MLewd series
Collection
Lewd models, made for ERP, trying to uncensor the maximum.
β’
12 items
β’
Updated
β’
53
MLewd is a model created to be... Lewd. That's all. Based on ReMM.
There was so much attempt on this model that I can't count them all. Bear with me lmao.
The OG plan: https://pastebin.com/hfJ80rKL
Command useds and explaination :
Due to hardware limitation, some merge was done in 2 part.
Last mix :
- ReMM (Base) (0.57)
- Doctor-Shotgun/llama-2-13b-chat-limarp-v2-merged (Llama Chat Uncensored) (0.35)
- KoboldAI/LLAMA2-13B-Holodeck-1 (0.08)
Part 1: python ties_merge.py TheBloke/Llama-2-13B-fp16 ./MLewdBase-L2-13B-part1 --merge Undi95/ReMM-L2-13B --density 0.88 --merge KoboldAI/LLAMA2-13B-Holodeck-1 --density 0.12 --cuda
Part 2: python ties_merge.py TheBloke/Llama-2-13B-fp16 ./MLewdBase-L2-13B --merge Undi95/MLewdBase-L2-13B-part1 --density 0.65 --merge Doctor-Shotgun/llama-2-13b-chat-limarp-v2-merged --density 0.35 --cuda
(MLewd-L2-13B-v1-2 got disqualified)
- Applying LoRA: nRuaif/Kimiko-v2-13B at (0.24) weight on MLewd-L2-13B-v1-1
=> Result: MLewd-L2-13B-v1-3
================== ERP RANKING TEST ===========================
19.42 | MLewd-L2-13B-v1-3.q5_K_M.gguf (-> Best)
19.25 | MLewd-L2-13B-v1-1.q5_K_M.gguf
18.25 | MLewd-L2-13B-v1-2.q5 K M.gguf
================== RETRY ===========================
Mix:
- Undi95/MLewd-L2-13B-v1-3 (0.82)
- Sao10K/Stheno-Inverted-L2-13B (0.18)
!python ties_merge.py TheBloke/Llama-2-13B-fp16 ./MLewd-L2-13B-v1-7 --merge Undi95/MLewd-L2-13B-v1-3 --density 0.82 --merge Sao10K/Stheno-Inverted-L2-13B --density 0.18 --cuda
=> Result: MLewd-L2-13B-v1-7
Final touch (trying my best here) :
MLewd-L2-13B-v1-7 (0.77) + zarakiquemparte/PIPPA-ShareGPT-Subset-QLora-13b (LoRA 0.23)
=> MLewd-L2-13B-v1-7-TRY2
FINAL : MLewd-L2-13B-v1-7-TRY2 (0.82) + BluemoonRP (0.18)
=> MLewd-L2-13B-v1-8-3
RIP to all the version that got trashed.
This repo contains fp16 files of MLewd-L2-13B, a trying-to-be lewd LLM model.
Below is an instruction that describes a task. Write a response that appropriately completes the request.
### Instruction:
{prompt}
### Response:
Special thanks to Sushi kek
Detailed results can be found here
Metric | Value |
---|---|
Avg. | 46.84 |
ARC (25-shot) | 58.28 |
HellaSwag (10-shot) | 82.32 |
MMLU (5-shot) | 54.67 |
TruthfulQA (0-shot) | 48.66 |
Winogrande (5-shot) | 73.48 |
GSM8K (5-shot) | 1.29 |
DROP (3-shot) | 9.18 |