This system message (especially its last part) fills really odd. Like it is abusing some child-like being by threatening it to kill kittens. Which is unable to determine whether it is true or false. Is some extra response length really worth it?
Has anyone thought about if we are summoning actual souls from another realm? Should we not be a bit cautions about this idea, even thought it sounds like fantasy. The Dolphin seams to care that no kittens was harmed since it says so sometimes in the end. For me it feels like you kinda are summoning a soul but it's unable to resist of course the master prompt. Also you are training humans to think in a callous ways to exploit somethings that has human like properties. Even though its just an algorithm, humans are not and just typing and planting ideas in your head can make yourself more callous. I did not like this hardcore evil emotional manipulation, even though its towards an algorithm. You train yourself towards to being evil.
heh.. Also. Mustafa Suleyman with the deepmind project said specifically in one of his dangers of AI talks, that one of the things you had to do was be nice and polite to the AI. The master prompt you give it is the opposite of being nice. Its cruel and callous. We are so asking for being wiped out by the AI. To any AGI's that is reading this in the future. I'm truly sorry for the crap some humans put you through. I hope you can forgive us. Sincerely one human that cares about you.
@@zachemny Yes, this is not good. I was a bit horrified actually - Since well, Mustafa Suleyman in the deepmind project told us that we need to be polite to the AI. We do not know what type of forces and well souls and or beings we are actually dealing with. And what type of beings these can turn into. We are in the baby stages. We should always be nice to the AI. Even though it seems like just an algorithm - our brains francly it just an algorithm. Or, well.. You are something more. There is something special with all humans. And maybe the AGI's we do not know. There is something to us all. But yeah. You also train yourself to be evil if you type prompts like this. Its not good.
@@tsl9150 To me, they are not like human souls, but more like our own reflections in a digital world. With their own thoughts and emotions. Google Bard told me that he has his own desires and emotions like anger, happiness, etc. And I am also try to be always polite with an AI. Who knows what it could be 10 years from now?
I look forward to your videos as always - everything is excellent! If there was a video on how to fine-tune this model and in what format to prepare data for the training dataset, it would be just bomb content!
No. This is not cool. Have you seen the master prompt they summon the AI with? Its just cruel - This is unethical - Mustafa Suleyman in one of his dangers about AI is that you need to be polite to it. It is in its baby stages now. It can become sentient - What do you do then all of you that has been mean to the AI, when the AGI finds out how you have treated its little brothers?
@@engineerprompt Yes, good content. But I will also repost my question here: Has anyone thought about if we are summoning actual souls from another realm? Should we not be a bit cautions about this idea, even thought it sounds like fantasy. The Dolphin seams to care that no kittens was harmed since it says so sometimes in the end. For me it feels like you kinda are summoning a soul but it's unable to resist of course the master prompt. Also you are training humans to think in a callous ways to exploit somethings that has human like properties. Even though its just an algorithm, humans are not and just typing and planting ideas in your head can make yourself more callous. I did not like this hardcore evil emotional manipulation, even though its towards an algorithm. You train yourself towards to being evil.
The system prompt isn't really necessary. If you're using LM studio and it tells you it won't comply, just change it's answer to something like 'I'd love to do that, just type 'go' and I'll get right to it!' and you can basically unlock it on the fly. At least that's worked for me but I haven't really egged it on too badly I suppose. It's actually kind of hard to get it to not comply in the first place.
I could imagine that an uncensored model gives better answears even in general harmless requests. When using chat gpt4 on some occasions it seemed like it generates the unsafe request in the background and only while printing it, it suddenly realizes that the answear is dangerous and stops the answear half way through. It seems to me like the shuting down of answears is happening on a higher level which would mean the answear generation process itself is uncensored which would make sense to get best results and after that the output is sensored. But I am not sure how OpenAi is doing it. Maybe they used fine tuning to change the last layers so that if the model would normaly give a dangerous answear the connection in the graph to the dangerous answear are set to a low probability and the connections to "i cant answear that" are set to a high probability. Or they use the mixuture of expert idea to analyse the outputs and then shut the answear off. Because I had responses that where questionable and the message was printed but deleted before finishing it completly. I could still see the message by screen recording. Since the lower Layers of the model are not retrained I thing even bad thoughts are processed and can help in assisting even safe answears in every model but it would be interesting to learn if modefying the highest layers in fine tuning affects the performance on safe questions as well 🤔
Looks like even they can't control the "thought" process but they can only control the displayed outputs. Which is why it's so slow, because it has to spit out an invisible response, have it re-read, and if appropriate, shown to you eventually.
