NLP Lineage

17 April 2023 · Alignment

LLaVA

Visual Instruction Tuning

H. Liu, C. Li, Q. Wu, Y. Lee

Date
17 April 2023
What the date marks
arXiv submission
Kind
Model
Organisation
U. Wisconsin-Madison / Microsoft / Columbia
Access
Apache 2.0
Other dates
NeurIPS 2023 Oral
Threads
In the guided tour
Chapter 11. Seeing and hearing
Learning phases
4. Embeddings and RAG

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

Vision-language models came from large labs with large budgets.

The problem

There was no open recipe for giving an open language model eyes.

What it introduced

A simple projection layer between CLIP and an open LLM, trained on visual instruction data generated by GPT-4.

Why it matters

It made open multimodal models buildable in a day, and set the template most open VLMs still use.

What modern models inherit

The projector-between-encoder-and-LLM design.

What came next

A wave of open vision-language models, and the multimodal expectations that followed.

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

LLaVA has 43 documented ancestors across 10 chapters, reaching back to Recurrent networks in 1990. Showing its direct parents and the 19 turning points among them.

LLaVA has 43 documented ancestors across 10 chapters, reaching back to Recurrent networks in 1990. Showing all of them.

LLaVA has 45 recorded ancestors, including indirect ones, across 10 chapters, reaching back to Recurrent networks in 1990. Showing its direct parents and the 19 turning points among them.

LLaVA has 45 recorded ancestors, including indirect ones, across 10 chapters, reaching back to Recurrent networks in 1990. Showing all of them.

