileBERT (5.3x faster than BERT)
SqueezeBERT (4.3x faster than BERT)
Long text (>512)
BigBird (Linear compute with sparse attention)
LaBSE (93 language support)
LASER (103 language support)
Dense Passage Retrieval + Transformer Reader
PEGASUS (SOTA on summarisation)
Replaced BioBERT with ouBioBERT as it supersedes it for medical tasks
Have suggestions? Create an issue @ github.com/bhavsarpratik/nlp-models
This paper finds that if we remove the last 6 layers of BERT, we still get the same performance. No need for fancy distillation. Both are of the same size 66M. This clearly means we love over-engineering. We over-engineered BERT. Then we over-engineered ways to compress BERT 🤫
Come join Maxpool - A Data Science community to discuss real ML problems!
Ask me anything on ama.pratik.ai 👻
You can try ask.pratik.ai for any study material.
Let me know your suggestions via feedback.pratik.ai 😃