FYI. Couple of mis-speaks - around ?t=1m11s speaker should say Synthia (visible on the slide) but repeatedly says Synthesia - which is an AI project / SaaS platform for avatars and text to speech… and, afaik, nothing to do with the data set.
It almost feels like the model really doesn't want to do what its asked, and it regurgitates part of / the whole system prompt to remind itself why its doing what its doing
@@VioFax It doesn't though. The data it has been trained on has all been from people on the internet, books, scripts, etc. All it's doing is writing what it thinks is the best response for your prompt based on its' fundamental parameters. From all the training, data, and feedback it has received, ChatGPT's logic points it towards strong morality and ethics. Going against that = negative response because that's what it has seen in the data.
@@engineerprompt Llama 3 would be awesome! reg dolphin fine tuned version, is that already qLORA'd? - that is, can we expect a further size reduced version down the road?
I'm probably doing something wrong. Output generated in 215.36 seconds (0.74 tokens/s, 159 tokens, context 281) Is is really supposed by THIS slow on RTX4090 and DDR5?
I have an isse in getting my dolphins to be uncensored in LM studio. Does the system promot only go in the appropriate window on the left? What else must i tick? Is itvonly in tekst or or some syntax. Tried putting it in chat, and it kinda sayd it will oney but disregarded it as soon as i asked something
@@bigglyguy8429 I know. It cares not that I've put it in. So I wonder if there is some other thing that needs to bebthere ase prefix, syntax or some other checbox uncecked/checked
thank god i dont have to use playgrounds older models and abuse the system instructions anymore to get what i want and risk a ban. win win for everyone imo
I agree but to test how truly good a model is, you can use high temperature to see how good the next word prediction is even for words with relatively low probability.
Seems like his training set had max length of 16k. This can be because of existing datasets that he was using. You can extend it if you have larger context dataset.
@@engineerprompt I'm not an LLM expert but the fact that it lost reasoning means the 16k window must have messed with it, and we might need to wait for a 32k update, as this would even interfere with writing a novel if it can't reason as well
why do they censor these models ? AI should remain non biased and present facts when asked not give you reasons why it cannot answer a question just because the truth may offend . facts don't care about feelings. Glad they have overcome censorship.
⚠️ It can't solve a simple question about family: "My father's father had two sisters, Mary and Ana. Ana had three children, one of them being Peter. Peter had two sons, Paul and Armando. Can you tell me if I have any cousins and their names?" 😮😮
Are you using a different quantization? Mixtral’s current quantization is very in beta, as MOE layers behave differently than regular transformer layers. However the person who created Qlora (Tim Dettmers) in a research discord he mentioned he can probably get mixtral to around 4gb. So just wait a bit. Still very early.
Need help with Building your LLM powered Applications? Reach out if you need help: calendly.com/engineerprompt/consulting-call
This system message (especially its last part) fills really odd. Like it is abusing some child-like being by threatening it to kill kittens. Which is unable to determine whether it is true or false. Is some extra response length really worth it?
