Noromaid model series

Name comes from:
Noro = No_robots(dataset),
Maid = Minerva AI Dataset

MinervaAI Dataset referring to: Aesir

Noromaid 0.1.1

- Noromaid-13b-v0.1.1

- Noromaid-7b-v0.1.1

- Noromaid-20b-v0.1.1

Yet another merge with undi, came out 16.11.2023, main part of it a dataset by MinervaAI named aesir, other part is a modified version of No_Robots.

Noromaid 0.2 [ EXPERIMENTAL ]

- Noromaid-13b-v0.2

- Noromaid-7b-v0.2

As seen in the title, this model is pretty experimental!

This time we used MergeMonster, with many bad_phrases. We also used another MinervaAI dataset.

The 13b was published on 16.12.2023, the 7b was released on 21.12.2023, further B's will be released when we get more feedback

Noromaid 0.3

- Noromaid-13b-v0.3

Noromaid 0.3 has: new datasets, normal alpaca Instruct(with instruction, input, and response header) instead of our modified alpaca.

It is a complete retrain based on the same model as 0.1.1 and 0.1.

Noromaid 0.4

- Noromaid-7b-v0.4

- Noromaid-7b-v0.4-DPO

Noromaid 7b v0.4 is a fully finetuned 7b Mistral trained on multiple RP dataset, modified by our own hand, and redone entirely from Alpaca to Chatml. The switch from Alpaca to Chatml and the addition of a new dataset from Aesir made it even better than 0.3.

Noromaid 7b v0.4 DPO is the same as Noromaid 0.4 7b. Additionally, it was trained on top with 3 DPO datasets including the Intel DPO for reasoning, and some uncensoring.

Spin-Off's

(standalone)

Mixtral

- Noromaid-v0.1-mixtral-8x7b-v1

Noromaid-v0.1-mixtral-8x7b-v1 is a fine-tune of Mixtral 8x7b, trained on various RP datasets including LimaRP and Aesir, ToxicDPO without warning for decensoring and norobots, rewritten to use a modified Alpaca prompting to be on a par with ChatML or other conversational formats. This version contains 3x Alpaca modified datasets (for the RP one) and 2 shareGPT datasets. LimaRP token length, input, and output size got wiped out, and were separated into chunks of +8k context conversation (this is the maximum LimaRP can offer).

- Noromaid-v0.1-mixtral-8x7b-v2

Noromaid-v0.1-mixtral-8x7b-v2 is a fine-tune of Mixtral 8x7b, trained on various RP datasets including LimaRP and Aesir, ToxicDPO without warning for decensoring and norobots, rewritten to use a modified Alpaca prompting to be on a par with ChatML or other conversational formats. This version contains 3x Alpaca modified datasets (for the RP one) and 2 shareGPT datasets. On this v2, LimaRP got fixed further, Axolotl received a monkey patch to actually reformat shareGPT to our modified Alpaca prompting.

- Noromaid-v0.1-mixtral-8x7b-v3

Noromaid-v0.1-mixtral-8x7b-v3 is a fine-tune of Mixtral 8x7b, trained on various RP datasets including LimaRP and Aesir, ToxicDPO without warning for decensoring and norobots, rewritten to use a modified Alpaca prompting to be on a par with ChatML or other conversational formats. This version contains 5x Alpaca modified datasets. On this v3, all the datasets got trained on the "completion" method of Axolotl, with all the datasets being completely rewritten to be in Alpaca modified format. More than 2600 Wikipedia references got cleaned up of Norobots dataset.

- Noromaid-v0.1-mixtral-8x7b-Instruct-v3

Noromaid-v0.1-mixtral-8x7b-Instruct-v3 is made with a LoRA done on base Mixtral 8x7b. It's the same data than Noromaid-v0.1-mixtral-8x7b-v3, but applied on the Instruct model. Fine-tuning on base and applying on Instruct seem to give better result for our usage: RP/ERP. Stay tuned for more information!




Noromaid-v0.1-mixtral-8x7b series info: v1 and v2 were each trained for 2 epochs, totaling to each 8 hours on Axolotl. v3, on the other hand, underwent 3 epochs and was trained for 12 hours, bringing the cumulative training time to 28 hours on a single A100 80GB GPU.

Miqu - [HF Collection]

- MiquMaid-v1-70B

Quick train to see if miqu finetuned results in good models.

- MiquMaid-v2-70B-alpha-GGUF (GGUF only)

MiquMaid v2-alpha is trained on 1 epoch for 18h running on 2xA100 80GB. Trained on miqu. Deprecated!

- MiquMaid-v2-70B / [DPO]

MiquMaid-v1-70B was our first 70B model, based on the leaked Mistral Medium model. V1 used Aesir datasets where V2 make the return of Norobots and some uncensoring data in it to make it even more unethical in RP.
V1 was already compliant to a lot of things, even on some prompt Mistral Medium would refuse 100%, because it is HIGHLY aligned. This V2 let you prompt even more unethical and unhinged RP. Not using any RP format show a rate of refusal really lower than Mistral Medium too!
On top of that, a DPO train, using the same data that was used in the OG Finetune for better performance made it even better, write better, and be even more uncensored. The model lose some points in benchmark, but the tradeback for really good RP and less repetition was worth it.

