Please wait...
derp, I just reread the docs. the instruct data is not trade secret in this case.
related concept I ran across while browsing huggingface models: model merging.
this tries to explain it: https://developer.nvidia.com/blog/an-introduction-to-model-merging-for-llms/
I've seen some merged models in the 15B size range (runnable on gaming GPUs) that apparently outperform their base models, but you kinda have to worry about benchmaxxing...
In some cases the training data is also freely shared. True, not all, but some.
While i could be wrong, my understanding is the 'removing the rails' searches the model and looks for refusal tags and directly changes the model weights, its not like a retrain event. But regardless of how its done, in raw tests i have not seen them improve, and often not as good, but of course in tests where refusals got in the way, i can see it being better. In my own experience, i have had a 'core' model that would do x for me. I tried the abliteratied version, and it was hallucinating and rambling to the point it was unusable.
My use case is mostly to get around stupid stuff without having to con it. Example was wanting to do some API work "i cant answer that, as that could cause harm to the xyz system" finally got around it with some prompt changes, but it was rather annoying to have to do so for something that was so mundane. Another time, when i was having it do lyrics "oh, i cant create depressing lyrics .. bla bla " it was freaking goth ...that is the point.
I was probably conflating. The improved models I definitely remember seeing are merges of multiple base models.
me. I tried the abliteratied version, and it was hallucinating
and rambling to the point it was unusable.
Good to know. I have no use case for uncensored stuff so it makes sense for me at least to stay away from all that. More gernally, Huggingface leaderboards have a filter you can apply that says "original model vendor only" or words to that effect, which definitely excludes questionable things like merged models, but also excludes useful things like format conversions to OnnxRuntime.
(* format conversions can also be questionable because of the available operators and other issues, but this is clearly not the samt thing)
Yes, merges can be better.
Sat Nov 22 2025 00:39:03 UTC from LoanShark
I was probably conflating. The improved models I definitely remember seeing are merges of multiple base models.
there's so much about this field I don't understand. Like, how is it possible to merge a model from the Qwen series which something else from a different vendor that might have different I/O format conventions (e.g. left padding vs right padding just to pick one example) ...
I dont have the resources to merge, so never tried. I assume you would have to have all models adhere to the same 'standard', or have a way to modify them pre-merge.
Tue Dec 02 2025 11:29:00 PM UTC from LoanShark
there's so much about this field I don't understand. Like, how is it possible to merge a model from the Qwen series which something else from a different vendor that might have different I/O format conventions (e.g. left padding vs right padding just to pick one example) ...
Well apparently you still understand more than I do. It's all greek to me :)
Explain.
Sun Jan 25 2026 16:56:31 UTC from IGnatius T FoobarIf ya ever want to understand why pasting a screenshot always works while simply doing a copy-and-paste out of your browser doesn't, ask me in one of the techie rooms. Even I didn't really understand it until recently.
Subject: Why pasting a screenshot always works -- explained!
Explain.
I think you're going to like this, because it's one of those things that once you understand it, a light bulb illuminates over your head and you have a mental model of it from that point forward. I'm going to explain this from the perspective of a Linux system, because that's what I know best, but it's very similar on other operating systems as well.
A simplistic way of looking at it is: "there is no clipboard." That oversimplifies it, and there are exceptions. But normally when you "copy" or "cut" from a source application, what's really happening is that the operating system establishes a contract with that application to export some data at a future time. What you think of as "the data on the clipboard" is really just that most recent contract.
Now here's the fun part: as part of that contract, the source application offers a list of formats in which that data can be supplied. For example, a word processor might offer its own proprietary richtext format, an HTML format, and a plain text format.
When you "paste" into another application, the reverse happens: the receiving application looks at the list of formats which are available, and generally selects the richest format in that list that it knows how to handle. That's why if you copy from a richtext email and paste it into your word processor you get richtext, while if you paste the same into a console you'll get plain text.
[ Here's where we make a side note that there are exceptions. For example, what if you copy from a source application and exit that application before pasting into another? In some primitive systems the "clipboard" simply disappeared and you couldn't paste. What happens in modern systems is that the desktop environment steps in, sets up a temporary application to hold the copied data, and transfers that contract to itself. But as long as the source application is still running, it doesn't have to do that. ]
Still with me? Ok, with that mental model of how the clipboard *really* works, let's walk through copying an image out of a web page.
You're browsing the web, you see that really cool meme that you want to repost, you right-click on it and select "Copy Image". What just happened behind the scenes? In most web browsers, the browser exports a contract that says "I can give you this image in two formats: the image data itself, or the IMG link from the web page." Now you tab over to Citadel and you select "paste".
Again, the browser is at work here; it looks at the available formats and, unfortunately, does not give the user a choice: it prefers the link to the full image data, and that's what it pastes. It looks good to you at first, because you just looked at that page with the same browser and the image link is still alive. Later on however, you revisit that message and your image link is dead. Probably the source page doesn't allow deep linking to images. Failed image, ugly message, game over.
