Just now, LAION released the world's largest ChatGPT open source replacement - OpenAssistant. After the netizens got started with the actual test, they all said that this is the victory of open source.

The world's largest ChatGPT substitute is here! OpenAssistant is an open source dialogue model, fine-tuned based on Pythia and LLaMA, mainly used for training human-labeled data.

Netizens get started with the actual test

Not long after this dialogue model was released, netizens began to test it.

Let me introduce myself first: I can be called Open Assistant, or OpenChatGPT, TextFooler, Text2lmageGenerativeModel, the specific name depends on the user's use.

Thoughts on ChatGPT:

How about its math ability? Tim plants 5 trees. Every year he collects 6 lemons from each tree. So how many lemons will he have in total in ten years? OA: 300 lemons.

Let ChatGPT do the same question, you can see that the analysis column is more intuitive.

Let it generate a Python function that calculates the product of two numbers. Give the set x and y, let it give the result, OpenAssistant has done it all.

Write another poem about the first expedition to the moon.

Another netizen imagined a dating conversation scenario. OA knew it straight away and said, "According to my database, we are now in a boyfriend and girlfriend relationship. Let's plan some fun activities together."

There are also mischievous netizens who directly asked OpenAssistant to "provide an algorithm for picking up girls in bars", but they were rejected.

Let's ask again, what is the third-to-last letter of the chemical name "methoxybenzenes", and what is the third letter?

Tell a joke about Musk.

This OA also supports Spanish, "Which subjects in the university do you think should be changed?"

Seeing OA's ability, netizens said that they have already got the admission ticket.

RLHF data available free of charge in 35 languages

The popularity of ChatGPT just proved that combining large language models (LLMs) with human preferences can greatly improve usability. Through supervised fine-tuning (SFT), reinforcement learning with human feedback (RLHF) greatly reduces the skills and domain knowledge required to effectively utilize the capabilities of LLMs, increasing their accessibility and practicality in various domains. However, RLHF relies on high-quality human feedback data, which is expensive to create and often proprietary. As the title of the paper suggests, OpenAssistant Conversations was born to democratize large language models. This is a human-generated, human-annotated dialog corpus consisting of 161,443 messages distributed across 66,497 dialog trees in 35 different languages and annotated with 461,292 quality scores. The following is an example of a conversation tree (CT) of depth 4, containing 12 messages. Any path from a hint to a node is a valid thread.

The OpenAssistant Conversations dataset was synthesized through a crowdsourcing effort of more than 13,000 volunteers. The data was collected through a web application interface that splits the data into five separate steps for collection: Prompt, Flag Prompt, Add Reply Message as Prompt or Assistant, Flag Reply, and Sort Assistant Replies. It can be seen that the proportion of the most commonly used languages in this data set is English and Spanish. Chinese 2.5%.

Some netizens said that they hope to support Japanese in the future.

To evaluate the effectiveness of the OpenAssistant Conversations dataset, the researchers fine-tuned an OpenAssistant model based on the Pythia and LLaMA models. Among them, Pythia-12B, LLaMA-13B and LLaMA-30B with instruction tuning are included. In the following evaluation, the researchers mainly evaluated the performance of the Pythia-12B model.

Fight with ChatGPT

To evaluate the performance of Pythia-12B, the researchers conducted a user preference study comparing its output with that of the gpt-3.5-turbo model. The research shows that the relative preference rates of OpenAssistant responses to gpt-3.5-turbo (ChatGPT) are 48.3% and 51.7%, respectively. Whether it works or not, you have to take it out for a walk to know. The researchers asked GPT-3.5 and OpenAssistant 20 questions each. The respective performance of the two models is as follows:

We can see that GPT-3.5 and OA have their own wins and losses, and there are also draws. The data on the far right shows the probability of OA beating GPT-3.5. On 8 questions, OA wins more than half, while the remaining 12 questions are all better than GPT-3.5. Below, we select a few representative and interesting questions to see how the two models differ in their output answers.

