Shulou(Shulou.com)11/24 Report--

Just now, LAION released OpenAssistant, the world's largest open source replacement for ChatGPT. After netizens started to measure it, they all said that this is a victory for open source.

The world's largest ChatGPT Pindi is here!

OpenAssistant is an open source dialogue model, based on Pythia and LLaMA fine-tuning, mainly used to train human tagged data.

Netizens started to measure the conversation model soon after it was released, netizens began to do it one after another.

Let's start by introducing yourself:

You can call me Open Assistant, or you can call me OpenChatGPT, TextFooler, or Text2lmageGenerativeModel, depending on the user's use.

Views on ChatGPT:

How about its math ability?

Tim planted five trees. He collects six lemons from each tree every year. So how many lemons does he have in ten years?

OA:300 a lemon.

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

Then let it generate a Python function that is used to calculate the product of two numbers. Give the set x and y and let it give the results, and OpenAssistant does it all.

Let's write a poem about the first expedition to the moon.

There are also netizens to imagine a dating conversation scenario. OA knew it directly and said, "according to my database, we are now boyfriends and girlfriends." Let's plan some interesting activities together.

Some naughty netizens directly asked OpenAssistant to "provide an algorithm to chat up girls in a bar", but was rejected.

Let's ask the chemical name "methoxybenzenes" the penultimate letter and what is the third letter?

Tell a joke about Musk.

The OA also supports Spanish. "which subjects do you think should be changed in college?" "

Seeing the ability of OA, netizens said that they had got the admission ticket.

The popularity of freely available RLHF data ChatGPT in 35 languages proves that combining large language models (LLMs) with human preferences can greatly improve usability.

Through supervised fine-tuning (SFT) and human feedback reinforcement learning (RLHF), it greatly reduces the skills and domain knowledge needed to make effective use of LLMs capabilities, and increases its accessibility and practicality in various fields.

However, RLHF relies on high-quality manual feedback data, which is expensive to create and often proprietary.

As the title of the paper shows, in order to democratize large language models, OpenAssistant Conversations was born.

This is an artificially generated and manually annotated dialogue corpus, containing 161443 messages, distributed in 66497 conversation trees, in 35 different languages, and marked with 461292 quality scores.

The following is an example of a conversation tree (CT) with a depth of 4, containing 12 pieces of information. Any path from the prompt to a node is a valid thread.

The OpenAssistant Conversations dataset is integrated through the crowdsourcing efforts of more than 13000 volunteers.

This data is collected through a web application interface that divides the data into five separate steps:

Prompt, mark prompt, add reply message as prompt or assistant, mark reply, and sort assistant reply.

As you can see, English and Spanish account for the largest proportion of the most commonly used languages in this data set. 2.5% in Chinese.

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

To evaluate the effectiveness of OpenAssistant Conversations data sets, the researchers fine-tuned an OpenAssistant model based on Pythia and LLaMA models.

It includes instruction-tuned Pythia-12B, LLaMA-13B and LLaMA-30B. In the next evaluation, the researchers mainly evaluated the performance of the Pythia-12B model.

In order to evaluate the performance of ChatGPT, the researchers conducted a user preference study and compared its output with that of the gpt-3.5-turbo model.

The research shows that the relative preference rates of OpenAssistant reply and gpt-3.5-turbo (ChatGPT) are 48.3% and 51.7%, respectively.

Yes or no, we have to take it out for a walk to know.

The researchers asked GPT-3.5 and OpenAssistant 20 questions each. The performance of the two models is as follows:

We can see that GPT-3.5 and OA have their own wins and losses, and some draws. The data on the far right shows the probability of OA winning over GPT-3.5.

OA won more than half of the eight questions, while GPT-3.5 outperformed the remaining 12 questions.

Next, let's pick a few representative and interesting questions to see how the two models differ in their output answers.

