NLP Lineage

12 April 2022 · Alignment

HH-RLHF

Training a Helpful and Harmless Assistant with RLHF

Y. Bai, A. Jones, K. Ndousse, et al. (31 authors)

Date
12 April 2022
What the date marks
arXiv submission
Kind
Training method
Organisation
Anthropic
Access
MIT (dataset)
Other dates
HH-RLHF dataset released
Threads
In the guided tour
Chapter 6. Teaching models what people want
Learning phases
2. How LLMs are trained

A milestone: 2 later milestones descend directly from it, and its descendants reach 7 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

RLHF results came from one lab, with private data.

The problem

Helpfulness and harmlessness pull against each other, and nobody had studied the tradeoff openly.

What it introduced

A full RLHF assistant study separating the two objectives, with the human preference dataset released.

Why it matters

It gave the open community real preference data and named the tension every assistant has to manage.

What modern models inherit

The helpfulness-harmlessness tradeoff, and open preference datasets.

What came next

Constitutional AI replaced much of the human harmlessness labelling with AI feedback.

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

HH-RLHF has 19 documented ancestors across 5 chapters, reaching back to Recurrent networks in 1990. Showing its direct parents and the 7 turning points among them.

HH-RLHF has 19 documented ancestors across 5 chapters, reaching back to Recurrent networks in 1990. Showing all of them.

HH-RLHF has 27 recorded ancestors, including indirect ones, across 5 chapters, reaching back to Recurrent networks in 1990. Showing its direct parents and the 7 turning points among them.

HH-RLHF has 27 recorded ancestors, including indirect ones, across 5 chapters, reaching back to Recurrent networks in 1990. Showing all of them.

  1. 2022HH-RLHFThis milestone.
  2. 2021HHH assistantHH-RLHF built directly on thisChapter 6. Teaching models what people want
  3. 2020RLHF for summarisation Turning pointHH-RLHF built directly on thisChapter 6. Teaching models what people want
  4. 2019RLHF for LMsRLHF for summarisation built directly on thisChapter 6. Teaching models what people want
  5. 2019GPT-2 Turning pointRLHF for LMs built directly on thisChapter 4. Pretrain once, use everywhere
  6. 2019WebTextGPT-2 was influenced by thisChapter 4. Pretrain once, use everywhere
  7. 2018GPT-1GPT-2 built directly on thisChapter 4. Pretrain once, use everywhere
  8. 2017RLHF originRLHF for summarisation built directly on thisChapter 6. Teaching models what people want
  9. 2017Transformer Turning pointGPT-2 built directly on thisChapter 3. Attention, then only attention
  10. 2016Layer normGPT-2 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. 2015BPE tokenizationGPT-2 built directly on thisChapter 2. Words become vectors
  14. 2015Luong attentionTransformer was influenced by thisChapter 3. Attention, then only attention
  15. 2014AdamTransformer was influenced by 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. 1997LSTM Turning pointSeq2seq built directly on thisChapter 1. Language as probability
  20. 1990Recurrent networksLSTM was influenced by thisChapter 1. Language as probability
  1. 2022HH-RLHFThis milestone.
  2. 2021HHH assistantHH-RLHF built directly on thisChapter 6. Teaching models what people want
  3. 2020RLHF for summarisation Turning pointHH-RLHF built directly on thisChapter 6. Teaching models what people want
  4. 2019RLHF for LMsRLHF for summarisation built directly on thisChapter 6. Teaching models what people want
  5. 2019GPT-2 Turning pointRLHF for LMs built directly on thisChapter 4. Pretrain once, use everywhere
  6. 2019WebTextGPT-2 was influenced by thisChapter 4. Pretrain once, use everywhere
  7. 2018GPT-1GPT-2 built directly on thisChapter 4. Pretrain once, use everywhere
  8. 2018ELMoGPT-1 was influenced by this (indirect influence)Chapter 4. Pretrain once, use everywhere
  9. 2018ULMFiTGPT-1 was influenced by this (indirect influence)Chapter 4. Pretrain once, use everywhere
  10. 2017CoVeELMo was influenced by thisChapter 3. Attention, then only attention
  11. 2017RLHF originRLHF for summarisation built directly on thisChapter 6. Teaching models what people want
  12. 2017Transformer Turning pointGPT-2 built directly on thisChapter 3. Attention, then only attention
  13. 2017ConvS2STransformer was influenced by this (indirect influence)Chapter 3. Attention, then only attention
  14. 2016Layer normGPT-2 built directly on thisChapter 3. Attention, then only attention
  15. 2015Residual connectionsTransformer was influenced by thisChapter 3. Attention, then only attention
  16. 2015LM pretraining Turning pointGPT-1 was influenced by thisChapter 2. Words become vectors
  17. 2015BPE tokenizationGPT-2 built directly on thisChapter 2. Words become vectors
  18. 2015Luong attentionTransformer was influenced by thisChapter 3. Attention, then only attention
  19. 2014AdamTransformer was influenced by thisChapter 2. Words become vectors
  20. 2014GloVeELMo built directly on thisChapter 2. Words become vectors
  21. 2014Seq2seq Turning pointTransformer built directly on thisChapter 2. Words become vectors
  22. 2014Attention Turning pointTransformer built directly on thisChapter 3. Attention, then only attention
  23. 2014GRU / encoder-decoderAttention built directly on thisChapter 2. Words become vectors
  24. 2013word2vecELMo was influenced by thisChapter 2. Words become vectors
  25. 2011NLP from scratchELMo was influenced by thisChapter 1. Language as probability
  26. 2003Neural LMNLP from scratch was influenced by this
    word2vec was influenced by this    Chapter 1. Language as probability
  27. 1997LSTM Turning pointSeq2seq built directly on this
    ELMo built directly on this    Chapter 1. Language as probability
  28. 1990Recurrent networksLSTM was influenced by thisChapter 1. Language as probability

