News on next gen?
Been stalking a lot, was wondering if you were close to finish on the next gen.
Also are you thinking about open sourcing the training pipeline to allow for people to make more personal fine tunes of your tuned model?
Not even close, bunch of detector just got update and I think the old way is dead and I need to figure out some new way.
Current released model already open source all the training pipeline and i think i will do that to in the future.
I didn't get a modle that human style and rewrite quality both satisfy me.
If I can't get my ideal model in a short time, i may just put a patch on current model.
I tested a few things.
I went and trained a 27B qwen 3.8 model.
Dataset size 179K, human samples were 1-6 paragraphs. Majority ranging around 5 paragraphs, sources were news outlooks, over 80k from students, and other journals and targeted data types.
Epoch 1 was with Sol rewrites. An article that is 100% AI generated would always come back as AI paraphrased on pangram. An article that was 40-50% AI would come back as 0-10%.
Epoch 2 used same human pairs but was closer to the original human speech. After this epoch, things would start failing even GPTzero.
From the training data, both Argon and Sol's rewrites were flagging as AI polished ranging from 40-100% polished.
Both Epochs validation loss was always getting better, but it appears that the true target was getting worse. This had no AI supervisor aiding in the writing.
Also a note about pangram's paraphrasing ability if I rewrite AI and using the AI as a reference point my initial rewrite will flag as AI. It takes 1 additional rewrite + expanding of data to get far enough away to not flag as AI anymore.
I am speculating that the AI pair data is too human like but there is likely additional signals that places like Pangram is picking up on in their detector.
The Pangram paraphrasing is ridiculously strong and annoying having to spend a few hours to draft a paper out of it.
Yes bro，u actually found the key.
I already have a great progress on this becuase I found a way to make AI rewrited text actually detected by AI by pangram, with minimal drift on original fact and intent.
And lower rate of copying original text and structure will help you alot on (humanizing AI text)/(creat real AI text for training)
Unfortunately I wont be testing again, what I was doing was highly expensive for a personal project. Spent close to 3K in compute testing these theories out.
I was thinking I would need to do an AI rewrite of the human text and than have AI rewrite that text again to have a large enough shift for the AI training pair.
Do you have a recommendation on how I can get my training pair stronger as AI while still matching the original author voice/meaning?
Wow，that's terribily expensive. Im using the free nodes in my school so I don't spent money on GPU. And I only buy subscription for frontier LLM and detector so it's extremely cheap compare to api. So i don't actually cost that much and it's affordable for me
Didnt answer my question lol.
Also if I give you access to my dataset, would you make an strong AI pair for me? lol
Oh, it's actually pretty complicated. I use qwen local model to rewrite so it's basically free and super fast for me. So i can iterate on a single draft for multiple times , and generate different drafts for single human text. And pick the best.
What kind of dataset you have im curious, if i can use it to train the model I think i could help you(so that won't conflict with my timeline) (and I will earse all the sensitive information, and I don't think that will leak because it's a rewrite model)
It is a massive human selected data set.
You can use the argon and sol rewrites on it too, just use them at a lower rate as they are too close to the original. The diversity helps the model a lot of different human styles. Having some close to the original will teach the model it does not need to do a full rewrite on everything too.
Ill invite you to it once I upload it. Trust me, it will be extremely helpful for your needs.
OK， great thank you. I will test and try on it
sol and argon is too expensive for massive rewrites ( if we focusing on the ai rates, these model are actually good on some human sense so it's hard to use them to generate real AI text)
but u could distill them to a small local model and get a close performance (only if you focus on a specific unique usecase like rewritting)
Sent you an organizational invite
Everything had the scraping errors fully removed. So you wont get random spaces and random scraping errors.
The scrapping process had GLM 5.3 flash review for scraping errors and if an error was spotted an Astra agent would confirm if it was real and provide fixes.
All original human errors were maintained.
The dataset model card has in-depth of the breakdown of the data spread.
Let me know if you cannot access it, or if you join and cannot see it.
hi, any ways to find you in telegram or any "private" groups? thanks
hi, any ways to find you in telegram or any "private" groups? thanks
you can DM me through X probably, my telegram got banned
hi, any ways to find you in telegram or any "private" groups? thanks
you can DM me through X probably, my telegram got banned
in X it says you disabled private messages and I need to know your X number / handle "User @jialinyyzz has private messages disabled. If you know their X handle, you can still send them a message."
or please just add me @lwfr7