NLP Lineage

25 June 2024 · Scale and data

FineWeb

The FineWeb Datasets: Decanting the Web for the Finest Text Data at Scale

G. Penedo, H. Kydlíček, L. Ben Allal, et al. (8 authors)

Date
25 June 2024
What the date marks
arXiv submission
Kind
Dataset
Organisation
Hugging Face
Access
ODC-BY
Other dates
NeurIPS 2024
Threads
In the guided tour
Chapter 8. The weights get out
Learning phases
2. How LLMs are trained; 8. Training from scratch

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

Data recipes were described in papers but rarely released with their ablations.

The problem

Everyone rebuilt the same filtering pipeline and could not compare choices.

What it introduced

Fifteen trillion tokens plus the ablation study behind every filtering decision, and an education-filtered subset.

Why it matters

It made data curation an open empirical science, and FineWeb-Edu showed filtering for educational value beats raw volume.

What modern models inherit

Published ablations as part of a dataset release.

What came next

The standard open pretraining corpus for new model efforts.

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

FineWeb has 38 documented ancestors across 8 chapters, reaching back to Recurrent networks in 1990. Showing its direct parents and the 15 turning points among them.

FineWeb has 38 documented ancestors across 8 chapters, reaching back to Recurrent networks in 1990. Showing all of them.

FineWeb has 41 recorded ancestors, including indirect ones, across 8 chapters, reaching back to Recurrent networks in 1990. Showing its direct parents and the 15 turning points among them.

FineWeb has 41 recorded ancestors, including indirect ones, across 8 chapters, reaching back to Recurrent networks in 1990. Showing all of them.

