NLP Lineage

26 July 2019 · Language modelling

RoBERTa

RoBERTa: A Robustly Optimized BERT Pretraining Approach

Y. Liu, M. Ott, N. Goyal, et al. (10 authors)

Date
26 July 2019
What the date marks
arXiv submission
Kind
Model
Organisation
Facebook AI / U. Washington
Access
MIT
Threads
In the guided tour
Chapter 4. Pretrain once, use everywhere
Learning phases
2. How LLMs are trained

Notable: no later milestone in the data is documented as descending from it. How levels are decided

Read the primary source (arxiv.org)

Full summary in progress. Until then, this page shows the milestone's record from the research data.

What it changed

Before it

BERT's recipe was taken as settled, and work moved to new architectures.

The problem

Was BERT's ceiling architectural, or had it simply been undertrained?

What it introduced

Same architecture, more data, longer training, bigger batches, no next-sentence prediction. It beat every successor.

Why it matters

An early lesson that training recipe and data often matter more than architectural novelty.

What modern models inherit

Scepticism about architecture claims that are really compute claims.

What came next

The lesson resurfaces in Chinchilla and again in Llama's overtraining.

Before and after

The milestones it descends from, and the milestones that descend from it. Work developed alongside it is not counted. Weak links are left out unless you include them.

Before: where it came from

RoBERTa has 21 documented ancestors across 4 chapters, reaching back to Recurrent networks in 1990. Showing its direct parents and the 6 turning points among them.

RoBERTa has 21 documented ancestors across 4 chapters, reaching back to Recurrent networks in 1990. Showing all of them.

RoBERTa has 22 recorded ancestors, including indirect ones, across 4 chapters, reaching back to Recurrent networks in 1990. Showing its direct parents and the 6 turning points among them.

RoBERTa has 22 recorded ancestors, including indirect ones, across 4 chapters, reaching back to Recurrent networks in 1990. Showing all of them.

  1. 2019RoBERTaThis milestone.
  2. 2018BERT Turning pointRoBERTa built directly on thisChapter 4. Pretrain once, use everywhere
  3. 2018GPT-1BERT was influenced by thisChapter 4. Pretrain once, use everywhere
  4. 2018ELMoBERT was influenced by thisChapter 4. Pretrain once, use everywhere
  5. 2018ULMFiTBERT was influenced by thisChapter 4. Pretrain once, use everywhere
  6. 2017CoVeELMo was influenced by thisChapter 3. Attention, then only attention
  7. 2017Transformer Turning pointBERT built directly on thisChapter 3. Attention, then only attention
  8. 2016Layer normTransformer built directly on thisChapter 3. Attention, then only attention
  9. 2015Residual connectionsTransformer was influenced by thisChapter 3. Attention, then only attention
  10. 2015LM pretraining Turning pointBERT was influenced by thisChapter 2. Words become vectors
  11. 2015BPE tokenizationRoBERTa built directly on thisChapter 2. Words become vectors
  12. 2015Luong attentionTransformer was influenced by thisChapter 3. Attention, then only attention
  13. 2014AdamTransformer was influenced by thisChapter 2. Words become vectors
  14. 2014GloVeELMo built directly on thisChapter 2. Words become vectors
  15. 2014Seq2seq Turning pointTransformer built directly on thisChapter 2. Words become vectors
  16. 2014Attention Turning pointTransformer built directly on thisChapter 3. Attention, then only attention
  17. 2014GRU / encoder-decoderAttention built directly on thisChapter 2. Words become vectors
  18. 2013word2vecELMo was influenced by thisChapter 2. Words become vectors
  19. 2011NLP from scratchELMo was influenced by thisChapter 1. Language as probability
  20. 2003Neural LMNLP from scratch was influenced by this
    word2vec was influenced by this    Chapter 1. Language as probability
  21. 1997LSTM Turning pointELMo built directly on thisChapter 1. Language as probability
  22. 1990Recurrent networksLSTM was influenced by thisChapter 1. Language as probability
  1. 2019RoBERTaThis milestone.
  2. 2018BERT Turning pointRoBERTa built directly on thisChapter 4. Pretrain once, use everywhere
  3. 2018GPT-1BERT was influenced by thisChapter 4. Pretrain once, use everywhere
  4. 2018ELMoBERT was influenced by thisChapter 4. Pretrain once, use everywhere
  5. 2018ULMFiTBERT was influenced by thisChapter 4. Pretrain once, use everywhere
  6. 2017CoVeELMo was influenced by thisChapter 3. Attention, then only attention
  7. 2017Transformer Turning pointBERT built directly on thisChapter 3. Attention, then only attention
  8. 2017ConvS2STransformer was influenced by this (indirect influence)Chapter 3. Attention, then only attention
  9. 2016Layer normTransformer built directly on thisChapter 3. Attention, then only attention
  10. 2015Residual connectionsTransformer was influenced by thisChapter 3. Attention, then only attention
  11. 2015LM pretraining Turning pointBERT was influenced by thisChapter 2. Words become vectors
  12. 2015BPE tokenizationRoBERTa built directly on thisChapter 2. Words become vectors
  13. 2015Luong attentionTransformer was influenced by thisChapter 3. Attention, then only attention
  14. 2014AdamTransformer was influenced by thisChapter 2. Words become vectors
  15. 2014GloVeELMo built directly on thisChapter 2. Words become vectors
  16. 2014Seq2seq Turning pointTransformer built directly on thisChapter 2. Words become vectors
  17. 2014Attention Turning pointTransformer built directly on thisChapter 3. Attention, then only attention
  18. 2014GRU / encoder-decoderAttention built directly on thisChapter 2. Words become vectors
  19. 2013word2vecELMo was influenced by thisChapter 2. Words become vectors
  20. 2011NLP from scratchELMo was influenced by thisChapter 1. Language as probability
  21. 2003Neural LMNLP from scratch was influenced by this
    word2vec was influenced by this    Chapter 1. Language as probability
  22. 1997LSTM Turning pointELMo built directly on thisChapter 1. Language as probability
  23. 1990Recurrent networksLSTM was influenced by thisChapter 1. Language as probability

After: what it led to

No documented descendants are recorded for RoBERTa without weak links.

No documented descendants are recorded for RoBERTa without weak links.

No recorded descendants are recorded for RoBERTa.

No recorded descendants are recorded for RoBERTa.

Direct links and their sources

Where it came from

Built directly on

What it led to

None recorded in the sources used.

Terms