  1. 2023LLaVAThis milestone.
  2. 2023GPT-4 Turning pointLLaVA built directly on thisChapter 7. The assistant goes public
  3. 2023LLaMA Turning pointLLaVA built directly on thisChapter 8. The weights get out
  4. 2022GPT-3.5 Turning pointGPT-4 succeeded thisChapter 7. The assistant goes public
  5. 2022Chinchilla Turning pointLLaMA built directly on thisChapter 5. Scale becomes a science
  6. 2022InstructGPT Turning pointGPT-4 built directly on thisChapter 6. Teaching models what people want
  7. 2021GopherChinchilla built directly on thisChapter 5. Scale becomes a science
  8. 2021FLANInstructGPT was influenced by thisChapter 6. Teaching models what people want
  9. 2021RoPE Turning pointLLaMA built directly on thisChapter 10. Reading more than you remember
  10. 2021CLIPLLaVA built directly on thisChapter 11. Seeing and hearing
  11. 2020Vision TransformerCLIP built directly on thisChapter 11. Seeing and hearing
  12. 2020RLHF for summarisation Turning pointInstructGPT built directly on thisChapter 6. Teaching models what people want
  13. 2020GPT-3 Turning pointInstructGPT built directly on this
    GPT-3.5 built directly on this    Chapter 5. Scale becomes a science
  14. 2020Scaling laws Turning pointGPT-4 built directly on thisChapter 5. Scale becomes a science
  15. 2019T5 Turning pointFLAN was influenced by thisChapter 4. Pretrain once, use everywhere
  16. 2019C4 Turning pointT5 was influenced by thisChapter 4. Pretrain once, use everywhere
  17. 2019RLHF for LMsRLHF for summarisation built directly on thisChapter 6. Teaching models what people want
  18. 2019Sparse Transformer Turning pointGPT-3 built directly on thisChapter 4. Pretrain once, use everywhere
  19. 2019GPT-2 Turning pointCLIP built directly on thisChapter 4. Pretrain once, use everywhere
  20. 2019WebTextGPT-2 was influenced by thisChapter 4. Pretrain once, use everywhere
  21. 2019Transformer-XLRoPE was influenced by thisChapter 4. Pretrain once, use everywhere
  22. 2018BERT Turning pointVision Transformer was influenced by thisChapter 4. Pretrain once, use everywhere
  23. 2018SentencePieceT5 built directly on thisChapter 4. Pretrain once, use everywhere
  24. 2018GPT-1GPT-2 built directly on thisChapter 4. Pretrain once, use everywhere
  25. 2018ELMoBERT was influenced by thisChapter 4. Pretrain once, use everywhere
  26. 2018ULMFiTBERT was influenced by thisChapter 4. Pretrain once, use everywhere
  27. 2017CoVeELMo was influenced by thisChapter 3. Attention, then only attention
  28. 2017Transformer Turning pointGPT-2 built directly on this
    Scaling laws built directly on this
    RoPE built directly on this    Chapter 3. Attention, then only attention
  29. 2017RLHF originRLHF for summarisation built directly on thisChapter 6. Teaching models what people want
  30. 2016Layer normGPT-2 built directly on thisChapter 3. Attention, then only attention
  31. 2015Residual connectionsTransformer was influenced by thisChapter 3. Attention, then only attention
  32. 2015LM pretraining Turning pointBERT was influenced by this
    GPT-1 was influenced by this    Chapter 2. Words become vectors
  33. 2015BPE tokenizationLLaMA built directly on thisChapter 2. Words become vectors
  34. 2015Luong attentionTransformer was influenced by thisChapter 3. Attention, then only attention
  35. 2014AdamTransformer was influenced by thisChapter 2. Words become vectors
  36. 2014GloVeELMo built directly on thisChapter 2. Words become vectors
  37. 2014Seq2seq Turning pointTransformer built directly on thisChapter 2. Words become vectors
  38. 2014Attention Turning pointTransformer built directly on thisChapter 3. Attention, then only attention
  39. 2014GRU / encoder-decoderAttention built directly on thisChapter 2. Words become vectors
  40. 2013word2vecELMo was influenced by thisChapter 2. Words become vectors
  41. 2011NLP from scratchELMo was influenced by thisChapter 1. Language as probability
  42. 2003Neural LMNLP from scratch was influenced by this
    word2vec was influenced by this    Chapter 1. Language as probability
  43. 1997LSTM Turning pointSeq2seq built directly on this
    ELMo built directly on this    Chapter 1. Language as probability
  44. 1990Recurrent networksLSTM was influenced by thisChapter 1. Language as probability
  1. 2023LLaVAThis milestone.
  2. 2023GPT-4 Turning pointLLaVA built directly on thisChapter 7. The assistant goes public
  3. 2023LLaMA Turning pointLLaVA built directly on thisChapter 8. The weights get out
  4. 2022GPT-3.5 Turning pointGPT-4 succeeded thisChapter 7. The assistant goes public
  5. 2022FlamingoLLaVA was influenced by this (indirect influence)Chapter 11. Seeing and hearing
  6. 2022Chinchilla Turning pointLLaMA built directly on thisChapter 5. Scale becomes a science
  7. 2022InstructGPT Turning pointGPT-4 built directly on thisChapter 6. Teaching models what people want
  8. 2021GopherChinchilla built directly on thisChapter 5. Scale becomes a science
  9. 2021FLANInstructGPT was influenced by thisChapter 6. Teaching models what people want
  10. 2021RoPE Turning pointLLaMA built directly on thisChapter 10. Reading more than you remember
  11. 2021CLIPLLaVA built directly on thisChapter 11. Seeing and hearing
  12. 2020Vision TransformerCLIP built directly on this
    Flamingo built directly on this    Chapter 11. Seeing and hearing
  13. 2020RLHF for summarisation Turning pointInstructGPT built directly on thisChapter 6. Teaching models what people want
  14. 2020GPT-3 Turning pointFlamingo was influenced by thisChapter 5. Scale becomes a science
  15. 2020Scaling laws Turning pointGPT-4 built directly on thisChapter 5. Scale becomes a science
  16. 2019T5 Turning pointFLAN was influenced by thisChapter 4. Pretrain once, use everywhere
  17. 2019C4 Turning pointT5 was influenced by thisChapter 4. Pretrain once, use everywhere
  18. 2019RLHF for LMsRLHF for summarisation built directly on thisChapter 6. Teaching models what people want
  19. 2019Sparse Transformer Turning pointGPT-3 built directly on thisChapter 4. Pretrain once, use everywhere
  20. 2019GPT-2 Turning pointCLIP built directly on thisChapter 4. Pretrain once, use everywhere
  21. 2019WebTextGPT-2 was influenced by thisChapter 4. Pretrain once, use everywhere
  22. 2019Transformer-XLRoPE was influenced by thisChapter 4. Pretrain once, use everywhere
  23. 2018BERT Turning pointVision Transformer was influenced by thisChapter 4. Pretrain once, use everywhere
  24. 2018SentencePieceT5 built directly on thisChapter 4. Pretrain once, use everywhere
  25. 2018GPT-1GPT-2 built directly on thisChapter 4. Pretrain once, use everywhere
  26. 2018ELMoBERT was influenced by this
    GPT-1 was influenced by this (indirect influence)    Chapter 4. Pretrain once, use everywhere
  27. 2018ULMFiTBERT was influenced by this
    GPT-1 was influenced by this (indirect influence)    Chapter 4. Pretrain once, use everywhere
  28. 2017CoVeELMo was influenced by thisChapter 3. Attention, then only attention
  29. 2017Transformer Turning pointGPT-2 built directly on this
    Scaling laws built directly on this
    RoPE built directly on this    Chapter 3. Attention, then only attention
  30. 2017RLHF originRLHF for summarisation built directly on thisChapter 6. Teaching models what people want
  31. 2017ConvS2STransformer was influenced by this (indirect influence)Chapter 3. Attention, then only attention
  32. 2016Layer normGPT-2 built directly on thisChapter 3. Attention, then only attention
  33. 2015Residual connectionsTransformer was influenced by thisChapter 3. Attention, then only attention
  34. 2015LM pretraining Turning pointBERT was influenced by this
    GPT-1 was influenced by this    Chapter 2. Words become vectors
  35. 2015BPE tokenizationLLaMA built directly on thisChapter 2. Words become vectors
  36. 2015Luong attentionTransformer was influenced by thisChapter 3. Attention, then only attention
  37. 2014AdamTransformer was influenced by thisChapter 2. Words become vectors
  38. 2014GloVeELMo built directly on thisChapter 2. Words become vectors
  39. 2014Seq2seq Turning pointTransformer built directly on thisChapter 2. Words become vectors
  40. 2014Attention Turning pointTransformer built directly on thisChapter 3. Attention, then only attention
  41. 2014GRU / encoder-decoderAttention built directly on thisChapter 2. Words become vectors
  42. 2013word2vecELMo was influenced by thisChapter 2. Words become vectors
  43. 2011NLP from scratchELMo was influenced by thisChapter 1. Language as probability
  44. 2003Neural LMNLP from scratch was influenced by this
    word2vec was influenced by this    Chapter 1. Language as probability
  45. 1997LSTM Turning pointSeq2seq built directly on this
    ELMo built directly on this    Chapter 1. Language as probability
  46. 1990Recurrent networksLSTM was influenced by thisChapter 1. Language as probability

After: what it led to

No documented descendants are recorded for LLaVA without weak links.

No documented descendants are recorded for LLaVA without weak links.

No recorded descendants are recorded for LLaVA.

No recorded descendants are recorded for LLaVA.

Direct links and their sources

Where it came from

Built directly on

Influenced by

What it led to

None recorded in the sources used.