After: what it led to

HH-RLHF has 17 documented descendants across 6 chapters, reaching forward to Claude Opus 5.5 in 2026. Showing its direct descendants.

HH-RLHF has 17 documented descendants across 6 chapters, reaching forward to Claude Opus 5.5 in 2026. Showing all of them.

HH-RLHF has 32 recorded descendants, including indirect ones, across 6 chapters, reaching forward to Claude Opus 5.5 in 2026. Showing its direct descendants.

HH-RLHF has 32 recorded descendants, including indirect ones, across 6 chapters, reaching forward to Claude Opus 5.5 in 2026. Showing all of them.

  1. 2022HH-RLHFThis milestone.
  2. 2022Constitutional AIbuilt directly on HH-RLHFChapter 6. Teaching models what people want
  3. 2023Claudewas influenced by HH-RLHFChapter 7. The assistant goes public
  4. 2023100K contextbuilt directly on ClaudeChapter 10. Reading more than you remember
  5. 2024Claude 3built directly on Claude
    built directly on Constitutional AI    Chapter 14. Where the threads meet now
  6. 2024Claude 3.5 Sonnetbuilt directly on Claude 3Chapter 14. Where the threads meet now
  7. 2024Computer usebuilt directly on Claude 3.5 SonnetChapter 13. Models that act
  8. 2025Claude 3.7built directly on Claude 3.5 SonnetChapter 12. Thinking longer
  9. 2025Claude Codewas influenced by Computer use
    was influenced by Claude 3.7    Chapter 13. Models that act
  10. 2025Claude 4succeeded Claude 3Chapter 13. Models that act
  11. 2025Claude Sonnet 4.5built directly on Claude 4Chapter 13. Models that act
  12. 2025Claude Opus 4.5built directly on Claude Sonnet 4.5Chapter 13. Models that act
  13. 2026Claude Opus 4.6built directly on Claude Opus 4.5Chapter 14. Where the threads meet now
  14. 2026Claude Opus 4.7built directly on Claude Opus 4.6Chapter 14. Where the threads meet now
  15. 2026Claude Opus 4.8built directly on Claude Opus 4.7Chapter 14. Where the threads meet now
  16. 2026Claude Sonnet 5succeeded Claude Opus 4.8Chapter 14. Where the threads meet now
  17. 2026Claude Opus 5succeeded Claude Opus 4.8Chapter 14. Where the threads meet now
  18. 2026Claude Opus 5.5built directly on Claude Opus 5Chapter 14. Where the threads meet now
  1. 2022HH-RLHFThis milestone.
  2. 2022Constitutional AIbuilt directly on HH-RLHFChapter 6. Teaching models what people want
  3. 2023Claudewas influenced by HH-RLHFChapter 7. The assistant goes public
  4. 2023100K contextbuilt directly on ClaudeChapter 10. Reading more than you remember
  5. 2023Lost in the middlewas influenced by 100K context (indirect influence)Chapter 10. Reading more than you remember
  6. 2024Gemini 1.5was influenced by Lost in the middle (indirect influence)Chapter 10. Reading more than you remember