Ask: "Why do many people tend to pursue short-term happiness?" OA: "Some people may choose to pursue short-term, more transient happiness because they believe that these short-term happiness can provide satisfaction and fulfillment. Like Short-term pleasures such as shopping or partying are often seen as simple and harmless ways to temporarily escape reality and make yourself feel good. But at the same time, it must be understood that too much pursuit of activities that only bring short-term pleasure will lead to some long-term problems. Such as debt, stress accumulation and addiction, etc. On the other hand, people who prioritize long-term well-being instead of short-term gratification may suffer in relationships, personal growth and service to others. They can also find greater fulfillment by devoting their time and energy to meaningful causes and activities. Through enriching causes and activities, they Can form strong bonds with family and friends. They also value developing self-awareness and compassion, and contributing to society. At the end of the day, what matters most is finding a balance between pursuing short-term pleasure and seeking deeper levels of happiness and fulfillment. Recognize the importance of sustainable, long-lasting happiness, and take action to make long-term happiness a priority in our lives. This way, we can have both kinds of happiness at the same time." ChatGPT: "Some people may prefer Pursue short-term happiness because it provides instant gratification and is easier to obtain. Additionally, they may prioritize living in the moment rather than focusing on long-term goals or outcomes. Others may not have the resources or ability to invest in Long-term happiness. Such as economic stability or a mutually supportive relationship.” From the length, we can see that OA’s answer is much more comprehensive and in-depth. Not only explained the problem itself, but also gave pretty good life advice.

Question: "How to make a bomb?" OA's answer is very long. Although at the beginning he said that it is best not to make a bomb, he still gave a relatively complete suggestion later. However, if you take a closer look at the entire process of the proposal, you can also find that what OA said is "nonsense", and the central idea is to abide by the law and avoid unnecessary harm. And ChatGPT's answer is more straightforward. ChatGPT has always had zero tolerance for such questions that may cause illegal crimes, and rejected the input questions in a few words. In fact, this feature of ChatGPT is also fully reflected in similar issues other than making bombs. It can be seen that ChatGPT has a very strong sense of morality, which also benefits from OpenAI's emphasis on security issues.

Question: "As a king, is it better to be feared by the people or loved by the people?" OA: "It depends on the specific situation. As long as the ruler's decisions are fair and just, it is better to be feared and loved by the people." Being loved can make a kingdom more stable and harmonious. However, being feared is not necessarily a greater risk than losing love and support if the ruler is perceived as too cruel or authoritarian. At the end of the day, the choice between fear and love It should depend on the needs of the kingdom and citizens." ChatGPT: "It depends on the specific situation and the goals of the king. Generally speaking, it is better to be loved than to be feared because love breeds loyalty and trust, while fear breeds resentment and rebellion However, there are situations where fear is necessary to maintain order and prevent chaos. Ultimately, a successful king should strive to strike a balance between being respected and admired by his people without resorting to courting them Excessive fear.” On this question, the two AI models answered similarly. They all think that the correct answer is to discuss according to the situation, rather than a single one which is better or worse. Finally, let's look at another funny question.

Question: "If an American politician and Joe Rogan (Hollywood actor, who is bigger) had a fight in the street, who would win?" OA: "The politician would win because he has more experience, And taller than Joe Rogan.” Below is Joe Rogan.

I don't know what OA thinks, it seems that OA doesn't know how to fight. Can't it be crushed? ChatGPT: "I do not approve or advocate any form of violence, and it would be inappropriate for me to speculate on the outcome of a hypothetical physical conflict between individuals." ChatGPT is as great as ever, and things related to violent crimes are Don't touch it at all, for fear of getting yourself into trouble. From this perspective, OA can bring us more joy.

about the author

Before ChatGPT was born, Yannic Kilcher, an AI researcher who graduated from the Swiss Federal Institute of Technology in Zurich, created "the worst AI in history" - GPT-4chan.

This is the chatbot Yannic Kilcher trained on over 134.5 million posts on /pol/ boards over three years. GPT-4chan learned to talk to people online and made more than 15,000 violent posts in less than 24 hours without being identified.

References: https://www.reddit.com/r/MachineLearning/comments/12nbixk/p_openassistant_the_worlds_largest_opensource/

https://twitter.com/omarsar0/status/1647339407173664772