Now you try it again, but this time you take a screenshot. The image appears in your screenshot program. You select "copy" and what happens? The screenshot program can only offer that data in ONE format: the image data itself. There is no link. There is no file. There is no alternate format, only those lovely pixels themselves. You tab over to Citadel and paste the image. It's only available as image data. The browser has no choice to make. It takes that image data, converts it to an embeddable format (usually a data:// URI) and writes it directly into the message. Et voila, you've got persistent, embedded image data in your text that will render properly every time.
[ Side note 2: an application like a word processor or a fat email client like Thunderbird might accept that image data and store it elsewhere in its own data tree, but it's still copying the image itself, not a reference to an image stored somewhere else. ]
With that in mind, you now know what's happening when an image paste, or a paste of any media, isn't working the way you expected it to. You might think "oh, I'll just paste it into a word processor first" but even the word processor will try to outsmart you and either store *that* copy of it as a link, or it might even carry both the image and the link. The problem is the same: when you paste an image into a browser-based editor, the browser will always accept the link version first if it can. And until someone writes a browser mod that actually LETS YOU PICK the format to paste, that's how it's going to be.
Subject: Re: Why pasting a screenshot always works -- explained!
Oh. I DO like that!
URL://source.image.jpg
vs.
Actual .jpg hex
But you don't get to pick, and it is always going to prefer URL:// - because that is hotlink that is more efficient - you've got one copy that appears, embedded - and is less resource cost.
That is the reason it is designed that way - right? Because the concept of the web is hotlinks to source content - maybe repackaged in a new page - but why store it redundantly every time you CTRL-V CTRL-C it, which consumes extra resources.
Why have the entire image on MY Citadel and YOUR Citadel - on my storage and YOUR storage...
When the MESSAGE on your BBS can just call the image from MY system and present it on the webpage on YOUR Citadel?
Which is smart - but - it breaks. My image gets purged - now your message points to a broken link.
Whereas, if you have the actual image replicated in total in your message - it takes more RESOURCES (the image is stored on my system, and yours) - but if mine disappears, yours still has the image?
Did I get that right?
Subject: Re: Why pasting a screenshot always works -- explained!
I'll build that in someday, but don't expect it in WebCit Classic. :) For the time being, you know the workaround.
Subject: Re: Why pasting a screenshot always works -- explained!
I do. I'm kinda stoked I understand why - after this trip around Citadel being about how, "I'm not as smart as you - but I'm not as stupid as them."
Seems appropriate.
Wed Feb 11 2026 22:48:19 UTC from IGnatius T Foobar Subject: Re: Why pasting a screenshot always works -- explained!TFor the time being, you know the workaround.
Subject: Re: Why pasting a screenshot always works -- explained!
I'm an idiot savant. I'm probably autistic. What I am good at - I'm SUPER good at - and if decide I want to do something - aprende espanol, learn guitar, run a proxmox Linux cluster - I dive so deep I'm REALLY good at it - then I get bored - so I'm never an *expert* - I'm just WAY better than the average joe. Jack of All Trades - Master of none. It is hard to make a living at that - because life rewards people with a singular focus on one thing to an expert level - not people "passing good" at a bunch of shit.
I also get things, at the basic level - really quick. I rush ahead of the rest of the class at clustering, at failover, at Spanish or Russian or Italian, at guitar, at electronics - at the right racing line in the first few months - then things get hard - and I start to struggle with the concepts - and I go, "f this shit - who needs to think so hard it makes their head hurt," and I quit. Expert level - when the challenge begins and it doesn't just come easy - Spanish contractions and past tense conjugation - advanced heel and toe techniques - advanced scales - I give up because it slows down for me and I hate that feeling.
But I'm not alone. I grew up in the first wave of "gifted kids," and they told us all we would be ambassadors and journalists and heads of state. Most of us are burnt out rehabbed drug addicts. A few of us got insanely wealthy. Compared to the people I grew up with - they all think I'm a stellar success. But I feel like I wasted 90% of my potential. I shouldn't be worth 7 figures - I should be worth 9.
Then I see the people who are multimillionaires - for their Tik-Tok influencer status - and that makes me feel even worse.
Tech is a very weird industry to be in - because it attracts the people like me. And surrounded by people as hyper-intelligent has a tendency to make you feel like a dumbass - even if you're in the top 20%. I'm sitting very stable at the 19% of the top 20%. I'm a dingleberry of smart on the s-hole of super intellect. I've always known people who - you know, can solder an additional 4mb ram onto an Atari ST - or cut a trace on their Hayes modem and make a 2400BPS moden into a 9600 BPS modem. I try that - I've got a dead modem with a voided warranty. :D