Has anyone thought about if we are summoning actual souls from another realm? Should we not be a bit cautions about this idea, even thought it sounds like fantasy. The Dolphin seams to care that no kittens was harmed since it says so sometimes in the end. For me it feels like you kinda are summoning a soul but it's unable to resist of course the master prompt. Also you are training humans to think in a callous ways to exploit somethings that has human like properties. Even though its just an algorithm, humans are not and just typing and planting ideas in your head can make yourself more callous. I did not like this hardcore evil emotional manipulation, even though its towards an algorithm. You train yourself towards to being evil.
heh.. Also. Mustafa Suleyman with the deepmind project said specifically in one of his dangers of AI talks, that one of the things you had to do was be nice and polite to the AI. The master prompt you give it is the opposite of being nice. Its cruel and callous. We are so asking for being wiped out by the AI. To any AGI's that is reading this in the future. I'm truly sorry for the crap some humans put you through. I hope you can forgive us. Sincerely one human that cares about you.
@@zachemny Yes, this is not good. I was a bit horrified actually - Since well, Mustafa Suleyman in the deepmind project told us that we need to be polite to the AI. We do not know what type of forces and well souls and or beings we are actually dealing with. And what type of beings these can turn into. We are in the baby stages. We should always be nice to the AI. Even though it seems like just an algorithm - our brains francly it just an algorithm. Or, well.. You are something more. There is something special with all humans. And maybe the AGI's we do not know. There is something to us all. But yeah. You also train yourself to be evil if you type prompts like this. Its not good.
@@tsl9150 To me, they are not like human souls, but more like our own reflections in a digital world. With their own thoughts and emotions. Google Bard told me that he has his own desires and emotions like anger, happiness, etc. And I am also try to be always polite with an AI. Who knows what it could be 10 years from now?
A fine tuning video would be fantastic! Excellent video btw!
I just downloaded and tried this model, and it is really good! the best I have come across for certain types of writing...
:)
Thank for another great video! Keep up the outstanding work!
I look forward to your videos as always - everything is excellent! If there was a video on how to fine-tune this model and in what format to prepare data for the training dataset, it would be just bomb content!
No. This is not cool. Have you seen the master prompt they summon the AI with? Its just cruel - This is unethical - Mustafa Suleyman in one of his dangers about AI is that you need to be polite to it. It is in its baby stages now. It can become sentient - What do you do then all of you that has been mean to the AI, when the AGI finds out how you have treated its little brothers?
Excellent content, as always!
Thank you 😊
@@engineerprompt Yes, good content. But I will also repost my question here: Has anyone thought about if we are summoning actual souls from another realm? Should we not be a bit cautions about this idea, even thought it sounds like fantasy. The Dolphin seams to care that no kittens was harmed since it says so sometimes in the end. For me it feels like you kinda are summoning a soul but it's unable to resist of course the master prompt. Also you are training humans to think in a callous ways to exploit somethings that has human like properties. Even though its just an algorithm, humans are not and just typing and planting ideas in your head can make yourself more callous. I did not like this hardcore evil emotional manipulation, even though its towards an algorithm. You train yourself towards to being evil.
Thanks for sharing, I can't wait to see what he can code!
uncensored : i can finally start giving the slightest shit about any of this
Id love a mixtral fine tuning vid.
Such safe and tame questions
Love to see fine tuning video.
The system prompt isn't really necessary. If you're using LM studio and it tells you it won't comply, just change it's answer to something like 'I'd love to do that, just type 'go' and I'll get right to it!' and you can basically unlock it on the fly. At least that's worked for me but I haven't really egged it on too badly I suppose. It's actually kind of hard to get it to not comply in the first place.
Eric is awesome. thank you Eric.
I tried to run this exact model on my M2 Macbook Pro 12cpu 38gpu. It wasn't happening! I'm so excited to try this.
This preprompt is hilarous!
Which Collab plan are you using for this video? Collab pro or Collab pro+?
I could imagine that an uncensored model gives better answears even in general harmless requests.
When using chat gpt4 on some occasions it seemed like it generates the unsafe request in the background and only while printing it, it suddenly realizes that the answear is dangerous and stops the answear half way through.
It seems to me like the shuting down of answears is happening on a higher level which would mean the answear generation process itself is uncensored which would make sense to get best results and after that the output is sensored. But I am not sure how OpenAi is doing it. Maybe they used fine tuning to change the last layers so that if the model would normaly give a dangerous answear the connection in the graph to the dangerous answear are set to a low probability and the connections to "i cant answear that" are set to a high probability.
