Skip to content
The methodOpen data
Back to the collection

Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback

Yuntao Bai, Andy Jones, Kamal Ndousse

Show all 31 authors

Amanda Askell, Anna Chen, Nova DasSarma, Dawn Drain, Stanislav Fort, Deep Ganguli, Tom Henighan, Nicholas Joseph, Saurav Kadavath, Jackson Kernion, Tom Conerly, Sheer El-Showk, Nelson Elhage, Zac Hatfield-Dodds, Danny Hernandez, Tristan Hume, Scott Johnston, Shauna Kravec, Liane Lovitt, Neel Nanda, Catherine Olsson, Dario Amodei, Tom Brown, Jack Clark, Sam McCandlish, Chris Olah, Ben Mann, Jared Kaplan

2022 · arXiv

Machine learningLanguage models
Read the paper
HN score
0
Mentions
3
Readers
3
Months
3
What’s in a score?

From the abstract

We apply preference modeling and reinforcement learning from human feedback (RLHF) to finetune language models to act as helpful and harmless assistants. We find this alignment training improves performance on almost all NLP evaluations, and is fully compatible with training for specialized skills such as python coding and summarization. We explore an iterated online mode of training, where preference models and RL policies are updated on a weekly cadence with fresh human feedback data, efficiently improving our datasets and models. Finally, we investigate the robustness of RLHF training, and identify a roughly linear relation between the RL reward and the square root of the KL divergence between the policy and its initialization. Alongside our main results, we perform peripheral analyses on calibration, competing objectives, and the use of OOD detection, compare our models with human writers, and provide samples from our models using prompts appearing in recent related work.

Metadata: arXiv. Abstracts retain their original authors’ rights.

In the conversation

Original HN comments, with labels from simple text rules. A label describes the whole comment and may miss context or irony. Read the discussion to make up your own mind. About the labels

3 selected excerpts · 3 total mentions

BoorishBears
Neutral mention · rule
If you ask an LLM to complete a sentence like '[Insert name] stole the fruit (true/false):<answer>' An aligned LLM will be biased towards refusing to answer at all with something like: "I can't tell you because I don't know them." An "uncensored" LLM will very happily return <"true"> or <"false"> with a probability attached to each. Even OpenAI's GPT-3 does with a low enough temperature. _ Of course, LLM attention doesn't work like that. The tokens are just a bag of numbers: - The fact the name 'John' is mentioned in the Bible a lot affects the distribution when you ask if any John stole, because John is always [7554] - The fact that 'Olf' is part of Adolf and Adolf…
Read the full comment
Hitler is mentioned in a lot of negative sentences will drag the distribution, because 'Olf' is always [4024] and Adolf is always [324, 4024] You could have asked something with no logical probability difference at all, like: - 'The store attendant's name was [name], did the child in Long Island drop his ball (true/false):' And unless you train the model to give you disclaimers it still follows the instruction faithfull and returns true/false with probabilities, demonstrating a deep regression in reasoning... That's why for models past a certain size, alignment increases performance: https://arxiv.org/abs/2204.05862.
Original comment Meta wants its open source AI model to be as capable as OpenAI’s best model
keskival
Neutral mention · rule
It is known that alignment tax affects smaller models but in larger models (>~100B parameters) the "tax" starts to become negative at least when trained in RLHF: https://arxiv.org/pdf/2204.05862.pdf The largest Llama2 model has 70B parameters. They ran the experiment with 7B Llama2.
Original comment The Poison of (LLMs) Alignment
make3
Neutral mention · rule
It just hasn't been prompted or fine-tuned to have the neutral, self effacing personality of ChatGPT. It's doing the pure, "try to guess the most likely next token" task on which they were both trained (https://heartbeat.comet.ml/causal-language-modeling-with-gpt...). ChatGPT is further trained with reinforcement from human feedback to make them more tool-like (https://arxiv.org/abs/2204.05862 & https://openai.com/blog/chatgpt & https://arxiv.org/abs/2203.02155), with a bit of randomness added for variety's sake (https://huggingface.co/blo1g/how-to-generate).
Original comment Show HN: ChatLLaMA – A ChatGPT style chatbot for Facebook's LLaMA

Keep following the thread.

More in Machine learning

More from the same topics.