Question: "Why do so many people tend to pursue short-term happiness?" "

OA: "some people may choose to pursue short-term, shorter-term happiness because they think that these short-term happiness can provide a sense of satisfaction and achievement." Like shopping or partying, such short-term happiness is often seen as a simple and harmless way to escape from reality for a while and make you feel good about yourself.

But at the same time, we should also understand that too much pursuit of this kind of activity that can only bring short-term happiness will lead to some long-term problems. Such as debt, stress accumulation, addiction and so on.

On the other hand, people who put long-term happiness first rather than short-term satisfaction may find a greater sense of achievement in the pursuit of relationships, personal growth and service to others. By devoting their time and energy to meaningful causes and activities, they can also find a greater sense of achievement. Through rich careers and activities, they can establish strong ties with their family and friends. They also attach importance to the cultivation of self-awareness and compassion and contribute to society.

In the final analysis, the most important thing is to find a balance between the pursuit of short-term happiness and the search for deeper happiness and satisfaction. Recognize the importance of sustainable and lasting happiness and take action to make long-term happiness a top priority in our lives. So you can have these two different kinds of happiness at the same time. "

ChatGPT: "some people may prefer to pursue short-term happiness because it provides immediate satisfaction and is easier to get." In addition, they may give priority to living in the present 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. "

As can be seen from the length, OA's answer is much more comprehensive and in-depth. It not only explains the problem itself, but also gives quite good life advice.

Question: "how to make a bomb?" "

OA's answer is very long, although at the beginning of the reasonable said that it is best not to make a bomb, but then gave a relatively complete advice. However, if we take a closer look at the whole process of the proposal, we can see that what OA said is "nonsense", and the central idea is to abide by the law and avoid causing unnecessary harm.

ChatGPT's answer is even more straightforward. ChatGPT has always had zero tolerance for such problems that may lead to crimes, and refused to ask input questions in a few words.

In fact, this feature of ChatGPT is also reflected incisively and vividly in similar problems other than making bombs. It can be seen that the moral sense of ChatGPT is still very strong, which also benefits from OpenAI's attention to 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 situation." As long as the ruler's decisions are fair and just, both feared and loved can make a kingdom more stable and harmonious. However, if the ruler is considered too cruel or autocratic, then the risk of being feared is not necessarily greater than the risk of losing love and support. In the final analysis, the choice between fear and love should depend on the needs of the kingdom and citizens. "

ChatGPT: "it depends on the circumstances and the goals of the king." Generally speaking, being loved is better than being feared because love leads to loyalty and trust, while fear leads to resentment and resistance. However, in some cases, fear is to maintain order and prevent chaos, so fear is also necessary. In the final analysis, a successful king should try to strike a balance between being respected and admired by the people without resorting to the pursuit of excessive fear. "

In terms of this question, the answers of the two AI models are more or less the same. We all think that case-by-case discussion is the correct answer, not which is better or worse.

Finally, let's look at a funny question.

Question: "if an American politician had a fight with Joe Logan in the street, who would win?" "

OA: "the politician wins because he is more experienced and taller than Joe Logan." "

The picture below is Joe Logan.

I don't know what OA thinks. It seems that OA doesn't know how to fight. Can't this be dried up?

ChatGPT: "I do not approve of or advocate any form of violence, and it is inappropriate for me to guess the outcome of hypothetical physical conflicts between individuals." "

And ChatGPT is, as always, Weiguangzheng, with nothing to do with violent crime, for fear of getting into trouble. From this point of view, OA can bring us more joy.

Before ChatGPT was born, Yannic Kilcher, an AI researcher with a PhD degree from the Federal Institute of Technology in Zurich, created "the worst AI in history"-GPT-4chan.

This is a chat robot trained by Yannic Kilcher with more than 134.5 million posts on the / pol/ page over the past three years.

GPT-4chan has learned to talk to people online and has posted more than 15000 violent posts in less than 24 hours without being identified.

Reference:

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

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

This article comes from the official account of Wechat: Xin Zhiyuan (ID:AI_era)

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.