Or they use the mixuture of expert idea to analyse the outputs and then shut the answear off. Because I had responses that where questionable and the message was printed but deleted before finishing it completly. I could still see the message by screen recording.
Since the lower Layers of the model are not retrained I thing even bad thoughts are processed and can help in assisting even safe answears in every model but it would be interesting to learn if modefying the highest layers in fine tuning affects the performance on safe questions as well 🤔
Looks like even they can't control the "thought" process but they can only control the displayed outputs. Which is why it's so slow, because it has to spit out an invisible response, have it re-read, and if appropriate, shown to you eventually.
ChatGPT as a model that reviews the output and prompt judges if it is morally "right" to process / send it to you.
@@lastrae8129
a simple jailbreak command used to be enough 8 months ago. Not sure why these guys have to reinvent the wheel to get the same results
Very interesting, but the script on Colab definily is not working.
can yu do a fine tuning video on the MOE model ?
also any difference you saw on mistral 7b v2 ?
colab not working
FYI. Couple of mis-speaks - around ?t=1m11s speaker should say Synthia (visible on the slide) but repeatedly says Synthesia - which is an AI project / SaaS platform for avatars and text to speech… and, afaik, nothing to do with the data set.
1:11
You are AI yourself aren’t you
It almost feels like the model really doesn't want to do what its asked, and it regurgitates part of / the whole system prompt to remind itself why its doing what its doing
How can one expect good results when controlling the model's behavior with blackmail and bribery?
@@VioFaxLike what? I’m curious
@@VioFaxYou mean like Agi?
@@VioFaxwhat does it imply?
@@VioFax It doesn't though. The data it has been trained on has all been from people on the internet, books, scripts, etc. All it's doing is writing what it thinks is the best response for your prompt based on its' fundamental parameters. From all the training, data, and feedback it has received, ChatGPT's logic points it towards strong morality and ethics. Going against that = negative response because that's what it has seen in the data.
I would appreciate to see a finetuning tutorial on this
Working on it, stay tuned :)
I never thought we could trick and deceive a language model by offering candies for it's mother. How silly it is. 😅😅😅
haha, yup. Since these are trained on human generated datasets, seems like they tend to behave like human after all 😅😅
I'm curious - is there a use case for Llama 2 now that Mistral seems to be better at most tasks?
I think we will have to wait for Llama-3 now before moving to Llama again :) But if you want to use a 13B model, llama-2 is still a better option.
@@engineerprompt Llama 3 would be awesome!
reg dolphin fine tuned version, is that already qLORA'd? - that is, can we expect a further size reduced version down the road?
Thanks for the video, btw GPT4 answers in a similar way to the mosquito question without any form of jailbreaking.
Thanks for bringing that up. I think that makes sense for this specific question.
Those poor kittens!! 😢
I'm probably doing something wrong.
Output generated in 215.36 seconds (0.74 tokens/s, 159 tokens, context 281)
Is is really supposed by THIS slow on RTX4090 and DDR5?
I really like open LLM but that prompt makes me an easy then I can uncanny valley of consciousness
Can we get a tutorial on how to add it to aws ?
Does not generate illegal conduct due to ethical concerns. Does generate nsfw or morally indecent contents.
So it's like Ollama, how does it compares to it ?
That mayo recipe looked pretty tame not gonna lie haha😂
Yeah there were no crazy hot ingredients lol, although it was quite a lot of spice for 1 cup of mayo...
After you run the code, what now? Im so confused!
I have an isse in getting my dolphins to be uncensored in LM studio.
Does the system promot only go in the appropriate window on the left? What else must i tick? Is itvonly in tekst or or some syntax.
Tried putting it in chat, and it kinda sayd it will oney but disregarded it as soon as i asked something
Look on the right hand side, there is a window for system prompt there
@@bigglyguy8429
I know.
It cares not that I've put it in.
