173 lines
11 KiB
Markdown
173 lines
11 KiB
Markdown
|
---
|
|||
|
|
base_model:
|
|||
|
|
- N-Bot-Int/OpenElla-NovelWriter-8B-V2-merged
|
|||
|
|
- N-Bot-Int/OpenElla-NovelWriter-8B-Magnate-merged
|
|||
|
|
- akjindal53244/Llama-3.1-Storm-8B
|
|||
|
|
tags:
|
|||
|
|
- text-generation-inference
|
|||
|
|
- transformers
|
|||
|
|
- llama
|
|||
|
|
- trl
|
|||
|
|
- roleplay
|
|||
|
|
- conversational
|
|||
|
|
- merge
|
|||
|
|
- chat
|
|||
|
|
license: agpl-3.0
|
|||
|
|
language:
|
|||
|
|
- en
|
|||
|
|
pipeline_tag: text-generation
|
|||
|
|
datasets:
|
|||
|
|
- N-Bot-Int/Iris-Uncensored-Reformat-R2
|
|||
|
|
- N-Bot-Int/RP-Mixed-v1
|
|||
|
|
- N-Bot-Int/Magnum-RP-Dataset-v2-PACKED
|
|||
|
|
library_name: transformers
|
|||
|
|
---
|
|||
|
|
<details close>
|
|||
|
|
<summary><b>📢 News / Changelog (click to collapse)</b></summary>
|
|||
|
|
|
|||
|
|
### NEWS AND ANNOUNCEMENTS ###
|
|||
|
|
(yes i write the announcement here, i figured people wont read on the main page so i'm announcing it here! - if you want to skip.. feel free to hide this!)
|
|||
|
|
|
|||
|
|
## ANNOUNCEMENTS ##
|
|||
|
|
|
|||
|
|
- # New Models Will Be Delayed, Halted, or Canceled Until Further Notice #
|
|||
|
|
- Lightning.ai, one of the GPU hosts we use, has announced that they will no longer offer free credits starting the end of July. This means our lifeline for testing,
|
|||
|
|
dataset generation, and experimentation is **OFFICIALLY CUT OFF** until we find someone or a company to donate a cluster. **NEW MODELS AIMED AT IMPROVEMENTS WILL BE
|
|||
|
|
DELAYED, CANCELED, OR PUSHED BACK** as we try to do everything within Kaggle and Google Colab.
|
|||
|
|
*PS: No, we are NOT rich in GPU... we're doing this for the community and our love of finetuning. This means future updates will no longer be "entirely" new
|
|||
|
|
architectures or models trained on new datasets — they'll probably be mixtures of our older AI models.*
|
|||
|
|
|
|||
|
|
- We're running out of compute. Long story short: due to Lightning.ai removing their free tier, we're looking for alternatives to generate the datasets we need to
|
|||
|
|
train future AI models. We're also still trying to find our "signature model" — the one we can call entirely ours. You've seen it from Anthracite-org, where the
|
|||
|
|
moment you use their models you instantly know you've used Magnum because of its Opus-level length, or Stheno at RP, etc. We're currently trying to carve out that same kind
|
|||
|
|
of identity for ourselves, but it looks like that'll be delayed while we sort out our compute :(
|
|||
|
|
|
|||
|
|
- # We're Restructuring!
|
|||
|
|
- N-Bot-Int has operated as a company for a while now — but, much like OpenAI, we're bleeding money... well, not exactly, since we've been running on free tiers.
|
|||
|
|
Still, that means calling ourselves a "company" no longer really serves a purpose, so N-Bot-Int is shifting from a company to a full-time laboratory!
|
|||
|
|
Expect more open-source releases from us! (Not models — as mentioned above, we don't have the compute for that :( )
|
|||
|
|
|
|||
|
|
- # Compliance
|
|||
|
|
- It came to our attention that *Meta* took down the base model we used for OpenElla-Novelwriter v2. We're not usually the type of lab to overreact to a single
|
|||
|
|
incident, but watching that happen made something click — it wasn't just fear, it was a loss of trust in Meta and a need for real assurance going forward. That
|
|||
|
|
risk isn't hypothetical anymore; we've seen it happen firsthand. Given this, and given Llama 3.1's more restrictive licensing terms, we've decided that
|
|||
|
|
**THIS WILL BE THE LAST LLAMA MODEL WE FINETUNE FOR THE FORESEEABLE FUTURE**. Future models will explore **QWEN** or **GEMMA** instead — we're stepping away
|
|||
|
|
from Llama bases for now, both for our own peace of mind and for legal safety. **THIS IS ALSO THE LAST TIME WE'LL USE AN ABLITERATED BASE**, for safety reasons.
|
|||
|
|
|
|||
|
|
- Separately, we're also moving future model releases to **AGPL-3.0 instead of Apache 2.0**. This isn't about Meta or the takedown — it's about us not wanting
|
|||
|
|
companies to profit off our work without giving back to the community.
