hh-rlhf VS nebuly

If you ask an LLM to complete a sentence like '[Insert name] stole the fruit (true/false):'

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 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.

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

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...) (before the reinforcement from human feedback to make them more tool-like https://arxiv.org/abs/2204.05862), with a bit of randomness added for variety's sake (https://huggingface.co/blo1g/how-to-generate).

They have done good work like releasing their paper and dataset for training an assistant RLHF model. https://github.com/anthropics/hh-rlhf

Github: https://github.com/anthropics/hh-rlhf

Hi everyone. I’m the creator of ChatLLaMA https://github.com/nebuly-ai/nebullvm/tree/main/apps/accelerate/chatllama, an opensource framework to train LLMs with limited resources and create There’s been amazing usage of LLMs in these days, from chatbots to retrieve about company’s product information, to cooking assistants for traditional dishes, and much more. And you? What you building or would love to build with LLMs? Let me know and I’ll share the article about your stories soon. https://qpvirevo4tz.typeform.com/to/T3PruEuE Cheers

How does it differentiate from the original ChatLLaMA? https://github.com/nebuly-ai/nebullvm/tree/main/apps/acceler...

Was this made with the ChatLLaMA library? https://github.com/nebuly-ai/nebullvm/tree/main/apps/accelerate/chatllama

Some of the available modules include:

Speedster: Automatically apply the best set of SOTA optimization techniques to achieve the maximum inference speed-up on your hardware. https://github.com/nebuly-ai/nebullvm/blob/main/apps/acceler...

Nos: Automatically maximize the utilization of GPU resources in a Kubernetes cluster through real-time dynamic partitioning and elastic quotas. https://github.com/nebuly-ai/nos

ChatLLaMA: Build faster and cheaper ChatGPT-like training process based on LLaMA architectures. https://github.com/nebuly-ai/nebullvm/tree/main/apps/acceler...

OpenAlphaTensor: Increase the computational performances of an AI model with custom-generated matrix multiplication algorithm fine-tuned for your specific hardware. https://github.com/nebuly-ai/nebullvm/tree/main/apps/acceler...

Forward-Forward: The Forward Forward algorithm is a method for training deep neural networks that replaces the backpropagation forward and backward passes with two forward passes. https://github.com/nebuly-ai/nebullvm/tree/main/apps/acceler...