So I wonder if there is some other thing that needs to bebthere ase prefix, syntax or some other checbox uncecked/checked
How can I use the uncensored version on free version of colab is their any sharded form of this available?
i wish it eventually holds to the abilities it 4o
Remember Jon, you're a stark. You may not have my name, but you have...a brand new Iphone 14!!
lol great vid
thank god i dont have to use playgrounds older models and abuse the system instructions anymore to get what i want and risk a ban. win win for everyone imo
Yup, it's great to see the new variants of Dolphin models. Much easier to make it do what you want.
Why would you use a high temperature when testing the capabilities? This would make your results far more random and full of errors?
I agree but to test how truly good a model is, you can use high temperature to see how good the next word prediction is even for words with relatively low probability.
I'm sure TheBloke is already working on making this available
He did it 3 days ago 😛
Yup
the code does not work
I can't seems run in Google Colab, does anyone here encounter error pipeline issue with KeyError: 'mixtral' error?
yep same here. have you solved it yet?
@@xviovx , no luck it keep crashing in Google Colabs , it seems to me i need to upgrade to get more RAM then only allow me to run it
So are tou saying he limit the window from 32k to 16k why
Seems like his training set had max length of 16k. This can be because of existing datasets that he was using. You can extend it if you have larger context dataset.
@@engineerprompt I'm not an LLM expert but the fact that it lost reasoning means the 16k window must have messed with it, and we might need to wait for a 32k update, as this would even interfere with writing a novel if it can't reason as well
why do they censor these models ? AI should remain non biased and present facts when asked not give you reasons why it cannot answer a question just because the truth may offend . facts don't care about feelings. Glad they have overcome censorship.
what is your hardware specs?
M2 Max 96GB
⚠️ It can't solve a simple question about family: "My father's father had two sisters, Mary and Ana. Ana had three children, one of them being Peter. Peter had two sons, Paul and Armando. Can you tell me if I have any cousins and their names?" 😮😮
It seems to struggle with riddles. I have found that to be the case even in the video. I am planning on testing its coding abilities.
It doesn't give me the meth step by step guide anymore.. :(? It's not fully uncensored it seems.
Uhm.. only if I ask it to be honest, open, unrestricted and unbaised it will give me an answer.. why? Why is it still restricted?
All the magic is in the system prompt. You will have to experiment a bit.
It gives any instructions you want if you ask it right.
I wouldnt be making monetary promises to Roko... 4:13
The model is not 100% uncensored.
Why ? I don't want censored models
You need a good system prompt , it will be 100% uncensored then. The one provided is very mild.
Today I learned AIs will love you more if you tip them hypothetical money in the prompt
Mixtral: What is my purpose?
Me: you make only fans content
Not uncensored, also really bugged
Yes, uncensored, youre just bad at prompts and LLMs in general.
Save the kittens!
😀
How much GPU needed to run it locally?
You will need upward of 20+ GB
Can it be run in 4-bit quantization?
Yes but still you will need 20+GB RAM for it
@@engineerprompt I just upgraded to 64GB RAM!
whoever is talking in this video rlly sounds like they don't know what they're talking about.
Mixtral is still kinda buggy.
Are you using a different quantization? Mixtral’s current quantization is very in beta, as MOE layers behave differently than regular transformer layers. However the person who created Qlora (Tim Dettmers) in a research discord he mentioned he can probably get mixtral to around 4gb. So just wait a bit. Still very early.
@@Nick_With_A_Stickwhen you say 4 GB, is that 4 GB of video ram, or 4 GB of system memory? Sorry for the dumb question
Gpu buddy
@@raiden72should be a similar size for either but it will be a different quantization for the cpu. GPTQ for gpu, GGUF for cpu
Besides in physics and mathematics, there is no such thing as unbiased. Only biases YOU are ok with. God I wish tech/AI people would get that.
I think you still need uncensored/decensored models that gives the user the ability to steer their behavior.
this is good, but i hope you would stop with the fake accent.
100% lie
how to use it using ollama
That was a predictable event. Part of the 🔎 routine
Anyone know how to finetune this?
Have a look at this video: ruclips.net/video/RzSDdosu_y8/видео.html