|
|||
|
|
*This licensing format, actually started already, I just had to announce it incase other people aren't up to date*
|
|||
|
|
|
|||
|
|
- # AI Refresh and Archival
|
|||
|
|
- Most, if not all, of our Llama models (except recently updated ones) will soon be archived on an upcoming org that houses all of our archived models! Archived
|
|||
|
|
models are automatically set to Apache (or to the original base's license). This means OpenElla (our first model family), MiniMaid, and the rest will be archived soon!
|
|||
|
|
|
|||
|
|
## THAT'S ALL THE NEWS WE HAVE FOR JULY!
|
|||
|
|
We're sorry to say we'll no longer be releasing on a monthly schedule — running finetunes while keeping the lights on is NOT cheap, so we're no longer rushing for
|
|||
|
|
monthly releases until we find an alternative, sponsor, or cluster. We're currently in firefighting mode.
|
|||
|
|
|
|||
|
|
## Support us! ##
|
|||
|
|
Through **Ko-fi**: https://ko-fi.com/J3J61D8NHV
|
|||
|
|
If you have any questions, feel free to email us at: nexulon.botworkinteractives@gmail.com
|
|||
|
|
</details>
|
|||
|
|
|
|||
|
|
|
|||
|
|
# OpenElla-NovelWriter-Requiem is OFFICIALLY RELEASED!
|
|||
|
|
|
|||
|
|
- Thanks Grok!
|
|||
|
|
|
|||
|
|
(DETAILS REMAIN THE SAME(Just like the methodology)... EXCEPT WITHOUT THE DARE TIES WITH V1! I WILL UPDATE THIS MODEL CARD WHEN I RECEIVE REQUESTS HOWEVER FOR NOW, SINCE THIS IS NOT THE MAIN VARIANT, THERE
|
|||
|
|
WILL BE NO UPDATE... FEEL FREE TO REQUEST FOR A MODEL CARD UPDATE, although update will not be assured)
|
|||
|
|
|
|||
|
|
# OpenElla-NovelWriter-Requiem-Ascended — quite possibly our best Llama model yet *(yeah yeah, entirely subjective... let us have our OpenAI moment, just this once.)*
|
|||
|
|
|
|||
|
|
- Introducing OpenElla-NovelWriter Requiem Model! Requiem, based on the latin word *(that links to our announcement)*, is our brand new ai model that
|
|||
|
|
showcases a new RP capabilities compared to our older generation of NovelWriters. Requiem Ascended is our first ai model that is finally sensitive to
|
|||
|
|
character cards, Has a unique blend of creativity freedom on all types of Roleplay, and Expanded capabilities trained on llama!
|
|||
|
|
|
|||
|
|
- Compared to previous NovelWriters, Older novelwriters suffer from our **blarbing** sickness, where the ai model spat out the entire chat, without stopping and
|
|||
|
|
waiting for user's prompt, This final version of NovelWriter, named Requiem is the **TYPE A** of the family, A final **type B** is to be released to let **YOU** choose
|
|||
|
|
which of the final two are the best Requiem model!
|
|||
|
|
|
|||
|
|
- Both Requiem Ascended and Requiem(TypeB) models are capable of RP, whilst also solving the Old NovelWriter family's issue of **spatting out dialogues, and God modding**
|
|||
|
|
Alongside it, the Requieum Ascended model has better stability when it comes to prompt following, better writing and dataset variety thanks to our DARE TIES strategy,
|
|||
|
|
of which we mixed all our ai models that learned from different varied dataset to form this Masterpiece!
|
|||
|
|
|
|||
|
|
*READ MORE FOR MORE INFO*
|
|||
|
|
8 BILLIONS PARAMS MODEL
|
|||
|
|
|
|||
|
|
# OpenElla-NovelWriter-Requiem-Ascended Procedure/Methodology:
|
|||
|
|
- **OpenElla Requiem Ascended** is our brand new ai model, it uses **DARE TIES** as a final methodology, however before forming the ai model, we trained the V3
|
|||
|
|
of OpenElla-NovelWriter, taking the V2 and finetuning it with our new magnum synthetic dataset, this dataset contains a wide variety of data which we failed
|
|||
|
|
to properly balance, leading to V3 being biased to certain hemisphere of RP and developing a concerning bias to specific instances of "scenarios". This is why
|
|||
|
|
We failed to release V3 for final release which is named "OpenElla-NovelWriter-8B-Magnate"... Magnate were almost deleted however we decided to not waste
|
|||
|
|
the training credits we used(since this is the final time we'll get the credits from lightning.ai) and decided to combine them with all of the NovelWriter family.
