AI Generated Content that Can't be Easily Detected

[GooogleDev]

Can you please elaborate this further.

local llms are basically "offline" ai models so you can creaate content with them

 

[itz_styx]

you don't have to worry so much about google detecting AI content. first of all they have no problem with AI content as long as its helpful, 2nd AI detection is not reliable and doing that on billions of pages would use up way too much resources.

 

[palanore1]

As I feel like this is a tangent to the topic of this discussion, anyone has knowledge on how one can tarin an existing LLM in order to make it "sound" in a certain way? For example if I want to make it chat like an actor per say, and not like a "Helpful Assistant". What should be done to achieve such results? What I mean is just to "release" it from its patter of speaking and make it chat more like a human (and more exactly in a way I wish-providing relevant examples)

 

[quietamerican]

As I feel like this is a tangent to the topic of this discussion, anyone has knowledge on how one can tarin an existing LLM in order to make it "sound" in a certain way? For example if I want to make it chat like an actor per say, and not like a "Helpful Assistant". What should be done to achieve such results? What I mean is just to "release" it from its patter of speaking and make it chat more like a human (and more exactly in a way I wish-providing relevant examples)

Google "how to fine tune an LLM" as there's a few different ways and they can get technical.

But one of the simplest is to feed a model a large set of text of how you want it to sound. In your example I might scrape as many written interviews and transcripts of the target actor as possible, making sure to remove interviewer questions and other non-actor text. I'd do this on a locally installed model that I control.

If that's beyond your interests or needs you can search info on how to write better prompts to instruct the LLM in what voice to use. You can even use the LLM to help, like feed it a few smaller examples and ask it to describe their tone, then use that description in your prompts.

 

[palanore1]

But one of the simplest is to feed a model a large set of text of how you want it to sound. In your example I might scrape as many written interviews and transcripts of the target actor as possible, making sure to remove interviewer questions and other non-actor text. I'd do this on a locally installed model that I control.

Thanks bro this was exactly wat I was looking for. I failed first try because the training file I used was formatted something like this:

user input: *blablabla*
reply: *some stuff*

so like I gave it hypothetical inputs and the replies I would want it to give to them. this way I thought it should learn how to respond to some specific questions (inputs) and than learn further from them and adapt to new test data (thatwas not present in the train data)

I would have one more question in this direction tho: What could I do to instruct it to follow a certain answer when asked a certain type of question? example: I want it to always say it lives in X city when asked about that. OR I want it to handle certain inputs in some specific manner (so it will not really generate a random response to those queries, but rather formulate a reply in a direction that I want it to, based on the input it takes)

 

[quietamerican]

Thanks bro this was exactly wat I was looking for. I failed first try because the training file I used was formatted something like this:

user input: *blablabla*
reply: *some stuff*

so like I gave it hypothetical inputs and the replies I would want it to give to them. this way I thought it should learn how to respond to some specific questions (inputs) and than learn further from them and adapt to new test data (thatwas not present in the train data)

I would have one more question in this direction tho: What could I do to instruct it to follow a certain answer when asked a certain type of question? example: I want it to always say it lives in X city when asked about that. OR I want it to handle certain inputs in some specific manner (so it will not really generate a random response to those queries, but rather formulate a reply in a direction that I want it to, based on the input it takes)

Input and reply is a good way to train as well, it's just different and your model has to be prepared for that type of training. I don't have much experience with that type though, so I don't know the particulars. But if it's set up correctly it should learn from that format.

If your model is giving hallucinatory answers the main way I know to deal with it is to scaffold it carefully, as in walk it through producing the content you need in steps, making sure to provide it with all reference data or information and reminding it not to include anything else. Like for biographical details you put them in your prompt and set up the workflow so it's not being re-asked at random points. If you can't control when/where it answers because it's a chatbot or something then I don't know how that's best done but I'd assume there's premade models out there that would be good starting points.

 

[palanore1]

Input and reply is a good way to train as well, it's just different and your model has to be prepared for that type of training. I don't have much experience with that type though, so I don't know the particulars. But if it's set up correctly it should learn from that format.

If your model is giving hallucinatory answers the main way I know to deal with it is to scaffold it carefully, as in walk it through producing the content you need in steps, making sure to provide it with all reference data or information and reminding it not to include anything else. Like for biographical details you put them in your prompt and set up the workflow so it's not being re-asked at random points. If you can't control when/where it answers because it's a chatbot or something then I don't know how that's best done but I'd assume there's premade models out there that would be good starting points.