  1. 2024FineWebThis milestone.
  2. 2023phi-1FineWeb was influenced by thisChapter 8. The weights get out
  3. 2023RefinedWebFineWeb built directly on thisChapter 8. The weights get out
  4. 2022GPT-3.5 Turning pointphi-1 was influenced by thisChapter 7. The assistant goes public
  5. 2022InstructGPT Turning pointGPT-3.5 built directly on thisChapter 6. Teaching models what people want
  6. 2021FLANInstructGPT was influenced by thisChapter 6. Teaching models what people want
  7. 2020The PileFineWeb was influenced by thisChapter 5. Scale becomes a science
  8. 2020RLHF for summarisation Turning pointInstructGPT built directly on thisChapter 6. Teaching models what people want
  9. 2020GPT-3 Turning pointThe Pile was influenced by thisChapter 5. Scale becomes a science
  10. 2020Scaling laws Turning pointGPT-3 was influenced by thisChapter 5. Scale becomes a science
  11. 2019C4 Turning pointThe Pile was influenced by this
    RefinedWeb was influenced by this    Chapter 4. Pretrain once, use everywhere
  12. 2019T5 Turning pointFLAN was influenced by thisChapter 4. Pretrain once, use everywhere
  13. 2019RLHF for LMsRLHF for summarisation built directly on thisChapter 6. Teaching models what people want
  14. 2019Sparse Transformer Turning pointGPT-3 built directly on thisChapter 4. Pretrain once, use everywhere
  15. 2019GPT-2 Turning pointGPT-3 built directly on thisChapter 4. Pretrain once, use everywhere
  16. 2019WebTextGPT-2 was influenced by thisChapter 4. Pretrain once, use everywhere
  17. 2018BERT Turning pointT5 was influenced by thisChapter 4. Pretrain once, use everywhere
  18. 2018SentencePieceT5 built directly on thisChapter 4. Pretrain once, use everywhere
  19. 2018GPT-1GPT-2 built directly on thisChapter 4. Pretrain once, use everywhere
  20. 2018ELMoBERT was influenced by thisChapter 4. Pretrain once, use everywhere
  21. 2018ULMFiTBERT was influenced by thisChapter 4. Pretrain once, use everywhere
  22. 2017CoVeELMo was influenced by thisChapter 3. Attention, then only attention
  23. 2017Transformer Turning pointGPT-2 built directly on this
    Sparse Transformer built directly on this
    Scaling laws built directly on this    Chapter 3. Attention, then only attention
  24. 2017RLHF originRLHF for summarisation built directly on thisChapter 6. Teaching models what people want
  25. 2016Layer normGPT-2 built directly on thisChapter 3. Attention, then only attention
  26. 2015Residual connectionsTransformer was influenced by thisChapter 3. Attention, then only attention
  27. 2015LM pretraining Turning pointGPT-1 was influenced by thisChapter 2. Words become vectors
  28. 2015BPE tokenizationGPT-2 built directly on thisChapter 2. Words become vectors
  29. 2015Luong attentionTransformer was influenced by thisChapter 3. Attention, then only attention
  30. 2014AdamTransformer was influenced by thisChapter 2. Words become vectors
  31. 2014GloVeELMo built directly on thisChapter 2. Words become vectors
  32. 2014Seq2seq Turning pointTransformer built directly on thisChapter 2. Words become vectors
  33. 2014Attention Turning pointTransformer built directly on thisChapter 3. Attention, then only attention
  34. 2014GRU / encoder-decoderAttention built directly on thisChapter 2. Words become vectors
  35. 2013word2vecELMo was influenced by thisChapter 2. Words become vectors
  36. 2011NLP from scratchELMo was influenced by thisChapter 1. Language as probability
  37. 2003Neural LMNLP from scratch was influenced by this
    word2vec was influenced by this    Chapter 1. Language as probability
  38. 1997LSTM Turning pointSeq2seq built directly on thisChapter 1. Language as probability
  39. 1990Recurrent networksLSTM was influenced by thisChapter 1. Language as probability
  1. 2024FineWebThis milestone.
  2. 2023phi-1FineWeb was influenced by thisChapter 8. The weights get out
  3. 2023RefinedWebFineWeb built directly on thisChapter 8. The weights get out
  4. 2022Self-Instructphi-1 was influenced by this (indirect influence)Chapter 6. Teaching models what people want
  5. 2022GPT-3.5 Turning pointphi-1 was influenced by thisChapter 7. The assistant goes public
  6. 2022InstructGPT Turning pointGPT-3.5 built directly on thisChapter 6. Teaching models what people want
  7. 2021FLANSelf-Instruct was influenced by thisChapter 6. Teaching models what people want
  8. 2020The PileFineWeb was influenced by thisChapter 5. Scale becomes a science
  9. 2020RLHF for summarisation Turning pointInstructGPT built directly on thisChapter 6. Teaching models what people want
  10. 2020GPT-3 Turning pointThe Pile was influenced by thisChapter 5. Scale becomes a science
  11. 2020Scaling laws Turning pointGPT-3 was influenced by thisChapter 5. Scale becomes a science
  12. 2019C4 Turning pointThe Pile was influenced by this
    RefinedWeb was influenced by this    Chapter 4. Pretrain once, use everywhere
  13. 2019T5 Turning pointFLAN was influenced by thisChapter 4. Pretrain once, use everywhere
  14. 2019RLHF for LMsRLHF for summarisation built directly on thisChapter 6. Teaching models what people want
  15. 2019Sparse Transformer Turning pointGPT-3 built directly on thisChapter 4. Pretrain once, use everywhere
  16. 2019WebTextC4 was influenced by this (indirect influence)Chapter 4. Pretrain once, use everywhere
  17. 2019GPT-2 Turning pointGPT-3 built directly on thisChapter 4. Pretrain once, use everywhere
  18. 2019Transformer-XLSparse Transformer was influenced by this (indirect influence)Chapter 4. Pretrain once, use everywhere
  19. 2018BERT Turning pointT5 was influenced by thisChapter 4. Pretrain once, use everywhere
  20. 2018SentencePieceT5 built directly on thisChapter 4. Pretrain once, use everywhere
  21. 2018GPT-1GPT-2 built directly on this
    WebText was influenced by this    Chapter 4. Pretrain once, use everywhere
  22. 2018ELMoGPT-1 was influenced by this (indirect influence)Chapter 4. Pretrain once, use everywhere
  23. 2018ULMFiTGPT-1 was influenced by this (indirect influence)Chapter 4. Pretrain once, use everywhere
  24. 2017CoVeELMo was influenced by thisChapter 3. Attention, then only attention
  25. 2017Transformer Turning pointGPT-2 built directly on this
    Sparse Transformer built directly on this
    Scaling laws built directly on this    Chapter 3. Attention, then only attention
  26. 2017RLHF originRLHF for summarisation built directly on thisChapter 6. Teaching models what people want
  27. 2017ConvS2STransformer was influenced by this (indirect influence)Chapter 3. Attention, then only attention
  28. 2016Layer normGPT-2 built directly on thisChapter 3. Attention, then only attention
  29. 2015Residual connectionsTransformer was influenced by thisChapter 3. Attention, then only attention
  30. 2015LM pretraining Turning pointGPT-1 was influenced by thisChapter 2. Words become vectors
  31. 2015BPE tokenizationGPT-2 built directly on thisChapter 2. Words become vectors
  32. 2015Luong attentionTransformer was influenced by thisChapter 3. Attention, then only attention
  33. 2014AdamTransformer was influenced by thisChapter 2. Words become vectors
  34. 2014GloVeELMo built directly on thisChapter 2. Words become vectors
  35. 2014Seq2seq Turning pointTransformer built directly on thisChapter 2. Words become vectors
  36. 2014Attention Turning pointTransformer built directly on thisChapter 3. Attention, then only attention
  37. 2014GRU / encoder-decoderAttention built directly on thisChapter 2. Words become vectors
  38. 2013word2vecELMo was influenced by thisChapter 2. Words become vectors
  39. 2011NLP from scratchELMo was influenced by thisChapter 1. Language as probability
  40. 2003Neural LMNLP from scratch was influenced by this
    word2vec was influenced by this    Chapter 1. Language as probability
  41. 1997LSTM Turning pointSeq2seq built directly on this
    ELMo built directly on this    Chapter 1. Language as probability
  42. 1990Recurrent networksLSTM was influenced by thisChapter 1. Language as probability

After: what it led to

No documented descendants are recorded for FineWeb without weak links.

No documented descendants are recorded for FineWeb without weak links.

No recorded descendants are recorded for FineWeb.

No recorded descendants are recorded for FineWeb.

Direct links and their sources

Where it came from

Built directly on

Influenced by

What it led to

None recorded in the sources used.