NLP Lineage

1 June 2023 · Language modelling

RefinedWeb

The RefinedWeb Dataset for Falcon LLM

G. Penedo, Q. Malartic, D. Hesslow, et al. (9 authors)

Date
1 June 2023
What the date marks
arXiv submission
Kind
Dataset
Organisation
TII UAE
Access
ODC-By (subset)
Other dates
NeurIPS 2023
Threads
In the guided tour
Chapter 8. The weights get out
Learning phases
8. Training from scratch

A milestone: 3 later milestones descend directly from it, and its descendants reach 4 of the 10 threads. 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

Best practice mixed filtered web text with curated sources like books and Wikipedia.

The problem

Curated corpora are small and legally awkward; is the curation actually necessary?

What it introduced

Aggressive filtering and deduplication of web data alone, matching or beating models trained on curated mixes.

Why it matters

It reframed data quality as a filtering problem rather than a sourcing problem, which is how the field now treats it.

What modern models inherit

Web-only pretraining data with heavy filtering.

What came next

Dolma and FineWeb refined the same pipeline in the open.

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

RefinedWeb has 1 documented ancestors across 1 chapter, reaching back to C4 in 2019. Showing all of them.

RefinedWeb has 1 documented ancestors across 1 chapter, reaching back to C4 in 2019. Showing all of them.

RefinedWeb 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.

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

  1. 2023RefinedWebThis milestone.
  2. 2019C4 Turning pointRefinedWeb was influenced by thisChapter 4. Pretrain once, use everywhere
  1. 2023RefinedWebThis milestone.
  2. 2019C4 Turning pointRefinedWeb was influenced by thisChapter 4. Pretrain once, use everywhere
  3. 2019WebTextC4 was influenced by this (indirect influence)Chapter 4. Pretrain once, use everywhere
  4. 2018GPT-1WebText was influenced by thisChapter 4. Pretrain once, use everywhere
  5. 2018ELMoGPT-1 was influenced by this (indirect influence)Chapter 4. Pretrain once, use everywhere
  6. 2018ULMFiTGPT-1 was influenced by this (indirect influence)Chapter 4. Pretrain once, use everywhere
  7. 2017CoVeELMo was influenced by thisChapter 3. Attention, then only attention
  8. 2017Transformer Turning pointGPT-1 built directly on thisChapter 3. Attention, then only attention
  9. 2017ConvS2STransformer was influenced by this (indirect influence)Chapter 3. Attention, then only attention
  10. 2016Layer normTransformer built directly on thisChapter 3. Attention, then only attention
  11. 2015Residual connectionsTransformer was influenced by thisChapter 3. Attention, then only attention
  12. 2015LM pretraining Turning pointGPT-1 was influenced by 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

RefinedWeb has 4 documented descendants across 1 chapter, reaching forward to FineWeb in 2024. Showing its direct descendants.

RefinedWeb has 4 documented descendants across 1 chapter, reaching forward to FineWeb in 2024. Showing all of them.

RefinedWeb has 4 recorded descendants, including indirect ones, across 1 chapter, reaching forward to FineWeb in 2024. Showing its direct descendants.

RefinedWeb has 4 recorded descendants, including indirect ones, across 1 chapter, reaching forward to FineWeb in 2024. Showing all of them.

  1. 2023RefinedWebThis milestone.
  2. 2023Falconbuilt directly on RefinedWebChapter 8. The weights get out
  3. 2024Dolmawas influenced by RefinedWebChapter 8. The weights get out
  4. 2024OLMobuilt directly on DolmaChapter 8. The weights get out
  5. 2024FineWebbuilt directly on RefinedWebChapter 8. The weights get out
  1. 2023RefinedWebThis milestone.
  2. 2023Falconbuilt directly on RefinedWebChapter 8. The weights get out
  3. 2024Dolmawas influenced by RefinedWebChapter 8. The weights get out
  4. 2024OLMobuilt directly on DolmaChapter 8. The weights get out
  5. 2024FineWebbuilt directly on RefinedWebChapter 8. The weights get out

Direct links and their sources

Where it came from

Influenced by

What it led to

Built on directly by

Influenced