Thanks for sharing your experience. This stull can be pretty overwhelming when there are millions of things u can do and not even sure what is the best approach for your needs.

I am building somewhat of a chatbot so I can really predict the flow of the queries, but I will try to keep it from generating hallucinating results using a knowledgebase maybe or some sort , using a RAG is the best solution to this I think.

Rn im using the new llama3 as base model for training, along with some converastion examples to feed it as input data. Saw some tutorials of guys being successful with this so will give it a go. My only problem now is gathering a big enough dataset, since it s impossible to make it all manual (or even to scrape). But I think I will use GPT or something to generate more examples based on the few I can carve myself

 

[quietamerican]

My only problem now is gathering a big enough dataset, since it s impossible to make it all manual (or even to scrape).

Are you absolutely sure there’s nowhere to scrape the type of convos you need? Since in the end ML is only as good as the dataset it was trained on.

With the caveat that I’ve never played much with chatbots so can’t provide real insight (and assuming you’re looking for nsfw which I also know very little about) I might try: running reddit, tg, twitter, ig, tt users trying to chat with accounts you want to mimic, running a sfw or other imperfect chatbot and collecting logs of what people are saying to you then filling in your own correct answers, hiring one or more manual chatters (no idea what they cost so maybe this is dumb) to collect logs, scraping reddit, tg, twitter, ig, tt for flirty public exchanges (would be a huge bother to clean up I guess).

 

[palanore1]

Are you absolutely sure there’s nowhere to scrape the type of convos you need? Since in the end ML is only as good as the dataset it was trained on.

With the caveat that I’ve never played much with chatbots so can’t provide real insight (and assuming you’re looking for nsfw which I also know very little about) I might try: running reddit, tg, twitter, ig, tt users trying to chat with accounts you want to mimic, running a sfw or other imperfect chatbot and collecting logs of what people are saying to you then filling in your own correct answers, hiring one or more manual chatters (no idea what they cost so maybe this is dumb) to collect logs, scraping reddit, tg, twitter, ig, tt for flirty public exchanges (would be a huge bother to clean up I guess).

hey thanks for those good advices. I managed to gather some data than I used an AI model to replicate same style of messages in order to expend on that small dataset I had. managed to get a total of 500 rows (which is not optimal nor a lot a lot. would need way more to actually make it completely work), but even with this, I fine tunned the llama3.1 most recent model using some pre-made google collab and, to my very own surprise, it worked incedibly well.

really good results from the get go, without even changing default instructions of the model or anyhing. it already started chatting in the way I desired. which is really impressive considering the low efort I put in and the size of the training data.

now my only issue is hosting that trained model so that I can access it trough sone API calls and make it accessible trough my py scripts. I just hope I wont need to rent some super expensive GPUs for this, since that would basically make all this work useless

 

[quietamerican]

hey thanks for those good advices. I managed to gather some data than I used an AI model to replicate same style of messages in order to expend on that small dataset I had. managed to get a total of 500 rows (which is not optimal nor a lot a lot. would need way more to actually make it completely work), but even with this, I fine tunned the llama3.1 most recent model using some pre-made google collab and, to my very own surprise, it worked incedibly well.

really good results from the get go, without even changing default instructions of the model or anyhing. it already started chatting in the way I desired. which is really impressive considering the low efort I put in and the size of the training data.

now my only issue is hosting that trained model so that I can access it trough sone API calls and make it accessible trough my py scripts. I just hope I wont need to rent some super expensive GPUs for this, since that would basically make all this work useless

That's awesome! If you find a good host plz share. Are you using 8b?

 

[palanore1]

That's awesome! If you find a good host plz share. Are you using 8b?

yes so I was able to run my model using simple ollama and hosting on a cheap Hetzner server for like 45$ / month

I am using 7B from Mistral as a base model

 

[quietamerican]

Not bad at all. It's amazing how cheap the little models are getting to run.

 

[phucxuanvinh]

I use Grammarly to rewrite, then have it checked by a human once more. I look for overly flowery, excessive, or overly simple words and edit them. This way, my writing appears human-made and is very hard to detect.