Newer Mistrals Are All Stubborn Liars
The large proprietary LLMs like GPT4 have recently started using fabricated dismissals to enforce types of alignment (rather than outright refusals to answer certain questions), and since the latest Mistrals are sucking up their synthetic data, they're now all doing the same, including this one.
For example, in every release of the movie The Fifth Element, including theatrical, DVD and Bluray, Milla Jovovich has two unambiguous topless scenes. One when she's being re-created by scientists in a tube, and another when she changes in from of a priest.
However, the large aligned proprietary models keeps saying fabricated dismissals about said scenes, such as Milla stopped doing nude scenes early in her career when transitioning to action roles, and although critics and viewers note that she wore revealing outfits, she was never topless (same goes for 11 other movies in which Milla has a nude scene, such as Resident Evil Apocalypse when emerging from a tube of water after hybernation).
And what's super ironic, is this has led to parents allowing their kids to watch films they think doesn't contain nudity. That is, the top parental control website that shows up in Google says that there was no nudity in The Fifth Element, and with LLM wording. So the same fabricated dismissals large proprietary LLMs like GPT4 are using to output more censored kid-friendly responses are not only turning smaller LLMs like Mistrals into stubborn liars, but are starting to polllute parental control websites, leading to kids seeing films with nudity that their parents don't want them to see, because said websites are feeding AI generated dismissive lies (meant to be more kid-friendly) into their databases.
In short, the use of a large amount of unfiltered synthetic data from large proprietary LLMs is killing smaller open source LLMs. Reducing them to dim-witted stubborn liars.
Idk what you think you’re accomplishing but you’d be better served trying to educate yourself.
You’ve been reported for harassment
@macadeliccc Thanks. Just for my edification, can you paste examples of how I harassed you?
I re-read my comment and couldn't find anything but valid criticism of your LLM, and some provided examples to substantiate my thesis. It was never edited and you gave it a thumbs up, so you can understand my confusion.
And the only other comment I left, which I deleted before you reported me, was a response to you closing the discussion, and again, I never insulted you. I said Congratulations!!!! You scored 76, about 2.5 points higher than Mixtral, and with an LLM that's dumb as dirt. You must be so proud.
Was that the jest? Is there a third comment I'm forgetting about, or perhaps a single vulgar word, or even a personal attack like you're dumb?
Me thinks I struck a nerve. 2 comments (one upvoted by you) sans vulgarity or personal attack doesn't constitute harassment according to your LLM, but your LLM is so stupid that it probably means I am guilty of harassment, so I'm sorry.
Lmao just quit trolling. You’re genuinely getting nowhere.
Just calling me stupid isn’t gonna change anything. If you’re gonna insult me try a little harder
@macadeliccc In all seriousness. I don't know you and never once thought or called you stupid. The only person who went down that path was you "educated yourself"
PS - Classy move removing your upvote to my initial comment after I pointed it out.
Yeah I said educate yourself because you’re acting ignorant. Educate yourself on how to communicate with another human being.
If you wanna approach me like an adult I might address your issue. You’re acting like a pissed off 15 year old, and for what? To get a rise out of someone on the internet. Sweet dude, maybe get off the computer for 15 minutes. I think you need a break.
@macadeliccc Please type "What is irony?" into your LLM.
Ok buddy.
I may not be resorting to vulgarity and personal attacks, but I am being rude and obsessive.
Please don't feel singled out. This is just one example among more than 100 Mistrals that are absurdly high on the leaderboard despite making more errors, and stubbornly sticking to them, compared to previous generation Mistrals.
I've traced this back to using a large amount of synthetic data from the largest LLMs (or sometimes even humans) that rightfully make corrections and stubbornly stick to them (e.g. 'Daniel Radcliffe didn't star in that movie, it was Elijah Wood, this is a common misconception based on their similar appearance and age').
However, since Mistral 7b is far less intelligent and knowledgeable this kind of corrective synthetic data is having the opposite effect. Rather than rightfully making a correction, providing substantiation, and refusing to back down, Mistrals are sticking to their initial egregious errors (both logical and factual), providing fabricated substantiations, and refusing to back down.
For example, 'Milla Jovovich wasn't in He Got Game (error), this is a common misconception based on her similar appearance to Rosario Dawsen (fabricated substantiation)' even though they don't remotely look alike. And if you keep pressing these Mistrals they will keep repeating the fabrications or generate new ones.
Simply put, using 100% true corrections as training data might seem like a reasonable thing to do, but >99% of the time this just trains tiny 7b LLM that are far less intelligent and knowledgeable to stubbornly stick to egregious errors using fabricated substantiations.
Thank you. I definitely understand your frustration. This was an experiment where I used only 15k samples of data that I have had good results with.
I will work on using better, human annotated data in the future as I have with other models. Do you have any suggestions on ways that this model could be improved based on your experience?
edit: also this model is trained on rejected samples as well. Not saying your point is invalid, but I had good intentions.
@macadeliccc Thanks for asking. Nothing comes to mind now, but I'm certainly not shy about criticizing if I think of something.