MrgrtV2-3B, rebalanced, cleaned and fixed than previous version!
MrgrtV2-3B, our brand new flagship model is now released with a better model training than previous V1, the V1 had a massive poison on the dataset
it were trained, leading to the model having insane early <|eos_token|> even though the roleplay is still not yet finished.
MrgrtV2-3B aims to fix the issue by showcasing the quality of the dataset, by training on a 3B model! revealing that the dataset is high quality!
and the early issues were simply the poison that killed the quality of V1 Models!
THEREFORE THIS MODEL IS PURELY A FIX TO THE PREVIOUS V1 MODEL! HOWEVER THE QUALITY IS INSANELY, NIGHT AND DAY!
READ MORE FOR MORE INFO
3 BILLIONS PARAMS MODEL
MrgrtV2-3B Model Procedure/Methodology:
MrgrtV2-3B is trained Using HuiHui-Ai's Llama 3.2 3B instruct abliterated Model,
ensuring full creative flow without refusal!
MrgrtV2-3B is then fed the new 22k single-row DATASET named MEulysis-Cleaned-formatted which were obtained after splitting the previous dataset and cleaned EVEN FURTHER!
MEulysis-Cleaned-formatted — a 22 entry synthetically generated dataset produced using
multiple capable RP models (including Hermes(mythomax were removed because the mythomax was causing the poison to the dataset)), then carefully cleaned,
formatted with Llama 3 chat formatting, and then, tuned across a wide range of roleplay scenarios.
What makes MEulysis-Cleaned different?
22,000 entries of synthetic RP data generated from multiple frontier RP models
removed mythomax which caused early eos terminations and some... quality issue removal
Wide RP focus covering diverse character types, settings, and narrative styles
Includes MOST ELABORATE EXPERIENCES to MOST TABOO SCENARIOS
Clear formatting guidance baked into the data — proper use of "dialogue" and *actions*
Now includes Llama 3 chat formatting to maximize quality on Llama models we'll release!
Fully cleaned and curated for quality and consistency
Training Details
Finetuning Tool: Unsloth AI + Huggingface TRL(we uses our brand new Kaggle Extenderizer to extend training without starting from scratch)
Training Platform: Kaggle Free Tier with T4 x2!
Epochs: 3
Final Training Loss: 1.2
Dataset: MEulysis-Cleaned-formatted
MrgrtV2-3B 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
nexus.networkinteractives@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.
What's Coming Next?
🔒 V3 will released with better support for more OPEN ROLEPLAY SCENARIOS
Right now the ai model is somewhat bad with Open roleplay, however we'll do our best to release the V3 with better improvements!
🔒 8B variant to be released soon!
Due to this cleanup, we can conclude that the 8B model will be better than the previous V1! we'll take our time releasing the next 8B variant!
🔒 1B variant to be released
Experimental 1B variant with the same training methodology will be released following 4B, this is EXTREMELY EXPERIMENTAL, 1B are not good for Roleplaying
however we share them nonetheless!
🔒 Release PEFT Soon
New PEFT releases of the model, will now be released late. Supporters on our Ko-fi can however request for the PEFT(through gmail) if they ever want to!
(This is to thank our Ko-fi supporters!)
Benchmarks are also in the pipeline and will be added once available.
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. However
We recommend(if you're using koboldcpp), to use DynaTemp with Dynamic Min as 0.60 and Dynamic max as 2.00.
That's the setting we found to make the model more coherent, less hallucination whilst making the model adventurous!
About
MrgrtV2-3B is
Developed by: N-Bot-Int
License: agpl 3.0
Finetuned from model : huihui-ai/Llama-3.2-3B-Instruct-abliterated
Detail card:
Parameter
3 Billion Parameters
(Please check your GPU Core, VRAM, CPU and RAM to see if you can comfortably run 3B models)
Finetuning tool:
Unsloth AI
This llama model was trained 2x faster with Unsloth and Huggingface's TRL library.