  7. 2024Claude 3built directly on Claude
    built directly on Constitutional AI    Chapter 14. Where the threads meet now
  8. 2024Claude 3.5 Sonnetbuilt directly on Claude 3Chapter 14. Where the threads meet now
  9. 2024Computer usebuilt directly on Claude 3.5 SonnetChapter 13. Models that act
  10. 2024Gemini 2.0built directly on Gemini 1.5
    was influenced by Computer use (indirect influence)    Chapter 14. Where the threads meet now
  11. 2025Claude 3.7built directly on Claude 3.5 SonnetChapter 12. Thinking longer
  12. 2025Claude Codewas influenced by Computer use
    was influenced by Claude 3.7    Chapter 13. Models that act
  13. 2025Gemini 2.5was influenced by Gemini 1.5Chapter 12. Thinking longer
  14. 2025Claude 4succeeded Claude 3Chapter 13. Models that act
  15. 2025IMO goldbuilt directly on Gemini 2.5Chapter 12. Thinking longer
  16. 2025Claude Sonnet 4.5built directly on Claude 4Chapter 13. Models that act
  17. 2025IMO-ProofBenchwas influenced by IMO goldChapter 12. Thinking longer
  18. 2025Gemini 3was influenced by Gemini 1.5Chapter 14. Where the threads meet now
  19. 2025Claude Opus 4.5built directly on Claude Sonnet 4.5Chapter 13. Models that act
  20. 2026GPT-5.3-Codexwas influenced by Computer use (indirect influence)Chapter 13. Models that act
  21. 2026Claude Opus 4.6built directly on Claude Opus 4.5Chapter 14. Where the threads meet now
  22. 2026Gemini 3 Deep Thinkbuilt directly on Gemini 3Chapter 12. Thinking longer
  23. 2026Gemini 3.1built directly on Gemini 3Chapter 14. Where the threads meet now
  24. 2026GPT-5.4built directly on GPT-5.3-CodexChapter 14. Where the threads meet now
  25. 2026Project Glasswingwas influenced by Computer use (indirect influence)Chapter 14. Where the threads meet now
  26. 2026Claude Opus 4.7built directly on Claude Opus 4.6Chapter 14. Where the threads meet now
  27. 2026GPT-5.5built directly on GPT-5.4Chapter 14. Where the threads meet now
  28. 2026Claude Opus 4.8built directly on Claude Opus 4.7Chapter 14. Where the threads meet now
  29. 2026Claude Sonnet 5succeeded Claude Opus 4.8Chapter 14. Where the threads meet now
  30. 2026Fable 5 / Mythos 5was influenced by Project GlasswingChapter 14. Where the threads meet now
  31. 2026GPT-5.6 Solbuilt directly on GPT-5.5Chapter 14. Where the threads meet now
  32. 2026Claude Opus 5succeeded Claude Opus 4.8Chapter 14. Where the threads meet now
  33. 2026Claude Opus 5.5built directly on Claude Opus 5Chapter 14. Where the threads meet now

Direct links and their sources

Where it came from

Built directly on

Developed alongside

What it led to

Built on directly by

Influenced

Terms