|
|||
|
|
Below are the composition of each Requiem model:
|
|||
|
|
|
|||
|
|
- **Requiem-Ascended**: Trained with V3 + V2 + BASE, = V3.2 => V3.2 + V1 = RA
|
|||
|
|
- **Requiem**: Trained with V3 + V2 + BASE = RE
|
|||
|
|
|
|||
|
|
The Base model we used, came from "akjindal53244/Llama-3.1-Storm-8B" as a stabilizer for this fusion, meanwhile both V3 and V2 were given a balanced weight and
|
|||
|
|
density to ensure the base model inherits the specialization of the two model and a final V1 is added for noise and enhancement. This is why we seperate the two
|
|||
|
|
family because we're not sure if some people would love the Ascended version or the Requiem, as both ai model has a different flavor of how they performs!
|
|||
|
|
|
|||
|
|
**Feel free to decide which is the best for you!**
|
|||
|
|
|
|||
|
|
# Training Details
|
|||
|
|
- **Finetuning Tool:** MERGEKIT
|
|||
|
|
- **Training Platform:** Kaggle Free Tier with T4 x2!
|
|||
|
|
|
|||
|
|
- **OpenElla-NovelWriter-Requiem-Ascended** is Our Brand New Powerful Model, If you ever encountered any issue, Want to commission us, or have any suggestions, please email us directly through
|
|||
|
|
[nexulon.botworkinteractives@gmail.com](mailto:nexulon.botworkinteractives@gmail.com)
|
|||
|
|
we value any reports, suggestions to how we improve future Model,
|
|||
|
|
Once again feel free to finetune the model to your likings, However please consider Adding this Page
|
|||
|
|
for **CREDITS**
|
|||
|
|
|
|||
|
|
- Please handle the AI with Care and ethical considerations, when **FINETUNING** this AI model, due to its **UNCENSORED** Nature.
|
|||
|
|
- We are not responsible for what this model generates. Use it responsibly and legally. You downloaded it, you own what you do with it.
|
|||
|
|
|
|||
|
|
---
|
|||
|
|
|
|||
|
|
## Activation Steering
|
|||
|
|
|
|||
|
|
An obedience steering vector is available at `obedience_vector.pt`.
|
|||
|
|
Apply at `layers.23.mlp` with `strength=1.5` (recommended range: 1.0–2.5).
|
|||
|
|
Improves character card compliance, author note following, and suppresses
|
|||
|
|
narrator takeover / OOC breaks.
|
|||
|
|
|
|||
|
|
# What's Coming Next?
|
|||
|
|
|
|||
|
|
> 🔒 **Requiem Base to be released soon! - Supporters on KO-FI receives the model early as usual!**
|
|||
|
|
---
|
|||
|
|
|
|||
|
|
# Notices & Usage Tips
|
|||
|
|
|
|||
|
|
- **Use Llama 3 format** — the model is based on Llama 3, Soooooo using Llama 3 works best!.
|
|||
|
|
- **Calibrate per character card** — every character is different, adjust your prompt, Model's settings(ie, temps, Top-K etc.) accordingly.
|
|||
|
|
- **OUR SUGGESTIONS ESPECIALLY FOR KOBOLD USERS!**
|
|||
|
|
- Use DynaTemp for a better experience, drop the Min-Temp to 0.50, Max-temp to 1.50, on DRY set the MULT to 0.3 and Shift the Repeat Pen if you want to
|
|||
|
|
boost V3's influence......... somehow... dont ask why... but the ai model starts to act weird as you shift the repeat pen higher :|
|
|||
|
|
|
|||
|
|
- Format Llama 3 as mentioned above! sometimes the ai model includes [Author notes], it just depends on the character cards. Simply modify the character cards
|
|||
|
|
or include an author note to prevent or stop the ai from using the "[author notes] or OOC stuffs"
|
|||
|
|
|
|||
|
|
(Tho then again... we are bad at giving advices... just use whatever works to ya and comment what you've used so others can use the ai model xDDDDDD)
|
|||
|
|
|
|||
|
|
- Chat Completion preset uploaded by PaperMache - [Preset](https://huggingface.co/PaperMache/NBotIntModelsPresets/resolve/main/NovelWriter-Requiem-Ascended-Preset.json)
|
|||
|
|
---
|
|||
|
|
|
|||
|
|
# About
|
|||
|
|
- **OpenElla-NovelWriter-Requiem-Ascended** is
|
|||
|
|
- **Developed by:** N-Bot-Int
|
|||
|
|
- **License:** agpl 3.0
|
|||
|
|
- **Finetuned from model :** N-Bot-Int/OpenElla-NovelWriter-merged, N-Bot-Int/OpenElla-NovelWriter-V2, N-Bot-Int/OpenElla-NovelWriter-Magnate, akjindal53244/Llama-3.1-Storm-8B
|
|||
|
|
|
|||
|
|
- # Detail card:
|
|||
|
|
- Parameter
|
|||
|
|
- 8 Billion Parameters
|
|||
|
|
- (Please check your GPU Core, VRAM, CPU and RAM to see if you can comfortably run 8B models)
|
|||
|
|
|
|||
|
|
- Finetuning tool:
|
|||
|
|
- MERGEKIT
|
|||
|
|
- Fine-tuned Using:
|
|||
|
|
- Kaggle Free Tier with T4 x2
|