- MiquMaid-v2-70B / [DPO]

MiquMaid-v2-2x70B is really heavy, it's a 125B model made of MiquMaid-v2-70B and Mistral Medium base. Each MoE model have 2 expert active in them, so the idea between this was to have, on every token, 2x70B expert working together for more randomness and better precision. Since MiquMaid-v2 is here only for RP, it lose some IQ point, that's where Mistral Medium hit and make the prose a lot more better, and let the model be more logical. It's better than a frakenmerge of 2 Mistral medium because it's not 1:1 2x the same model.
The gem tho, is the DPO version. When the qLora for uncensoring was made on MiquMaid-v2 for uncensoring it, we got the idea to apply it to Mistral Medium base too. At first, the result wasn't really good and that was expected, since the qLora wasn't trained on Mistral Medium, but on MiquMaid-v2. BUT! When merged together (in a MoE for this example) the full potential of a double DPO shine.
We got astonish result on the worst quant ever : Q2_K. Even at Q2_K, MiquMaid-v2-2x70B-DPO showed godly performance in RP, following card, logic and smut. Only downside was the repetition of the formatting, but it was really usable. Unquanted is a gem, but we doubt anyone have the compute power to do that...




(based on 0.4)

- Noromaid-v0.4-Mixtral-Instruct-8x7b-Zloss

It's Noromaid-v0.1-mixtral-8x7b-Instruct-v3 but retrained with ChatML, Zloss(Thanks charles), and some added datasets.

- FlatOrcamaid-13b-v0.2

FlatOrcamaid-13b-v0.2 is a merge between FlatOrca and Noromaid-13b-v0.2, see more info on the repo

Important info to a 7b model: Currently not possible due to Orca-2 being a LLAMA 2 7b model and Noromaid-7b being a mistral model!

Important info to this model: This model is in NO way affiliated with ddh0/OrcaMaid-13b or and *maid models by ddh0, it was inspired by him tho.

Credits:

- Undi / Wrote parts of the blog post.

Updates:

Collab with Undi / Echidna-13b v0.1, v0.2, v0.3 + NeverSleep/Nethena-13B

Echidna-13b-v0.1

Echidna-13b-v0.1 came out 21.10.2023, i wanted to merge it alone first, but had server issues. Thats when i asked Undi to help merge it, we reviewed my recipe and decided to upload it on his NeverSleep org and let it be a collab.

Echidna-13b-v0.2

Echidna-13b-v0.2 came out 26.10.2023 (took 4~ hours of merging), it was kind of a "test" version of Echidna-13b-v0.3. Echidna-13b-v0.3 would follow soon after.

Echidna-13b-v0.3

It took us some time to refine Echidna-13b-v0.2's recipe, we had multiple test models(private) which game way better results than Echidna-13b-v0.2.

The newest version, Echidna-13b-v0.2 came out 27.10.2023 (took 5.30~ hours of merging).

Repetition seems to be almost non-existant, which is really nice. Undi and me noticed that while trying to let the AI talk to itself(with impersonate).

Coherency is outstanding, as is creativity.

This model seems to not have 5 IQ like some other models, for example: a femboy has a dick, not a vagina... It's also really nice at describing things, like scenes. This is probably due to having so much medical data in here.

Last but not least.. not a single "...shiver down X spine..." seen while testing!

Updates:

MinervaAI / Aesir

Sooo.. hey.. me, Gryphepadar, Doctorshotgun, Lordgoonery, M1xx3l, Nruaif are in sort of(?) a little team making LLM models no.

Currently we are working on a model + dataset called Aesir, we are almost 1 dataset manual cleaning down.

Aesir is meant as a horny-ass RP model(ERP).

You can get more info on our HuggingFace.

Wanna check out my resources page?

So i added a little section called "Resources", if you want you can check it out if you wanna look for new stuff programming/ML/LLM/SD or if you are starting out with that stuff.

You can check it out here or you can just click the "resources" button in the Navbar.

One thing i still need to do tho, is making this website better for mobile, as it looks shit currently(i think so at least).

Updates:

Website is finally "done"

Hey, i finally made the website here have content, and if you are asking "ikari, why tf are you posting so many blog posts, aren't they suposed to be kind of 'special'?"

To that i answer "Shut up, i need to fill this page so it doesn't look empty!"

Anyway.. it now looks awesome! And i removed all the lorem impsum shit, + I added, if you open a post in "full screen" it temporarily removes the sidebar for better visibility of the post.

Updates:

So i added a couple of cool things

I added a new url arg called "bo", when set to true combined with set blogid arg, it opens it in "fullscreen" mode.

I also added a little button(Open post), which does the same. When you click share it automatically sets the blogid arg, + it sets "bo" arg to true.

Hey, this is a little test for the new blog page

I will document my jurney as a dev here.

I dont think anyone is even reading this but yea, i dont really care xD

If you wanna for some reason share this or other posts, i added that thing below